AGRICULTURAL PRODUCTIVITY AND CLIMATE ADAPTATION: INSIGHTS FROM
PASTURE TECHNOLOGY AND THE HUMAN WELFARE IMPACTS OF RISING
TEMPERATURES

By

Jose María Martínez

A DISSERTATION

Submitted to
Michigan State University
in partial fulfillment of the requirements
for the degree of

Agricultural, Food, and Resource Economics—Doctor of Philosophy

2025

ABSTRACT

Productive sustainability and climate adaptation strategies in agriculture and development are key

goals of global public policy. Understanding how agricultural technology can boost productivity

in sectors like cattle ranching, and how rising temperatures directly and indirectly affect human

well-being is essential for shaping effective policies in both developed and developing countries.

The first chapter examines the impact of introduced pastures on lowland ranching productivity in

Latin America. The final two chapters evaluate how rising temperatures influence human capital

accumulation and crop yields in Colombia and the US, respectively, and how students’ and farmers’

behaviors modify these effects.

The first chapter examines three critical questions about introduced pasture adoption in Colom-

bian lowland ranching. First, it estimates the impact of using Brachiaria grass, a high-nutrient grass

forage variety, on farm-level carrying capacity and livestock revenues. Second, it assesses whether

combining complementary practices like fertilization and weed control boosts the impact of using

these pastures. Third, it evaluates the potential reduction in land demands for cattle ranching at-

tributable to introduced grasses. The study follows an instrumental variable framework, combining

a cross-sectional representative sample of Colombian lowland ranchers, historical data on violence

and weather, and spatial information on the location of R&D centers that disseminate knowledge

about intensive grazing systems. The use of pastures significantly increases farms’ carrying ca-

pacities, although at levels below those suggested by agronomic trials. In addition, our analysis

suggests that gains from productivity are more significant when combined with complementary

practices and that, in the absence of these introduced pastures, reaching similar output levels would

have potentially required twice as much land as currently used.

The second chapter studies the impact of exposure to increasing temperatures on human capital

accumulation, with particular attention to the potentially differentiated effects between rural and

urban settings. Combining rich microdata from a high-stakes, national-level high-school exit

exam in Colombia (Saber 11) between 2014 and 2019 with weather-station-level information, we

use a fixed-effect estimation to measure the effect of increasing average temperatures on student

success. We find that a 1°C significantly decreases math and overall test scores in urban settings,

while there is an apparent positive effect among rural students. Leveraging time-use data, we find

evidence of individuals in rural areas responding to increases in temperature by reallocating time

towards off-farm activities, which are human capital intensive. Finally, we use information from the

announcement of a national scholarship program that introduced exogenous variation in the stakes

of Saber 11, to find that the impact of the scholarship policy decreases as temperatures increase.

Our analysis underscores how the effectiveness of incentive-based policies may be hindered in the

context of a warming world.

The third chapter explores the impacts of heat on crop yields and crop production decisions

among sorghum farmers in Kansas, U.S. This analysis addresses a two-part question often over-

looked in climate econometrics. Specifically, we examine whether (a) the unbalanced nature of yield

panel data likely reflects strategic behavior by farmers when deciding whether to grow sorghum,

and (b) whether this leads to estimation bias from sample selection. Our results suggest that (1) a

farm’s decision to produce sorghum in a given year is likely a climate-sensitive response, indicating

an underlying climate adaptation strategy, and (2) accounting for this response as a driver of sample

selection results in lower estimates of heat impacts on yields compared to the existing literature.

Our results highlight the need to further consider how farm data discontinuity in long-term panels

can affect the consistency of traditional fixed-effect estimates.

In conclusion, livestock farming still needs to boost productivity, and the first chapter of-

fers insights into how agricultural technologies, like introduced pastures, can help achieve this.

Meanwhile, the second and third chapters emphasize the importance of accurately assessing the

impacts of rising temperatures on human welfare—specifically in terms of human capital and food

production–and contribute to the broader policy discussion on global climate adaptation strategies.

Copyright by
JOSE MARÍA MARTÍNEZ
2025

To my wife and my sister.
I love you.

v

ACKNOWLEDGEMENTS

My work and achievements would not have been possible without the outstanding support I received

from mentors, colleagues, staff, friends, and family throughout my Ph.D. journey. I am forever

grateful.

First, I would like to thank my advisor, Mywish K. Maredia. Her mentorship and sup-

port—extending even before my time at MSU—have been crucial to my growth as a researcher.

More importantly, Mywish has taught me by example what it means to be an outstanding human

being. Thank you for believing in me and for your unwavering guidance through the wild ride of

graduate school.

I am grateful to Songqing Jin, Saweda Liverpool-Tasie, David Ortega, Jeff Wooldridge, Scott

Swinton, Eduardo Nakasone, Brent Ross, and Tom Reardon for fostering a supportive and intellec-

tually stimulating environment throughout my studies. I would also like to especially thank Nicky

Mason and Jamie Bloom for their exceptional support and patience in helping me navigate the

day-to-day challenges of graduate school life. I truly appreciate every effort you made to help me

succeed at AFRE.

I am deeply grateful to my family for their unwavering love, support, and belief in me. Your

care, sacrifices, and understanding made this journey possible. Also, a special thank you to my

dear friends and colleagues: thank you for every thoughtful discussion, every cup of coffee, and

every laugh.

Finally, I gratefully acknowledge the financial support provided by the Sustainable Michigan

Endowed Project (SMEP) and the Seevers Graduate Scholars Program.

vi

TABLE OF CONTENTS

CHAPTER 1

Introduction .

IMPACT OF INTRODUCED PASTURES IN COLOMBIAN
1
LOWLAND RANCHING . . . . . . . . . . . . . . . . . . . . . . . . .
1
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
4
. 12
. 15
. 23
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
. 28
. 38
. 45

.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

.
1.1
. .
1.2 Methods . .
.
.
.
1.3 Data .
.
.
.
.
1.4 Results .
.
1.5 Discussion .
.
.
1.6 Conclusion .
.
Figures and Tables
. .
REFERENCES .
.
.
.
APPENDIX .

.
.
. .
.
.
.
.
.
.
.
.
. .
. .
.
.

.
.

.

CHAPTER 2

.

.

.

.

Introduction .

SWEATING BULLETS: HEAT, HIGH-STAKES EVALUATIONS,
AND THE ROLE OF INCENTIVES . . . . . . . . . . . . . . . . . . . 59
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
2.1
2.2 Theoretical Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. 62
2.3 Background and Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
2.4 Methods . .
. 70
2.5 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. 89
2.6 Conclusion . .
. 91
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
Figures and Tables
. .
REFERENCES .
. 106
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
OPTIMIZATION PROBLEM . . . . . . . . . . . . . . . . . . . 112
APPENDIX 2A
ADDITIONAL FIGURES AND TABLES . . . . . . . . . . . . 114
APPENDIX 2B

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. .
. .
. .

. .

. .

CHAPTER 3

.
.
. .

.
. .

Introduction .

HEAT, YIELDS, AND INCIDENTAL TRUNCATION . . . . . . . . .

.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. 146
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
3.1
. 149
3.2 Methods . .
Identification strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
3.3
3.4 Data .
. 154
3.5 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
. 162
3.6 Conclusion . .
. 164
.
Figures and Tables
. 172
. .
REFERENCES .
. 176
.
.
.
APPENDIX .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. .
. .
. .
.
.

. .

. .

. .

. .

.

vii

CHAPTER 1

IMPACT OF INTRODUCED PASTURES IN COLOMBIAN LOWLAND RANCHING

1.1

Introduction

Cattle production in Latin America is at a crossroads. Accounting for the largest share of

agricultural land, the sector is highly vulnerable to climate change, faces a need for scaling up

efforts to increase its productivity, and is in dire need of mitigating its greenhouse gas (GHG)

emissions (Arango et al., 2020; Bravo-Parra, 2021; Tapasco, LeCoq, Ruden, Rivas, & Ortiz, 2019).

In the case of lowland ranching, which is largely allocated in marginal lands, the challenge of

increasing intensification has met the difficulties of providing high-nutrient feed to animals under

poor soil conditions (Burkart, 1975; Lane, 1981; Toledo & Nores, 1986). In this context, the use of

introduced pastures1 has been promoted as a first-choice solution, as these have been able to deliver

sustainable production (Bogaerts et al., 2017), with agronomic trials suggesting they can assist in

increasing productivity (Holmann, Argel, & Pérez, 2008; Rivas & Holmann, 2004), facilitating

intensification (González-Quintero et al., 2021; Mazzetto, Feigl, Schils, Cerri, & Cerri, 2015),

reducing animal-based GHG emissions (Congio et al., 2021; Gerssen-Gondelach et al., 2017), and

serving as carbon sinks (Fisher et al., 1994; Hyman et al., 2022).

Despite well-documented evidence on introduced pasture adoption levels (ISPC, 2018), the

factors behind this adoption and its field-level impacts remain unexplored (Baltenweck et al.,

2020). Efforts from both National and International Agricultural Research Systems have focused

extensively on the development of production technologies (Vaccaro, 1997), while major limitations

remain in understanding the institutional and individual factors that shape the landscape of cattle

ranching and how farmers make decisions about technological adoption at the farm level (Enciso,

Triana, Diaz, & Burkart, 2019). On the other hand, most studies focused on the effect of using

introduced pastures are either agronomic experiments related to gains in cattle live weight (Holmann

et al., 2008; Pérez-López & Afanador-Tellez, 2017; Rivas & Holmann, 2004), ex-ante evaluations

1These grass forages varieties (mostly from the Brachiaria or Megathyrsus genera) are not native to Latin America,
and are either selected or improved germplasm lines from materials first brought into the continent from Africa (ISPC,
2018)

1

(Enciso, Charry, Rincón, & Burkart, 2021), and aggregate economic valuations (Fuglie, Peters, &

Burkart, 2021), but there are no studies assessing this impact with representative farm-level data.

As the literature finds that agronomic analysis can deliver estimates of the effect of using improved

practices that may not be achievable in real-world farm conditions (Laajaj, Macours, Masso, Thuita,

& Vanlauwe, 2020), providing impact estimates based on farm-level data is of the utmost relevance

for effective policymaking.

What is the impact of introduced pasture adoption on productivity? How does it affect revenues?

What does this impact translate to for a traditionally land-intensive sector? Taking Colombia as

a study case, we build a novel dataset combining information from a cross-sectional farm-level

survey, weather data, and official information on violent events and illegal shocks to transport and

trade experienced at the Municipality level. We use linear regression and optimal instrumental

variables methods to provide meaningful estimates of the effect of introduced pastures on carrying

capacity (i.e., kilograms of livestock per ha) and livestock-related revenues.

Among our sampled ranchers, the adoption rate of introduced pastures is estimated to be 66%

of pasture acreage. Our findings indicate that the distance to centers of technological dissemination

is significantly and negatively correlated with the adoption of introduced pasture varieties, whereas

education and other technological endowments are positively correlated. Also, introduced pasture

adoption negatively correlates with the number of violent guerrilla incursions that occurred at

the Municipality level. The latter is also the case for the number of illegal checkpoints enforced

by guerillas and other criminal groups. Our estimates suggest that introduced pastures matter

for productivity, with an average increase of 1.34% in carrying capacity for each additional 1%

increase in the share of introduced grass varieties. However, the effect is seemingly heterogeneous,

conditional on different pasture manegement practices, with potentially larger gains in carrying

capacity for those implementing practices such as weed control or fertilization. In addition, we find

a strong effect on sales: for each additional 1% of pasture adoption, there is an expected increase

of 2.05% in livestock-related revenues, on average.

Early introduction, dissemination, and adoption of African-native pastures in the Americas

2

started in the nineteenth century. Yet, in the absence of renovation schemes or proper management

practices, these pastures were rapidly naturalized, and their nutritional superiority decayed (Cam-

puzano, Triana-Angel, & Burkart, 2022; Van Ausdal, 2009), which further pushed livestock as a

large mechanism of deforestation and rapid degradation of soils (Dale, 1981; Faminow, 1997). The

first selected lines of introduced pastures (e.g., Brachiaria decumbens) were brought into South

America in the early 1950s, although modern selected varieties with better adaptation were released

in the 1990s (ISPC, 2018), which were the result of large scales of cross-validation through the

region (Rao, Zeigler, Vera, & Sarkarung, 1993; Toledo & Nores, 1986). In the case of Colombia,

pasture seed traders were established in the early 1970s (Hurtado, Diaz, Enciso, Gallo, & Burkart,

2020), along with research on the selection and dissemination of introduced materials following

the establishment of the tropical forages germplasm bank at the International Center for Tropical

Agriculture (CIAT), its subsequent Brachiaria improvement program in the late 1980s (Fuglie et

al., 2021), and transversal efforts from the Colombian NARS through the Colombian Agricultural

Institute and the Colombian Corporation of Agricultural Research AGROSAVIA (formerly Cor-

poica). The drivers of adoption remain unexplored, though, with breeders acknowledging a need to

better understand its dynamics for successfully improving strategies to link research with farmers

(Enciso et al., 2019; Peters, Lascano, Roothaert, & de Haan, 2003).

Several studies have explored the effects of the adoption of introduced pasture on livestock

productivity and farmers’ incomes. Based on agronomic trials controlling cattle feed, Holmann

et al. (2008) find that a switch from native savanna to Brachiaria grasses can increase animal live

weight gains (kg/ha/year) by between 40 and 70 percent. In addition, given that introduced pastures

are high-nutrient and are tolerant to several biotic and abiotic stressors in comparison to native

savannas, studies reveal that they can significantly increase productive efficiency of ranching in the

Americas (Herrero et al., 2013; Jank, Barrios, Do Valle, Simeão, & Alves, 2014; Mazzetto et al.,

2015; Pérez-López & Afanador-Tellez, 2017). Based on these findings, ex-ante analyses project

that a transition from native grasslands to introduced pastures can boost farmers’ revenues (Enciso

et al., 2021), especially when complementary practices like fertilization are enforced. Moreover,

3

when used in combination with trees for forage and shadow provision, they can bring significant

co-benefits in terms of GHG mitigation in addition to economic gains (Sandoval, Florez, Enciso

Valencia, Sotelo Cabrera, & Burkart, 2023). However, as mentioned earlier, these studies rely on

agronomic evaluations and non-probabilistic samples, which may lead to large estimated impacts

from pasture adoption that are unlikely to replicate in real-world field settings.

Our contribution is threefold. First, we contribute to the literature on pasture adoption, showing

that closeness to technology dissemination centers matters. Since these centers are not sellers

of seeds, this reveals that there is an important transaction cost to be considered for accessing

information about these pasture varieties. Moreover, we study how the history of violent groups’

attacks on population centers and shocks to trade and transportation affect observed technology

adoption rates, highlighting how shocks to infrastructure and value chains can affect technology

dissemination. Second, we benchmark an econometric approach based on representative data to

provide consistent estimates of the effect of introduced pastures on carrying capacity vis à vis

common methods in the field.

In line with the literature, our impact estimates fall short from

those suggested by agronomic trials, which ultimately would suggest the presence an upward bias

in expected gains measures in the currently available literature (e.g., ex-ante studies). Third, we

extend the analysis of impact directly into welfare metrics. Converse to previous studies that relied

on agronomic trials’ impact estimates to derive welfare implications, our framework allows us to

directly test for any relationships between adoption and livestock-related revenues.

The remainder of the manuscript is organized as follows: Section 1.2 introduces our theoretical

framework and empirical strategy, and discusses what covariates are included in the analysis.

Section 1.3 presents the data and descriptive statistics. Results are summarized in Section 1.4,

while Section 1.5 concludes.

1.2 Methods

1.2.1 Theoretical background and technology of interest

We consider an agricultural household model where a farmer aims to maximize their profits,

conditional on resource endowments, time allocation, and production technology constraints. Given

4

the strong historical frictions in Colombian rural land and labor markets (UPRA, 2016), this

allows for an intertwining between consumption and production decisions, including technological

adoption (Bardhan & Udry, 1999). We are interested in the farmer’s decision about what share of

their pastures will be allocated to introduced varieties.

Lowland ranching in Latin America is based on foraging systems, which follow one of (or

a combination of) two production schemes, namely extensive grazing on native savannas (i.e.,

traditional system) or intensive grazing based on sown, introduced pastures (i.e., improved system).

Traditional production largely relies on savanna burning practices, which increase the availability

of digestible plants for cattle, but at the cost of a high incidence of forest clearing (Fisher, Lascano,

Vera, & Rippstein, 1992). However, the low nutrient levels of native savannas lead to low carrying

capacities, with averages ranging between 0.2 and 0.5 animal units per hectare (90 and 225 kg

of livestock/ha) in Colombia. Meanwhile, introduced pastures work as a perennial crop of sown

grasses that provide high-nutrient content to grazing animals, allowing average carrying capacities

ranging between 1.5 and 4 animal units per hectare (675 and 1800 kg of livestock/ha) (Hyman et al.,

2022), although evaluations of pasture systems based on commercially available varieties suggest

that average long-term carrying capacities are in the lower end of this range (Vera & Hoyos Garcés,

2019). Therefore, these figures point to an average increase in carrying capacity between 200%

and 650% upon transitioning from a traditional system to an improved system based on introduced

pastures.2

To summarize, switching from native savannas to an introduced pasture system is expected

to increase carrying capacity by providing more biomass with higher nutrient levels in the same

amount of land. Under profit maximization, farmers’ use of introduced pastures leads to either

an increase in the stock of animals or a reduction in the amount of land dedicated to pastures.

Nevertheless, ranchers may not necessarily attach themselves to one system exclusively. For

instance, Ospina, Rusch, Pezo, Casanoves, and Sinclair (2012) find that in seasonally dry climates,

2Calculated as a simple percent increase. For instance, assuming 0.5 and 1.5 animal units per ha for the traditional
and improved systems, respectively, the percentage increase is Δ×100%=((1.5-0.5)/0.5)×100%=200% . Under an
assumption of linearity in elasticity, this suggests that each additional 1% in the share of introduced pastures should
lead to at least a 2% increase in productivity.

5

semi-natural grasslands can be competitive to sown, introduced pastures, especially when tree

covers and fertilization schemes are absent. Ultimately, the impact on productivity from the use

of introduced pastures can be captured through the observed weights of livestock per hectare (i.e.,

the reported carrying capacity of the farm). Our setting acknowledges that the level of introduced

pasture adoption is likely endogenous with productivity.

1.2.2

Identification strategy

We are interested in determining the elasticity of cattle productivity (and revenues) with respect

to the adoption of introduced pastures. To model this, let 𝑆𝑖 𝑗 𝑘 be the share of introduced pastures of

farmer 𝑖 in Department 𝑗 and Municipality 𝑘, m𝑖 𝑗 𝑘 and G𝑘 vectors of farm- and Municipality-level

covariates, 𝚪 and 𝜼 vectors of unknown parameters, and 𝛾 𝑗 a Department-level unobserved effect.

With our outcome of interest 𝑦𝑖 𝑗 𝑘 set as the carrying capacity of the farm in the last year, we are

interested in fitting a model that follows

log(𝑦𝑖 𝑗 𝑘 ) = 𝜏𝑆𝑖 𝑗 𝑘 + m𝑖 𝑗 𝑘 𝚪 + G𝑘 𝜼 + 𝛾 𝑗 + 𝑒𝑖 𝑗 𝑘 ,

(1.1)

where 𝑒𝑖 𝑗 𝑘 is an error term, and 𝜏 is the effect of interest.

In general, the decision to adopt a new technology (and the extent of adoption) is likely correlated

with unobservables also affecting observed yields. For instance, farmers with greater (unobserved)

abilities may be more likely to achieve larger-than-average yields out of varietal adoption, which

in turn will make them more likely to adopt (Larochelle, Alwang, Norton, Katungi, & Labarta,

2015). In the case of lowland regions, especially in a setting like rural Colombia where property

rights are poorly defined (UPRA, 2016), ranching also serves a purpose of signaling land tenure and

control (Triana Ángel, Burkart, & Pazos Cárdenas, 2024). Hence, unobserved factors like a latent

need to signal land ownership can hinder the adoption of introduced pastures as these tend to make

cattle concentrate (i.e., potentially reducing animal roaming). Therefore, such an unobserved factor

could lead to a downward bias in linear regression,3 and ordinary least squares (OLS) estimates

of coefficients in (1) are likely inconsistent (Angrist & Pischke, 2008). To overcome the latter,

we follow an instrumental variables approach, using plausibly exogenous sources of variation that

3A situation that could lead to a similar outcome is one where the treatment effects are heterogeneous, and one’s
identification strategy retrieves a local average treatment effect (LATE) larger than the average treatment effect (ATE).

6

can directly affect adoption, but that only affect yields through their effect on adoption (exclusion

restriction). We will expand on this in the following subsection.

While selection-on-observables approaches, such as those leveraging Oster (2019) bounds, are

valuable for addressing endogeneity concerns in some contexts, they present notable challenges in

the case of livestock adoption. Unlike settings where adoption decisions can be largely explained by

observable factors (e.g., soil characteristics in direct planting decisions, as in Assunção, Bragança,

and Hemsley (2019)), pasture adoption is heavily influenced by unobserved farmer characteristics,

such as managerial ability, risk aversion, and long-term investment capacity. These factors introduce

biases that may not be fully accounted for through an extensive set of observable controls.

Additionally, in our context, financial constraints and land tenure security—two key determi-

nants of technology adoption—are not fully captured by available data, making it challenging to

ensure that all relevant confounders are addressed. Given these limitations, an instrumental vari-

able (IV) approach provides a more robust strategy to mitigate endogeneity concerns by leveraging

exogenous variation in information access and conflict-related disruptions. Nonetheless, we ac-

knowledge that no identification strategy is without limitations, and we recognize the importance of

sensitivity analyses in strengthening causal claims. Future research could explore complementary

approaches, such as structural modeling or quasi-experimental designs with longitudinal farm and

policy data, to further validate these findings.

Unlike other studies that treat adoption as a binary treatment, we focus on the adoption of

introduced pastures a fractional-response variable. The latter follows that most sampled farmers

are adopters, but their extent of adoption (i.e., the share of pastures on introduced varieties) reveals

an important degree of variation. The latter also has important implications in our empirical

approach to address endogeneity concerns. As proven by Xu (2021), there is a drawback to using

approaches like two-stage least squares (2SLS) when addressing limited endogenous explanatory

variables: waving the non-linear nature of the true conditional mean in the first stage will render a

problem of artificially weak instrumentation. The latter means that, in a case like ours, measures

like the calculated effective F statistic (Montiel & Pflueger, 2013) are potentially misleading after

7

2SLS. Therefore, we further follow Xu (2021) and Wooldridge (2010) and use a two-step optimal

instrument approach, estimating a first-stage model with fractional regression by quasi-maximum

likelihood (QMLE) (Papke & Wooldridge, 1996), and then using the non-linear prediction as

an instrument for both IV and control function estimation. Measuring the effect of introduced

pastures on revenues is straightforward, as we can estimate a model like that of equation (1.1) with

log-revenues as a response variable, also by IV methods.

In addition, we account for possible differences in the effect due to pasture management by

fitting an extended specification that follows

log(𝑦𝑖 𝑗 𝑘 ) = 𝜏𝑆𝑖 𝑗 𝑘 + 𝜌1(𝑆𝑖 𝑗 𝑘 × 𝑊𝐶𝑖 𝑗 𝑘 ) + 𝜌2(𝑆𝑖 𝑗 𝑘 × 𝐹𝑖 𝑗 𝑘 ) + 𝜌3(𝑆𝑖 𝑗 𝑘 × 𝑊𝐶𝐹𝑖 𝑗 𝑘 )

+ (cid:164)m𝑖 𝑗 𝑘 (cid:164)𝚪 + G𝑘 𝜼 + 𝛾 𝑗 + 𝑒𝑖 𝑗 𝑘 ,

(1.2)

where 𝑊𝐶𝑖, 𝐹𝑖, and 𝑊𝐶𝐹𝑖 are mutually exclusive binary variables indicating, respectively, whether

the farmer manages their pastures only with weed control, only with fertilization, or both. Each

separate binary variable is added into the vector of remaining covariates (cid:164)m𝑖 𝑗 𝑘 . This flexible

structure allows to measure how the impact of introduced pastures may vary across possible groups

of complementary practices adopted in the farm, but it also implies a potential issue of bias in

estimating 𝜌1, 𝜌2, and 𝜌3 by OLS Therefore, for the case of estimating equation (1.2), we follow a

control function approach as it is a straightforward strategy to account for the potential endogeneity

of 𝑆 and its interactions with a single added covariate (Wooldridge, 2015). In addition, we provide

bootstrap standard errors to account for regression error (Wooldridge, 2010, 2015).

A caveat from this heterogeneity analysis is worth noting. The use of pasture-complementary

practices like fertilization, is considered a potentially endogenous decision that can impact pro-

ductivity (Kassie, Jaleta, Shiferaw, Mmbando, & Mekuria, 2013; Martinez, Labarta, & Gonzalez,

2023; Teklewold, Kassie, & Shiferaw, 2013). However, any adoption or feasible impact is condi-

tional to a prior decision on the adoption of introduced pastures. The latter follows that the use of

fertilization on native savanna leads to relatively insignificant increases of this grass’s availability

(Fisher et al., 1992), a fact that is also reflected in our sample, where the use of fertilization and

weed control only occurs among those using introduced pastures. Nevertheless, we acknowledge

8

this potential limitation and interpret any heterogeneity as a difference in the average impact of the

use of introduced pastures, but not as an impact from the use of complementary practices.

1.2.3 Sources of exogenous variation and threats to validity

1.2.3.1 Proximity to centers of R&D and use of other technologies

We consider the distance from the farm to agricultural R&D centers (i.e., CIAT or AGROSAVIA)

with extensive focus on ranching systems as an exogenous variation affecting the access to infor-

mation about technological innovations like introduced pastures. On the one hand, CIAT is a

not-for-profit center that develops and deploys agricultural technologies,4 while AGROSAVIA is

a non-profit, decentralized public entity, focused on scientific research, technology adaptation,

transfer, and advice.5 These institutions provide technological solutions and recommendations

with unrestricted access for the population, through open access to gene banks, data-and-research

repositories, and demonstration plots.6 Nevertheless, Colombia has historically experienced bot-

tlenecks in its public extension system, including a lack of funding and disconnectedness between

extensionists and scientists (Enciso, Triana, Díaz, & Burkart, 2022; Rodríguez, Ramírez-Gómez,

Aguilar-Gallegos, & Aguilar-Ávila, 2016). This institutional gap suggests that proximity to R&D

centers is most relevant as a proxy for direct access to the pasture-related knowledge and technolo-

gies, rather than a broad-based exposure to all livestock management recommendations.7

A potential threat to the validity of this exogenous variation arises from whether farmers are

strategically locating themselves close to these R&D centers. Yet, land is an essentially fixed

asset, which, combined with the high frictions of the Colombian agricultural land market (UPRA,

4See https://www.cgiar.org/research/center/alliance-bioversity-ciat/
5See https://www.agrosavia.co/en/what-we-do
6Demonstration plots have been a key component of the dissemination strategy, with selected sites established
in collaboration with local research stations, producer associations, and select lead farmers to showcase the benefits
of improved pastures. However, the geographic placement of these plots has been largely opportunistic rather than
systematically distributed, reducing concerns that they correlate with broader agricultural infrastructure or productivity-
enhancing factors.

7We also acknowledge that complementary practices such as animal health interventions could theoretically be
influenced by proximity to R&D centers. However, practices such as cattle vaccination are unlikely to be a major
confounding factor. While vaccinations for diseases like Foot-and-Mouth Disease, Brucellosis, and Rabies are crucial
for cattle productivity, these vaccines are already widely adopted in Colombia, with compliance rates exceeding 98%
according to data from the National Federation of Cattle Ranchers (FEDEGAN). Given this near-universal adoption,
it is unlikely that proximity to an R&D center significantly alters vaccination behavior in a way that would bias our
estimates.

9

2016), makes it unlikely that farmers can credibly choose their location strategically close to these

technology dissemination hotspots. Further, it could be argued that these centers provide access

to other means of technology, like improved cattle breeds, which in fact increase the demand for

introduced pastures. Nonetheless, although these institutions do some research on improved cattle

breeds, these efforts are only applicable to the Andean region, which is outside our area of study.

Therefore, we do not foresee any reverse causality. Finally, it could also be argued that this measure

of distance could be partially capturing a proximity to other kinds of infrastructure that could affect

productivity (e.g., roads, ports, input markets). However, our data analysis suggests that (a) farms’

distance to R&D centers is, for instance, uncorrelated with the distance to the closest agricultural

input market,8 and (b) when considering both these variables as factors affecting adoption, only

the distance to R&D centers remains statistically significant, and with a virtually unchanged partial

effect.

In summary, having the presence of CIAT and AGROSAVIA centers as a shock of knowledge

dissemination outside of the farmers’ decision-making, is arguably a weak assumption. Therefore,

with these centers continuously providing recommendations for intensification based on pastures,

we assume that the proximity to these centers affects observed cattle intensity only through its effect

on pasture adoption. Finally, as technologies and practices usually follow joint patterns of adoption

(Kassie et al., 2013; Martinez, Labarta, Gonzalez, & Lopera, 2021; Teklewold et al., 2013), we use

an index of ownership of farming durable goods (e.g., water pumps, sprayers, harrows, or tractors)

as a factor influencing adoption only.

1.2.3.2 Historical (indirect) shocks of violence

In addition, we use historical Municipality-level data on guerilla violent events and the presence

of illegal checkpoints by criminal groups as shock variables affecting access to introduced pastures

and their use but not cattle intensity. We consider these variables as exogenous shocks to the

historical seed availability and expansion in the adoption of introduced pastures. In our application,

we bring in official information from the Colombian National Center of Historical Memory (Centro

8In this case, a farm-level linear regression of the distance to the closest R&D on the distance to the closest ag-input

market and Department-level dummies, yields a statistically insignificant partial correlation of 0.0083 (Table 1A.1).

10

Nacional de Memoria Histórica, 2015) on the total number of attacks on the population that occurred

in the Municipalities during the 1988-2012 period,9, and from the CEDE Municipality Panel Data

(Acevedo & Bornacelly, 2014) on the number of illegal checkpoints set up by criminal groups

during the 2002-2016 period.10

Guerrilla attacks are concentrated in the cabeceras (downtowns) of each Municipality, which

are also the center of input trade in each location. However, these attacks are not targeted directly on

civilian (individual or group) property. Instead, they account for impacts on civilian infrastructure

resulting from attacks where the main targets were military (or police). Hence, it could be argued

that while these attacks may have affected a wider set of socioeconomic outcomes in the short-term,

their long-term effect may be of prominence in the stability of markets—thus in the distribution

of technologies. Conversely, other forms of violence (e.g., massacres) that are directly targeted

to civilians are more likely to affect property rights and human capital in the long run. This is

consistent with some correlational evidence from our data, where there is no significant correlation

between attacks on population and massacres.11 Likewise, illegal checkpoints imply a direct shock

to transportation and inter-municipal trade, which follow the same rationale as a historical plausibly

exogenous shock.

The period of violence considered for these variables largely coincide with the period where

most introduced grasses have been released and commercialized, namely 1987-2013 (Enciso et al.,

2019). Combined, we expect increases in these variables to reflect both weaker value chains and

a lower willingness to invest in technologies, which ultimately lead to lower rates of introduced

9These are defined as “[“[A]ny transitory military operation that consists of a temporary penetration of the territory
and that seeks to devastate adversaries and their material and symbolic environment, enhancing the devastating effect
of military action with the use of unconventional weapons and the attack against civilian targets. The temporary
nature of the penetration of the territory does not mean that it is ephemeral, but implies the deployment of a large
armed contingent with the capacity to sustain a medium-duration action, which is why it should not be confused
with disturbance (or baiting) or an attack that exclusively impacts a military target” (Centro Nacional de Memoria
Histórica, 2015; Grupo de Memoria Histórica, 2013). These kinds of attacks ceased in 2012 following the beginning
of peace negotiations between the Colombian Government and FARC.

10This measure combines reported illegal checkpoints set by guerrillas, gangs, and paramilitary groups to commit

land piracy and other criminal acts.

11With a Municipality-level regression of the number of massacres on the number of attacks on populations and
Department-level controls and robust standard errors, we retrieve a partial correlation of 0.326 and a 95% CI [-
0.581;1.235]. Using weights to account for the number of observations per Municipality yields a partial correlation of
0.174 and a 95% CI [-0.771;1.119].

11

pasture adoption.

1.2.4 Additional covariates

As control variables (or covariates) in our model, we include several farm and farmer-level

characteristics shown by previous studies to be correlated with technology adoption decisions,

including farm size, access to credit, membership in a farmers’ association, education of the

household head, house- and farm-assets, and whether the household head is female. Educational

attainment by the head of the household is connected to an increased ability to access (and process)

knowledge and new information, hence increasing technological adoption. However, at increasing

levels, education can be a path to off-farm income that can reduce the odds of dependence on

farming (Martey, Al-Hassan, & Kuwornu, 2012). Meanwhile, evidence on gender gaps suggests

that women face tighter restrictions to access credit and tend to possess fewer financial endowments

(Ouma, Jagwe, Obare, & Abele, 2010), hence adding a control for whether the head of the household

is female can reveal if there are important differences in pasture adoption or cattle yields to account

for. Finally, we use the average maximum daily temperature and average daily precipitation in

the year as variables potentially affecting productivity. In the context of the low-altitude, tropical

production system of our study area, increases in temperature can negatively affect productivity

due to the heat stress on cattle (Bravo-Parra, 2021), while additional rain can favor animal health

and weight gain (Ospina et al., 2012).

1.3 Data

We use data collected by CIAT between October 2016 and early February 2017. The data is

representative of beef and dual-purpose production for the Caribbean and Eastern Plains – Amazon

Piedmont regions.

It contains information on 1,021 farmers distributed across the departments

of Arauca, Casanare, Meta, Vichada (Eastern Plains), Caquetá, Guaviare, Putumayo (Amazon

Piedmont), Bolívar, Cesar, Córdoba, La Guajira, Magdalena, and Sucre (Caribbean). The sampling

followed a two-stage clustered procedure adjusted by potential intraclass correlation, selecting a

group of primary sampling units (PSU) and then defining the number of households to be randomly

12

sampled within each PSU12. Figure 1.1 presents the spatial distribution of the sample and the

location of CIAT and AGROSAVIA Research Centers with programs of influence to our regions

of interest.

In addition, we retrieved weather information from open-access databases by the

Colombian Institute of Hydrology, Meteorology, and Environmental Studies (IDEAM) for the

2016 agricultural season, and violence data from Centro Nacional de Memoria Histórica (2015).

1.3.1 Descriptive statistics

We summarize the variables used in this study in Table 1.1. An average farm reports around

536 kilograms of livestock per hectare. Hence, relative to an animal unit (roughly 450 kg), an

average rancher has slightly above an animal unit per hectare, although there is considerable

dispersion. Meanwhile, the average livestock-related revenues over the previous year were COP

684,415 (approximately USD 225 at 2016 prices), also with a considerable degree of variance.

The average share of pastures with introduced varieties among sampled farmers is 78%, while the

sample overall adoption (i.e., total acreage with introduced pastures over total acreage of pastures)

for the lowlands is 66%. Reflecting the spatial distribution displayed in figure 1.1, the average farm

is located within 226.2 km from a technological dissemination center. This distance is calculated

as the road distance from the farm to the closest R&D center.13

The largest share of farmers only does weed control in their grasslands (46%), while 39% do

no pasture management. On the other hand, 7% of farmers only do fertilization as a practice of

pasture management, while the remaining 8% do both fertilization and weed control. Participation

in farmers’ associations is low, with a reported level of 22 percent. Low access to specialized

knowledge and network support can constrain the odds of improving cattle production systems.

Access to credit is only reported in 22% of the farms, further revealing that households may not

find means to overcome liquidity constraints in the short-run. This is consistent with the results

12We use the unweighted sample for the regression analysis due to unavailability of weight data for the Caribbean

part of the dataset, but account for cluster correlation among observations.

13For each farm, we use DIVA-GIS open-access data on Colombia’s road network to calculate the path distance
from each farm to CIAT and each of the AGROSAVIA centers that have a research program in livestock, and then
keep the minimum distance. This approach will leave us with 994 observations for our main specification, due to
the exclusion of farms that are significantly away from any primary and secondary road, especially those from Puerto
Leguizamo, Putumayo (southernmost cluster of observations in Figure 1) as this Municipality is only accessible via
boat or airplane. Thus, later we test the robustness of our analysis using geodesic distances for all 1,021 datapoints.

13

in Mejía, Diaz, Enciso, and Burkart (2021), showing how most agricultural credit institutions are

clustered in the Andean region. The average number of household durable goods is 5.94 (out of

11), the figure is 3.81 (out of 9) for farming durable goods. Hence, the average farm is relatively

more endowed with household wealth than technological wealth.

Farm sizes vary widely, ranging from 1.5 to 7,000 hectares (roughly 3.7 to 17,300 acres),

with an average size of 200 hectares (495 acres). Nevertheless, it is worth noting that introduced

pasture adoption and the use of complementary practices are not exclusive to certain scales of

production. Table 1.2 disaggregates some of these measures by quintiles of farm size and reveals

a certain degree of homogeneity in terms of pasture adoption and use of complementary practices.

Conversely, there is an apparent correlation between productivity (and revenues, both at per-hectare

levels) and farm size, which emphasizes the need to control for this attribute in our impact analysis.

Circling back to Table 1.1, in terms of education, we found that a majority of heads of household

have completed primary education (71%), while only 15% report having completed secondary

education. Tertiary education was found in only 6% of the household heads, whereas the share of

those without any level of education is 8%. We use the latter as the base category to test whether

increasing achieved education systematically increases adoption. The participation of females as

household heads is reported in only 7% of the cases, initially suggesting that the sector will vastly

reflect male decision-making. Across the 45 municipalities, on average, at least one attack on the

population led by guerrillas, paramilitary groups, or other violent actors occurred in the 1988-2012

period. On the other hand, the average municipality experienced 1.64 illegal checkpoints between

2002 and 2016. Average precipitation across municipalities was 5.66 mm/day, with a low of 1.43

mm/day and a maximum of 16 mm/day. Finally, the 2016 average daily maximum temperature was

31.78°C (89.2°F), so the average farm and cattle are exposed to rather high levels of heat stress,

which can negatively affect cattle productivity (Bravo-Parra, 2021).

14

1.4 Results

1.4.1 First-stage results

We summarize the average partial effects’ estimates for our first-stage analysis in Table 1.3.

First, without additional controls, we find a statistically significant correlation between the share

of introduced pastures and (a) the distance to technology dissemination centers, (b) the number of

attacks on the population that occurred in the Municipality between 1988-2012, and (c) the number

of illegal checkpoints that were present at the Municipality (columns one and three). These are

robust to the addition of further factors (columns two and four). On average, and following an OLS

(QMLE) approach, a ten-percent increase in the distance between the farm and the closest research

and dissemination center is connected to a reduction of approximately 1.15 (1.23) percentage

points in the share of adoption of introduced pastures, which highlights the relevance of CIAT and

AGROSAVIA research centers in the cattle ranching system of Colombia. Also, on average, each

additional attack on the population registered in the 1988-2012 period is connected to a reduction

of pasture adoption of 4.28 (3.08) percentage points, while an additional illegal checkpoint relates

to an average decrease in adoption of 2.86 (2.57) percentage points. These findings go in line

with the literature on the relationship between land use and conflict: disruptions in access to

intensification technologies (i.e., introduced pastures) will lead ranchers to rely to a greater extent

on extensive traditional grazing, with the latter ultimately leading to increased rates of deforestation

(Bautista-Cespedes, Willemen, Castro-Nunez, & Groen, 2021).

We find no statistical differences attributable to educational status of the household head,

access to credit, a farmers’ association membership, or having a female-led household. Our

results suggest that household wealth, as captured in the household assets index, seems unrelated

to adoption. Conversely, technological endowment proxies reveal a relevant relationship with

adoption intensity. The average effect from an additional durable farming asset is connected to

an additional 2.1 (1.96) percentage points in the share of introduced pastures from each additional

durable farming asset, which is significant at the 1% level. Finally, we note that the (log) farm size

appears statistically significantly correlated with the adoption of pastures, with an increase of 10%

15

of the farm size related to a decrease of 0.17 (0.13) percentage points of adoption.

1.4.2 Effects of introduced pasture adoption on productivity

We observe a strong unconditional correlation between the adoption of introduced pastures and

the farm’s carrying capacity (Figure 1.2), which we further disaggregate in our main impact analysis

results in Table 1.4. Using only Department-level controls (column 1), this simple correlation

suggests that for each additional one percentage point increase in the adoption of introduced pastures,

the number of animals that can be sustained per hectare increases by 0.377%. Equivalently, due

to the linearity of the estimate, this translates into an increase of 37.7% in carrying capacity by

transitioning from native savannas to introduced pastures. Adding further controls in the OLS

specification (Column 2) reveals this correlation to be robust to the inclusion of further covariates.

After including farm-, farmer- and Municipality-level controls, our estimation suggests that the

transition from native savannas into introduced pastures implies an average increase of 33.68% in

livestock weight per hectare.

As mentioned in our methods section, we acknowledge the likely endogeneity of the adoption

of pastures and our measure of yield, so we extend our analysis with instrumental variable methods.

In columns three, four, and five, we report the 2SLS, two-step optimal IV, and control function (CF)

estimates for equation (1.1), respectively. All three estimates for the effect of introduced pastures

on carrying capacity are significantly larger than that of OLS–in line with our expectation of a

presence of downward bias–and statistically equivalent to each other. In both 2SLS and optimal IV,

the robust regression endogeneity tests (Wooldridge, 1995) suggest that there was an underlying

endogeneity problem in the case of OLS, which is further confirmed by the statistical significance

of the control function coefficient estimate. Resembling the findings of Xu (2021), we see that

despite having similar estimates and precision levels, the 2SLS approach would suggest a potential

problem of weak instruments, unlike the two-step optimal IV approach, which renders an effective

F statistic well above the thumb-rule of 23 from Montiel and Pflueger (2013).

We will focus on estimates using the optimal IV and control function approaches (the latter in

parentheses). Our estimates suggest that carrying cacpaccity strongly responds to the adoption of

16

introduced pastures. On average, for each additional 1% point of adoption, cattle yields increase by

1.34% (1.44%). Per the linearity of the model, this implies that transitioning from native savannas

to introduced pastures can more than double productivity, i.e., it would increase carrying capacity

by 134 (144) percent on average. While the sizeable difference between OLS and optimal IV (or

CF) estimates may be concerning at face value, note that the latter estimates are still below the

expected increase in carrying capacity suggested by the agronomic literature, which is in the order

of 200 percent and above (Fisher et al., 1992; Vera & Hoyos Garcés, 2019). Therefore, these initial

results are of high relevance on two contrasting points. First, they highlight that a correlational

analysis of the effect of introduced pastures on carrying capacities can significantly underestimate

their impact and thus their potential for intensification of cattle systems in lowlands.14 Second, our

analysis also indicates that the on-farm impact of introduced pastures is still below to that suggested

by agronomic trials. The latter goes in line with the literature which suggests that while these trials

are rigorous, their external validity remains a concern (Laajaj et al., 2020). Our findings further

emphasize the need for farm-level analysis impacts of agricultural technologies to inform policy

decisions.

Finally, we detect a significant difference in productivity between male and female-led reported

households. On average, when households report a female head, this is linked to a decrease in

carrying capacity by 20.89% (23.23%), likely reflecting the limitations in access to production

resources and implications of structural gender discrimination (Gumucio, Mora-Benard, Twyman,

& Hernández-Ceballos, 2016). In addition, we find that being member of a farmers’ association is

related to an increase of 14.41 (15.16) percent in carrying capacities, pointing to the relevance of

ranchers’ networking in improving the productivity of cattle systems in lowlands. The magnitude

and significant effect reveal that more institutional work should be put into the national research

system to tackle these disparities. From the Municipality-level variation, we detect a statistically

14It is also likely that the marginal adopters identified by the instrument (i.e., those whose adoption decisions are
influenced by exposure to information through proximity to dissemination centers or conflict disruption) may experience
much larger productivity gains than the average adopter–i.e., as mentioned in Footnote 3, we may be identifying a
LATE larger than the ATE. This could occur if these marginal adopters previously faced significant knowledge or
resource constraints that limited their ability to transition to improved pasture systems, making the returns to adoption
particularly high for them.

17

significant effect from the average daily precipitation registered during the last agricultural season,

although only for the optimal IV specification. We find that an increase in precipitation of 1 mm/day

is correlated with an increase in livestock weight per hectare of 2.8%.

1.4.2.1 Testing for identification assumptions and robustness

We begin by providing correlational evidence in support for our exclusion restriction assump-

tions. First, we find a statistically insignificant correlation between the farm’s distance to the closest

R&D dissemination center and the distance to the closest input market (Table 1A.1). Additionally,

we evaluate whether the proximity to this infrastructure affects adoption by adding it as a regressor

in our first-stage specification (Table 1A.2). We find that this new distance variable holds no

significant correlation with our adoption measure, while our main inference remains unchanged.

Furthermore, we run Poisson regressions where the dependent variables were: (a) the number

of complementary practices adopted on the farm, and (b) a farm assets index, while including

both distance to the closest R&D center and distance to the closest input market as covariates

(Table 1A.3). We find that proximity to R&D centers has no significant effect on the number of

complementary practices adopted, but proximity to input markets significantly affects farm asset

accumulation.15 Therefore, these suggest that R&D centers do not systematically drive broader

agricultural intensification beyond their specific role in knowledge transfer for improved pasture

adoption.16

Meanwhile, in the case of our violence variables we evaluate whether these variables are

correlated with key socioeconomic variables such as household wealth or education levels (Table

1A.4), and do not find any statistically significant association. The latter further supports our

argument that their primary channel of influence is through information availability rather than

direct effects on productivity. This is expected, as our selection of conflict measures is deliberate,

15We further include specifications that add farmers’ self-reported distance to the closest extension services in
Tables REF, finding that our inference regarding the distance to R&D dissemination centers does not change. However,
this additional testing is only available for a subset of the sample where the additional covariate is available.

16Another potential robustness check would consider a placebo test using the distances to non-cattle-focused
CGIAR centers in the country. Yet, to our knowledge, there are no alternative CGIAR institutions in Colombia,
hence implementing this test within our dataset is unfeasible. Future research could consider using a geolocation of
non-cattle-focused demonstration plots established by CIAT to define a comparable placebo.

18

focusing on likely disruptions of civilian life as a collateral consequence, rather than violence that

directly alters land tenure or governance structures in the long run.17

Now, in terms of the estimation strategy, although the optimal IV approach relies on a first-

stage analysis that warrants a non-linear estimation, some concerns may arise as to whether one is

incurring in identifying an effect off a non-linearity alone. We provide additional evidence that, in

our analysis, instrument strength is not caused by identification from a non-linearity (Table 1A.5),

showing that in the absence of exogenous sources of variation, the problem of weak instrumentation

would reemerge even with a non-linear first stage. Therefore, as in Xu (2021), we confirm that

it is the combination of exogenous variation and a proper conditional mean that circumvents the

problem of artificially weak instrumentation.

Furthermore, we evaluate the sensitivity of our estimates considering three separate analyses

using each of our main sources of exogenous variation (i.e., distance to R&D, attacks on populations,

and illegal checkpoints), and summarize the results in Tables 1A.6 and 1A.7. Our inference remains

unchanged in such a case: the partial average effects from first stage analyses are within a standard

error from those of our main specification, and so is the case for our estimates of impact of

introduced pastures on carrying capacity. Moreover, using each source of exogenous variation

separately under our optimal-IV approach, we reject a null hypothesis of weak instruments across

all specifications. We further test whether using a geodesic distance to R&D, which allows us to

use the full sample, has any significant effect on our main inference (see Tables 1A.8, 1A.9, and

1A.10), but find qualitatively equivalent results.

A potential concern with using a log-linear model lies in the non-negative nature of the carrying

capacity measure, atop a potential endogeneity problem. To that purpose, we present a brief

summary of the case of an exponential model estimated by Poisson QMLE in Table 1A.11, whose

estimates are directly comparable to those from Table 1.4. Consistent with our findings from OLS

and IV estimation under a log-linear model, we encounter significant differences between Poisson

17Future studies in this context could follow an alternative and perhaps more robust testing, having dependent
variables related to farm infrastructure (which our data, unfortunately, does not possess), to assess whether other
sources of violence (e.g., massacres) are likely correlated with them but not our preferred measures of conflict.

19

and IV-Poisson estimates (either relying on linear or non-linear estimates in the first step of the

analysis). Without acknowledging the likely endogenous nature of the context, estimates would

point to an increase in carrying capacity of 21.9% by transitioning from native savanna to introduced

pasture, yet this figure goes up to 127.8 (122.8) percent under an IV (optimal IV) approach. Finally,

we also test whether our impact estimates are sensitive to excluding non-adopters (Table 1A.12),

finding that all estimates are statistically equivalent to those of our baseline estimates, so the effect

does not depend solely on the intensive margin.

1.4.2.2 Land use implications

Our estimates have important implications for land use in lowland ranching and its relationship

with the adoption of introduced pastures. Leaving aside equilibrium effects, we can calculate a

counterfactual measure of productivity for users of introduced pastures had they not adopted these

varieties and then project how much more land would be required to keep the same level of output

without using introduced pastures. Following similar approximations done in Larochelle et al.

(2015) and Larochelle and Alwang (2022) for the case of improved bean varieties and welfare

measures in Sub-Saharan Africa, we use our optimal IV estimate18 of 𝜏 from Equation (1.1) to

derive a counterfactual yield measure (again in kg/ha), such that

𝑦𝐶𝐹 = 𝑦 × exp [− ˆ𝜏𝑆] ,

(1.3)

where 𝑦 and 𝑆 are the observed levels of productivity and pasture adoption. The average counter-

factual additional need for land is calculated straightforwardly as

¯𝑦𝐶𝐹 − ¯𝑦
¯𝑦
The average observed carrying capacity (yield) among adopters of pastures is 545.28 kg/ha,

(1.4)

𝜓 =

.

while we predict an average counterfactual yield of 192.25 kg/ha had they not adopted these pastures.

To put these figures in perspective, in the absence of introduced pastures, land use in pastures would

have had to increase by 183% to reach the same level of production, which would have potentially

led to a dramatic increase in deforestation. These findings further highlight how introduced pasture

varieties have been key to increasing productive efficiency (Herrero et al., 2013; Jank et al., 2014)

18More specifically, we use the estimate of 𝜏 from Poisson IV with optimal instruments mentioned above, since

using directly the estimates from linear regression would be a violation of Jensen’s inequality.

20

and are a crucial element to achieving sustainable intensification (González-Quintero et al., 2021;

Mazzetto et al., 2015).

1.4.3 Heterogeneity by pasture management

Following the literature, we check for differences in the effect of productivity across different

kinds of pasture management, fitting the model in equation (1.2) by OLS and CF. We present

the average partial effects of the share of introduced pastures conditional on different pasture

management schemes in Figure 1.3. Under an OLS framework, our estimates would suggest that

without pasture management, there are no effects from the adoption of introduced pastures on

cattle intensity, with statistically significant effects only when pasture management practice occurs.

Nevertheless, after accounting for the endogeneity in pasture adoption via a control function

approach, our analysis suggests that there are significant gains in yields regardless of whether

pasture management practices take place, but that they do increase the impact of pastures19. In

terms of point estimates, a 1% point increase in the use of introduced pastures alone will increase

productivity by 1.26%. However, when a farm’s pastures are managed by making weed control only,

a 1 percentage point increase in the adoption of introduced pastures implies an increase of 1.54%

in cattle intensity. The figure is 1.87% for the case of only fertilization and 1.53% whenever both

practices occur together. Nevertheless, caution should be advised when arguing different levels of

impact based on our control function estimates; note that we cannot reject the hypothesis that these

heterogeneous effects are statistically equivalent to one another, evidenced by their overlapping

confidence intervals.20

Overall, our results hold even when considering different settings of pasture management

observed at the farm level.

It is worth highlighting that our analysis builds upon the working

assumption that practices like fertilization and weed control are taken as given. Nevertheless,

among pasture adopters, these practices may reflect an underlying decision-making process by

19The associated point estimates are summarized in Table 1A.13.
20Having a point estimate of combined practices being lower than that of only using fertilization has a likely
agronomic driver. Specifically, common weed control practices usually remove some plants that serve as an important
soil coverage, which, combined with fertilization, can lead to accelerated soil degradation and thus limit the average
gains from using introduced pastures.

21

the farmer that could benefit from economic modeling, as suggested by previous studies on other

crops21 (Kassie et al., 2013; Martinez et al., 2021; Teklewold et al., 2013). While addressing the

dual endogeneity of introduced pasture adoption and pasture management practices is beyond the

scope of this paper, it will likely be of the utmost crucial for future studies on the adoption of

sustainable livestock systems, which typically integrate various practices.

1.4.4

Introduced pasture adoption and revenues per hectare

We plot the relationship of revenues per hectare with the share of introduced pastures in Figure

1.4, revealing that without any controls there is a relevant correlation, although slightly noisier as

noted by the wider confidence intervals than those in the case of carrying capacity. However, just

as in the case of yields, we must consider a potential problem of endogeneity from self-selection if

trying to estimate an impact. We replicate this analysis with revenues as the response variable in

Table 1.5. Under an OLS framework, we detect a robust correlation between revenues per ha and

introduced pastures, even after considering additional controls, such that each additional 1% point

in adoption implies an average 0.31% increase in revenues per ha. However, we find that considering

potential endogeneity significantly alters our inference. Under an optimal IV approach, the average

response goes up to 2.05%.

In transition terms, switching from native savannas to introduced

pastures has the potential to increase average revenue per hectare by roughly 205%. Unfortunately,

due to a systematic lack of recordkeeping of costs among surveyed households, we have no available

information on costs to provide an analysis based on profit rather than revenues. A back-of-the-

envelope calculation based on data from Enciso et al. (2021) would suggest that going from native

grasslands to improved pastures with standard management would increase costs per hectare by

130% in the year when pastures are established, and by 45% in subsequent years. Therefore,

in reference to a potential increase in revenues in the order of 205%, profits per hectare could

potentially increase by a significant margin (up to 75%). However, given the lack of proper cost

data in our sample, a specific figure cannot be ascertained from our statistical analysis.

21For example, an ordered regression analysis on the number of pastures observed in the farm (Table 1A.14) suggests
that factors like membership in a farmers’ association, education, and access to credit are correlated with the adoption
of these complementary practices, resembling the results of Martinez et al. (2021) for complementary technologies in
rice farming.

22

We can derive further relevant insights when exploring possible heterogeneities across various

groups of complementary practices (Figure 1.5) for the impact of introduced pastures on rev-

enues.22 For instance, note that under an OLS specification, transitioning into an improved grazing

system would only lead to a statistically significant gain in revenues among farmers that only use

weed controls. Conversely, there appear to be significant gains in revenue across all groups of

complementary-practice adoption from the use of introduced pastures, even in the case of no other

complements. However, we are cautious of any back of the envelope calculations for each case,

given both the unavailability of cost information at the farm-level or any representative average

figures per each specific practice. Therefore, yet again, these projected increases in revenue do not

necessarily imply a direct increase in profit.

1.5 Discussion

The sustainability of ranching based on foraging systems remains at the center of debates

surrounding their land use and productive efficiency. Yet, impacts from using improved systems

based on introduced pastures remained largely unexplored (Baltenweck et al., 2020). Our research

on the adoption of introduced pastures in ranching systems examined whether using these materials

increases farm-level productivity at the farm level, based on a case study for two macro-regions

of Colombia: the Caribbean and Eastern Plains – Amazon Piedmont. Our results suggest that the

adoption of introduced pastures is strongly correlated with the proximity to technological research

and dissemination centers, other technological endowments, education, farm size, and historical

events of violence at the Municipality level. Also, our findings indicate a statistically significant

gain in carrying capacity from using introduced pastures. We find a gain in productivity when

introduced pastures are implemented, although gains in productivity are larger when weed control or

fertilization regimes take place. Our estimates of potential gains in productivity (carrying capacity)

from adoption are below those suggested by agronomic trials (Fisher et al., 1992; Hyman et al.,

2022). Nonetheless, they coincide with the literature that points to proper pasture management as

a necessary action to guarantee the largest average increases in productivity (González-Quintero et

22The associated point estimates are summarized in Table 1A.15.

23

al., 2021; Pérez-López & Afanador-Tellez, 2017). In addition, our findings point to a significant

reduction in ranching’s land demand, which is attributable to improved systems based on introduced

pastures.

Our findings are crucial for understanding current conditions and shaping the future of ranching

technology research in the Latin American lowlands, particularly in Colombia. First, they highlight

the historical role of the National and International Agricultural Research Systems in enhancing

ranching productivity through the selection and introduction of higher-nutrient pasture varieties in

the seed system. Second, they reveal how past violence and conflict have constrained the agricultural

sector’s potential, with our study showing that each additional violent attack on the population

and each additional illegal checkpoint experienced at the Municipality level decreased the share

of introduced pastures by an average of 3.08 and 2.57 percentage points, respectively. Third, the

significant productivity increase from adopting introduced pastures suggests that strategies aimed at

increasing perceived prices through product differentiation, by signaling the use of more sustainable

production (Charry, Narjes, Enciso, Peters, & Burkart, 2019), could incentivize the adoption of

low-emission grazing systems with introduced pastures. Fourth, the notable productivity gains from

well-managed pastures imply that farmers’ benefits could be coupled with social co-benefits, such

as ground carbon storage (Fisher et al., 1994; Hyman et al., 2022). This underscores the desirability

of programs aiming at intensification (Arango et al., 2020; Tapasco et al., 2019), which farmers

have shown a willingness to pursue with better technical support systems (Carriazo, Labarta, &

Escobedo, 2020). Finally, additional productivity per hectare of land suggests that pasture-based

intensification programs could effectively reduce the deforestation pressure, contributing positively

to sustainability goals.

Production solely under introduced pastures is not a silver bullet for achieving optimal pro-

ductivity levels, particularly in seasonally dry climates, where maintaining a share of grasses as

semi-natural pastures or savannas can be more strategic (Ospina et al., 2012). However, given

their widespread adoption in Colombia and Latin America (ISPC, 2018), introduced pastures are

arguably more appealing as a primary public policy target vis à vis other intensive schemes like

24

silvopastoral systems. Although the latter may offer the most desirable traits in terms of circularity

and climate mitigation (Murgueitio, Chará, Barahona, Cuartas, & Naranjo, 2014; Parodi et al.,

2022; Silva-Parra, Trujillo-González, & Brevik, 2021), their adoption remains dramatically low,

with roughly 55,000 ha in LAC after years of regional promotion (Chará et al., 2019). In contrast,

introduced pastures have not only achieved broader adoption, but have also proven successful for

intensification and sustainable, climate-adapted production in Latin America (Bogaerts et al., 2017;

Mazzetto et al., 2015), and have been historically (and within our sample) used by farmers of all

scales (Campuzano et al., 2022; Van Ausdal, 2009).

Notwithstanding these potential gains from the use of introduced pastures, Latin America still

needs larger efforts at the national scale toward sustainable intensification in ranching. In our case

of Colombia, the regions under analysis are the ones putting the largest pressure on the agricultural

frontier and forests, particularly in the Amazon Piedmont and Eastern Plains. Nevertheless, associ-

ation membership and access to credit services among producers remain severely low, highlighting

the unmet demands of knowledge (Carriazo et al., 2020) and credit (Mejía et al., 2021) in this popu-

lation. Although the so-called burger connection23 has not yet impacted deforestation in Colombia

(Dávalos, Holmes, Rodríguez, & Armenteras, 2014), the anticipated entry of Colombian beef into

the US beef market (Vargas-Cuéllar, 2022) could trigger a significant demand shock. If unmanaged,

this could lead to increased land pressures (Jank et al., 2014) and substantial deforestation, similar

to what has occurred in Brazil over the past five years (McCoy & Ledur, 2022). This surge in

demand might also boost the incentives for illegal production, which is already harming Colombian

forests (Murillo-Sandoval et al., 2023).

1.6 Conclusion

Using cross-sectional, geographical, weather, and historical data, we investigate the factors

driving the adoption of introduced pastures in Colombian lowland ranching. By leveraging historical

information on violence and conflict events and a measure of proximity to centers of technological

R&D and promotion, we implement an instrumental variable approach to measure the impact of

23This relates to the connection between national-or regional-level demands for beef and observed rates of defor-

estation (McAlpine, Etter, Fearnside, Seabrook, & Laurance, 2009).

25

introduced pasture adoption on land-carrying capacity (kg of livestock per ha) and livestock-related

revenues.

Our results reveal that proximity to research dissemination significantly influences pasture

adoption, underscoring the vital role that National and International Agricultural Research Systems

play in reaching intensification goals. Additionally, historical violence and alterations to transit

and transportation, which disrupt supply chains, also affect the observed levels of adoption. More

importantly, we provide impact estimates suggesting that transitioning from native savannas to

introduced pastures can potantially increase the weight-carrying capacity of ranching by more than

a 100%, with even larger gains when complementary practices like weed control and fertilization

are in place. Likewise, livestock-related revenues are positively and significantly increased with

adoption. While we find that correlational approaches are likely underestimating the true impact

from the use of these pastures, we still detect productivity gains below those suggested by agronomic

trials, revealing that the latter are unlikely to materialize under real farm conditions.

While we provide important insights on how introduced pastures affect productivity and factors

associated with welfare, our findings could be expanded in at least two areas. First, future studies

should focus on determining a representative on-farm measure of GHG emissions and its rela-

tionship with grazing systems based on introduced pastures. Given the low levels of spontaneous

adoption of alternative intensification methods like silvopastoral systems, enhancing and intensify-

ing grazing systems with introduced pastures could significantly improve the sector’s sustainability

and help reduce the pressure on the agricultural frontier driven by the increasing global demand for

beef. Additionally, detailed information about animal weight gain and associated costs is needed

to fully assess the impact on profitability.

Second, our analysis is limited by the cross-sectional nature of the main dataset, so future

studies should focus on longitudinal data. Incorporating time variability would enable the analysis

of spillover effects and provide clearer identification strategies of the impacts of adopting intro-

duced pastures. In addition, having access to more recent, representative information is crucial

for understanding how the violence de-escalation and peacebuilding efforts influence ranching

26

intensification. Our main dataset was collected in 2016-2017, shortly after the Colombian Peace

Accords, which may have introduced a structural change in value chains, seed systems, and the

decision-making processes of agricultural households. Future research should utilize more recent

data to determine if the observed impacts are not merely an artifact of this structural change.

27

Figures and Tables

Figure 1.1 Location of surveyed farms, CIAT, and AGROSAVIA Research Centers with relevant
dissemination of cattle ranching technologies.

28

Figure 1.2 Relationship between the share of introduced pastures and cattle yield. Linear
prediction and 95% confidence intervals.

29

Figure 1.3 Ordinary least squares (OLS) and Control function (CF) estimates of heterogeneity of
the average effect of adopting an additional 1% of introduced pastures on cattle yield by pasture
management practices used in the farm (95% CI).

30

Figure 1.4 Relationship between the share of introduced pastures and livestock-related revenue.
Linear prediction and 95% confidence intervals.

31

Figure 1.5 Ordinary least squares (OLS) and Control function (CF) estimates of heterogeneity of
the average effect of adopting an additional 1% of introduced pastures on livestock-related revenue
by pasture management practices used in the farm (95% CI).

32

Table 1.1 Descriptive statistics of the sample and Municipality-level variables.

Farm-Level Variables (N=1,021)
Bovine kilograms per hectare
Revenue per hectare
Share of pastures with introduced varieties

Distance to closest dissemination center
Distance to closest input market
Grass management
No management
Only weed control
Only fertilization
Fertilization and weed control
Associativity and credit
Member of a farmers’ association
Access to credit
Assets Indices
Household assets index (Total)
Farming assets index (Total)
Other controls
Size of the farm
Female household head
Education level of HH: No education
Education level of HH: Primary
Education level of HH: Secondary
Education level of HH: Tertiary
Municipality-level variables (N=45)
Attacks on population (Total, 1988-2012)
Illegal checkpoints (Total, 2002-2016)
Precipitation (avg. daily, 2016)
Maximum temperature (avg. daily, 2016)

Unit
Kg/ha
COP(1,000)/ha
Avg./farm(a)
Pooled Avg.(b)
Km
Km

Mean
536.98
684.4
0.78
0.66
226.2
17.9

S.D.
478.80
1,531.7
0.32

Min
11.67
0.4
0

Max
4,648.89
25,864.8
1

139.3
16.1

6.2
0.9

551.5
89.9

1 = Yes
1 = Yes
1 = Yes
1 = Yes

1 = Yes
1 = Yes

Count
Count

Ha
1 = Yes
1 = Yes
1 = Yes
1 = Yes
1 = Yes

Count
Count
mm/day
Celsius

0.39
0.46
0.07
0.08

0.22
0.22

5.97
3.81

200.91
0.07
0.08
0.71
0.15
0.06

1.09
1.64
5.66
31.78

0.49
0.50
0.26
0.27

0.41
0.42

2.40
2.02

467.08
0.26
0.26
0.45
0.36
0.24

0
0
0
0

0
0

0
0

1.5
0
0
0
0
0

1.55
3.09
3.35
3.36

0
0
1.43
18.37

1
1
1
1

1
1

11
9

7,000
1
1
1
1
1

5
17
16
35.37

Note — (a) This reports the average per farm share of pastures with introduced varieties. (b) Pooled average sums the total in-sample
area with introduced pastures and divides it by the total in-sample area in all pastures.

33

Table 1.2 Mean values of pasture adoption, complementary practices, and outcomes of interest by quintiles of farm size.

Variables
Observations
Bovine kilograms per hectare
Revenue per hectare
Share of pastures with introduced varieties
Does fertilization
Does weed control

Unit
𝑁
Kg/ha
COP(1,000)/ha
Avg./farm(a)
1 = Yes
1 = Yes

Quintile 1 Quintile 2 Quintile 3 Quintile 4 Quintile 5
215
438.24
658.92
0.78
0.11
0.69

208
909.95
1,189.62
0.79
0.20
0.34

198
342.89
528.24
0.73
0.18
0.62

206
561.02
571.07
0.77
0.13
0.47

194
423.48
480.25
0.84
0.15
0.58

Note — (a) This reports the average per farm share of pastures with introduced varieties.

34

Table 1.3 First-stage analysis of factors associated with the share of introduced pastures in the
farm.

Variables

(1)

(2)

(3)

(4)

Log-Distance to closest dissemination center(a)

Attacks on population(a)

Illegal checkpoints(a)

Log-Area of the farm(a)

Education of HH (Primary)

Education of HH (Secondary)

Education of HH (Tertiary)

Female HH (1 = Yes)

Access to Credit (1 = Yes)

Household Assets Index

Farm Assets Index

-0.1215***
(0.039)
-0.0469***
(0.018)
-0.0309***
(0.011)

-0.1320***
(0.042)
-0.0355**
(0.016)
-0.0279***
(0.010)

-0.1157***
(0.038)
-0.0428**
(0.018)
-0.0286**
(0.012)
-0.0179**
(0.009)
0.0571
(0.041)
0.0383
(0.049)
-0.0079
(0.060)
-0.0180
(0.038)
0.0083
(0.024)
0.0061
(0.005)
0.0210***
(0.007)

-0.1234***
(0.041)
-0.0308**
(0.015)
-0.0257**
(0.010)
-0.0139*
(0.008)
0.0557
(0.036)
0.0317
(0.039)
0.0011
(0.052)
-0.0193
(0.036)
0.0044
(0.024)
0.0051
(0.005)
0.0196***
(0.006)

Observations
Method
Dept Controls
Cluster (village) standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

994
OLS
Yes

994
OLS
Yes

994
QMLE
Yes

994
QMLE
Yes

Notes — Response variable is the share of pastures with introduced varieties. (a) Regression included the variable and its square,
and the table reports the full average partial effect.

35

Table 1.4 Effect of introduced pasture adoption on cattle yield.

Variables

(1)

(2)

(3)

(4)

(5)

Share of Improved Pastures

Log-Area of the farm(a)

Education of HH (Primary)

Education of HH (Secondary)

Education of HH (Tertiary)

Female HH (1 = Yes)

Member of Farmers’ Assoc. (1 = Yes)

Access to Credit (1 = Yes)

Average precip. in the past year

Avg. max. temp. in the past year

Control Function (𝑣)

0.3770***
(0.091)

0.3368***
(0.072)
-0.2782***
(0.020)
0.0881
(0.074)
0.1217
(0.092)
0.1292
(0.116)
-0.2507***
(0.078)
0.1295**
(0.052)
-0.0765
(0.047)
0.0186
(0.012)
-0.0157
(0.013)

1.3701***
(0.261)
-0.2721***
(0.020)
0.0004
(0.091)
0.0383
(0.109)
0.1042
(0.128)
-0.2079**
(0.088)
0.1444**
(0.058)
-0.0839
(0.054)
0.0282**
(0.013)
0.0005
(0.016)

1.3470***
(0.247)
-0.2722***
(0.020)
0.0024
(0.091)
0.0402
(0.108)
0.1048
(0.127)
-0.2089**
(0.088)
0.1441**
(0.057)
-0.0837
(0.054)
0.0280**
(0.013)
0.0001
(0.015)

994
OLS
Yes

994
OLS
Yes

Observations
Method
Dept Controls
Effective F
Montiel-Pflueger: 5% Bias Critical value
Montiel-Pflueger: 10% Bias Critical value
Robust reg. endogeneity Test (p-value)
Cluster (village) standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

994
2SLS
Yes
6.664
24.38
14.15
0.000

994
Opt-IV
Yes
53.99
37.42
23.11
0.000

1.4456***
(0.303)
-0.2698***
(0.021)
-0.0119
(0.098)
0.0334
(0.120)
0.0839
(0.132)
-0.2312**
(0.091)
0.1516**
(0.059)
-0.0788
(0.055)
0.0123
(0.012)
-0.0187
(0.015)
-1.205***
(0.313)

994
CF
Yes

Notes — Response variable is (log) Kilograms of livestock per hectare. Two-step control function (CF) regression estimates report
cluster bootstrapped standard errors. (a) Regression included the variable and its square, and the table reports the full average partial
effect.

36

Table 1.5 Effect of introduced pasture adoption on livestock-related revenues per hectare.

Variables

(1)

(2)

(3)

(4)

(5)

Share of Improved Pastures

Log-Area of the farm(a)

Education of HH (Primary)

Education of HH (Secondary)

Education of HH (Tertiary)

Female HH (1 = Yes)

Member of Farmers’ Assoc. (1 = Yes)

Access to Credit (1 = Yes)

Average precipitation in the past year

Avg. max. temp. in the past year

Control Function (𝑣)

0.4376***
(0.165)

0.3081**
(0.135)
-0.6613***
(0.046)
0.2409
(0.170)
0.3778*
(0.229)
0.4429*
(0.240)
-0.1992
(0.161)
0.1731
(0.121)
-0.0158
(0.108)
0.0460
(0.033)
-0.0147
(0.034)

1.9849***
(0.568)
-0.6436***
(0.050)
0.0312
(0.207)
0.1639
(0.264)
0.3553
(0.291)
-0.0999
(0.175)
0.1729
(0.130)
-0.0410
(0.113)
0.0696*
(0.036)
0.0136
(0.034)

2.0551***
(0.563)
-0.6428***
(0.050)
0.0224
(0.208)
0.1550
(0.265)
0.3516
(0.294)
-0.0957
(0.178)
0.1729
(0.131)
-0.0421
(0.114)
0.0706**
(0.036)
0.0148
(0.034)

786
OLS
Yes

786
OLS
Yes

Observations
Method
Dept Controls
Effective F Statistic
Montiel-Pflueger: 5% Bias Critical value
Montiel-Pflueger: 10% Bias Critical value
Robust reg. endogeneity test (p-value)
Cluster (village) standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

786
2SLS
Yes
6.208
24.71
14.33
0.0010

786
Opt-IV
Yes
43.81
37.42
23.11
0.0002

2.2676***
(0.705)
-0.6446***
(0.052)
0.0624
(0.222)
0.2102
(0.282)
0.3528
(0.300)
-0.1734
(0.185)
0.2216*
(0.134)
-0.0325
(0.117)
0.0315
(0.034)
-0.0255
(0.038)
-2.1363***
(0.726)

786
CF
Yes

Notes — Response variable is (log) Revenue from livestock-related sales per hectare. Two-step control function (CF) regression
estimates report cluster bootstrapped standard errors. (a) Regression included the variable and its square, and the table reports the
full average partial effect.

37

REFERENCES

Acevedo, K. M., & Bornacelly, I. D. (2014, July). Panel Municipal del CEDE (Documentos CEDE
(Retrieved from

No. 12223). Universidad de los Andes, Facultad de Economía, CEDE.
https://ideas.repec.org/p/col/000089/012223.html)

Angrist, J., & Pischke, J. S. (2008). Mostly Harmless Econometrics: An Empiricist’s Companion.

Princeton University Press.

Arango, J., Ruden, A., Martinez-Baron, D., Loboguerrero, A. M., Berndt, A., Chacón, M., . . .
Chirinda, N. (2020). Ambition meets reality: achieving GHG emission reduction targets in
the livestock sector of Latin America. Frontiers in Sustainable Food Systems, 4, 65. doi:
10.3389/fsufs.2020.00065

Assunção, J., Bragança, A., & Hemsley, P.

(2019). Geographic heterogeneity and technology
adoption: evidence from brazil. Land Economics, 95(4), 599–616. doi: 10.3368/le.95.4.599

Baltenweck, I., Cherney, D., Duncan, A., Eldermire, E., Lwoga, E. T., Labarta, R., . . . Teufel,
N. (2020). A scoping review of feed interventions and livelihoods of small-scale livestock
keepers. Nature Plants, 6(10), 1242–1249. doi: 10.1038/s41477-020-00786-w

Bardhan, P. K., & Udry, C. (1999). Development Microeconomics. Oxford: Oxford University

Press.

Bautista-Cespedes, O. V., Willemen, L., Castro-Nunez, A., & Groen, T. A. (2021). The effects of
armed conflict on forest cover changes across temporal and spatial scales in the Colombian
Amazon. Regional Environmental Change, 21(3), 70. doi: 10.1007/s10113-021-01770-6

Bogaerts, M., Cirhigiri, L., Robinson, I., Rodkin, M., Hajjar, R., Costa Junior, C., & Newton,
P. (2017). Climate change mitigation through intensified pasture management: Estimating
greenhouse gas emissions on cattle farms in the Brazilian Amazon. Journal of Cleaner
Production, 162, 1539–1550. doi: 10.1016/j.jclepro.2017.06.130

Bravo-Parra, A. (2021). Cadenas sostenibles ante un clima cambiante: La ganadería en Colombia.

Bonn, Germany: Deutsche Gesellschaft für Internationale Zusammenarbeit (GIZ).

Burkart, A. (1975). Evolution of grasses and grasslands in South America. TAXON, 24(1), 53–66.

doi: 10.2307/1219001

Campuzano, L., Triana-Angel, N., & Burkart, S. (2022). Cattle ranching in Colombia: a monolithic
industry? Historia Ambiental Latinoamericana y Caribeña (HALAC) Revista de la Solcha,
12(2), 81–106. doi: 10.32991/2237-2717.2022v12i2.p81-106

Carriazo, F., Labarta, R., & Escobedo, F. J. (2020). Incentivizing sustainable rangeland practices
and policies in Colombia’s Orinoco region. Land Use Policy, 95, 104203. doi: 10.1016/

38

j.landusepol.2019.104203

Centro Nacional de Memoria Histórica. (2015). Bases de datos - Basta ya! Bogotá, Colombia.

Charry, A., Narjes, M., Enciso, K., Peters, M., & Burkart, S. (2019). Sustainable intensification
of beef production in Colombia—Chances for product differentiation and price premiums.
Agricultural and Food Economics, 7(1), 22. doi: 10.1186/s40100-019-0143-7

Chará, J., Rivera, J., Barahona, R., Murgueitio, E., Calle, Z., & Giraldo, C. (2019). Intensive
silvopastoral systems with Leucaena leucocephala in Latin America. Tropical Grasslands-
Forrajes Tropicales, 7(4), 259–266. doi: 10.17138/tgft(7)259-266

Congio, G. F. D. S., Bannink, A., Mayorga Mogollón, O. L., Hristov, A. N., Jaurena, G., Gonda,
H., . . . Astigarraga, L. (2021). Enteric methane mitigation strategies for ruminant livestock
systems in the Latin America and Caribbean region: A meta-analysis. Journal of Cleaner
Production, 312, 127693. doi: 10.1016/j.jclepro.2021.127693

Dale, I. J. (1981). Parthenium weed in the Americas: A report on the ecology of Parthenium
hysterophorus in South, Central, and North America. Australian Weeds, 1(1), 8–14.

Dávalos, L. M., Holmes, J. S., Rodríguez, N., & Armenteras, D. (2014). Demand for beef is
unrelated to pasture expansion in northwestern Amazonia. Biological Conservation, 170,
64–73. doi: 10.1016/j.biocon.2013.12.018

Enciso, K., Charry, A., Rincón, A., & Burkart, S. (2021). Ex-ante evaluation of economic impacts of
adopting improved forages in the colombian orinoquía. Frontiers in Environmental Science,
9, 673481. doi: 10.3389/fenvs.2021.673481

Enciso, K., Triana, N., Diaz, M., & Burkart, S.

Informe preliminar: Limitantes y
oportunidades del proceso de adopción y difusión de tecnologías forrajeras en Colombia.
Cali, Colombia:
(Retrieved from
International Center for Tropical Agriculture (CIAT).
https://hdl.handle.net/10568/106848)

(2019).

Enciso, K., Triana, N., Díaz, M., & Burkart, S. (2022). On (Dis)connections and transformations:
the role of the agricultural innovation system in the adoption of improved forages in colombia.
Frontiers in Sustainable Food Systems, 5, 741057. doi: 10.3389/fsufs.2021.741057

Faminow, M. D. (1997). The Brazilian cattle sector: Status, prospects and controversies. Canadian
Journal of Agricultural Economics/Revue canadienne d’agroeconomie, 45(3), 179–199.

Fisher, M. J., Lascano, C. E., Vera, R. R., & Rippstein, G. (1992). Integrating the native savanna
resource with improved pastures. In CIAT (Ed.), Pastures for the Tropical Lowlands: CIAT’s
Contribution. Cali Colombia: International Center for Tropical Agriculture (CIAT).

Fisher, M. J., Rao, I. M., Ayarza, M. A., Lascano, C. E., Sanz, J. I., Thomas, R. J., & Vera, R. R.

39

(1994). Carbon storage by introduced deep-rooted grasses in the South American savannas.
Nature, 371(6494), 236–238. doi: 10.1038/371236a0

Fuglie, K., Peters, M., & Burkart, S. (2021). The extent and economic significance of cultivated
forage crops in developing countries. Frontiers in Sustainable Food Systems, 5, 712136. doi:
10.3389/fsufs.2021.712136

Gerssen-Gondelach, S. J., Lauwerijssen, R. B., Havlík, P., Herrero, M., Valin, H., Faaij, A. P., &
Wicke, B. (2017). Intensification pathways for beef and dairy cattle production systems:
Impacts on GHG emissions, land occupation and land use change. Agriculture, Ecosystems
& Environment, 240, 135–147. doi: 10.1016/j.agee.2017.02.012

González-Quintero, R., Bolívar-Vergara, D. M., Chirinda, N., Arango, J., Pantevez, H., Barahona-
Rosales, R., & Sánchez-Pinzón, M. S.
(2021). Environmental impact of primary beef
production chain in Colombia: Carbon footprint, non-renewable energy and land use using
Life Cycle Assessment. Science of The Total Environment, 773, 145573. doi: 10.1016/
j.scitotenv.2021.145573

Grupo de Memoria Histórica. (2013). ¡Basta ya! Colombia: Memorias de guerra y dignidad.

Bogotá, Colombia: Imprenta Nacional.

Gumucio, T., Mora-Benard, M., Twyman, J., & Hernández-Ceballos, M. (2016). Género en la
ganadería: Consideraciones iniciales para la incorporación de una perspectiva de género
en la investigación de la ganadería en Colombia y Costa Rica. Cali, Colombia: CCAFS
Working Paper No. 159. Retrieved from https://hdl.handle.net/10568/73258

Herrero, M., Havlík, P., Valin, H., Notenbaert, A., Rufino, M. C., Thornton, P. K., . . . Obersteiner,
M. (2013). Biomass use, production, feed efficiencies, and greenhouse gas emissions from
global livestock systems. Proceedings of the National Academy of Sciences, 110(52), 20888–
20893. doi: 10.1073/pnas.1308149110

Holmann, F., Argel, P., & Pérez, E. (2008). Impact from the adoption of improved forages in
smallholder farms in central america. Cali, Colombia: ILRI and CIAT Working Paper No.
208. (Retrieved from https://hdl.handle.net/10568/54442)

Hurtado, A., Diaz, M., Enciso, K., Gallo, I., & Burkart, S. (2020). El sistema de semillas para
la ganadería en Colombia. Cali, Colombia: Alliance of Bioversity International and CIAT.
(Retrieved from https://hdl.handle.net/10568/113045)

Hyman, G., Castro, A., Da Silva, M., Arango, M., Bernal, J., Pérez, O., & Rao, I. M. (2022).
Soil carbon storage potential of acid soils of Colombia’s Eastern High Plains. Frontiers in
Sustainable Food Systems, 6, 954017. doi: 10.3389/fsufs.2022.954017

ISPC. (2018). Tropical forages and the diffusion of Brachiaria cultivars in Latin America. Rome,

Italy: CGIAR Independent Science and Partnership Council. Brief No. 70.

40

Jank, L., Barrios, S. C., Do Valle, C. B., Simeão, R. M., & Alves, G. F. (2014). The value of
improved pastures to Brazilian beef production. Crop and Pasture Science, 65(11), 1132.
doi: 10.1071/CP13319

Kassie, M., Jaleta, M., Shiferaw, B., Mmbando, F., & Mekuria, M. (2013). Adoption of interrelated
sustainable agricultural practices in smallholder systems: Evidence from rural Tanzania.
Technological Forecasting and Social Change, 80(3), 525–540. doi: 10.1016/j.techfore
.2012.08.007

Laajaj, R., Macours, K., Masso, C., Thuita, M., & Vanlauwe, B. (2020). Reconciling yield gains
in agronomic trials with returns under African smallholder conditions. Scientific Reports,
10(1), 14286. doi: 10.1038/s41598-020-71155-y

Lane, I. R. (1981). The use of cultivated pastures for intensive animal production in developing
countries. BSAP Occasional Publication, 4, 105–143. doi: 10.1017/S0263967X00000501

Larochelle, C., & Alwang, J. (2022).

Impacts of improved bean varieties adoption on dietary
diversity and food security in Rwanda. The European Journal of Development Research,
34(2), 1144–1166. doi: 10.1057/s41287-021-00376-2

Larochelle, C., Alwang, J., Norton, G. W., Katungi, E., & Labarta, R. A. (2015). Impacts of
improved bean varieties on poverty and food security in Uganda and Rwanda. In T. S. Walker
& J. Alwang (Eds.), Crop improvement, adoption and impact of improved varieties in food
crops in Sub-Saharan Africa. Boston, MA: CGIAR and CAB International.

Martey, E., Al-Hassan, R., & Kuwornu, J. (2012). Commercialization of smallholder agriculture
in Ghana: A Tobit regression analysis. African Journal of Agricultural Research, 7(14),
2131-2141.

Martinez, J. M., Labarta, R. A., & Gonzalez, C.

Impacts of the joint adoption of
improved varieties and chemical fertilizers on rice productivity in Bolivia: implications for
Global Food Systems. Frontiers in Sustainable Food Systems, 7, 1194930. doi: 10.3389/
fsufs.2023.1194930

(2023).

Martinez, J. M., Labarta, R. A., Gonzalez, C., & Lopera, D. C. (2021). Joint adoption of rice
technologies among Bolivian farmers. Agricultural and Resource Economics Review, 50(2),
252–272. doi: 10.1017/age.2021.9

Mazzetto, A., Feigl, B., Schils, R., Cerri, C., & Cerri, C. (2015). Improved pasture and herd
management to reduce greenhouse gas emissions from a Brazilian beef production system.
Livestock Science, 175, 101–112. doi: 10.1016/j.livsci.2015.02.014

McAlpine, C., Etter, A., Fearnside, P., Seabrook, L., & Laurance, W. (2009). Increasing world
consumption of beef as a driver of regional and global change: A call for policy action based
on evidence from Queensland (Australia), Colombia and Brazil. Global Environmental

41

Change, 19(1), 21–33. doi: 10.1016/j.gloenvcha.2008.10.008

McCoy, T., & Ledur, J. (2022, April). How Americans’ love of beef is helping destroy the Amazon

rainforest. The Washington Post.

Mejía, D., Diaz, M., Enciso, K., & Burkart, S. (2021). What is the relationship between agricultural
credit, cattle, and deforestation? A spatial analysis in Colombia. Cali, Colombia: CIAT
Policy Brief No. 61. Retrieved from https://hdl.handle.net/10568/116087

Montiel, J. L., & Pflueger, C. (2013). A robust test for weak instruments. Journal of Business &

Economic Statistics, 31(3), 358–369. doi: 10.1080/00401706.2013.806694

Murgueitio, E., Chará, J., Barahona, R., Cuartas, C., & Naranjo, J.

(2014). Los sistemas sil-
vopastorales intensivos (SSPI), herramienta de mitigación y adaptación al cambio climático.
Tropical and Subtrpical Agroecosystems, 13(3), 501-507.

Murillo-Sandoval, P. J., Kilbride, J., Tellman, E., Wrathall, D., Van Den Hoek, J., & Kennedy, R. E.
(2023). The post-conflict expansion of coca farming and illicit cattle ranching in Colombia.
Scientific Reports, 13(1), 1965. doi: 10.1038/s41598-023-28918-0

Ospina, S., Rusch, G. M., Pezo, D., Casanoves, F., & Sinclair, F. L. (2012). More stable productivity
of semi-natural grasslands than sown pastures in a seasonally dry climate. PLoS ONE, 7(5),
e35555. doi: 10.1371/journal.pone.0035555

Oster, E. (2019). Unobservable selection and coefficient stability: theory and evidence. Journal of
Business & Economic Statistics, 37(2), 187–204. doi: 10.1080/07350015.2016.1227711

Ouma, E., Jagwe, J., Obare, G. A., & Abele, S. (2010). Determinants of smallholder farmers’
participation in banana markets in Central Africa: the role of transaction costs. Agricultural
Economics, 41(2), 111–122. doi: 10.1111/j.1574-0862.2009.00429.x

Papke, L. E., & Wooldridge, J. M. (1996). Econometric methods for fractional response variables
with an application to 401(K) plan participation rates. Journal of Applied Econometrics,
11(6), 619–632. doi: 10.1002/(SICI)1099-1255(199611)11:6<619::AID-JAE418>3.0.CO;2
-1

Parodi, A., Valencia-Salazar, S., Loboguerrero, A. M., Martínez-Barón, D., Murgueitio, E., &
Vázquez-Rowe, I. (2022). The sustainable transformation of the Colombian cattle sector:
Assessing its circularity. PLOS Climate, 1(10), e0000074. doi: 10.1371/journal.pclm
.0000074

Peters, M., Lascano, C., Roothaert, R., & de Haan, N. (2003). Linking research on forage germplasm
to farmers: the pathway to increased adoption—a ciat, ilri and iita perspective. Field Crops
Research, 84(1), 179-188. doi: https://doi.org/10.1016/S0378-4290(03)00149-7

42

Pérez-López, O., & Afanador-Tellez, G. (2017). Comportamiento agronomico y nutricional de
genotipos de Brachiaria spp. manejados con fertilizacion nitrogenada solos y asociados con
Pueraria phaseoloides en condiciones de la altillanura colombiana. Revista de la Facultad
de Medicina Veterinaria y de Zootecnia, 64(3). doi: 10.15446/rfmvz.v64n3.68694

Rao, I. M., Zeigler, R. S., Vera, R., & Sarkarung, S. (1993). Selection and breeding for acid-soil

tolerance in crops. BioScience, 43(7), 454–465.

Rivas, L., & Holmann, F. (2004). Impacto de la adopción de híbridos de Brachiarias resistentes al
salivazo: Colombia, México, y Centroamérica. Cali, Colombia: CIAT Working Paper No.
195. Retrieved from https://hdl.handle.net/10568/1747

Rodríguez, H., Ramírez-Gómez, C. J., Aguilar-Gallegos, N., & Aguilar-Ávila, J. (2016). Network
analysis of knowledge building on rural extension in Colombia. Agronomía Colombiana,
34(3), 393–402. doi: 10.15446/agron.colomb.v34n3.58500

Sandoval, D. F., Florez, J. F., Enciso Valencia, K. J., Sotelo Cabrera, M. E., & Burkart, S. (2023).
Economic-environmental assessment of silvo-pastoral systems in Colombia: An ecosystem
service perspective. Heliyon, 9(8), e19082. doi: 10.1016/j.heliyon.2023.e19082

Silva-Parra, A., Trujillo-González, J. M., & Brevik, E. C. (2021). Greenhouse gas balance and
mitigation potential of agricultural systems in Colombia: A systematic analysis. Greenhouse
Gases: Science and Technology, 11(3), 554–572. doi: 10.1002/ghg.2066

Tapasco, J., LeCoq, J. F., Ruden, A., Rivas, J. S., & Ortiz, J. (2019). The livestock sector in
Colombia: toward a program to facilitate large-scale adoption of mitigation and adaptation
practices. Frontiers in Sustainable Food Systems, 3, 61. doi: 10.3389/fsufs.2019.00061

Teklewold, H., Kassie, M., & Shiferaw, B. (2013). Adoption of multiple sustainable agricultural
practices in rural ethiopia. Journal of Agricultural Economics, 64(3), 597–623. doi: 10.1111/
1477-9552.12011

Toledo, J. M., & Nores, G. A. (1986). Tropical pasture technology for marginal lands of tropical
America. Outlook on Agriculture, 15(1), 2–9. doi: 10.1177/003072708601500101

Triana Ángel, N., Burkart, S., & Pazos Cárdenas, M. (2024). Land, cows, and bullets: an untold
history of cattle ranching during the armed conflict in Colombia, 1980–2010. Frontiers in
Sustainable Food Systems, 8, 1374861. doi: 10.3389/fsufs.2024.1374861

UPRA. (2016). Mercado de tierras rurales productivas en Colombia: Caracterización, marco
conceptual, jurídico e institucional. Bogotá, Colombia: Unidad de Planificación Rural y
Agropecuaria, Ministerio de Agricultura.

Vaccaro, L. (1997). Constraints and opportunities for livestock development in mixed farming
systems in tropical Latin America and the Caribbean. Outlook on Agriculture, 26(4), 227–

43

234. doi: 10.1177/003072709702600404

Van Ausdal, S. (2009). Pasture, profit, and power: An environmental history of cattle ranching in

Colombia, 1850–1950. Geoforum, 40(5), 707–719. doi: 10.1016/j.geoforum.2008.09.012

Vargas-Cuéllar, J. P.

(2022, March). Exportaciones de carne bovina cambian rumbo e irían a

mercado de Estados Unidos. La República.

Vera, R. R., & Hoyos Garcés, P. (2019). Long-term beef production from pastures established with
and without annual crops compared with native savanna in the Eastern Plains of Colombia:
A compilation and analysis of on-farm results 1979–2016. Tropical Grasslands-Forrajes
Tropicales, 7(1), 1–13. doi: 10.17138/tgft(7)1-13

Wooldridge, J. M. (1995). Score diagnostics for linear models estimated by two stage least squares.
In G. S. Maddala, P. C. B. Phillips, & T. N. Srinivasan (Eds.), Advances in econometrics and
quantitative economics: Essays in honor of Professor C. R. Rao. Oxford: Blackwell.

Wooldridge, J. M.

(2010). Econometric analysis of cross section and panel data (2nd ed.).

Cambridge, MA: MIT Press.

Wooldridge, J. M. (2015). Control function methods in applied econometrics. Journal of Human

Resources, 50(2), 420–445. doi: 10.3368/jhr.50.2.420

Xu, R. (2021). On the instrument functional form with a binary endogenous explanatory variable.

Economics Letters, 206, 109993. doi: 10.1016/j.econlet.2021.109993

44

APPENDIX

Table 1A.1 Testing for distance to R&D centers and its relationship with the distance to input
markets and extension services.

Variables

(1)

(2)

Distance to closest input market

0.0084
(0.025)

Distance to closest extension services

Observations(a)
R-squared
Method
Dept Controls
Cluster (village) standard errors in parentheses.
*** p<0.01, ** p<0.05, * p<0.1

993
0.781
OLS
Yes

0.0457
(0.037)
0.0056
(0.040)

611
0.682
OLS
Yes

Notes — Response variable is the road distance from the farm to the closest R&D dissemination center. (a) Reduced number of
observations in Column 2 responds to data limitations in the sample. We only have available data on the distance to extension
services for Eastern Plains and Amazon Piedmont.

45

Table 1A.2 Alternate first-stage analysis of factors associated with the share of introduced pastures
in the farm, including distance to closest input market and closest extension services.

Variables

(1)

(2)

Log-Distance to closest dissemination center(a)

Log-Distance to closest input market(a)

Log-Distance to closest extension services(a)

Attacks on population(a)

Illegal checkpoints(a)

Log-Area of the farm(a))

Education of HH (Primary)

Education of HH (Secondary)

Education of HH (Tertiary)

Female HH (1 = Yes)

Member of a Farmer’s Association (1 = Yes)

Access to Credit (1 = Yes)

Household Assets Index

Farm Assets Index

-0.1196***
(0.041)
0.0106
(0.013)

-0.0310**
(0.015)
-0.0270***
(0.010)
-0.0153*
(0.008)
0.0525
(0.036)
0.0287
(0.039)
-0.0024
(0.052)
-0.0176
(0.036)
-0.0417
(0.026)
0.0039
(0.023)
0.0056
(0.005)
0.0200***
(0.006)

-0.1151*
(0.063)
0.0653
(0.047)
-0.0511
(0.046)
-0.0200
(0.014)
-0.0302***
(0.010)
-0.0441***
(0.011)
-0.0205
(0.044)
-0.0370
(0.053)
-0.0843
(0.070)
-0.0154
(0.040)
-0.0374
(0.039)
0.0513*
(0.030)
0.0239***
(0.007)
0.0130
(0.008)

Observations
Method
Dept Controls
Cluster (village) standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

993
QMLE
Yes

611
QMLE
Yes

Notes — Response variable is the share of pastures with introduced varieties. (a) Regression included the variable and its square,
and the table reports the full average partial effect.

46

Table 1A.3 Additional testing of distance to R&D and distance to alternate infrastructures in
broader agricultural intensification.

Variables

(1)

(2)

Comp.Pract. Comp.Pract.

(3)
Farm Assets

(4)
Farm Assets

Log-Distance to closest dissemination center(a)

Log-Distance to closest input market(a)

Log-Distance to closest extension services(a)

) Log-Area of the farm(a)

Observations(b)
Method
Dept Controls
Cluster (village) standard errors in parentheses.
*** p<0.01, ** p<0.05, * p<0.1

-0.0294
(0.065)
-0.0453
(0.030)

-0.0069
(0.023)

993
Poisson
Yes

-0.0742
(0.086)
0.1548**
(0.077)
-0.193***
(0.074)
-0.0167
(0.023)

611
Poisson
Yes

-0.0633
(0.043)
-0.0545***
(0.019)

0.1397***
(0.013)

993
Poisson
Yes

-0.1053
(0.069)
0.0015
(0.046)
-0.0734
(0.045
0.1490***
(0.022)

611
Poisson
Yes

Notes — Reporting semi-elasticity estimates from Poisson regression. Response variables are (1-2) the number of complementary
practices for pasture management and (3-4) the number of durable farming assets in the farm. (a) Regression included the variable
and its square, and the table reports the full average partial effect. (b) Reduced number of observations in Column 2 responds to
data limitations in the sample. We only have available data on the distance to extension services for Eastern Plains and Amazon
Piedmont.

Table 1A.4 Additional testing of selected violence variables and correlation with measures of
wealth and education.

(1)
House Asset Index

(2)
High-Ed (1 = Yes)

Variables

Attacks on population(a)

Illegal checkpoints(a)

Log-Area of the farm(a)

-0.0179
(0.020)
-0.0043
(0.013)
0.0117
(0.009)

Observations
Method
Dept Controls
Cluster (village) standard errors in parentheses.
*** p<0.01, ** p<0.05, * p<0.1

994
Poisson
Yes

-0.0158
(0.021)
-0.0192
(0.016)
0.0157*
(0.008)

994
OLS
Yes

Notes — Reporting semi-elasticity estimates for Poisson regression and average partial effects for ordinary least squares (OLS).
Response variables are (1) the number of complementary house valuable assets and (2) an index variable equal to one when the head
of the household has completed either secondary or tertiary education, and zero otherwise. (a) Regression included the variable
and its square, and the table reports the full average partial effect.

47

Table 1A.5 Additional testing on instrument strength by non-linearity.

Variables

Share of Introduced Pastures

Log-Area of the farm(a)

Education of HH (Primary)

Education of HH (Secondary)

Education of HH (Tertiary)

Female HH (1 = Yes)

Average precipitation in the past year

Average maximum temperature in the past year

Observations
Method
Dept Controls
Effective F Statistic
Montiel-Pflueger: 5% Bias Critical value
Montiel-Pflueger: 10% Bias Critical value
Robust reg. endogeneity test (p-value)
Cluster (village) standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

(1)

(2)

Share of IP Log-Kg/Ha

-0.0033
(0.008)
0.0936**
(0.038)
0.0812**
(0.035)
0.0430
(0.046)
-0.0245
(0.038)

1,021
QMLE
Yes

1.6768**
(0.721)
-0.2667***
(0.020)
-0.0347
(0.120)
-0.0230
(0.139)
0.1017
(0.139)
-0.1924*
(0.101)
0.0354**
(0.018)
-0.0010
(0.016)

1,021
Opt-IV
Yes
10.56
37.42
23.11
0.0369

Notes — (a) Regression included the variable and its square, and the table reports the full average partial effect.

48

Table 1A.6 Alternate first-stage analysis of factors associated with the share of introduced pastures
in the farm, using each main source of exogenous variation separately.

Variables

(1)

(2)

(3)

Log-Distance to closest dissemination center(a)

-0.1297***
(0.039)

Attacks on population(a)

Illegal checkpoints(a)

Log-Area of the farm(a)

Education of HH (Primary)

Education of HH (Secondary)

Education of HH (Tertiary)

Female HH (1 = Yes)

Member of Farmers’ Association (1 = Yes)

Access to Credit (1 = Yes)

Household Assets Index

Farm Assets Index

Observations
Method
Dept Controls

-0.0327**
(0.014)

-0.0110
(0.008)
0.0773**
(0.038)
0.0592
(0.038)
0.0235
(0.049)
-0.0115
(0.036)
-0.0301
(0.025)
0.0060
(0.024)
0.0070
(0.005)
0.0183***
(0.006)

-0.0294***
(0.008)
-0.0106
(0.008)
0.0735*
(0.038)
0.0593
(0.037)
0.0106
(0.052)
-0.0263
(0.037)
-0.0269
(0.025)
0.0056
(0.024)
0.0074
(0.005)
0.0182***
(0.006)

-0.0174**
(0.009)
0.0613
(0.038)
0.0379
(0.039)
-0.0082
(0.053)
-0.0272
(0.037)
-0.0396
(0.026)
0.0088
(0.024)
0.0058
(0.005)
0.0193***
(0.006)

994
QMLE
Yes

1,021
QMLE
Yes

1,021
QMLE
Yes

Notes — Response variable is the share of pastures with introduced varieties. (a) Regression included the variable and its square,
and the table reports the full average partial effect.

49

Table 1A.7 Alternate estimation of the effect of introduced pasture adoption on cattle yield, using
each main source of exogenous variation separately.

Variables

(1)

(2)

(3)

Share of Improved Pastures

Log-Area of the farm(a)

Education of HH (Primary)

Education of HH (Secondary)

Education of HH (Tertiary)

Female HH (1 = Yes)

Member of Farmers’ Association (1 = Yes)

Access to Credit (1 = Yes)

Average precipitation in the past year

Average maximum temperature in the past year

Source of exogenous variation(b)
Observations
Method
Dept Controls
Effective F Statistic
Montiel-Pflueger: 5% Bias Critical value
Montiel-Pflueger: 5% Bias Critical value
Robust reg. endogeneity test (p-value)
Cluster (village) standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

1.8097***
(0.348)
-0.2695***
(0.021)
-0.0369
(0.102)
0.0028
(0.119)
0.0936
(0.141)
-0.1897*
(0.099)
0.1507**
(0.063)
-0.0871
(0.060)
0.0322**
(0.015)
0.0074
(0.017)

1.5705***
(0.403)
-0.2709***
(0.020)
-0.0166
(0.100)
0.0221
(0.117)
0.0994
(0.134)
-0.1996**
(0.093)
0.1473**
(0.060)
-0.0854
(0.056)
0.0300**
(0.015)
0.0036
(0.017)

1.7259***
(0.344)
-0.2700***
(0.021)
-0.0298
(0.102)
0.0096
(0.120)
0.0956
(0.139)
-0.1932**
(0.097)
0.1495**
(0.063)
-0.0865
(0.059)
0.0315**
(0.015)
0.0061
(0.017)

Distance to R&D Attacks on Pop.

994
Opt-IV
Yes
35.8
37.42
23.11
7.83e-08

994
Opt-IV
Yes
27.08
37.42
23.11
0.00158

Illegal Checkp.
994
Opt-IV
Yes
41.24
37.42
23.11
3.81e-06

Notes — Response variable is (log) Kilograms of livestock per hectare. (a) Regression included the variable and its square, and the
table reports the full average partial effect. (b) Describes the source of exogenous variation used in the first-stage analysis. The
optimal IV estimation uses the non-linear conditional mean from the first-stage QMLE analysis.

50

Table 1A.8 Alternate first-stage analysis of factors associated with the share of introduced pastures
in the farm.

Variables

(1)

(2)

(3)

(4)

Distance to closest dissemination center (10 km)(a,b)

Attacks on population (1988-2012)(a)

Illegal checkpoints (2002-2016) (a)

-0.0888**
-0.04
-0.0522***
-0.017
-0.0375***
-0.012

-0.0876**
-0.034
-0.0410***
-0.015
-0.0334***
-0.01

-0.0759*
-0.039
-0.0492***
-0.017
-0.0336**
-0.013
-0.0166**
-0.008
0.0678*
-0.04
0.0606
-0.048
0.0192
-0.06
-0.0131
-0.037
-0.032
-0.026
0.012
-0.023
0.0068
-0.005
0.0193***
-0.007

-0.0727**
-0.033
-0.0370**
-0.014
-0.0296***
-0.011
-0.0128
-0.008
0.0622*
-0.035
0.0488
-0.037
0.0234
-0.048
-0.0154
-0.035
-0.0381
-0.025
0.01
-0.023
0.0059
-0.005
0.0178***
-0.006

Log-Area of the farm(a)

Education of HH (Primary)

Education of HH (Secondary)

Education of HH (Tertiary)

Female HH (1 = Yes)

Member of Farmers’ Association (1 = Yes)

Access to Credit (1 = Yes)

Household Assets Index

Farm Assets Index

Observations
Method
Dept Controls

1,021
OLS
No

1,021
OLS
Yes

1,021
QMLE
No

1,021
QMLE
Yes

Notes — Response variable is the portion of pastures with introduced varieties. (a) Regression included the variable and its square,
and the table reports the full average partial effect. (b) Using distance variable based on the geodesic distance from farm to closest
dissemination center on a per 10 kilometer scale.

51

Table 1A.9 Alternate estimation of the effect of introduced pasture adoption on cattle yield.

Variables

(1)

(2)

(3)

(4)

(5)

Share of Introduced Pastures

Log-Area of the farm(a)

Education of HH (Primary)

Education of HH (Secondary)

Education of HH (Tertiary)

Female HH (1 = Yes)

Average precipitation in the past year

Avg. max. temp. in the past year

Control Function (𝑣)

0.3580***
(0.091)

0.3386***
(0.072)
-0.2718***
(0.019)
0.0917
(0.074)
0.1113
(0.090)
0.1452
(0.115)
-0.2443***
(0.077)
0.0166
(0.012)
-0.0113
(0.012)

1.0865***
(0.252)
-0.2690***
(0.019)
0.0211
(0.085)
0.0362
(0.102)
0.1209
(0.121)
-0.2153***
(0.082)
0.0271**
(0.013)
-0.0056
(0.013)

1.0846***
(0.276)
-0.2690***
(0.019)
0.0212
(0.083)
0.0364
(0.104)
0.1210
(0.125)
-0.2154**
(0.085)
0.0271**
(0.013)
-0.0056
(0.014)

1,021
OLS
Yes

1,021
OLS
Yes

Observations
Method
Dept Controls
Effective F
Montiel-Pflueger: 5% Bias Critical value
Montiel-Pflueger: 10% Bias Critical value
Robust reg. endogeneity Test (p-value)
Cluster (village) standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

1,021
2SLS
Yes
6.619
25
14.37
0.001

1,021
Opt-IV
Yes
60.92
37.42
23.11
0.0005

1.1477***
(0.296)
-0.2680***
(0.020)
0.0128
(0.085)
0.0316
(0.106)
0.1065
(0.125)
-0.2307***
(0.087)
0.0145
(0.012)
-0.0169
(0.013)
-0.8893***
(0.297)

1,021
CF
Yes

Notes — Response variable is (log) Kilograms of livestock per hectare. Two-step optimal IV (Opt-IV) and control function (CF)
regression estimates report cluster bootstrapped standard errors. (a) Regression included the variable and its square, and the table
reports the full average partial effect.

52

Table 1A.10 Alternate estimation of the effect of introduced pasture adoption on livestock-related
revenues per hectare.

Variables

(1)

(2)

(3)

(4)

(5)

Share of Introduced Pastures

Log-Area of the farm(a)

Education of HH (Primary)

Education of HH (Secondary)

Education of HH (Tertiary)

Female HH (1 = Yes)

Average precipitation in the past year

Avg. max. temp. in the past year

Control Function (𝑣)

0.3988**
(0.164)

0.3167**
(0.134)
-0.6610***
(0.045)
0.2611
(0.169)
0.3806*
(0.224)
0.4783**
(0.239)
-0.2046
(0.157)
0.0430
(0.032)
-0.0054
(0.032)

1.7840***
(0.520)
-0.6502***
(0.047)
0.0688
(0.202)
0.1867
(0.255)
0.3984
(0.283)
-0.1123
(0.167)
0.0669*
(0.035)
0.0133
(0.031)

1.8617***
(0.571)
-0.6496***
(0.047)
0.0587
(0.211)
0.1764
(0.271)
0.3941
(0.292)
-0.1074
(0.177)
0.0682*
(0.036)
0.0143
(0.033)

796
OLS
Yes

796
OLS
Yes

Observations
Method
Dept Controls
Effective F Statistic
Montiel-Pflueger: 5% Bias Critical value
Montiel-Pflueger: 10% Bias Critical value
Robust reg. endogeneity test (p-value)
Cluster (village) standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

796
2SLS
Yes
5.549
24.71
14.2
0.002

796
Opt-IV
Yes
42.54
37.42
23.11
0.001

1.9932***
(0.608)
-0.6524***
(0.047)
0.1066
(0.212)
0.2291
(0.274)
0.4088
(0.297)
-0.1875
(0.176)
0.0343
(0.033)
-0.0201
(0.034)
-1.8302***
(0.632)

796
CF
Yes

Notes — Response variable is (log) Revenue from livestock-related sales per hectare. Two-step optimal IV (Opt-IV) and control
function (CF) regression estimates report cluster bootstrapped standard errors. (a) Regression included the variable and its square,
and the table reports the full average partial effect.

53

Table 1A.11 Effect of introduced pasture adoption on cattle yield with exponential model.

Variables

(1)

(2)

(3)

(4)

0.2198***
(0.084)

Share of Improved Pastures

Log-Area of the farm(a)

Education of HH (Primary)

Education of HH (Secondary)

Education of HH (Tertiary)

Female HH (1 = Yes)

Average precipitation in the past year

Average maximum temperature in the past year

0.2180***
(0.076)
-0.2657***
(0.024)
0.0819
(0.084)
0.0725
(0.107)
0.1841
(0.125)
-0.2774***
(0.073)
0.0014
(0.016)
0.0017
(0.016)

1.2789***
(0.278)
-0.2896***
(0.020)
0.0797
(0.076)
0.0888
(0.104)
0.1477
(0.115)
-0.3055***
(0.075)
-0.0044
(0.013)
-0.0129
(0.016)

1.2283***
(0.262)
-0.2726***
(0.020)
0.0158
(0.079)
0.0489
(0.104)
0.1453
(0.113)
-0.2816***
(0.074)
-0.0062
(0.013)
-0.0113
(0.016)

1,021

Observations
Method
Dept Controls
Cluster (village) standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1
Notes — Response variable is Kilograms of livestock per hectare. IV Poisson with optimal instrument (Opt-IV Poisson) regression
estimates report cluster bootstrapped standard errors.

Poisson QMLE Poisson QMLE

IV Poisson Opt-IV Poisson

1,021

1,021

1,021

Yes

Yes

Yes

Yes

54

Table 1A.12 Effect of introduced pasture adoption on cattle yield with sample limited to adopters (only including farms with 𝑆𝑖 > 0).

Variables

(1)

(2)

(3)

(4)

(5)

0.5181***
(0.101)

Share of Improved Pastures

Log-Area of the farm

Education of HH (Primary)

Education of HH (Secondary)

Education of HH (Tertiary)

Female HH (1 = Yes)

Member of Farmers’ Association (1 = Yes)

Access to Credit (1 = Yes)

Average precipitation in the past year

Average maximum temperature in the past year

Control Function (v)

0.4259***
(0.083)
-0.2634***
(0.020)
0.0791
(0.075)
0.1238
(0.094)
0.1174
(0.120)
-0.2817***
(0.081)
0.1373***
(0.053)
-0.0610
(0.048)
0.0116
(0.012)
-0.0204
(0.014)

1.5579***
(0.351)
-0.2495***
(0.020)
0.0360
(0.088)
0.0914
(0.110)
0.0924
(0.126)
-0.2633***
(0.089)
0.1432**
(0.058)
-0.0816
(0.055)
0.0188
(0.013)
-0.0054
(0.015)

1.3984***
(0.303)
-0.2515***
(0.020)
0.0420
(0.085)
0.0959
(0.106)
0.0959
(0.123)
-0.2659***
(0.087)
0.1424**
(0.057)
-0.0787
(0.053)
0.0178
(0.013)
-0.0075
(0.015)

1.4791***
(0.402)
-0.2488***
(0.021)
0.0315
(0.088)
0.0947
(0.110)
0.0761
(0.126)
-0.2807***
(0.096)
0.1502**
(0.061)
-0.0777
(0.054)
0.0075
(0.012)
-0.0218
(0.015)
-1.1435**
(0.412)

974
OLS
Yes

974
OLS
Yes

Observations
Method
Dept Controls
Effective F
Montiel-Pflueger: 5% Bias Critical value
Montiel-Pflueger: 10% Bias Critical value
Robust reg. endogeneity Test (p-value)
Cluster (village) standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1
Notes — Response variable is (log) Kilograms of livestock per hectare. Two-step optimal IV (Opt-IV) and control function (CF)
regression estimates report cluster bootstrapped standard errors. (a) Regression included the variable and its square, and the table
reports the full average partial effect.

974
Opt-IV
Yes
44.43
37.42
23.11
0.0003

974
2SLS
Yes
5.38
25.94
15.11
0.0002

974
CF
Yes

55

Table 1A.13 Heterogeneous effect of introduced pastures on cattle yield by complementary
management practice.

(1)

(2)

Variables

Pasture

Pasture + Weed control

Pasture + Fertilization

Pasture + Weed control + Fertilization

Control function (𝑣)

0.1590
(0.108)
0.4305***
(0.111)
0.6998**
(0.323)
0.5695*
(0.319)

Observations
Method
Dept Controls
Other Controls
Cluster (village) standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

994
OLS
Yes
Yes

1.2665***
(0.320)
1.5464***
(0.308)
1.8771***
(0.495)
1.6374***
(0.431)
-1.2055***
(0.316)

994
CF
Yes
Yes

Notes — Response variable is (log) Kilograms of livestock per hectare. Control function (CF) regression estimates report cluster
bootstrapped standard errors.

56

Table 1A.14 Ordered Logit regression and marginal effects on the number of pasture management
practices

Variables

Log-Area of the farm

Education of HH (Primary)

Education of HH (Secondary)

Education of HH (Tertiary)

Female HH (1 = Yes)

Member of Farmers’ Association (1 = Yes)

Access to Credit (1 = Yes)

Household Assets Index

Farm Assets Index

𝜆1

𝜆2

(1)
Coeff.

(2)
Pr(𝑘=0)

(3)
Pr(𝑘=1)

(4)
Pr(𝑘=2)

0.0106
(0.009)
-0.0681*
(0.037)
-0.0778*
(0.045)
-0.1667***
(0.060)
0.0013
(0.038)
-0.0582***
(0.022)
-0.0392*
(0.021)
-0.0066
(0.005)
-0.0126*
(0.006)

-0.0043
(0.004)
0.0278*
(0.016)
0.0268*
(0.015)
0.0167
(0.022)
-0.0005
(0.015)
0.0172**
(0.007)
0.0130**
(0.006)
0.0026
(0.002)
0.0051*
(0.003)

-0.0063
(0.006)
0.0403*
(0.022)
0.0510
(0.033)
0.1499*
(0.077)
-0.0008
(0.022)
0.0410**
(0.017)
0.0263
(0.016)
0.0039
(0.003)
0.0075*
(0.004)

-0.0907
(0.080)
0.5952*
(0.330)
0.6495*
(0.368)
1.4901**
(0.580)
-0.0114
(0.323)
0.5211***
(0.196)
0.3469*
(0.198)
0.0559
(0.041)
0.1078*
(0.056)
-1.0937**
(0.528)
3.0112***
(0.521)

Dept Controls
Observations
Cluster (village) standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

Yes
947

Yes
947

Yes
947

Yes
947

Notes — The response variable is the number of pasture management practices used on the farm. Column 1 reports the coefficient
estimates from Ordered Logit regression. Columns 2-4 report the average marginal effects of covariates on the probability of
adopting 𝑘 pasture management practices. Estimation is restricted to farms with a non-zero level of introduced pasture adoption.
Coefficients 𝜆𝑘 are the estimates of cut-off points that partition the predicted values on the extent of adoption.

57

Table 1A.15 Heterogeneous effect of introduced pastures on livestock-related revenues by
complementary management practice.

(1)

(2)

Variables

Pasture

Pasture + Weed control

Pasture + Fertilization

Pasture + Weed control + Fertilization

Control function (𝑣)

0.0950
(0.182)
0.6784***
(0.223)
0.1532
(0.592)
-0.1944
(0.479)

Observations
Method
Dept Controls
Other Controls
Cluster (village) standard errors in parentheses
*** p<0.01, ** p<0.05, * p<0.1

786
OLS
Yes
Yes

2.1057***
(0.715)
2.6836***
(0.710)
2.2690**
(0.992)
1.7292**
(0.834)
-2.1822***
(0.717)

786
CF
Yes
Yes

Notes — Response variable is (log) Revenue from livestock-related sales per hectare. Control function (CF) regression estimates
report cluster bootstrapped standard errors.

58

CHAPTER 2

SWEATING BULLETS: HEAT, HIGH-STAKES EVALUATIONS, AND THE ROLE OF
INCENTIVES

2.1

Introduction

Although climate change effects are widely studied with a focus on the long run and on

monetary measurements of its impacts (Carleton & Hsiang, 2016; Lenton et al., 2023), temperature

and environmental changes are becoming increasingly evident in the short run. Hence, analyses

that properly measure how harsher weather conditions currently affect diverse social and economic

dimensions become of the highest priority (Carleton et al., 2022; Dell, Jones, & Olken, 2014;

Hsiang & Kopp, 2018). Human capital accumulation has recently become a central topic in this

regard (Garg, Jagnani, & Taraz, 2020; Park, Behrer, & Goodman, 2021; Park, Goodman, Hurwitz,

& Smith, 2020). Nevertheless, most of these findings focus on cases of temperate or subtropical

weather such as the United States or China, where temperature shocks can be highly seasonal, and

thus have a higher relevance if they occur specifically during the period in which students are being

evaluated. The case of teenagers and young adults in hot tropical countries, where exposure to

increasingly high temperatures throughout the year is becoming the rule instead of a more frequent

seasonal event, remains poorly explored.

What are the impacts of long-term exposure to increasing temperatures on students’ performance

in high-stakes evaluations? How does a hotter environment affect the cost of exerting effort in these

high-stakes tests? Using a model of human capital accumulation with temperature stress, we derive

three main testable hypotheses: (1) increases in heat exposures negatively affect human capital

accumulation, (2) positive shocks to expected income favor the accumulation of human capital,

and (3) exerting effort is costly when experiencing hotter temperatures. Combining five years of

information from a national semi-annual high school exit exam in Colombia (Saber 11) and daily

weather station data on temperature and precipitation, we use fixed-effects estimation to measure

the effect of temperature on exam scores. Also, we exploit a large-scale scholarship program

(i.e., Ser Pilo Paga, henceforth SPP) as a source of exogenous variation in the stakes of the exam,

59

allowing us to set up a quasi-experimental design to test whether temperature increases make it

costly to exert effort in preparation for a high-stakes exam.

We find that temperature increases have significant effects on Saber 11 scores, especially for

students in urban areas. Our estimates suggest that a 1°C increase in the average daily maximum

temperature experienced in the year before the exam reduces scores by at least 2% of a standard

deviation. Relative to days with temperatures between 20-25°C, each additional day of exposure

to temperatures between 30-35°C causes a Math score reduction of 0.05% of a standard deviation,

while an additional day over 35°C reduces this score by 0.11% of a standard deviation. Slightly

lower effects are found for Spanish and total scores. Interestingly, we find that higher temperatures

slightly improve students’ performance in rural areas. Leveraging detailed time-use data, we provide

suggestive evidence that higher temperatures lead to an increase in the time spent by youths on

activities complementary to studying. We also find that temperature increases lead to a reallocation

in the time spent by adults on off-farm labor, an activity that is more human-capital intensive.

Combining variation in temperatures and in the exam stakes resulting from SPP, we find that for

each additional percentage point of exposure to the program, average Math scores increase between

3.03% and 6.12% of a standard deviation, with no noticeable effects on Spanish scores. Among

students in urban areas, increases in SPP exposure imply lower Math score gains for each additional

increase in the average maximum temperature experienced in the year before the exam. Putting

these effects into perspective, our results suggest that an interquartile change in the exposure to SPP

leads to an 11.9% increase in absolute terms in the impact of temperature on exam performance.

Therefore, our findings suggest that the additional effort exerted by students as a response to the

program led to a stronger negative effect of temperature on scores.

While this subject is still an emerging literature, and most studies focus on the short-term

effects of temperature shocks (e.g., those occurring the day of the exam or the week before it)

on score performance, some well-defined patterns are worth highlighting. Graff-Zivin, Hsiang,

and Neidell (2018) uses same-day weather variation to measure the effect of temperature in the

performance of reading and Math tasks in the National Longitudinal Survey of the Young - 1979

60

cohort, with evidence suggesting that only “high-complexity” tasks like Math are affected by these

shocks. Similar results for low-stakes examinations were found for China in Zhang, Chen, and

Zhang (2024).

In the context of high-stakes examinations in Korea, Cho (2017) also finds a significant effect

of same-day temperature shock on Math scores but not on reading. Higher temperatures during the

day of the exam have also been found to decrease exam scores in China (Graff-Zivin, Song, Tang, &

Zhang, 2020), the United States (Park, 2022), and at the world level based on PISA tests (Park et al.,

2021). In the context of South America, Li and Patel (2021) find for Brazil that the effects of higher

temperatures on high-stakes exam scores may not be substantive. Conversely, Hoffmann, Pulido,

and Vera-Cossio (2023) finds that exposure to high temperatures in Colombia during the week

before a Saber 11 test day negatively affects scores; on average, an additional hour of exposition to

temperatures above 32°C reduces total scores by 0.11% of a standard deviation. To our knowledge,

only the works in Park et al. (2020), for the United States, and Garg et al. (2020), for India, exploit

the temperature in the year before the exam as potentially affecting final scores, finding that days

above 30°C can reduce scores by 0.05% and 0.3% of a standard deviation compared to days with

temperatures between 15-17°C, respectively.

Our contribution is, hence, threefold. First, we contribute to the literature on the effects of heat

on high-stakes exam performance showing that longer-term exposure to increasing temperatures

has negative significant effects, thus suggesting that heat affects not only exam performance but

also human capital accumulation (Garg et al., 2020; Graff-Zivin et al., 2018; Park et al., 2021).

Second, we study the effect of temperature on students’ scores in a tropical country, where the range

of variation of temperature is particularly narrow and, therefore, the adaptation strategies available

to students may differ (Helo Sarmiento, 2023). Understanding these dynamics is of the utmost

importance for countries for which, as in the case of Colombia, the use of adaptation strategies

like air conditioning is severely low (under 5% of buildings (Statista Research Department, 2023)).

Third, and more importantly, using an exogenous shock to the stakes of the exam, we provide quasi-

experimental evidence revealing that exerting effort becomes increasingly costly as individuals

61

are exposed to increasing temperature levels and, therefore, provide evidence of an adaptation

mechanism elusive so far in the literature (Burke et al., 2016). Moreover, we provide correlational

evidence of a mechanism of opportunity costs that supports both the predictions of our theoretical

model and our causal analysis.

The remainder of our paper proceeds as follows. In Section 2.2, we set up a theoretical model

revealing the structural relationship between exposure to heat and the accumulation of human

capital and the role of incentives in this relationship. Then, in Section 2.3, we briefly discuss Ser

Pilo Paga, its relevance for our analysis, and we summarize our data. Section 2.4 presents our

identification strategies in detail. We present and discuss our main results and robustness checks in

Section 2.5. We conclude with Section 2.6, where we also provide directions for future research.

2.2 Theoretical Framework

We start by setting a static model where human capital accumulation is sensitive to temperature

stress. We have two goals with our model. First, we want to show how changes in the incentives

to exert effort may interact with temperature and, as a result, affect the accumulation of human

capital. Second, we use the framework to provide an interpretation of the effects being recovered

by our empirical analysis and the mechanisms through which they take place1

We begin by defining a human capital production function ℎ that follows:

ℎ = 𝑧 𝑓 (𝑒, 𝑡),

(2.1)

where 𝑧 is the individual’s cognitive skill or ability, 𝑒 is their effort level, 𝑡 is the temperature, and 𝑓 is

a function assumed to be increasing in effort ( 𝑓𝑒 > 0), decreasing in temperature ( 𝑓𝑡 < 0), and with

negative second and cross derivatives ( 𝑓𝑒𝑒 < 0, 𝑓𝑡𝑡 < 0, and 𝑓𝑒𝑡 < 0). Individuals take temperature

and their cognitive skills as given, with the latter being potentially affected by exogenous parental

investments not included in our model. Conversely, their effort is chosen endogenously to solve a

maximization problem (described below). In our empirical analysis, we proxy human capital with

1We closely follow Park (2022), although we have two main differences with his model: (a) we add cognitive
skills, which are not only important determinants of academic performance but also useful for interpretations in the
presence of heterogeneous effects by socioeconomic status; and (b), given our focus on the incentives to accumulate
human capital and their interaction with temperature stress, we derive comparative statics not only for temperature but
also for the returns to human capital.

62

students’ performance in the Colombian high-school exit exam to be described below.

Individuals derive instant utility according to 𝑈 (𝑐, 𝑔(𝑒, 𝑡)), which depends positively on the

consumption of a composite good 𝑐, whose price we normalize to one, and negatively on the

effective effort 𝑔. The latter is a function increasing in all of its arguments (𝑔𝑒 > 0, 𝑔𝑡 > 0), and

with positive second and cross derivatives (𝑔𝑒𝑒 > 0, 𝑔𝑡𝑡 > 0, and 𝑔𝑒𝑡 > 0). This last assumption

is consistent with the medical literature suggesting that effort is more physically taxing at higher

temperatures (Wendt, Van Loon, & Marken Lichtenbelt, 2007).

Given that there is only one period, individuals consume all their labor income (i.e., there are no

savings). Assuming no unemployment and that wages 𝑤 are determined in a perfectly competitive

labor market, we can write the individual’s problem as follows:

where the restriction corresponds to the standard budget constraint. The latter means that the agent

max
𝑐,𝑒

𝑈 (𝑐, 𝑔(𝑒, 𝑡))

s.t. 𝑤ℎ = 𝑐,

(2.2)

chooses its consumption and effort levels to maximize utility subject to its labor income. Plugging

equation (2.1) in the above problem, we can derive the first-order condition
𝜕ℎ
𝜕𝑒∗ = 𝑤𝑧

𝜕𝑈
𝜕 𝑓
which states that, at the optimum, the agent equalizes the marginal benefit of effort (increase in

𝜕 𝑓
𝜕𝑒∗ = −

𝜕𝑔
𝜕𝑒∗

𝜕𝑈
𝜕𝑔

𝜕𝑈
𝜕ℎ

𝑤

,

disposable income) to its marginal cost (increase in disutility).

Setting up our analysis with specific functions 𝑓 , 𝑔, and 𝑈, we can derive closed-form solutions

for the optimal levels of effort and human capital. In particular, we assume the following:

𝑓 (𝑒, 𝑡) =

,

0 < 𝛼 < 1,

𝛽 > 0;

𝑒𝛼
𝑡 𝛽

𝑔(𝑒, 𝑡) = 𝑒𝛾𝑡 𝜁 ,

𝛾 > 1,

0 < 𝜁 < 1;

𝑈 (𝑐, 𝑔(𝑒, 𝑡)) = 𝑐 − 𝑔(𝑒, 𝑡) = 𝑤𝑧 𝑓 (𝑒, 𝑡) − 𝑔(𝑒, 𝑡).

All of these functional forms satisfy the assumptions above. The stated restrictions ensure

the objective function in (2.2) is well-behaved (first- and second-order conditions hold).2 Solving

the maximization problem, it is straightforward to show that the optimal level of effort takes the

2One of the key assumptions required for the objective function to be concave in the effort is that 𝛾 > 𝛼, this is,

the disutility cost of effort increases faster than its effect on human capital (see Appendix 2A).

63

following form:

𝑒∗ =

(cid:16) 𝛼𝑤𝑧
𝛾𝑡 𝛽+𝜁

(cid:17) 1
𝛾− 𝛼 .

(2.3)

From this expression is clear that effort is increasing in wages and cognitive skills, while it

decreases with temperature. The first two results are standard in the economics literature (see, for

example, Chadi, De Pinto, and Schultze (2019)). The third one is consistent with recent empirical

evidence, including our empirical work with time-use data below, suggesting that temperature

affects the time allocated to working and studying (Alberto, Jiao, & Zhang, 2021; Graff-Zivin &

Neidell, 2014).

Plugging the expression for the optimal effort in the human capital production function and

linearizing, we can get an equation akin to what is usually estimated in the empirical literature,

including our analysis below.3 Concretely,

(cid:20)

ˆℎ =

−𝛽 −

(cid:21)

𝛼(𝛽 + 𝜁)
𝛾 − 𝛼

ˆ𝑡 +

ˆ𝑤 +

(cid:21)

(cid:20) 𝛼
𝛾 − 𝛼
𝛼(𝛽 + 𝜁)
𝛾 − 𝛼

(cid:20) 𝛾
𝛾 − 𝛼

(cid:21) (cid:20) 𝛼

𝛾 − 𝛼

(cid:21)

(cid:21)

ˆ𝑧

ˆ𝑡 ˆ𝑤 +

(cid:20)

1
2

−𝛽 −

𝛼(𝛽 + 𝜁)
𝛾 − 𝛼

(cid:21) (cid:20) 𝛾

(cid:21)

𝛾 − 𝛼

ˆ𝑡 ˆ𝑧 + 𝑅,

(2.4)

(cid:20)

1
2

+

−𝛽 −

where ˆ𝑥 is the percent deviation of 𝑥 with respect to the population (or historical) mean and 𝑅 is a

residual term composed, among others, of the second-order own effects of each of the variables.4

As can be seen, a positive deviation in cognitive skills or wages is expected to increase human

capital. Conversely, higher-than-average temperatures are expected to decrease it. As in Park

(2022), the effect of temperature operates through two channels: a direct effect summarized by the

(structural) elasticity of human capital to temperature (𝛽) and an indirect one operating through

endogenous changes in effort resulting from temperature increases, namely 𝛼(𝛽 + 𝜁)/(𝛾 − 𝛼). In

most empirical setups, individuals’ efforts are unobserved or poorly measured. Nevertheless, to

identify this effect in (2.4) all we require is a shock to (expected) wages, namely ˆ𝑤, which in our

empirical application corresponds to a policy intervention to be explained in the following section.

Given our interest in the interaction between incentives and changes in temperature stress, our

expression for optimal human capital includes the cross-effect of wages and temperature. Now,

3The non-linearized expression for human capital and the full mathematical derivation of the optimization problem

can be found in Appendix 2A.

4By writing the variables in deviations we account for controls such as geography or time fixed-effects, which are

usual in empirical work, including ours.

64

notice that in our model an increase in expected wages incentivizes individuals’ efforts, which is

expected to increase the accumulation of human capital (first-order effect). However, exerting effort

also exposes individuals to the effects of temperature and associated reductions in human capital

(interaction effect). This last quantity is precisely what is captured by the fourth term in (2.4),

which is indeed negative. As we will show in our empirical analysis, we find evidence in favor of

this prediction.

Another effect of interest corresponds to the interaction between cognitive ability and tempera-

ture (fifth term in (2.4)). Our framework implies that this interaction should be negative as long as

individuals’ cognitive skills incentivize the investment in human capital. This is, given that individ-

uals from better socioeconomic backgrounds end up exerting more effort, they are further exposed

to temperature stress. We provide suggestive empirical evidence in favor of this implication.

2.3 Background and Data

2.3.1 Policy intervention: Ser Pilo Paga

In 2014, the Colombian Ministry of Education launched the Ser Pilo Paga program (“Being a

Good Student Pays Off”) to promote access to higher education. SPP was a scholarship initiative

that provided financial support to high-achieving students from low-income backgrounds, enabling

them to attend top universities in the country. The program covered the total cost of tuition to attend

a four-year (or five-year) degree-granting program at any high-quality university in Colombia. The

program provided individual funding to about 40,000 students, with 10,000 new students chosen

each year from Colombia’s top decile of the population in the high-school exit exam (Bernal &

Penney, 2019; Londoño-Vélez, Rodríguez, & Sánchez, 2020; Londoño-Vélez, Rodriguez, Sánchez,

& Álvarez-Arango, 2023; Medina et al., 2018).

Given the program’s objectives, applicants had to meet two criteria based on merit and need

to be eligible. First, they must score at or above the 90th percentile on the national standardized

high-school exit test (Saber 11). Second, they must come from disadvantaged households, defined

as those below the median of the wealth distribution in Colombia according to their SISBEN score,

the socioeconomic index used by the Colombian Government to target social programs. Finally, to

65

receive the award, students needed to secure admission to a high-quality accredited university in

the country.5

There is now abundant literature evaluating the SPP program and its effects on access to higher

education, graduation, learning, and wages in the short-and the medium-run (Bernal & Penney,

2019; Laajaj, Moya, & Sánchez, 2022; Londoño-Vélez, 2022; Londoño-Vélez et al., 2020, 2023).

For instance, SPP increased the likelihood of immediate access to higher education among eligible

students by eliminating the gap in access by socioeconomic strata among the top performers in the

Saber 11 (Londoño-Vélez et al., 2020). Relatedly, Bernal and Penney (2019) and Laajaj et al. (2022)

argue that SPP increased effort as measured by improvements in the test scores of students with

higher exposures to the program. On the negative side, other studies highlight that SPP allowed the

concentration of resources in private universities, the financial weakening of public universities,

and the reproduction of regional and social inequalities (Mora & Ruiz, 2019).

Instead of evaluating the effects of SPP on students’ performance in the Saber 11 test, which

others have already done (Bernal & Penney, 2019; Laajaj et al., 2022), we ask the following question:

Did the increased effort resulting from SPP lead to a larger effect of temperature on students’ test

scores? This question speaks to two broad and important topics in the education and climate change

literature. First, our results provide evidence of the interaction between environmental factors and

education policies, in our case, a large and generous scholarship program. Second, we show that

temperature stress increases the cost of effort and, therefore, that further increases in temperature

from climate change may end up reducing effort levels in equilibrium. To the best of our knowledge,

we are the first to provide causal evidence on this adaptation behavior.

2.3.2 Data

Our outcome of interest is the students’ results in the Colombian high school exit exam known as

Saber 11. This test has been mandatory since 1980 (compliance rate is around 97%) and is designed

to support universities and all other higher education institutions in their admission processes.

5As defined by the Operating Regulations of SPP by the Colombian Ministry of Education, in a case in which the
number of eligible students exceeds the funds allotted to the program, priority would be given to those with the lowest
SISBEN scores.

66

Saber 11 is administered semiannually on mostly fixed dates6 (usually in March and August) by the

Colombian Institute for the Evaluation of Education (ICFES). Because of methodological changes

in the test that limit the comparability of the scores across time, we use the information from the

period between 2014 (August) to 2019 (March), with each calendar year having information on

several hundreds of thousands of students. Since Saber 11 serves as a comprehensive evaluation of

knowledge and competencies acquired in school, we take it as a proxy variable of human capital

accumulation.

For weather, we use station-level data on daily maximum temperatures and total precipitation

from the Colombian Institute of Hydrology, Meteorology, and Environmental Studies (IDEAM)–

the national official meteorological data. This information has been available since 1970, but we

restrict our analysis to the period 2013-2019. In total, we use temperature data from 143 weather

stations located all over the country. Following Auffhammer and Kellogg (2011), we impute

missing daily observations with those from the five nearest weather stations with non-missing

data. We assign to each Municipality the data for the nearest weather station with information for

the full period, which allows us to address endogeneity in the temporal coverage and quality of

the data (Auffhammer, Hsiang, Schlenker, & Sobel, 2013). Matching the weather stations to the

Municipalities has the additional advantage of making it clear the source of variation of the weather

variables in the analysis, something that is less straightforward in reanalysis or gridded data (Dell

et al., 2014).7 Figure 2B.1 in Appendix 2B shows the location of the weather stations used in our

analysis, including those used in the imputation procedure only.8

As will be clear below, our empirical results point to important differences between students in

urban and rural areas. To explore the mechanisms behind these findings, we use information on

6Although the dates of the test change from year to year, each of the two annual tests occurs on a single day, within

the same time frame.

7In Table 2B.4 in the Appendix 2B, we show that our results are robust to dropping from the sample Municipalities

matched to weather stations farther than 25-kilometer or that relied more heavily on the imputation procedure.

8The correlation between the temperature variable in our main analysis and the ones obtained using Copernicus’
ERA5 is 0.74. We confirmed that our point estimates based on IDEAM data were within the 95% confidence intervals
for baseline estimates using ERA5. Nonetheless, estimates using weather-station data revealed higher precision, hence,
we chose them as the preferred data source for the entire analysis.

67

time use from the Colombian Longitudinal Survey (ELCA), available for 2010, 2013, and 2016.9

In particular, we employ information for youths between 10 and 16 years of age in urban and rural

areas, available in the last two waves of the survey. The question for the youth asks how much time

they spent on activities such as watching TV, reading, or doing homework on a typical weekday,

with the answer options being categorical (nothing, less than one hour, between 1 and 2 hours,

etc.).

To consider potential mechanisms driven by rural non-farm employment, we also consider

time-use data on the household heads and their spouses in rural areas from the three waves of

ELCA. In this case, the question asks for the exact time (hours and minutes) spent on activities such

as working on farms owned by the household, working for other households, or doing household

chores, with the reference period being a typical weekday in the week before the survey. In all

our analyses, we focus on a repeated cross-sectional approach to keep consistency with our main

results using students’ scores.10

Figures 2.1 and 2.2 show the average daily maximum temperature for the period 1981-2010,

which is a measure of climate (Dell et al., 2014), and the average standardized Saber 11 score in

Math for the exams between 2014-II and 2019-I, respectively. Three results stand out: Firstly, there

are high cross-sectional variations in the maximum temperature across counties, with an important

number having temperatures as low as 12°C and some as high as 32°C. Secondly, the results for

the Math section of the Saber 11 test also show substantial spatial variation. And thirdly, as also

evidenced in Figure 2.3, there is a remarkable correlation between temperatures and test scores,

with most high-performing counties concentrated in the relatively colder parts of the country.

We present the descriptive statistics of the data for our analysis in Table 2.1.11 Each specific

9ELCA data come from a probabilistic, stratified, multistage, and clustered sample, with information on 10,800
households (6,000 in urban areas and 4,800 in rural areas). The urban sample is representative of low- and middle-
income households at the national level and for five regions (Bogota, Central, Eastern, Atlantic, and Pacific). The rural
sample is representative of small farmers in four micro-regions (Mid-Atlantic, Coffee growing region, Cundiboyacense
Plateau, and Center-East).

10The data on time use in the ELCA have been used previously in the economics literature. In particular, Fernández,
Ibáñez, and Peña (2014) study how the Colombian armed conflict affects time-use patterns among rural households
using the first wave of the ELCA.

11The descriptive statistics for ELCA data are available in Appendix 2B, Table 2B.8.

68

area evaluated by Saber 11 allows for a score between 0 and 100,12 while the total score (as defined

by ICFES) will be between 0 and 500. On average, students in urban areas consistently outscore

those in rural areas by 4.8, 4.4, and 22.9 points in Math, Spanish, and total scores, respectively.13

Based on the test theoretical standard deviations for each area (10 points) and the total score (50

points), these amount to a difference roughly equivalent to 50% of a standard deviation. Most of the

evaluated students (55%) are female, while 6% are reported as part of an ethnic minority. In terms

of the education of the student’s mother, 29% have completed only primary education, while 33%

and 23% have completed their studies in secondary and postsecondary education, respectively.

Consistent with the climate normal in Figure 2.1, the average daily maximum temperature

experienced at the Municipalities in the calendar year before the exam date was 26.6°C, while

the median was at 27.8°C. These are close to the measures of maximum temperature on the days

when the exam occurred, which have a mean of 26.87 and a median of 28 degrees Celsius. In

terms of temperature experienced by weather bins, on average, municipalities experienced 6 days

with maximum temperatures of 15°C or below, while the number of days with temperatures over

35°C was 30.7. There is a considerable level of exposition to heat at the Municipality level, with

an average of 115.2 days experiencing peak temperatures between 30 and 35°C. In addition, the

average daily precipitation during the calendar year before the exam date was 5.52 millimeters per

day, while it was 4.66 millimeters during the exam days. Finally, in terms of plausibly exogenous

policy shocks, there were between 8 and 9 beneficiaries of the first round of SPP per Municipality,

although 50 percent of these Municipalities only experienced up to two beneficiaries.

12Saber 11 includes five areas of evaluation: Math, Spanish, natural sciences, social competencies, and English
(as a foreign language). We limit our main analysis to Math and Spanish (i.e., reading) since these areas are the most
widely studied in the literature, thus allowing for a direct comparison of effects relative to the ones in previous work.
13We classify a municipality as urban if the share of the population in the urban area was above the national average

for most of the years during 2004-2013.

69

2.4 Methods

2.4.1

Identification strategy

To study the average effect of heat on students’ test scores, we follow a linear unobserved effects

model:

100 × 𝑦𝑖 𝑗 𝑐𝑡 = 𝛽𝑇𝑇𝑐𝑡 + 𝛽𝑃𝑃𝑐𝑡 + 𝑋𝑖 𝑗 𝑐𝑡𝜃 + 𝑎 𝑗 + 𝑎𝑠𝑡 + 𝑢𝑖 𝑗 𝑐𝑡,

(2.5)

where 𝑦𝑖 𝑗 𝑐𝑡 is the standardized score in Math (or Spanish, or total) in the Saber 11 test for student

𝑖 in school 𝑗 and Municipality 𝑐 at time 𝑡 (semester in our case). Here 𝑋𝑖 𝑗 𝑐𝑡 is a vector of student

characteristics (e.g., age and gender), while 𝑎 𝑗 and 𝑎𝑠𝑡 are fixed effects that capture unobserved

heterogeneities at the school and State-time levels, and 𝑒𝑖 𝑗 𝑐𝑡 is a set of idiosyncratic shocks that

affect students performance in the test. The variable of interest is 𝑇𝑐𝑡, which is the average daily

maximum temperature (in °C) experienced during the year before the test date. As an additional

control, we include 𝑃𝑐𝑡, which measures the precipitation level in the same period as 𝑇𝑐𝑡. With

this variable, we aim to control for disruptions in learning caused by above-average rain levels.

The coefficient of interest is 𝛽𝑇 , which measures the effect on the student’s scores of experiencing

an increase in temperature. Specifically, a 1°C increase from the school average would change

scores by 𝛽𝑇 % of a standard deviation. We follow a fixed effects framework, thus allowing

for arbitrary dependence between the unobserved effects (𝑎 𝑗 , 𝑎𝑠𝑡) and the observed explanatory

variables (Wooldridge, 2010). Given that our temperature variable varies at the weather-station

level, following the recommendations in Angrist and Pischke (2008), MacKinnon, Nielsen, and

Webb (2022), and Abadie, Athey, Imbens, and Wooldridge (2023), we cluster our standard errors at

that level to allow for arbitrary correlation between schools (and students) within the area covered

by the station.

The identification assumption behind equation (2.5) is that conditional on students’ demographic

characteristics and school and State-time fixed effects, experiencing a hotter calendar year before

the test is not systematically correlated with other determinants of the student’s scores (Angrist &

Pischke, 2008). In other words, temperatures are as good as random conditional on the controls

(Dell et al., 2014). Following Hsiang (2016), we implement a series of measures to rule out several

70

threats to the identification: First, we control for precipitations to account for other potentially

correlated geophysical processes. Second, we compute temperature and precipitation using a

constant group of weather stations, which rules out any correlation between our weather data and

time-varying socioeconomic conditions. Third, we provide evidence suggesting that temperature

does not seem to have any noticeable effects on our sample composition (see next subsection).

Following Hsiang (2016), we also estimate the following flexible specification using our daily

weather data:

100 × 𝑦𝑖 𝑗 𝑐𝑡 =

𝐾
∑︁

𝛽𝑇,𝑘 𝐷𝐷 𝑘

𝑐𝑡 + 𝛽𝑃𝑃𝑐𝑡 + 𝑋𝑖 𝑗 𝑐𝑡𝜃 + 𝑎 𝑗 + 𝑎𝑠𝑡 + 𝑢𝑖 𝑗 𝑐𝑡,

(2.6)

where 𝐷𝐷 𝑘

𝑐𝑡 is the number of days in the period of interest where the maximum temperature in

𝑘=1

°C was in the interval 𝑘. When estimating this specification, we use the following temperature

intervals: [minimum, 15), [15, 20), [20, 25), [25, 30), [30, 35), and [35, maximum). The baseline

category throughout the analysis is the number of days between 20 and 25°C; thus, 𝛽𝑇,𝑘 measures

the effect of being exposed to one additional day with a maximum temperature in the 𝑘th bin with

respect to the omitted category.

One of the main goals of this paper is to investigate how changes in the incentives to achieve

higher scores in the Saber 11 interact with higher temperatures. For that, we evaluate how the effect

of heat varies for students exposed to SPP. In particular, we estimate the following specification

100 × 𝑦𝑖 𝑗 𝑐𝑡 = 𝛽𝑇𝑇𝑐𝑡 + 𝛽𝑇,𝑆𝑃𝑃 (𝑇𝑐𝑡 × 𝑆𝑃𝑃𝑐 × 𝑃𝑜𝑠𝑡𝑡) + 𝛽𝑆𝑃𝑃 (𝑆𝑃𝑃𝑐 × 𝑃𝑜𝑠𝑡𝑡)

+ 𝛽𝑃𝑃𝑐𝑡 + 𝑋𝑖 𝑗 𝑐𝑡𝜃 + 𝑎 𝑗 + 𝑎𝑠𝑡 + 𝑢𝑖 𝑗 𝑐𝑡,

(2.7)

where 𝑆𝑃𝑃𝑐 is the share of test-takers in Municipality 𝑐 that were awarded SPP in the first wave of

the program, and 𝑃𝑜𝑠𝑡𝑡 corresponds to a dummy for the post-SPP years. Notice then the interaction

between 𝑆𝑃𝑃𝑐 and 𝑃𝑜𝑠𝑡𝑡 captures the exposure to SPP originated only by the first cohort of winners,

which were selected based on pre-determined household characteristics (SISBEN wealth index)

and realized performance in the test at the moment of the program’s announcement. The coefficient

of interest is 𝛽𝑇,𝑆𝑃𝑃, which corresponds to the differential effect of temperature for students with

higher exposures to SPP.14 From our conceptual framework, we expect this coefficient to be negative

14Notice that any average differences in students’ scores associated with 𝑆𝑃𝑃𝑐 are already captured by the school

71

as it maps directly to the fourth term of equation (2.4).

In addition to the identifying assumption for the temperature’s effect previously discussed,

equation (2.7) requires that the evolution in the students’ scores for different levels of exposure to

SPP would have been the same in the counterfactual without the program, this is, that there are

parallel trends for the groups defined in terms of 𝑆𝑃𝑃𝑐 (Angrist & Pischke, 2008). As we already

mentioned, we compute this exposure variable using only the first cohort of the program, which

ensures we do not capture endogenous responses in the number of winners. Now, to explore the

possibility that the SPP’s exposure is capturing the impact of other policies or social programs,

the next subsection provides evidence on pre-trends for several dimensions arguably relevant to

students’ performance in the test.

2.4.2 Testing for identification assumptions

In our analysis, we attempt to exploit two sources of exogenous variation, namely the variations

in Municipality-level temperature and the exposure to SPP. Thus, we need to evaluate whether

(a) the weather correlates with attributes from students,15 and (b) the shock from SPP has some

correlation with pre-intervention attributes at the Municipality level. For the first one, we set up a

linear model:

𝑋𝑖 𝑗 𝑐𝑡 = 𝜂𝑇𝑇𝑐𝑡 + 𝜂𝑃𝑃𝑐𝑡 + 𝑎 𝑗 + 𝑎𝑠𝑡 + 𝑢𝑖 𝑗 𝑐𝑡,

from which we would expect that fixed-effect estimates for 𝜂𝑇 and 𝜂𝑃 are statistical zeroes. In other

words, we are checking if weather shocks are directly connected to attributes of individuals that can

also affect their performance in the test. We estimate this specification for six attributes: age, female

student, dummies of mother’s educational attainment–primary, secondary, or postsecondary-, and

whether the student is from an ethnic minority. Our results suggest that there is no evidence

of any statistically significant correlation between weather shocks (either from temperature or

precipitation) with students’ attributes (see Appendix 2B, Table 2B.1). Moreover, the estimated

coefficients are very close to zero, further providing support for using these weather shocks as

fixed effects (𝑎 𝑗 ). Similarly, the post-SPP indicator (𝑃𝑜𝑠𝑡𝑡 ) is captured by the State-time fixed effects (𝑎𝑠𝑡 ).

15A large literature in economics has relied on the assumption that temperature is exogenous conditional of unit
and time fixed effects (see Dell et al. (2014) and Hsiang (2016)). For the sake of completeness, we provide evidence
in favor of it in our context.

72

exogenous variations directly affecting students’ learning and performance.

We now consider the implicit assumption of parallel trends underlying our empirical strategy

of using the announcement of SPP. The assumption of parallel trends is theoretically untestable

and the practice of pretesting has recently faced considerable scrutiny (Roth, 2022). Nonetheless,

results from pretesting can provide relevant information on potential pre-treatment differences

that can be empirically considered to still achieve meaningful difference-in-difference estimates

(Bilinski & Hatfield, 2018). Following the literature on the economics of education, we focus on

school characteristics that have been thought to be important in explaining students’ performance.16

Given that these variables are also potentially endogenous to policy changes, controlling for them

allows us to account for the possibility that SPP assignment rules can turn other social programs

into confounders. Therefore, we proceed to analyze any potential differences in pre-trends at the

Municipality level based on the exposure to SPP, using pre-treatment information available from

the Municipality panel data from the Center of Studies on Economic Development (CEDE) of

Universidad de los Andes (Colombia).

With yearly information for the 2010-2019 period on six key variables (i.e., number of schools,

population of high-school age, the number of high-school students, total Saber 11 test takers, total

educational administrative personnel, and the total number of teachers), we follow an event-study

approach. We regress each of these variables on a saturated model where we interact all year

dummies (except 2014, the base year of our analysis) with the exposure to SPP in the first round.

Figure 2B.2 of Appendix 2B summarizes the point estimates and 95% confidence intervals. We

find evidence of a potential violation of the parallel trends assumption at the Municipality level on

variables like the total number of teachers and total high-school students. Meanwhile, there is a

weakly significant pre-trend correlation between larger exposures to SPP and the number of Saber

11 takers in Municipalities.17

16As discussed in Hanushek (2006) and Handel and Hanushek (2022), the effect of school resources on students’
test scores seems to be modest at best. However, recent papers have found evidence of positive effects for specific
inputs such as teachers (Chetty, Friedman, & Rockoff, 2014a, 2014b) or even school resources when using exogenous
variation for identification (Jackson, Johnson, & Persico, 2016).

17Notice that finding pre-trends in these Municipality characteristics does not necessarily imply that the same is
happening with students’ scores, the outcome of interest in our empirical analysis. As mentioned earlier, the literature

73

Pretesting at the student level comes with a challenge in data comparability for Saber 11 scores

available before 2014. In order to consolidate Colombia’s National Standardized Evaluation System

(SNEE by its acronym in Spanish), ICFES introduced a significant adjustment of the Saber 11 test,

changing it from a per-subject evaluation to an assessment of generic competencies. For instance,

the pre-2014 Math scores reflect a standardized outcome on a math-only set of questions, while the

Math scores from 2014 and later years reflect competency in both math and quantitative reasoning

questions (ICFES, 2019). Likewise, pre-2014 Spanish scores reflect outcomes on language-only

questions, while subsequent years’ Spanish scores reflect competency in critical reading (including

language and philosophy).18 This difference is also acknowledged in other studies relying on Saber

11 data, including those of Laajaj et al. (2022), Londoño-Vélez et al. (2020), and Posso, Saravia,

and Uribe (2023). In this sense, while score comparability pre- and post-2014 is unfeasible, we

can use data from 2010-2013 to evaluate whether there is evidence of potential pre-trends. We find

statistically significant deviations in Math and Spanish scores for the case of rural areas, but no

systematic patterns or sizable effects in either urban or rural areas that could threaten the validity

of our analysis (Appendix 2B, Figure 2B.3).

To further test our identification assumptions, we follow Pei, Pischke, and Schwandt (2019)

and take students’ comparable attributes to assess whether potential violations of a parallel trends

assumption may occur at the individual level, based on exposure to SPP. Our event-study estimates

suggest that there are no pre-treatment differences in key variables such as eligibility for SPP by

SISBEN category, whether the student is currently working, reported monthly family income, and

number of people living at the student’s dwelling (Appendix 2B, Figure 2B.4). Moreover, we find

that there are no statistically significant changes in the after-treatment period, except for eligibility

based on SPP. Therefore, these results provide evidence in favor of our approach, which focuses on

exposure to the first (unannounced) wave of SPP, since the following waves of the program may

finds weak effects of school resources on students’ performance (Handel & Hanushek, 2022; Hanushek, 2006).

18This also extends to other areas like physics, chemistry, and biology, which reported separate subject-specific
scores before 2014, but were combined into a single competence (and score) in Natural Science afterwards. Further-
more, in addition to the change in the evaluation structure, pre-2014 tests included two additional tests in the same time
window: an in-depth test and an interdisciplinary test. Therefore, there is a high likelihood that the time allocation in
effort will differ, which will exacerbate the issue of non-comparable scores between periods.

74

have been influenced by a systematic and strategic behavior of students requesting an evaluation

from local authorities to be included in SISBEN (Londoño-Vélez et al., 2020).

Nonetheless, following our findings on Municipality-level variation, we follow an approach in

the style of Bilinski and Hatfield (2018), acknowledging the potential differences in pre-trends and

setting “one step up” specifications for our main regressions. This is, we further consider speci-

fications that allow for heterogeneous trends based on Municipality-level differences as measured

in 2014, and confirm whether there are significant differences between estimates from this more

complex specification to those of simple fixed-effects estimates. Controlling for trends along these

dimensions, we expect that our estimates will account for other factors or policies that may affect

the evolution of students’ scores with different exposures to SPP, thus strengthening our causal

interpretation.

2.5 Results and Discussion

2.5.1 Temperature and test scores

We present the results from our baseline regression on weather shocks during the year before

the exam date in Figure 2.4 (with all regression summaries’ in Appendix 2B, Table 2B.2). When

focusing on the estimates for the total population, evidence points to insignificant correlations

between observed scores and increases in the average daily maximum temperature and average

precipitation in the year before the Saber 11 exam. The direction of the coefficient estimates is

similar in Math, Spanish, or total scores, but there are important differences in estimations when

comparing urban and rural areas,19 despite observing similar levels of exposure to temperature

levels between these kinds of Municipalities.20

While exposure to a 1°C increase in the average maximum temperature reduces the average

Math score by 2.89% of a standard deviation for urban students, the coefficient estimate is positive

for rural students. Similarly, while the same temperature increase has no statistically significant

19For reference, we further report the estimates of our baseline specification for the remaining subjects that are
evaluated in Saber 11, namely Natural Sciences, Social Sciences, and English as a Second Language (Appendix 2B,
Figure 2B.5). These estimates align with our main analysis, especially in the stark differences between urban and rural
settings.

20See Appendix 2B Figure 2B.6.

75

effect on Spanish scores for rural students, it reduces urban students’ scores by 2.28% of a standard

deviation. In terms of total scores, each additional one-degree increase decreases the result of urban

students by 2.51% of standard deviation, but it is connected to an increase of 1.43% of a standard

deviation in rural settings. On the other hand, each additional 1 mm/day in precipitation leads to a

reduction of 0.85% and 0.65% of a standard deviation in Math and overall scores, respectively, for

urban students. We do not detect any significant effect from precipitation in rural areas.21 Based

on the projected temperature increases for Colombia over the century, which under RCP 4.5 are

1.5-2.5 °C by mid-century (Alarcón Hincapié & Pabón Caicedo, 2013), the estimated coefficients

from this baseline specification are of economic significance.22

As evidenced by the results Figure 2.5 (Appendix 2B, Table 2B.5), year-long weather effect

estimates are robust to the addition of weather exposure during the day of the exam, suggesting that

our estimates capture the effects on learning–thus in human capital accumulation–and not only a

worsening on performance during the test. Again, we find no significant shocks at the national level,

but marked differences persist between urban and rural areas. For urban students, our calculations

still suggest a reduction in scores (-2.7%, -2.11%, and -2.58% of a standard deviation, respectively,

in Math, Spanish, and total) from a 1°C increase in the average maximum temperature over the

previous year, and an additional millimeter of average daily precipitation still points to a statistically

significant reduction of scores (-0.89% and -0.63% of a standard deviation in Math and total scores,

respectively). Although numerically different, these estimates are statistically equivalent to those

of our baseline estimates. Additional specifications including only the effect of test-day weather

(Appendix 2B, Table 2B.6) provide estimates of similar magnitudes and directions to those from

Figure 2.5 (Appendix 2B, Table 2B.5).23

Next, we study the effects of exposure (by number of days) across different temperature bins

21Our findings are robust to stringent clustering, even when allowing for unrestricted correlation between observa-

tions within clusters of weather stations (Appendix 2B, Table 2B.3).

22Furthermore, our estimates are arguably conservative, given that world-level emissions remain at tracking levels

of RCP 8.5 (Field, Barros, & Intergovernmental Panel on Climate Change, 2014).

23In addition, in Appendix 2B, Figure 2B.7, we test the robustness of the year-long AMT effect but instead using the
average maximum temperature during the week before the exam, which does not change our inference or implications
about the prevalence of longer-term exposure to increasing temperatures.

76

over the calendar year before the exam in Figure 2.6 (Appendix 2B, Table 2B.7). For urban areas’

students, on average, and relative to maximum temperatures between 20 and 25°C, one additional

day exposed to temperatures between 30 and 35°C implies a reduction of 0.055% of a standard

deviation in Math, and reductions of 0.066% and 0.047% of a standard deviation in Spanish and

overall scores. Likewise, one additional day exposed to temperatures above 35°C negatively affects

scores, implying reductions amounting to 0.11%, 0.095%, and 0.099% of standard deviations in

Math, Spanish, and total scores, respectively. Interestingly, it turns out that our point estimates for

the total score fall between 0.05% of a standard deviation in the United States (Park et al., 2020)

and 0.25-0.3% of a standard deviation in India (Garg et al., 2020) for each additional day with

temperatures above 30°C.24

Overall, our findings–specifically those for students in urban areas–provide results in line with

those in the literature. Nevertheless, these results highlight an important empirical difference

between studies in this line depending on the weather variable used in the analysis. For example,

studies like Graff-Zivin et al. (2020), Graff-Zivin et al. (2018), and Park (2022) focus their attention

on the exposure of test takers to high temperatures during exam days. Meanwhile, our findings

are more in line with those like Park et al. (2020) and Garg et al. (2020), who analyze the effects

of exposure to high temperatures on test scores focusing on long-run (e.g., past-year) measures of

exposure.

In line with the literature, we further consider a specification that allows for differences in

the effect of increasing temperatures conditional on the long-run historical average maximum

temperature experienced at the Municipality level. We present the key results in Figure 2.7,

again revealing a marked difference between urban and rural settings. We find that increasing

temperatures have a significant effect on the scores of urban students in Municipalities with long-

run temperatures under 28°C. Moreover, our results suggest that the negative effects of increasing

temperatures are of larger magnitudes among students in colder Municipalities. Conversely, we do

24In addition, following Helo Sarmiento (2023), we test for the robustness of these effects setting up a narrower
variation, with bins increasing by 2°C (Appendix 2B, Figure 2B.8), and pooling exposures to days below 17°C and
above 27°. We find quantitatively equivalent results, although weakly significant.

77

not detect any statistically significant effects for relatively high-heat Municipalities (above 28°C).

All these results go in line with recent evidence highlighting the differentiated effect of increasing

temperatures on cognitive performance between relatively cold and hot areas (Krebs, 2024), and

point to adaptation playing a role in this setting (Barreca, Clay, Deschenes, Greenstone, & Shapiro,

2016; Burke et al., 2016).

Effects of extreme temperature events on the same day of (or week leading to) high-stakes exams

are nonetheless relevant to evidence the need for investments in adaptation strategies (e.g., change

of exam timing in the season). However, our findings, when taken together with Garg et al. (2020),

suggest that developing countries with mostly hot tropical weather are more profoundly affected

in an increasingly warming world. Countries like Colombia and India, where the adoption of air

conditioning remains under 5% of buildings (Statista Research Department, 2023; Sung, 2022), and

that will potentially experience disproportionally harsher climatic conditions (Lenton et al., 2023),

are bound to experience considerable losses in human capital unless bold policy interventions take

place.

Conversely, our estimates for past-year temperature effects on the performance of students

in rural areas suggest that different mechanisms might play a significant role. For instance,

there is a high incidence of child labor in rural Colombia, which is estimated up to 20% for

children of ages between 15-17 by 2017 (Ramoni-Perazzi, Orlandoni-Merli, Castillo-Paredes, &

Peña Guillén, 2021). Hence, increases in average temperatures could be connected to adjustments

in time allocations that could–in theory, following Alberto et al. (2021) and Graff-Zivin and Neidell

(2014)–lead these rural students to allocate more time to study, since they would likely dedicate more

time to indoors activities.25 Below, we provide suggestive evidence in favor of this explanation.

While the estimates from Hoffmann et al. (2023) for short-term weather exposure to temperature

are qualitatively similar to our estimates of exposure to exam-day weather shocks,26 our findings

25Another related explanation is that as higher temperatures decrease agricultural productivity in Colombia
(Melo León et al., 2024), this could potentially lead to a reduction in the opportunity cost of studying (see Shah
and Steinberg (2021) for related evidence from India).

26Point estimates from Hoffmann et al. (2023), however, are statistically different from zero for total scores, unlike
ours (see Appendix 2B, Table 2B.6), that suggests the presence of an impact from heat only on the performance in
Math for urban students. It should be noted, however, that our results are not directly comparable to theirs, given that

78

support our main hypothesis: increases in heat lead to exposures that negatively affect the learning

process and ultimately affect students’ performance in high-stakes evaluations.

2.5.1.1 Temperature and time use

From our analysis so far, it follows that the effect of temperature on students’ performance in

Saber 11 varies markedly between urban and rural areas–negative for the former and slightly positive

for the latter. Here we attempt to uncover the mechanisms behind these heterogeneous effects. In

particular, we argue that temperature increases may incentivize students in rural areas to spend

more time in activities complementary to studying, with the underlying force being a reallocation

of labor towards off-farm labor where the returns to human capital are potentially larger (Agrawal

& Agrawal, 2019; Jolliffe, 2004). Rural non-farm employment accounts for over 50% of rural

households in Asia and Latin America, and it has been identified to work as a response from

these households to external shocks on production (Davis, Winters, Reardon, & Stamoulis, 2009;

Haggblade, Hazell, & Reardon, 2007; Reardon, Stamoulis, & Pingali, 2007). Hence, increases

in temperature would potentially push for a reduction of time allocated to on-farm labor, and

instead transitioning into rural non-farm employment. Moreover, evidence from Africa and Asia

suggests that as food production and processing tasks become further laborious or comparatively

less efficient, there is an additional push on spouses to allocate more time in rural non-farm (or

even urban) jobs (Reardon et al., 2019).

To test these hypotheses, we use data from ELCA to estimate specifications that follow:

𝑦𝑖𝑐𝑡 = 𝛽𝑇𝑇𝑐𝑡 + 𝛽𝑃𝑃𝑐𝑡 + 𝑋𝑖𝑐𝑡𝜃 + 𝑎𝑐 + 𝑎𝑚 + 𝑢𝑖𝑐𝑡,

where 𝑦𝑖𝑐𝑡 is a time-use indicator for individual 𝑖 in Municipality 𝑐 at period 𝑡, 𝑋𝑖𝑐𝑡 is a vector of

individual characteristics (age and gender), 𝑎𝑖 and 𝑎𝑚 are Municipality and month-by-year fixed-

effects, respectively, and 𝑒𝑖𝑐𝑡 is a set of unobservables that affect time-use patterns. We estimate

the above expression with two subsamples: First, we employ data for youths 10 to 16 years of age

in the last two waves of the ELCA. For them, we know how much time they spend on activities

we focus on the period 2014-2019 to ensure test scores are fully comparable given changes in the test methodology as
discussed previously. Meanwhile, Hoffmann et al. (2023) use data for the period 2009-2019 instead.

79

such as watching TV or reading for fun on a typical weekday.27. Given that these variables are

reported in bins, we construct indicators equal to one if the youth spends one hour or more on each

of the activities of interest. Second, we use data from the household heads and their spouses for all

the years in the survey. In this case, we have information about the time (hours and minutes) spent

on activities such as working on farms owned by the household, working on other households’

farms in agricultural and non-agricultural activities (off-farm labor), and doing housework. With

this information, we can compute the share of time each household member spends on different

activities.

As before, we are interested in the effect of temperature on the outcome 𝑦𝑖𝑐𝑡, which here corre-

sponds to time-use indicators. Given the reference period of the time-use variables (typical weekday

in the week before the survey), we recompute our temperature variable (𝑇𝑐𝑡) for each household

to be the average daily maximum temperature in the reference week excluding weekends.28 For

consistency, we also recompute our precipitation variable (𝑃𝑐𝑡) to match this reference period.

Hence, the identification assumption in the current setup is that temperature is as good as random

conditional on the controls (Dell et al., 2014), and thus, we include Municipality and month-by-year

fixed effects in the specification above. Given that the time-use indicators correspond to flows,

we match the timing of our weather variables to that of the time-use variables. In the appendix,

however, we show that our results are qualitatively similar when we compute 𝑇𝑐𝑡 and 𝑃𝑐𝑡 for the

year before the survey (see Appendix 2B, Tables 2B.9 and 2B.10).

Table 2.2 shows the results using information for the youths, reporting separately the cases

of rural and urban samples in Panels A and B, respectively. As can be seen, we find that a

higher temperature leads to a statistically significant increase in the percentage of individuals in

rural areas that spend one hour or more reading. We do not find evidence for changes in any

of the other activities, such as watching TV, spending time with parents or siblings, or doing

homework. Interestingly, we do not find any statistically significant effects for individuals in the

27None of the options corresponds to on-farm or off-farm labor, which means that we are not able to test for changes

in labor supply for this population group.

28Notice that the reference week is potentially household-specific. To simplify the notation, the time index 𝑡 can be

thought to correspond to the reference week instead of the semester as in equations (2.5)-(2.7).

80

urban areas, including reading. Overall, our results suggest that students in rural and urban areas

react differently to temperature stress, with the latter increasing the time spent on an activity

potentially complementary to studying.

Even though the results so far point to changes in time use consistent with improvements

in students’ academic performance in rural areas, it is unclear what mechanisms are behind our

findings. One possibility is that temperature stress incentivizes off-farm work, which may have

higher returns to education overall. To test for this possibility, we estimate the model above with

time-use data for adults in rural areas. We report the results for this exercise in Table 2.3, with panel

A having the estimates for the household heads and panel B for their spouses.29 As can be seen,

our results suggest that an increase in temperature leads to a reallocation towards off-farm labor.

In particular, we find that household heads (spouses) decrease (increase) the time spent working

on farms owned by the household (other households) when temperatures are higher. The latter

goes in line with findings across Africa and Asia, pointing to exogenous shocks in production and

productive efficiency, including temperature, as a push for reallocating from on-farm labor out to

non-farm activities (Colmer, 2021; Haggblade et al., 2007; Reardon et al., 2019). In summary, we

interpret these findings as suggestive of a shift in labor allocation in favor of activities potentially

more intensive in human capital for rural settings.30

2.5.2

Incentives and temperature stress

Following our conceptual framework, exerting effort is costly as it exposes individuals to

the effects of increasing temperatures. We use the exposure to SPP as an exogenous shock that

increases expected wages, thus leading to potential increases in test scores, but with negative

interaction effects with temperature.

In table 2.4, we summarize our estimates for equation (2.7). Besides a few small changes

in point estimates related to precipitation, estimates of the first-order effects from the average

29We estimate different models for these two groups given the stark differences in the activities each of them seems

to engage in, as reported in ELCA data (Appendix 2B Table 2B.8).

30Further disaggregation at quarterly-specific temperature effects aligns with our working hypothesis (Appendix
2B, Table 2B.11). With the largest share of students taking Saber 11 in August of every year, positive and statistically
significant effects are found in the first and second quarters (September-February), which are coincidentally those of
the main agricultural season of Colombia.

81

maximum temperature in the year before the exam remain statistically equivalent to those of our

baseline specifications (Appendix 2B, Tables 2B.2 and 2B.5). Meanwhile, the first-order effects of

SPP are consistently positive across all scores, yet only statistically different from zero for Math

(all populations) and total scores (national level and urban students). On average, an additional

1 percentage point increase in the exposure to SPP at the Municipality level led to an increase

of 3.03% of a standard deviation in Math scores across all students. We estimate this effect to

be 6.12% and 2.12% of a standard deviation for urban and rural students, respectively. For total

scores, we detect an average increase of 1.43% and 3.19% of a standard deviation for all students

and urban students, respectively, due to an additional 1 percentage point exposure to SPP.31 These

coefficients are aligned with the evidence in Bernal and Penney (2019) and Laajaj et al. (2022).

All but one of the estimated interaction effects follow the expected sign from our theoretical

model, which is a piece of initial evidence in favor of our hypothesis of unobserved costs of exerting

effort in a warming climate. Nonetheless, we only find this effect to be statistically significant in

the case of Math scores of urban students, which is precisely where SPP had the strongest effect on

performance. On average, when a 1 percentage point increase in exposure to SPP takes place along

with a 1°C increase in the average maximum temperature over the year before the Saber 11 test,

there is a reduction of 0.108% of a standard deviation in the Math section.32 To put these numbers

into perspective, our point estimates in column (2) suggest the temperature effect for students in

municipalities with exposures in the 25th and 75th percentiles are -2.267 and -2.537, a 11.9%

increase when going from the former to the latter (i.e., an interquartile change).33

Now, we can interpret the previous results in an alternative way. Notice that the negative

coefficient for the interaction between temperature and the exposure to SPP implies that heat

31We also find similar results for other subjects evaluated in Saber 11 (Appendix 2B, Table 2B.13), showing again
that SPP had a seemingly stronger effect in increasing scores among urban students, but mainly for the areas of Social
Science and English as a Second Language. Likewise, the interaction between SPP and increasing temperatures is only
significant in the case of urban areas but is now centered on Social Science scores.

32Based on the findings of Graff-Zivin et al. (2018), effects from heat shocks are more likely to be of larger
magnitudes in Math tests, as thermal insults primarily affect the performance of brain functions for tasks that are more
complex than the average–which is arguably the case of Math vis-à-vis other fields in standardized evaluations.

33This number can be calculated as [-0.108×(2.5-0)]/[-2.267]=11.9%, where 2.5 and 0 are the 75th and 25th

percentiles of SPP exposure, respectively.

82

reduced the effect on test scores of being exposed to the program. Using the descriptive statistics

in Table 2.1 and the estimated coefficients in Table 2.4, it follows that increasing the temperature in

a Municipality by a standard deviation reduces the effect of being exposed to SPP on Math scores

by around 10.9%. Our estimates, then, suggest that the effectiveness of social and educational

programs may be compromised by environmental factors such as temperature and, therefore, that

climate change can have important consequences for the design and implementation of educational

policy.

As discussed in the previous section, our estimation includes interactions of yearly dummies

with the Municipality-level dummies of quartiles of (a) the population of high-school age in

the municipality, (b) the total administrative personnel, and (c) the total number of teachers, at

2014 values (our available pre-treatment period). We evaluate the robustness of our findings by

comparing this “one step up” fixed effects estimation (Bilinski & Hatfield, 2018), with a standard

fixed-effects estimation excluding said interactions. Estimates from usual fixed-effects estimation

are statistically equivalent, although slightly less precise, under this alternate specification (see

Appendix 2B, Table 2B.12), thus providing evidence in favor of our overall empirical strategy.34

2.5.2.1 SPP targeted population as a source of heterogeneity

Based on the results above, and following our discussion about SPP, it is important to explore

whether these interaction effects between temperature and incentives on effort vary across specific

populations. While the previous findings indicate that SPP had larger direct and second-order

(interaction) effects among urban students, we now focus our attention on who among those urban

students was more largely impacted by the increased exposure to changes in temperature. Based

on the eligibility criteria, we explore two main heterogeneities of interest, namely (a) the students’

mothers’ educational attainment, and (b) the reported strata of the house where a student’s family

34We further explored an alternate specification to that of Equation (2.6), which considers how the SPP policy
shock affected the effect of being exposed to bins of temperature. We consider a specification such that the coefficients
of interest are interactions between exposure to temperature bins and SPP. These would capture how, on average, the
exposure to SPP alters the effect of experiencing an additional day with average maximum temperatures in the 𝑘-th
bin, relative to days with an AMT between 20 and 25°C. We present the estimates of these interactions in Appendix 2B
Figure 2B.9, with results mostly in line with our main findings. In brief, evidence suggests the SPP shock seemingly
improves the favorable effect of experiencing additional colder days.

83

lives.35

Parents’ educational attainments are well known to be significant determinants of their off-

spring’s own educational achievements and human capital accumulation (Björklund & Salvanes,

2011). It is such gaps that programs like SPP attempt to potentially address, by providing a window

of opportunity to those in less favored backgrounds. In this sense, and in line with our conceptual

framework, we would predict two stark differences: (a) exposition to SPP should have a larger effect

among exam takers whose mothers have lower levels of education–i.e., a more significant change

in incentives; and, as a result of this (b) the interaction effects of temperature and SPP exposure

should also be larger (in absolute value) for students whose mothers have lower educational levels.

To this purpose, we estimate our model in equation (2.7) on two sub-samples of urban exam-takers:

those whose mothers have up to primary education, and those with secondary or post-secondary

education.36

We summarize this first heterogeneity analysis in Panel A of Table 2.5. As suggested, we

observe that there are larger effects from exposure to SPP for exam-takers with mothers that have

lower educational attainment. Ultimately, as this effect is connected to a change in incentives, our

finding implies that SPP successfully promoted additional effort among those from less favored

backgrounds. Likewise, we see that most interaction effects between temperature and SPP exposure

are of the expected sign, although only two of them are statistically significant: for Math and total

scores, for students with mothers that have lower levels of education, as hypothesized. In addition,

notice that there are some interesting differences in the effect of temperature increases. While all

estimates are of the expected sign, they seem larger for those with better household educational

backgrounds–again consistent with our framework. We will return to this in the next subsection for

our mechanisms analysis.

Since the enforcement of Colombian Law 142 of 1994, the country has followed a system of so-

35Our focus on urban students is also motivated by the fact that the socioeconomic stratum has more variation in

urban settings where population density is higher.

36We also do these estimations for both the overall population and for rural students alone (see Appendix 2B,
Tables 2B.14 and 2B.15), but here we focus on urban students following our previous discussion on the potentially
differentiated mechanisms at play between urban and rural populations.

84

cioeconomic stratification that assigns a stratum number from 1 to 6 for a residence, which increases

with the quality of the residence. The system is widely used as a proxy income, although it has also

been found to serve as a potential source of social discrimination (Bogliacino, Jiménez Lozano, &

Reyes, 2018). We implement it in our analysis to separate the sample of exam takers between two

groups: those coming from households with potentially less-favored backgrounds (strata 1, 2, and

3) and those of potentially more favorable backgrounds (strata 4, 5, and 6). As seen in Panel B of

Table 2.5, we find that exposure to SPP had statistically significant effects among exam takers from

potentially less-favored backgrounds, which again follows the line of our conceptual framework.

Similarly, we find that the interaction between temperature and exposure to SPP is only significant

in the case of Math for students from less-favored backgrounds.

Overall, we observe a consistent result among urban exam-takers. For starters, the effects of

temperature and exposure to SPP are systematically larger and statistically significant in Saber 11

Math scores than in other sections. On the one hand, these findings are in line with the literature

suggesting that heat insults have larger effects on test performance and knowledge accumulation

related to high-complexity tasks, which is arguably the case with Math. On the other hand,

given that admissions to high-return post-secondary programs (like STEM or STEM-related fields

(Kinsler & Pavan, 2015)) usually give a larger weight to Math scores, the shock on incentives from

SPP was expected to have a larger impact on Math scores than in others sections, which is what we

find. Finally, this combination of shocks leads to important interaction effects again concentrated

in Math, mainly changing the behavior of eligible students.

2.5.2.2 Sex-based heterogeneities

Recent literature on climate change impacts reveals that these are sex-differentiated, including

dimensions of well-being (Goh, 2012), health (Navas-Martín et al., 2022), and time allocation in

labor (Lee, Haile, Seymour, & Azzarri, 2021), specifically among adults. Hence, as evidence

related to sex-differentiated impacts among youths is scarce, we further expand our analysis to

explore heterogeneities between male and female students. We summarize our results in Table 2.6.

Our results indicate that, at the overall level (Panel A), increases in temperature appear to only have

85

a statistically significant (and negative) effect on male students’ scores. Conversely, increasing

exposure to SPP seemingly has a positive and significant effect on female students, whereas its

interaction with SPP is not statistically significant in any case. Nonetheless, important differences

arise when comparing urban and rural settings, as reported in Panels B and C, respectively. Among

urban students, we find that increases in temperature negatively affect most scores of both male

and female students. In addition, and consistent with our main specification results from Table

2.4, we detect a significant effect from the exposure to SPP on Math and total scores regardless of

the student sex, while there is no apparent effect on Spanish scores. However, our results point to

a statistically significant interaction between increasing temperatures and SPP exposure only for

female students, suggesting that an effort-adjusting mechanism is more prominent among women.

Finally, when focusing on the case of rural students (Panel C), we find substantial differences

between male and female students. First, we note that increasing temperatures have no apparent

effects on male exam takers, but there is a positive and statistically significant effect on female

students’ Math, Spanish, and total scores. Thus, the unexpected positive effect from our baseline

analysis seems likely driven by behavioral responses to increases in temperature among young

females. However, the effect of increasing exposure levels to SPP is only significant for Math

scores of female students, while its interaction with increasing temperatures is statistically zero

overall. The stark sex-based differences in the effects of increasing temperatures (and incentives

to study) on observed scores warrant future research that addresses the institutional and behavioral

mechanisms driving these results.

2.5.2.3 Potential mechanism

Up to this point, we exploit the exogenous shock of SPP since, on average, it is tentatively

increasing the expected wages or, equivalently, reducing the opportunity costs of accessing tertiary

education (Londoño-Vélez et al., 2020). Consistent with this interpretation, we should expect

that temperature increases have a larger effect on those who are likely to have lower opportunity

costs for exerting effort. Access and completion of higher education are strongly correlated with

a student’s mother’s educational attainment (Guzman Ruiz et al., 2009) and household income

86

(Londoño-Vélez et al., 2020). Based on our available information, we can test whether temperature

increases have increasingly negative effects on students’ outcomes as their mother’s educational

attainment increases or as the reported stratum of the house is higher, which would be consistent

with our theoretical model and empirical results so far.

For the first case, we consider a specification that follows

100 × 𝑦𝑖 𝑗 𝑐𝑡 =

4
∑︁

𝛽𝐸 𝑀 (𝑘),𝑇 (𝑇𝑐𝑡 × 𝐸 𝑀𝑘 ) +

4
∑︁

𝛽𝐸 𝑀 (𝑘) 𝐸 𝑀𝑘 + 𝛽𝑃𝑃𝑐𝑡 + 𝑎 𝑗 + 𝑎𝑠𝑡 + 𝑢𝑖 𝑗 𝑐𝑡,

𝑘=1
where 𝑦𝑖 𝑗 𝑐𝑡 is the standardized score of interest, and 𝐸 𝑀𝑘 is an index variable of whether the

𝑘=1

student’s mother has maximum education attainment at level 𝑘. Namely, we have that 𝑘 goes

from 1 to 4, following if the mother has no educational attainment, primary education, secondary

education, or postsecondary education. We present our fixed-effect estimates in Table 2.7, taking

the no educational attainment cases as the base category (for results including coefficients for

dummy variables alone, see Appendix 2B, Table 2B.16).

All estimates are consistent with our rationale. To begin with, the first-order coefficient estimates

for the effect of temperature increases on test performance and those of our baseline estimates

remain in the same ballpark. Also, on average, when mothers’ educational attainments are larger,

so are their children’s test results–up to 51% of a standard deviation when compared to students

whose mothers report no educational attainments. Finally, we find that all estimates of the second-

order correlations of temperature increase relative to mothers’ schooling are of the expected sign

(negative), and all are increasingly negative with said education level. Moreover, this correlation is

found across all subjects, at the aggregate level, and in urban and rural settings.

Finally, we also estimate an alternative specification following

100 × 𝑦𝑖 𝑗 𝑐𝑡 =

6
∑︁

𝛽𝐻𝑆(𝑙),𝑇 (𝑇𝑐𝑡 × 𝐻𝑆𝑙) +

6
∑︁

𝛽𝐻𝑆(𝑙) 𝐻𝑆𝑙 + 𝛽𝑃𝑃𝑐𝑡 + 𝑎 𝑗 + 𝑎𝑠𝑡 + 𝑢𝑖 𝑗 𝑐𝑡,

𝑙=1
where 𝐻𝑆𝑙 is an index variable of whether the exam-taker dwelling is classified as stratum 𝑙,

𝑙=1

according to the Colombian residential strata system. In Table 2.8, we summarize fixed-effects

estimates under this alternate specification, with strata 1 (lowest in the scale of dwelling quality)

as the base category.37 In line with all our results so far, we observe that temperature increases

37Also see Appendix 2B, Table 2B.17, for a detailed summary table including coefficients for dummies of each

stratum.

87

have a larger impact on the scores of students coming from potentially more favored backgrounds.

For instance, take again the case of Math scores of urban exam-takers. While an increase of 1°C

in the average maximum temperature during the year leading to the exam day is connected to a

reduction of 2.08% of a standard deviation for those living in a stratum 1 residence, the magnitude

will increase by up to 1.38 percentage points–i.e., a rough total effect of 3.45% of a standard

deviation–for those in a stratum 6 dwelling.

Ultimately, we see that although the direct effects of heat increases may suggest potentially

different dynamics between rural and urban settings, the differentiated effects of increased heat

suggest a highly consistent mechanism across all individuals in terms of opportunity cost. For

those coming from a less favored background (lower mothers’ educational attainment or lower

dwelling stratum), the opportunity cost of exerting effort in preparation for an exam like Saber 11 is

comparatively high since their baseline expectations to access higher education are low. Conversely,

those from more favored backgrounds will perceive a lower opportunity cost for exerting additional

effort in studying, leading to a larger exposure to the effects of weather on average. It should be

stressed that we do not expect the last two specifications to have a causal interpretation. For that, we

would need exogenous variation in the mother’s education attainments and residence stratum, which

is beyond the scope of this paper. However, our estimates of 𝛽𝐸 𝑀 (𝑘),𝑇 and 𝛽𝐻𝑆(𝑙) are informative

of the mechanisms behind our findings using the variation created by SPP.

2.5.2.4 Additional robustness checks

A final concern about our policy effect of interest arises from potential differences between

SPP-exposed and SPP-unexposed Municipalities. For instance, it could be possible that the effect

is largely driven by those Municipalities with the largest shares of beneficiaries. Likewise, it could

also be the case that the effect is essentially driven by the absolute difference between untreated

Municipalities and those in the intensive margin. We extend our analysis to address these critical

points.

Based on the positive skew in the distribution of exposure to SPP across Municipalities (Ap-

pendix 2B, Figure 2B.10), we recalculate our estimates for equation (2.7) by systematically ex-

88

cluding those locations with the top one-, five-, and ten- percent levels of exposure to SPP (see

Appendix 2B, Tables 2B.18, 2B.19, and 2B.20, respectively). In all three specifications, we find,

yet again, that the largest effect of SPP is found among urban students, especially for the subject

of Mathematics, where we also find evidence of effort adjustment as temperatures increase with

coefficients statistically equivalent. Furthermore, we find that limiting the sample to those Mu-

nicipalities with non-zero exposure to SPP does not change our inference (Appendix 2B, Table

2B.21). Finally, we estimate a variation of our main specification treating the exposure to SPP as

binary–i.e., approaching an average treatment effect on SPP-exposed Municipalities. We find that

this approximation would result in qualitatively similar estimates (Appendix 2B, Table 2B.22) but

with much larger effects. Following our discussion so far, we believe this further strengthens our

preferred approach, as we base our analysis on more conservative estimates that better consider the

differentiated levels of exposure experienced across Municipalities.

2.6 Conclusion

We set up a theoretical model for human capital accumulation that is sensible to exogenous

weather shocks, which suggests that heat has an overall negative effect on students’ academic

performance. Also, our model states that increases in effort should improve academic performance,

but at the cost of higher exposure to thermal insults that can–as an interaction effect–reduce the

performance of students. We use information from a National high-school exit exam in Colombia

(Saber 11) between 2014-2019, precipitation and temperature data at the weather station level, and

information on a policy intervention of a scholarship program in 2015 that serves as an exogenous

shock to test these hypotheses.

Based on linear unobserved effect models and using fixed-effect regression methods, we demon-

strate that 1°C increases in the average maximum temperature during the year before taking a

high-stakes exam significantly reduce Math, Spanish, and total scores among students in urban

settings. Our findings of negative shocks from increases in average temperature are consistent with

the theoretical propositions and empirical results in the literature (Garg et al., 2020; Graff-Zivin et

al., 2018; Park et al., 2021, 2020). Nevertheless, we highlight that longer-term exposures (e.g., over

89

a calendar year) to increasing temperatures, may be of higher relevance than short-term exposures

(e.g., exam day or week before the exam) for countries with predominantly tropical weather.

We provide quasi-experimental evidence supporting the hypothesis that exerting effort in high-

stakes evaluations is increasingly costly for students as they are further exposed to weather shocks.

While government programs like Ser Pilo Paga create incentives that increase students’ outcomes,

the efficacy of these interventions may be compromised under higher temperatures. We also

evaluate a potential underlying mechanism supporting these findings, namely from the side of

opportunity costs to access tertiary education (and likely expected wages). Students whose mothers

achieved greater educational attainments, or who resided in houses of a higher socioeconomic

stratum, have better scores in high-stakes exams but exhibit larger negative shocks from exposition

to increased temperatures.

Among the limitations of our research, we encounter unexpected results from the effects of

increasing temperatures on rural areas–namely, a positive effect in scores. We hypothesize that

time reallocation in response to higher temperatures (Alberto et al., 2021; Graff-Zivin & Neidell,

2014) could play an important role in settings where child labor is high. We test this hypothesis

with time-use data for youths and their parents in Colombia, finding a reallocation towards activities

complementary to study (for youths) and non-farm labor (for household heads and spouses). While

these results follow our suggested mechanism, future research should explore these findings in

more detail and for other contexts. Furthermore, we encourage research trying to understand the

interaction between weather and social policies, which we believe will open the door to a more

realistic evaluation of the impacts of climate change.

90

Figures and Tables

Figure 2.1 Municipality-level historical average of daily maximum temperature, 1981–2010.

Notes — Figure displays the average of the daily maximum temperature (°C) at the Municipality level for the period 1981-2010
(i.e., climate normal). Raw weather data is publicly available from the Colombian Institute of Hydrology, Meteorology, and
Environmental Studies (IDEAM).

91

Figure 2.2 Municipality-level average Saber 11 standardized score in Math, 2014-II–2019-I.

Notes — The figure shows the Municipality-level average standardized Saber 11 score in Math for the tests between 2014-II and
2019-I. Student-level information is publicly available from the Colombian Institute for the Evaluation of Education (ICFES).

92

Figure 2.3 Correlation of Municipality-level historical average of daily maximum temperature
(1981–2010) and Saber 11 standardized scores (2014-II–2019-I).

Notes — Reporting Pooled-OLS coefficient estimates (and 95% confidence intervals) for dummy variables of the long-run average
maximum temperature (AMT) experienced in the Municipality, captured in temperature bins. The response variable is 100𝑦, where
𝑦 is the standardized score of the specific subject (Math, Spanish) or total test score. Estimates capture the difference in scores for
Municipalities with a long-run AMT in the bin relative to Municipalities that experienced a long-run AMT between 20 and 24°C.
Exam-taker attribute controls added to the regression include age (and its square), dummy variables for the maximum educational
attainment of the mother, dwelling stratum classification, whether the test taker is female, self-identification as part of an ethnic
minority, and whether the exam taker lives and studies in different Municipalities.

93

Figure 2.4 Effect of the past-year average maximum temperature on Saber 11 scores,
2014-II–2019-I.

Notes — Reporting Fixed-Effects estimation of the effect (and 95% confidence interval) from a 1°C increase in the average daily
maximum temperature at the Municipality level during the calendar before the Saber 11 test on the scores achieved by students
between 2014-II and 2019-I. The response variable is 100𝑦, where 𝑦 is the standardized score of the specific subject (Math, Spanish)
or total test score. Avg. maximum temperature captures the average daily maximum temperature (in °C) during the calendar year
before the exam within each round. Regressions include controls of precipitation at the Municipality level for the calendar year
before the exam date. Attribute controls added to the regression include age (and its square), dummy variables for the maximum
educational attainment of the mother, dwelling stratum classification, whether the test taker is female, self-identification as part of
an ethnic minority, and whether the exam taker lives and studies in different Municipalities.

94

Figure 2.5 Effect of the past-year average maximum and exam date maximum temperatures on
Saber 11 scores, 2014-II–2019-I.

Notes — Reporting Fixed-Effects estimation of the effect (and confidence interval) from 1°C increases in the exam-day maximum
temperature and the average daily maximum temperature at the Municipality level during the calendar before the Saber 11 test on
the scores achieved by students between 2014-II and 2019-I. The response variable is 100𝑦, where 𝑦 is the standardized score of the
specific subject (Math, Spanish) or total test score. Avg. maximum temperature captures the average daily maximum temperature
(in °C) during the calendar year before the exam within each round. Regressions include controls of precipitation at the Municipality
level for the calendar year before the exam date. Attribute controls added to the regression include age (and its square), dummy
variables for the maximum educational attainment of the mother, dwelling stratum classification, whether the test taker is female,
self-identification as part of an ethnic minority, and whether the exam taker lives and studies in different Municipalities.

95

Figure 2.6 Effect of the past-year number of days exposure to maximum temperatures on Saber 11
scores, 2014-II–2019-I.

Notes — Reporting Fixed-Effects estimation of the effect (and confidence interval) from 1 additional day exposition to maximum
temperatures in the range relative to days with maximum temperatures between 20°and 25°C at the Municipality level during
the calendar before the Saber 11 test on the scores achieved by students between 2014-II and 2019-I. The response variable is
100𝑦, where 𝑦 is the standardized score of the specific subject (Math, Spanish) or total test score. Regressions include controls of
precipitation at the Municipality level for the calendar year before the exam date. Attribute controls added to the regression include
age (and its square), dummy variables for the maximum educational attainment of the mother, dwelling stratum classification,
whether the test taker is female, self-identification as part of an ethnic minority, and whether the exam taker lives and studies in
different Municipalities.

96

Figure 2.7 Effect of the past-year average maximum temperature on Saber 11 scores by groups of
Long-Run Municipality daily maximum temperatures, 2014-II–2019-I.

Notes — Reporting Fixed-Effects estimation of the effect (and confidence interval) from 1°C increases in the average daily maximum
temperature at the Municipality level during the calendar before the Saber 11 test on the scores achieved by students between 2014-II
and 2019-I. The response variable is 100𝑦, where 𝑦 is the standardized score of the specific subject (Math, Spanish) or total test
score. Avg. maximum temperature captures the average daily maximum temperature (in °C) during the calendar year before the
exam within each round. Regressions include controls of precipitation at the Municipality level for the calendar year before the
exam date. Attribute controls added to the regression include age (and its square), dummy variables for the maximum educational
attainment of the mother, dwelling stratum classification, whether the test taker is female, self-identification as part of an ethnic
minority, and whether the exam taker lives and studies in different Municipalities.

97

Table 2.1 Descriptive statistics of the sample and Municipality-level variables.

Student-Level Information (𝑁=2,202,917)

Mean Median

S.D.

Saber 11 scores
Mathematics
Urban
Rural

Spanish

Urban
Rural

Total

Urban
Rural

Independent variables
Age
Female (1=Yes)
Mother: Complete primary (1=Yes)
Mother: Complete secondary (1=Yes)
Mother: Complete postsecondary (1=Yes)
Ethnic minority (1=Yes)
Lives and studies in different Municipalities (1=Yes)
House Strata: Low (1 = Residence stratum is 1, 2, or 3)
House Strata: High (1 = Residence stratum is 4, 5, or 6)
Municipality-Level Information (𝑁=1,052)

Ser Pilo Paga (SPP) 2015
Number of beneficiaries
Exposure to SPP
Weather variables
Average maximum temperature (°C)
Maximum temperature (°C, exam days)
Long-run average maximum temperature above 30°C (1=Yes)
Mean precipitation (mm/day)
Precipitation (mm, exam days)
Days per temperature bin
DD < 15
DD 15 − 20
DD 20 − 25
DD 25 − 30
DD 30 − 35
DD ≥ 35

52.2
53.7
48.9
53.1
54.5
50.1
261.5
268.7
245.8

17.29
0.55
0.29
0.33
0.23
0.06
0.03
0.925
0.075

8.61
0.017

26.66
26.87
0.34
5.52
4.66

6.0
72.2
67.1
74.8
115.2
30.7

52
52
49
53
54
50
258
266
242

17.08
1
0
0
0
0
0
1
0

2
0.012

27.81
28.00
0
4.35
0.30

0.0
0.0
8.0
23.1
36.1
0.0

11.5
11.6
10.8
9.7
9.7
9.1
47.3
47.6
42.6

0.99
0.5
0.45
0.47
0.42
0.24
0.17
0.26
0.26

59.27
0.02

6.19
6.77
0.47
3.97
9.89

20.5
124.2
104.7
96.2
122.7
58.0

Notes — Student-level information is publicly available from the Colombian Institute for the Evaluation of Education (ICFES). The
total number of test takers in urban Municipalities is 1,517,000, while the remaining 685,917 are located in rural Municipalities.
Age is calculated as the exact number of years from date of birth to date of exam. Female and Ethnic minority dummies are based
on self-identification from test takers. Mothers’ education attainment dummies are mutually exclusive and only consider schooling
levels fully completed. Anonymized information of beneficiaries of Ser Pilo Paga (SPP) was provided by the Colombian Ministry
of Education, which can be matched at the school level. The number of beneficiaries relates only to the number of assigned
scholarships within the Municipality in the first round of SPP. The variable Exposure to SPP is the ratio of beneficiaries in the
Municipality over the total number of test-takers within that first round period of SPP. Raw weather data is publicly available from
the Colombian Institute of Hydrology, Meteorology, and Environmental Studies (IDEAM).

98

Table 2.2 Effect of the reference-week average maximum temperature on time use for youths (10 to 16 years of age), ELCA (2013 and
2016).

Variables

(1)
TV

(2)
Computer

(3)
Parents

Panel A: Rural
(4)
Play

(5)
Chores

(6)
Exercise

(7)
Read

(8)
Homework

Avg. maximum temperature (AMT)

Mean precipitation

-0.013
(0.012)
-0.001
(0.002)

0.002
(0.007)
-0.001
(0.001)

-0.004
(0.013)
0.002
(0.002)

-0.014
(0.009)
-0.003
(0.002)

0.012
(0.011)
-0.004**
(0.002)

0.011
(0.009)
-0.002
(0.002)

0.018*
(0.009)
-0.001
(0.001)

Observations
𝑅2

3,197
0.046

3,197
0.061

3,197
0.063

3,197
0.070

3,197
0.070

3,197
0.081

3,197
0.037

0.001
(0.009)
0.000
(0.002)

3,197
0.047

Variables

Avg. maximum temperature (AMT)

Mean precipitation

Observations
𝑅2
Month-by-year FE
Municipality FE

(1)
TV

(2)
Computer

(3)
Parents

Panel B: Urban
(4)
Play

(5)
Chores

(6)
Exercise

(7)
Read

(8)
Homework

0.006
(0.010)
0.002
(0.001)

3,122
0.065
Yes
Yes

0.018
(0.014)
-0.001
(0.002)

3,122
0.098
Yes
Yes

-0.002
(0.011)
0.002
(0.002)

-0.009
(0.013)
0.001
(0.002)

-0.004
(0.011)
0.002
(0.002)

0.012
(0.010)
0.001
(0.002)

-0.007
(0.007)
-0.000
(0.001)

3,122
0.060
Yes
Yes

3,122
0.076
Yes
Yes

3,122
0.084
Yes
Yes

3,122
0.109
Yes
Yes

3,122
0.062
Yes
Yes

0.003
(0.009)
0.000
(0.002)

3,122
0.068
Yes
Yes

Notes — Cluster (Municipality level) robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. The dependent
variables are dummies equal to one if the respondent spent one hour or more in each of the activities mentioned in the column
header. Avg. maximum temperature captures the average of the daily maximum temperature (in °C) during the week before the
survey for each household (the reference period for the time-use variables). Mean precipitation is the average daily precipitation
(mm/day) also for the same reference period. Attribute controls added to the regression include age dummies and a dummy variable
for whether the individual is female.

99

Table 2.3 Effect of the reference-week average maximum temperature on time use for household
heads and spouses, ELCA (2010, 2013, and 2016).

Variables

Own-farm Off-farm Commuting Chores

(1)

Panel A: Household Heads
(2)

(3)

(4)

Avg. maximum temperature (AMT)

Mean precipitation

Observations
𝑅2

-0.620*
(0.321)
-0.014
(0.049)

12,082
0.127

0.355
(0.397)
0.018
(0.055)

12,082
0.195

0.014
(0.049)
-0.013
(0.008)

12,082
0.065

(5)
Leisure

0.011
(0.261)
0.023
(0.051)

0.253
(0.181)
-0.019
(0.026)

12,082
0.575

12,082
0.133

Variables

Avg. maximum temperature (AMT)

Mean precipitation

Observations
𝑅2
Month-by-year FE
Municipality FE

Panel B: Spouses

(1)

(2)

(3)

(4)

Own-farm Off-farm Commuting Chores

-0.104
(0.111)
-0.013
(0.029)

9,091
0.171
Yes
Yes

0.285**
(0.131)
0.057**
(0.022)

9,091
0.135
Yes
Yes

0.033
(0.021)
0.000
(0.005)

9,091
0.067
Yes
Yes

-0.009
(0.293)
-0.006
(0.051)

9,091
0.262
Yes
Yes

(5)
Leisure

-0.195
(0.352)
-0.038
(0.038)

9,091
0.118
Yes
Yes

Notes — Cluster (Municipality level) robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. The dependent
variables are the percentages of time spent on each activity mentioned in the column header. Avg. maximum temperature captures
the average daily maximum temperature (in °C) during the week before the survey for each household (the reference period for
the time-use variables). Mean precipitation is the average daily precipitation (mm/day) for the same reference period. Attribute
controls added to the regression include age and a dummy variable for whether the individual is female.

100

Table 2.4 Heterogeneous effect of the past-year average maximum temperature on Saber 11 scores by exposure to Ser Pilo Paga,
2014-II–2019-I.

Variables

Avg. maximum temperature (AMT)

Exposure to SPP × Post

AMT × Exposure to SPP × Post

(1)
Math

(2)
Math

(3)
Math

(4)
Spanish

(5)
Spanish

(6)
Spanish

(7)
Total

(8)
Total

(9)
Total

-0.511
(0.709)
3.032***
(0.995)
-0.0123
(0.0360)

-2.267***
(0.843)
6.127***
(1.812)
-0.108**
(0.0504)

1.423*
(0.774)
2.126*
(1.099)
-0.0111
(0.0395)

-0.553
(0.581)
0.0923
(0.630)
-0.00647
(0.0232)

-2.033***
(0.727)
0.376
(1.102)
0.0231
(0.0379)

0.891
(0.626)
-0.0699
(0.690)
-0.0237
(0.0239)

-0.268
(0.604)
1.457**
(0.722)
-0.00936
(0.0256)

-2.182***
(0.546)
3.216**
(1.284)
-0.0356
(0.0393)

1.467**
(0.684)
0.785
(0.820)
-0.0141
(0.0280)

Mean precipitation

Mean precipitation × Exposure to SPP × Post

Observations
𝑅2
School FE
Time-State FE
Attribute Controls
Heterogeneous Trends(a)
Sample

-0.354
(0.234)
-0.0970
(0.0691)

2,175,278
0.392
Yes
Yes
Yes
Yes
All

0.135
(0.381)
-0.235
(0.148)

1,506,211
0.384
Yes
Yes
Yes
Yes
Urban

-0.0226
(0.203)
-0.0342
(0.0712)

669,066
0.337
Yes
Yes
Yes
Yes
Rural

-0.536
(0.200)
-0.0173
(0.0428)

0.0944
(0.407)
0.0525
(0.0978)

-0.00988
(0.167)
-0.0318
(0.0528)

-0.268
(0.172)
-0.0581
(0.0530)

0.00478
(0.277)
-0.166
(0.105)

2,175,278
0.334
Yes
Yes
Yes
Yes
All

1,506,211
0.312
Yes
Yes
Yes
Yes
Urban

669,066
0.284
Yes
Yes
Yes
Yes
Rural

2,175,278
0.442
Yes
Yes
Yes
Yes
All

1,506,211
0.430
Yes
Yes
Yes
Yes
Urban

(0.159)
-0.0144
(0.0564)

669,066
0.371
Yes
Yes
Yes
Yes
Rural

Notes — Cluster (weather station level) robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. The response variable
is 100𝑦, where 𝑦 is the standardized score of the specific subject (Math, Spanish) or total test score. Avg. maximum temperature
captures the average daily maximum temperature (in °C) during the calendar year before the exam within each round. Mean
precipitation is the average daily precipitation (mm/day) for the calendar year before the exam date. Attribute controls added to the
regression include age (and its square), dummy variables for the maximum educational attainment of the mother, dwelling stratum
classification, whether the test taker is female, self-identification as part of an ethnic minority, and whether the exam taker lives
and studies in different Municipalities. (a) Trend heterogeneity is allowed by adding interactions between period (year-semester)
dummies and quartiles of Municipality-level variables (namely Total High-School Students, Total Number of School Teachers, and
Total Number of School Administrative Personnel).

101

Table 2.5 Heterogeneous effect of the past-year average maximum temperature on Saber 11 scores
by exposure to Ser Pilo Paga in Urban population under specific house strata and mother’s
educational attainment levels, 2014-II–2019-I.

Variables

Avg. Maximum Temperature

Exposure to SPP × Post

AMT × Exposure to SPP × Post

Mother’s Education
Observations
𝑅2
School FE
Time-State FE
Heterogeneous Trends(a)
Additional Controls

Variables

Avg. Maximum Temperature (AMT)

Exposure to SPP × Post

AMT × Exposure to SPP × Post

House Strata
Observations
𝑅2
School FE
Time-State FE
Heterogeneous Trends(a)
Additional Controls

Panel A: Differences by Mother’s Educational Attainment
Spanish
Math

Total Score

(1)

(2)

(3)

(4)

(5)

(6)

-1.570**
(0.756)
7.834***
(2.008)
-0.158***
(0.0545)

Low
550,045
0.256
Yes
Yes
Yes
Yes

-2.610*
(1.337)
4.233**
(2.000)
-0.0602
(0.0604)

High
955,949
0.386
Yes
Yes
Yes
Yes

-1.170
(1.057)
2.689*
(1.457)
-0.0408
(0.0427)

Low
550,045
0.205
Yes
Yes
Yes
Yes

-2.463**
(0.938)
-1.549
(1.602)
0.0749
(0.0550)

High
955,949
0.301
Yes
Yes
Yes
Yes

-0.917
(0.724)
5.713***
(1.474)
-0.108**
(0.0427)

Low
550,045
0.278
Yes
Yes
Yes
Yes

-2.855***
(0.878)
0.969
(1.671)
0.0251
(0.0520)

High
955,949
0.427
Yes
Yes
Yes
Yes

Math

Panel B: Differences by House Strata
Spanish

Total Score

(1)

(2)

(3)

(4)

(5)

(6)

-2.319***
(0.789)
6.532***
(1.794)
-0.110**
(0.0494)

Low
1,354,424
0.320
Yes
Yes
Yes
Yes

0.144
(3.642)
0.566
(4.303)
-0.0496
(0.136)

High
150,960
0.485
Yes
Yes
Yes
Yes

-2.062***
(0.707)
0.833
(1.050)
0.0114
(0.0360)

Low
1,354,424
0.263
Yes
Yes
Yes
Yes

-1.525
(2.112)
-2.911
(3.001)
0.136
(0.125)

High
150,960
0.370
Yes
Yes
Yes
Yes

-2.102***
(0.529)
3.754***
(1.263)
-0.0489
(0.0386)

Low
1,354,424
0.357
Yes
Yes
Yes
Yes

-2.234
(2.527)
-3.157
(2.878)
0.151
(0.109)

High
150,960
0.531
Yes
Yes
Yes
Yes

Notes — Cluster (weather station level) robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. The response
variable is 100𝑦, where 𝑦 is the standardized score of the specific subject (Math, Spanish) or total test score. Avg. maximum
temperature captures the average of the daily maximum temperature (in °C) during the calendar year before the exam within each
round. All regressions include mean precipitation (not reported), measured as the average daily precipitation (mm/day) for the
calendar year before the exam date. Attribute controls added to the regression include age (and its square), dummy variables for the
maximum educational attainment of the mother, dwelling stratum classification, whether the test taker is female, self-identification
as part of an ethnic minority, and whether the exam taker lives and studies in different Municipalities. (a) Trend heterogeneity is
allowed by adding interactions between period (year-semester) dummies and quartiles of Municipality-level variables (namely Total
High-School Students, Total Number of School Teachers, and Total Number of School Administrative Personnel).

102

Table 2.6 Heterogeneous effect of the past-year average maximum temperature on Saber 11 scores
by exposure to Ser Pilo Paga and specific sex of the student, 2014-II–2019-I.

Variables

(1)

(2)

(3)

(4)

Math

Panel A: Overall
Spanish

Avg. Maximum Temperature (AMT)

Exposure to SPP × Post

AMT × Exposure to SPP × Post

Sex of the Student
Observations
𝑅2

-1.664**
(0.689)
1.361
(1.165)
0.0419
(0.0406)

Male
986,944
0.383

0.421
(0.835)
4.275***
(0.988)
-0.0503
(0.0365)

Female
1,188,130
0.387

-1.128*
(0.607)
0.0551
(0.746)
-0.0236
(0.0260)

Male
986,944
0.338

-0.0116
(0.651)
0.0349
(0.769)
0.0104
(0.0290)

Female
1,188,130
0.341

Math

Panel B: Urban
Spanish

Variables

(1)

(2)

(3)

(4)

Total Score

(5)

(6)

-1.101*
(0.615)
0.909
(0.820)
-0.00164
(0.0278)

Male
986,944
0.436

0.450
(0.663)
1.835**
(0.777)
-0.0134
(0.0288)

Female
1,188,130
0.450

Total Score

(5)

(6)

Avg. Maximum Temperature (AMT)

Exposure to SPP × Post

AMT × Exposure to SPP × Post

Sex of the Student
Observations
𝑅2

-2.881***
(0.961)
5.377***
(1.841)
-0.0755
(0.0478)

Male
685,568
0.366

-1.558
(0.949)
6.573***
(2.227)
-0.125*
(0.0682)

Female
820,545
0.384

-2.080**
(0.896)
0.665
(1.244)
0.00178
(0.0418)

Male
685,568
0.307

-1.957***
(0.706)
-0.0927
(1.447)
0.0501
(0.0499)

-2.717***
(0.636)
3.352***
(1.232)
-0.0398
(0.0336)

Female
820,545
0.324

Male
685,568
0.414

-1.631**
(0.619)
3.028*
(1.685)
-0.0287
(0.0555)

Female
820,545
0.444

Variables

(1)

(2)

(3)

(4)

Math

Panel C: Rural
Spanish

Avg. Maximum Temperature (AMT)

Exposure to SPP × Post

AMT × Exposure to SPP × Post

Sex of the Student
Observations
𝑅2
School FE
Time-State FE
Heterogeneous Trends(a)
Additional Controls

-0.104
(0.803)
0.417
(1.289)
0.0345
(0.0494)

Male
301,374
0.348
Yes
Yes
Yes
Yes

2.557***
(0.920)
3.475***
(1.130)
-0.0466
(0.0391)

Female
367,585
0.320
Yes
Yes
Yes
Yes

-0.0932
(0.703)
-0.0612
(0.890)
-0.0497
(0.0334)

Male
301,374
0.303
Yes
Yes
Yes
Yes

1.812**
(0.698)
-0.0953
(0.952)
-0.00375
(0.0350)

Female
367,585
0.281
Yes
Yes
Yes
Yes

Total Score

(5)

(6)

0.308
(0.737)
0.124
(1.016)
-0.0113
(0.0378)

Male
301,374
0.387
Yes
Yes
Yes
Yes

2.437***
(0.744)
1.282
(0.858)
-0.0159
(0.0301)

Female
367,585
0.365
Yes
Yes
Yes
Yes

Notes — Cluster (weather station level) robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. The response
variable is 100𝑦, where 𝑦 is the standardized score of the specific subject (Math, Spanish) or total test score. Avg. maximum
temperature captures the average daily maximum temperature (in °C) during the calendar year before the exam within each round.
Mean precipitation is the average daily precipitation (mm/day) for the calendar year before the exam date. Attribute controls
added to the regression include age (and its square), dummy variables for dwelling stratum classification, self-identification as
part of an ethnic minority, and whether the exam taker lives and studies in different Municipalities. (a) Trend heterogeneity is
allowed by adding interactions between period (year-semester) dummies and quartiles of Municipality-level variables (namely Total
High-School Students, Total Number of School Teachers, and Total Number of School Administrative Personnel).

103

Table 2.7 Heterogeneous effect of past-year average maximum temperature on Saber 11 scores by education of the mother,
2014-II–2019-I.

Variables

Avg. maximum temperature (AMT)

AMT × Mother: Complete primary

AMT × Mother: Complete secondary

AMT × Mother: Complete tertiary

Observations
𝑅2
School FE
Time-State FE
Attribute Controls
Sample

(1)
Math

(2)
Math

(3)
Math

(4)
Spanish

(5)
Spanish

(6)
Spanish

(7)
Total

(8)
Total

(9)
Total

-0.308
(0.782)
-0.154***
(0.0542)
-0.392***
(0.0604)
-0.526***
(0.143)

2,175,363
0.392
Yes
Yes
Yes
All

-2.563**
(1.143)
-0.166***
(0.0596)
-0.350***
(0.0594)
-0.495***
(0.130)

1,506,211
0.384
Yes
Yes
Yes
Urban

1.592**
(0.766)
-0.100*
(0.0563)
-0.498***
(0.0978)
-0.718***
(0.139)

669,150
0.337
Yes
Yes
Yes
Rural

-0.145
(0.622)
-0.143***
(0.0455)
-0.384***
(0.0529)
-0.512***
(0.102)

2,175,363
0.334
Yes
Yes
Yes
All

-1.946***
(0.701)
-0.183***
(0.0477)
-0.361***
(0.0551)
-0.499***
(0.0737)

1,506,211
0.312
Yes
Yes
Yes
Urban

1.050*
(0.611)
-0.0701
(0.0520)
-0.460***
(0.0885)
-0.671***
(0.124)

669,150
0.284
Yes
Yes
Yes
Rural

0.0722
(0.647)
-0.166***
(0.0518)
-0.412***
(0.0540)
-0.575***
(0.117)

2,175,363
0.442
Yes
Yes
Yes
All

-2.138***
(0.714)
-0.200***
(0.0570)
-0.395***
(0.0611)
-0.560***
(0.109)

1,506,211
0.430
Yes
Yes
Yes
Urban

1.640**
(0.665)
-0.0971**
(0.0478)
-0.471***
(0.0878)
-0.730***
(0.125)

669,150
0.371
Yes
Yes
Yes
Rural

Notes — Cluster (weather station level) robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. The response
variable is 100𝑦, where 𝑦 is the standardized score of the specific subject (Math, Spanish) or total test score. Avg. maximum
temperature captures the average of the daily maximum temperature (in °C) during the calendar year before the exam within each
round. All regressions include mean precipitation (not reported), measured as the average daily precipitation (mm/day) for the
calendar year before the exam date. Attribute controls added to the regression include age (and its square), dummy variables for the
maximum educational attainment of the mother, dwelling stratum classification, whether the test taker is female, self-identification
as part of an ethnic minority, and whether the exam taker lives and studies in different Municipalities.

104

Table 2.8 Heterogeneous effect of past-year average maximum temperature on Saber 11 scores by house strata level, 2014-II–2019-I.

Variables

Avg. maximum temperature (AMT)

AMT × House Strata: Level 2

AMT × House Strata: Level 3

AMT × House Strata: Level 4

AMT × House Strata: Level 5

AMT × House Strata: Level 6

Observations
𝑅2
School FE
Time-State FE
Attribute Controls
Sample

(1)
Math

(2)
Math

(3)
Math

(4)
Spanish

(5)
Spanish

(6)
Spanish

(7)
Total

(8)
Total

(9)
Total

-0.183
(0.766)
-0.564***
(0.0571)
-0.878***
(0.119)
-1.392***
(0.200)
-1.374***
(0.249)
-1.619***
(0.279)

2,175,363
0.392
Yes
Yes
Yes
All

-2.311**
(1.146)
-0.554***
(0.0650)
-0.786***
(0.145)
-1.165***
(0.203)
-1.185***
(0.275)
-1.376***
(0.263)

1,506,211
0.384
Yes
Yes
Yes
Urban

1.591**
(0.760)
-0.533***
(0.0826)
-0.986***
(0.182)
-1.559***
(0.347)
-0.619**
(0.312)
-0.990**
(0.477)

669,150
0.337
Yes
Yes
Yes
Rural

-0.0721
(0.613)
-0.461***
(0.0571)
-0.794***
(0.110)
-1.165***
(0.163)
-1.211***
(0.184)
-1.338***
(0.209)

2,175,363
0.334
Yes
Yes
Yes
All

-1.813**
(0.691)
-0.441***
(0.0651)
-0.669***
(0.136)
-0.913***
(0.157)
-0.941***
(0.181)
-1.108***
(0.208)

1,506,211
0.312
Yes
Yes
Yes
Urban

1.045*
(0.592)
-0.439***
(0.0761)
-1.046***
(0.171)
-1.454***
(0.399)
-1.036***
(0.295)
-0.365
(0.370)

669,150
0.284
Yes
Yes
Yes
Rural

0.149
(0.632)
-0.504***
(0.0598)
-0.843***
(0.117)
-1.339***
(0.180)
-1.347***
(0.224)
-1.550***
(0.244)

2,175,363
0.442
Yes
Yes
Yes
All

-1.964***
(0.711)
-0.512***
(0.0695)
-0.755***
(0.145)
-1.124***
(0.178)
-1.134***
(0.240)
-1.324***
(0.231)

1,506,211
0.430
Yes
Yes
Yes
Urban

1.617**
(0.651)
-0.448***
(0.0735)
-0.995***
(0.169)
-1.442***
(0.351)
-0.796***
(0.292)
-0.627*
(0.327)

669,150
0.372
Yes
Yes
Yes
Rural

Notes — Cluster (weather station level) robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. The response variable is 100𝑦, where
𝑦 is the standardized score of the specific subject (Math, Spanish) or total test score. Avg. maximum temperature captures the average of the daily
maximum temperature (in °C) during the calendar year before the exam within each round. All regressions include mean precipitation (not reported),
measured as the average daily precipitation (mm/day) for the calendar year before the exam date. Attribute controls added to the regression include
age (and its square), dummy variables for the maximum educational attainment of the mother, dwelling stratum classification, whether the test taker is
female, self-identification as part of an ethnic minority, and whether the exam-taker lives and studies in different Municipalities.

105

REFERENCES

Abadie, A., Athey, S., Imbens, G., & Wooldridge, J. (2023). When should you adjust standard errors
for clustering? The Quarterly Journal of Economics, 138(1), 1–35. doi: 10.1093/qje/qjac038

Agrawal, T., & Agrawal, A. (2019). Who gains more from education? A comparative analysis
of business, farm and wage workers in India. The Journal of Development Studies, 55(6),
1081–1098. doi: 10.1080/00220388.2018.1443209

Alarcón Hincapié, J. C., & Pabón Caicedo, J. D. (2013). El Cambio Climático y la Distribución
Espacial de las Formaciones Vegetales en Colombia. Colombia Forestal, 16, 171–185.

Alberto, I. C., Jiao, Y., & Zhang, X.

(2021). Too hot or too cold to study? The effect of
temperature on student time allocation. Economics of Education Review, 84, 102152. doi:
10.1016/j.econedurev.2021.102152

Angrist, J., & Pischke, J.-S. (2008). Mostly Harmless Econometrics: An Empiricist’s Companion.

Princeton University Press.

Auffhammer, M., Hsiang, S., Schlenker, W., & Sobel, A. (2013). Using weather data and climate
model output in economic analyses of climate change. Review of Environmental Economics
and Policy, 7(2), 181–198. doi: 10.1093/reep/ret016

Auffhammer, M., & Kellogg, R. (2011). Clearing the air? The effects of gasoline content regulation
on air quality. American Economic Review, 101(6), 2687–2722. doi: 10.1257/aer.101.6.2687

Barreca, A., Clay, K., Deschenes, O., Greenstone, M., & Shapiro, J. (2016). Adapting to cli-
mate change: The remarkable decline in the US temperature-mortality relationship over the
twentieth century. Journal of Political Economy, 124(1), 105–159. doi: 10.1086/684582

Bernal, G., & Penney, J.

(2019). Scholarships and student effort: Evidence from Colombia’s
Ser Pilo Paga program. Economics of Education Review, 72, 121–130. doi: 10.1016/
j.econedurev.2019.04.008

Bilinski, A., & Hatfield, L. A. (2018). Nothing to see here? Non-inferiority approaches to parallel

trends and other model assumptions. doi: 10.48550/ARXIV.1805.03273

Björklund, A., & Salvanes, K. G. (2011). Education and Family Background. In Handbook of the
Economics of Education (Vol. 3, pp. 201–247). Elsevier. doi: 10.1016/B978-0-444-53429
-3.00003-X

Bogliacino, F., Jiménez Lozano, L., & Reyes, D.

(2018). Socioeconomic stratification and
stereotyping: lab-in-the-field evidence from Colombia. International Review of Economics,
65(1), 77–118. doi: 10.1007/s12232-017-0285-4

106

Burke, M., Craxton, M., Kolstad, C., Onda, C., Allcott, H., Baker, E., . . . others (2016). Op-
portunities for advances in climate change economics. Science, 352(6283), 292–293. doi:
10.1126/science.aad9634

Carleton, T., & Hsiang, S. (2016). Social and economic impacts of climate. Science, 353(6304),

aad9837. doi: 10.1126/science.aad9837

Carleton, T., Jina, A., Delgado, M., Greenstone, M., Houser, T., Hsiang, S., . . . others (2022).
Valuing the global mortality consequences of climate change accounting for adaptation costs
and benefits. The Quarterly Journal of Economics, 137(4), 2037–2105. doi: 10.1093/qje/
qjac020

Chadi, A., De Pinto, M., & Schultze, G. (2019). Young, gifted and lazy? The role of ability
and labor market prospects in student effort decisions. Economics of Education Review, 72,
66–79. doi: 10.1016/j.econedurev.2019.04.004

Chetty, R., Friedman, J., & Rockoff, J. (2014a). Measuring the impacts of teachers I: Evaluating
bias in teacher value-added estimates. American Economic Review, 104(9), 2593–2632. doi:
10.1257/aer.104.9.2593

Chetty, R., Friedman, J., & Rockoff, J. (2014b). Measuring the impacts of teachers II: Teacher
value-added and student outcomes in adulthood. American Economic Review, 104(9), 2633–
2679. doi: 10.1257/aer.104.9.2633

Cho, H. (2017). The effects of summer heat on academic achievement: A cohort analysis. Journal of
Environmental Economics and Management, 83, 185–196. doi: 10.1016/j.jeem.2017.03.005

Colmer, J.

(2021). Temperature, labor reallocation, and industrial production: Evidence from
India. American Economic Journal: Applied Economics, 13(4), 101–124. doi: 10.2139/
ssrn.3900866

Davis, B., Winters, P., Reardon, T., & Stamoulis, K.

(2009). Rural nonfarm employment and
farming: household-level linkages. Agricultural Economics, 40(2), 119–123. doi: 10.1111/
j.1574-0862.2009.00374.x

Dell, M., Jones, B., & Olken, B.

(2014). What do we learn from the weather? The new
climate-economy literature. Journal of Economic Literature, 52(3), 740–98. doi: 10.1257/
jel.52.3.740

Fernández, M., Ibáñez, A. M., & Peña, X.

(2014). Adjusting the labour supply to mitigate
violent shocks: Evidence from rural Colombia. The Journal of Development Studies, 50(8),
1135–1155. doi: 10.1080/00220388.2014.919384

Field, C. B., Barros, V. R., & Intergovernmental Panel on Climate Change (Eds.). (2014). Climate
change 2014: impacts, adaptation, and vulnerability: Working Group II contribution to the

107

fifth assessment report of the Intergovernmental Panel on Climate Change. New York, NY:
Cambridge University Press.

Garg, T., Jagnani, M., & Taraz, V. (2020). Temperature and human capital in India. Journal of the
Association of Environmental and Resource Economists, 7(6), 1113–1150. doi: 10.1086/
710066

Goh, A.

(2012).

A literature review of

the gender-differentiated impacts of cli-
mate change on women’s and men’s assets and well-being in developing coun-
tries. Washington, DC: CGIAR CAPRi. Working Paper No. 106.
(Retrieved from
https://ebrary.ifpri.org/digital/collection/p15738coll2/id/127247)

Graff-Zivin, J., Hsiang, S., & Neidell, M. (2018). Temperature and human capital in the short
and long run. Journal of the Association of Environmental and Resource Economists, 5(1),
77–105. doi: 10.1086/694177

Graff-Zivin, J., & Neidell, M. (2014). Temperature and the allocation of time: Implications for

climate change. Journal of Labor Economics, 32(1), 1–26. doi: 10.1086/671766

Graff-Zivin, J., Song, Y., Tang, Q., & Zhang, P. (2020). Temperature and high-stakes cognitive
performance: Evidence from the national college entrance examination in China. Journal
of Environmental Economics and Management, 104, 102365. doi: 10.1016/j.jeem.2020
.102365

Guzman Ruiz, C., Duran Muriel, D. M., Franco Gallego, J., Castaño Velez, E., Gallon Gomez, S.,
& Gomez Portilla, K. (2009). Desercion estudiantil en la educacion superior colombiana:
Metodologia de seguimiento, diagnóstico y elementos para su prevención. Bogotá: Ministerio
de Educacion Nacional. (OCLC: 777064247)

Haggblade, S., Hazell, P. B., & Reardon, T. (2007). Transforming the rural nonfarm economy:
International Food Policy Research

Opportunities and threats in the developing world.
Institute.

Handel, D., & Hanushek, E. (2022, dec). US school finance: Resources and outcomes. NBER

Working Paper, w30769. doi: 10.3386/w30769

Hanushek, E. (2006). School resources. In E. Hanushek & F. Welch (Eds.), Handbook of the

economics of education (Vol. 2, pp. 865–908). Elsevier.

Helo Sarmiento, J. (2023). Into the tropics: Temperature, mortality, and access to health care
in Colombia. Journal of Environmental Economics and Management, 119, 102796. doi:
10.1016/j.jeem.2023.102796

Hoffmann, B., Pulido, X., & Vera-Cossio, D. A. (2023). The unequal effect of temperature on test
scores: Evidence from Colombia. (Inter-American Development Bank. Discussion Paper

108

IDB-DP-1000.) doi: 10.18235/0004832

Hsiang, S. (2016). Climate econometrics. Annual Review of Resource Economics, 8, 43–75. doi:

10.3386/w22181

Hsiang, S., & Kopp, R.

(2018). An economist’s guide to climate change science. Journal of

Economic Perspectives, 32(4), 3–32. doi: 10.3386/w25189

ICFES.

(2019). Documentación del examen Saber 11. Report; Instituto Colombiano para la

Evaluación de la Educación.

Jackson, K., Johnson, R., & Persico, C. (2016). The effects of school spending on educational
and economic outcomes: Evidence from school finance reforms. Quarterly Journal of
Economics, 131(1), 157–218. doi: 10.1093/qje/qjv036

Jolliffe, D. (2004). The impact of education in rural Ghana: Examining household labor allocation
and returns on and off the farm. Journal of Development Economics, 73(1), 287–314. doi:
10.1016/j.jdeveco.2003.02.002

Kinsler, J., & Pavan, R. (2015). The Specificity of General Human Capital: Evidence from College
Major Choice. Journal of Labor Economics, 33(4), 933–972. doi: 10.1086/681206

Krebs, B. (2024). Temperature and Cognitive Performance: Evidence from Mental Arithmetic
Training. Environmental and Resource Economics, 87(7), 2035–2065. doi: 10.1007/s10640
-024-00881-y

Laajaj, R., Moya, A., & Sánchez, F. (2022). Equality of opportunity and human capital accumula-
tion: Motivational effect of a nationwide scholarship in Colombia. Journal of Development
Economics, 154, 102754. doi: 10.1016/j.jdeveco.2021.102754

Lee, Y., Haile, B., Seymour, G., & Azzarri, C.

(2021). The heat never bothered me anyway:
Gender-specific response of agricultural labor to climatic shocks in Tanzania. Applied
Economic Perspectives and Policy, 43(2), 732–749. doi: 10.1002/aepp.13153

Lenton, T. M., Xu, C., Abrams, J. F., Ghadiali, A., Loriani, S., Sakschewski, B., . . . Scheffer, M.
(2023). Quantifying the human cost of global warming. Nature Sustainability, 6, 1237–1247.
doi: 10.1038/s41893-023-01132-6

Li, X., & Patel, P. (2021). Weather and high-stakes exam performance: Evidence from student-level
administrative data in Brazil. Economics Letters, 199, 109698. doi: 10.1016/j.econlet.2020
.109698

Londoño-Vélez, J.

(2022). The impact of diversity on perceptions of income distribution and
preferences for redistribution. Journal of Public Economics, 214, 104732. doi: 10.1016/
j.jpubeco.2022.104732

109

Londoño-Vélez, J., Rodríguez, C., & Sánchez, F. (2020). Upstream and downstream impacts of col-
lege merit-based financial aid for low-income students: Ser Pilo Paga in Colombia. American
Economic Journal: Economic Policy, 12(2), 193–227. doi: 10.1257/pol.20180131

Londoño-Vélez, J., Rodriguez, C., Sánchez, F., & Álvarez-Arango, L. (2023, September). Financial
aid and social mobility: Evidence from Colombia’s Ser Pilo Paga. NBER Working Paper,
31737. doi: 10.3386/w31737

MacKinnon, J. G., Nielsen, M. Ø., & Webb, M. (2022). Cluster-robust inference: A guide to
empirical practice. Journal of Econometrics, 232(2), 272–299. doi: 10.1016/j.jeconom
.2022.04.001

Medina, P., Ariza, N., Navas, P., Rojas, F., Parody, G., Valdivia, J. A., . . . Penagos, J. F. (2018).
An unintended effect of financing the university education of the most brilliant and poorest
Colombian students: The case of the intervention of the Ser Pilo Paga program. Complexity,
2018, 1–9. doi: 10.1155/2018/3528206

Melo León, S. F., Romero Otálora, G. D., Buitrago, D. A., Riveros, L. C., Díaz Giraldo, C., & Roa-
Ortiz, S. A. (2024). Estimación de los efectos de la variabilidad climática en la producción
agrícola en Colombia. e-CUCBA, 11(21), 133–141. doi: 10.32870/e-cucba.vi21.330

Mora, A., & Ruiz, L. M. (2019). “Ser pilo no paga”: Privatización, desigualdad y desfinanciamiento
de la universidad pública en Colombia. Ciencia Política, 14(27), 115–142. doi: 10.15446/
cp.v14n27.73369

Navas-Martín, M., López-Bueno, J., Ascaso-Sánchez, M., Sarmiento-Suárez, R., Follos, F., Vellón,
(2022). Gender differences in adaptation to heat in Spain (1983–2018).

J., . . . Díaz, J.
Environmental Research, 215, 113986. doi: 10.1016/j.envres.2022.113986

Park, J. (2022). Hot temperature and high stakes performance. Journal of Human Resources,

57(2), 400–434. doi: 10.3368/jhr.57.2.0618-9535R3

Park, J., Behrer, P., & Goodman, J.

(2021). Learning is inhibited by heat exposure, both in-
ternationally and within the United States. Nature Human Behaviour, 5(1), 19–27. doi:
10.1038/s41562-020-00959-9

Park, J., Goodman, J., Hurwitz, M., & Smith, J. (2020). Heat and learning. American Economic

Journal: Economic Policy, 12(2), 306–39. doi: 10.1257/pol.20180612

Pei, Z., Pischke, J.-S., & Schwandt, H. (2019). Poorly measured confounders are more useful on
the left than on the right. Journal of Business & Economic Statistics, 37(2), 205–216. doi:
10.1080/07350015.2018.1462710

Posso, C., Saravia, E., & Uribe, P. (2023). Acing the test: Educational effects of the SaberEs
test preparation program in Colombia. Economics of Education Review, 97, 102459. doi:

110

10.1016/j.econedurev.2023.102459

Ramoni-Perazzi, J., Orlandoni-Merli, G., Castillo-Paredes, L., & Peña Guillén, J. A. (2021). Child
labor in Colombia: factors affecting the selection of economic activity. Revista de Economía
del Rosario, 24(2). doi: 10.12804/revistas.urosario.edu.co/economia/a.9086

Reardon, T., Echeverria, R., Berdegué, J., Minten, B., Liverpool-Tasie, S., Tschirley, D., & Zilber-
man, D. (2019). Rapid transformation of food systems in developing regions: Highlighting the
role of agricultural research & innovations. Agricultural Systems, 172, 47-59. (Agricultural
research for rural prosperity: Rethinking the pathways) doi: 10.1016/j.agsy.2018.01.022

Reardon, T., Stamoulis, K., & Pingali, P.

(2007). Rural nonfarm employment in developing
countries in an era of globalization. Agricultural Economics, 37(s1), 173-183. doi: 10.1111/
j.1574-0862.2007.00243.x

Roth, J.

(2022). Pretest with caution: Event-study estimates after testing for parallel trends.

American Economic Review: Insights, 4(3), 305–322. doi: 10.1257/aeri.20210236

Shah, M., & Steinberg, B. (2021). Workfare and human capital investment: Evidence from India.
Journal of Human Resources, 56(2), 380–405. doi: 10.3368/jhr.56.2.1117-9201R2

Statista Research Department.

lected household appliances in Colombia in 2018.
https://www.statista.com/statistics/1007425/household-penetration-appliances-colombia/)

(2023, March 31st). Household penetration rate of se-
(Retrieved from

Statista.

Sung, S. (2022, February 14th). Household penetration rate of home appliances in India in 2018.
Statista. (Retrieved from https://www.statista.com/statistics/370635/household-penetration-
home-appliances-india/)

Wendt, D., Van Loon, L., & Marken Lichtenbelt, W. (2007). Thermoregulation during exercise
in the heat: Strategies for maintaining health and performance. Sports Medicine, 37(8),
669–682. doi: 10.2165/00007256-200737080-00002

Wooldridge, J. M.

(2010). Econometric Analysis of Cross Section and Panel Data (2nd ed.).

Cambridge, Mass: MIT Press.

Zhang, X., Chen, X., & Zhang, X. (2024). Temperature and low-stakes cognitive performance.
Journal of the Association of Environmental and Resource Economists, 11(1), 75-96. doi:
10.1086/726007

111

APPENDIX 2A

OPTIMIZATION PROBLEM

We assume functional forms such that

𝑓 (𝑒, 𝑡) =

𝑒𝛼
𝑡 𝛽

,

0 < 𝛼 < 1,

𝛽 > 0;

𝑔(𝑒, 𝑡) = 𝑒𝛾𝑡 𝜁 ,

𝛾 > 1,

0 < 𝜁 < 1;

Hence, the individual’s problem is reduced to

𝑈 (𝑐, 𝑔(𝑒, 𝑡)) = 𝑐 − 𝑔(𝑒, 𝑡) = 𝑤𝑧 𝑓 (𝑒, 𝑡) − 𝑔(𝑒, 𝑡).

max
𝑒

𝑤𝑧

𝑒𝛼
𝑡 𝛽 − 𝑒𝛾𝑡 𝜁 ,

where the first-order condition (FOC) implies that

𝛼𝑤𝑧

𝑒𝛼−1
𝑡 𝛽 − 𝛾𝑒𝛾−1𝑡 𝜁 = 0,

while the second-order condition for maximization

𝛼(𝛼 − 1)𝑤𝑧

𝑒𝛼−2
𝑡 𝛽 − 𝛾(𝛾 − 1)𝑒𝛾−2𝑡 𝜁 < 0

is satisfied when 0 < 𝛼 ≤ 1 and 𝛾 > 1, and thus 𝛾 > 𝛼; hence, we must assume that the marginal

cost of effort increases faster than its marginal benefit.

From the FOC we can recover the optimal level of effort, 𝑒∗, which follows

𝑒∗ =

(cid:18) 𝛼𝑤𝑧
𝛾𝑡 𝛽+𝜁

(cid:19) 1
𝛾− 𝛼

.

Therefore, the production of human capital under an optimal effort level is

ℎ∗(𝑡, 𝑤, 𝑧) = 𝑧

(𝑒∗)𝛼
𝑡 𝛽

Finally, linearizing this expression with a second-order Taylor expansion, we have that

= 𝑡−𝛽−

𝛼(𝛽+𝜁 )
𝛾− 𝛼 𝑤

𝛼
𝛾− 𝛼 𝑧

𝛾
𝛾− 𝛼

(cid:19) 𝛼
𝛾− 𝛼

.

(cid:18) 𝛼
𝛾

ℎ∗(𝑡, 𝑤, 𝑧) = ℎ∗(𝑡′, 𝑤′, 𝑧′) +

(cid:19)

(cid:18) 𝑡 − 𝑡′
𝑡′

ℎ∗(𝑡′, 𝑤′, 𝑧′)

(cid:20)

−𝛽 −

(cid:21)

𝛼(𝛽 + 𝜁)
𝛾 − 𝛼
(cid:19)

+

ℎ∗(𝑡′, 𝑤′, 𝑧′)

(cid:19)

(cid:18) 𝑤 − 𝑤′
𝑤′
(cid:18) 𝑡 − 𝑡′
𝑡′
(cid:18) 𝑡 − 𝑡′
𝑡′

1
2

1
2

+

+

(cid:19)

(cid:19) (cid:18) 𝑤 − 𝑤′
𝑤′
(cid:19) (cid:18) 𝑧 − 𝑧′
𝑧′

(cid:21)

(cid:20) 𝛼
𝛾 − 𝛼

+

(cid:18) 𝑧 − 𝑧′
𝑧′

ℎ∗(𝑡′, 𝑤′, 𝑧′)

(cid:20)

−𝛽 −

(cid:19)

ℎ∗(𝑡′, 𝑤′, 𝑧′)

(cid:20)

−𝛽 −

ℎ∗(𝑡′, 𝑤′, 𝑧′)

(cid:20) 𝛾
𝛾 − 𝛼
(cid:21)

(cid:21) (cid:20) 𝛼

𝛾 − 𝛼
(cid:21) (cid:20) 𝛾

𝛾 − 𝛼

(cid:21)

(cid:21)

+ 𝑅2(𝑡, 𝑧, 𝑤),

𝛼(𝛽 + 𝜁)
𝛾 − 𝛼
𝛼(𝛽 + 𝜁)
𝛾 − 𝛼

where 𝑅2(𝑡, 𝑧, 𝑤) includes the cross-effect between wages and cognitive abilities and other own-

and cross-effects of second and higher orders that are approximate to zero. Then, the percent

112

deviation of human capital with respect to the mean follows:
(cid:20) 𝛼
𝛾 − 𝛼

𝛼(𝛽 + 𝜁)
𝛾 − 𝛼

(cid:20) 𝛾
𝛾 − 𝛼

−𝛽 −

ˆℎ =

ˆ𝑤 +

ˆ𝑡 +

ˆ𝑧 +

1
2

(cid:21)

(cid:21)

(cid:20)

(cid:20)

(cid:21)

−𝛽 −

𝛼(𝛽 + 𝜁)
𝛾 − 𝛼

(cid:21) (cid:20) 𝛼

(cid:21)

𝛾 − 𝛼

ˆ𝑡 ˆ𝑤

where ˆ𝑥 = (𝑥 − 𝑥′)/𝑥′, for 𝑥 ∈ {ℎ, 𝑤, 𝑡, 𝑧}.

(cid:20)

1
2

+

−𝛽 −

𝛼(𝛽 + 𝜁)
𝛾 − 𝛼

(cid:21) (cid:20) 𝛾

(cid:21)

𝛾 − 𝛼

ˆ𝑡 ˆ𝑧 + 𝑅,

113

APPENDIX 2B

ADDITIONAL FIGURES AND TABLES

Figure 2B.1 Spatial distribution of the weather stations used in the analysis.

Notes — Figure displays the weather stations used in our analysis (blue) and those included in the imputation procedure. Raw
weather data is publicly available from the Colombian Institute of Hydrology, Meteorology, and Environmental Studies (IDEAM).

114

Figure 2B.2 Pre-trend testing of exposure to Ser Pilo Paga on Municipality-level variables,
2010-2019.

Notes — Municipality-level information is publicly available from the Center for Studies on Economic Development (CEDE) of
Universidad de los Andes, Colombia. Graphs report the coefficient estimates (and 95% CI) from the regression of the outcome
of interest (e.g., the number of schools in the Municipality) on interactions between the exposure to Ser Pilo Paga (SPP) and year
dummies. Results are based on linear regression with State-Year and Municipality-level fixed effects. Excluded interaction for the
year 2014 (coefficient set at zero) takes into account the baseline of Saber 11 data, for which no Municipality was exposed to SPP.

115

Figure 2B.3 Pre-trend testing of exposure to Ser Pilo Paga on Math and Spanish scores,
2010-2013.

Notes — Student-level scores are publicly available from Instituto Colombiano para la Evaluación de la Educación (ICFES). Graphs
report the coefficient estimates (and 95% CI) from the regression of the outcome of interest, namely 100𝑦 with 𝑦 the standardized
Math or Spanish score, on interactions between the exposure to Ser Pilo Paga (SPP) and year dummies. Results are based on
linear regression with State-Year and Municipality-level fixed effects. Interaction for the year 2013 is excluded as a base-category
comparison. The analysis allowed for trend heterogeneity by adding interactions between period (year) dummies and quartiles of
Municipality-level variables (namely Total High-School Students, Total Number of School Teachers, and Total Number of School
Administrative Personnel).

116

Figure 2B.4 Pre-trend testing of exposure to Ser Pilo Paga on Student-level variables, 2010-2016.

Notes — Student-level information is publicly available from Instituto Colombiano para la Evaluación de la Educación (ICFES).
Selected covariates are only available for the years 2010-2016. Graphs report the coefficient estimates (and 95% CI) from the
regression of the outcome of interest on interactions between the exposure to Ser Pilo Paga (SPP) and year dummies. Eligibility
to SPP by SISBEN indicates whether the student reported being in the lower categories (1 and 2) of the SISBEN scores, making a
student eligible for the program on a need basis. Family income is a binary variable of whether the student reports a monthly family
income above 2 monthly minimum wages (MW). Results are based on linear regression with State-Year and Municipality-level fixed
effects. Excluded interaction for the year 2014 (coefficient set at zero) takes into account the baseline of Saber 11 data, for which
no Municipality was exposed to SPP.

117

Figure 2B.5 Effect of the past-year average maximum temperature on Saber 11 scores for
additional subjects, 2014-II–2019-I.

Notes — Reporting Fixed-Effects estimation of the effect (and 95% confidence interval) from a 1°C increase in the average daily
maximum temperature at the Municipality level during the calendar before the Saber 11 test on the scores achieved by students
between 2014-II and 2019-I. The response variable is 100𝑦, where 𝑦 is the standardized score of the specific subject, namely Natural
Science, Social Science, or English as a Second Language. Avg. maximum temperature captures the average daily maximum
temperature (in °C) during the calendar year before the exam within each round. Regressions include controls of precipitation at the
Municipality level for the calendar year before the exam date. Attribute controls added to the regression include age (and its square),
dummy variables for the maximum educational attainment of the mother, dwelling stratum classification, whether the test taker is
female, self-identification as part of an ethnic minority, and whether the exam taker lives and studies in different Municipalities.

118

Figure 2B.6 Average days of exposure to bins of maximum temperature in the year before Saber
11 during the period of analysis across Municipalities, 2014-II–2019-I.

Notes — The figure reports the average number of days a Municipality, during the year before the Saber 11 test, has been exposed
to days with maximum temperatures within a given bin between 2014-II and 2019-I.

119

Figure 2B.7 Effect of the past-year average maximum and exam week maximum temperatures on
Saber 11 scores, 2014-II–2019-I.

Notes — Reporting Fixed-Effects estimation of the effect (and confidence interval) from 1°C increases in the exam-week maximum
temperature and the average daily maximum temperature at the Municipality level during the calendar before the Saber 11 test on
the scores achieved by students between 2014-II and 2019-I. The response variable is 100𝑦, where 𝑦 is the standardized score of the
specific subject (Math, Spanish) or total test score. Avg. maximum temperature captures the average daily maximum temperature
(in °C) during the calendar year before the exam within each round. Regressions include controls of precipitation at the Municipality
level for the calendar year before the exam date. Attribute controls added to the regression include age (and its square), dummy
variables for the maximum educational attainment of the mother, dwelling stratum classification, whether the test taker is female,
self-identification as part of an ethnic minority, and whether the exam taker lives and studies in different Municipalities.

120

Figure 2B.8 Effect of the past-year number of days exposure to maximum temperatures with
alternate bins on Saber 11 scores, 2014-II–2019-I.

Notes — Reporting Fixed-Effects estimation of the effect (and confidence intervals, 90% and 95%) from 1 additional day exposition
to maximum temperatures in the range relative to days with maximum temperatures between 21°and 23°C at the Municipality level
during the calendar before the Saber 11 test on the scores achieved by students between 2014-II and 2019-I. The response variable
is 100𝑦, where 𝑦 is the standardized score of the specific subject (Math, Spanish) or total test score. Regressions include controls of
precipitation at the Municipality level for the calendar year before the exam date. Attribute controls added to the regression include
age (and its square), dummy variables for the maximum educational attainment of the mother, dwelling stratum classification,
whether the test taker is female, self-identification as part of an ethnic minority, and whether the exam taker lives and studies in
different Municipalities.

121

Figure 2B.9 Interaction effects of the past-year number of days exposure to maximum
temperatures and exposure to Ser Pilo Paga on Saber 11 scores, 2014-II–2019-I.

Notes — Reporting Fixed-Effects estimation of the average interaction effect (and confidence intervals, 90% and 95%) from 1
additional day exposition to maximum temperatures in the range and exposure to Ser Pilo Paga, relative to days with maximum
temperatures between 21°and 23°C at the Municipality level during the calendar before the Saber 11 test on the scores achieved by
students between 2014-II and 2019-I. The response variable is 100𝑦, where 𝑦 is the standardized score of the specific subject (Math,
Spanish) or total test score. Regressions include precipitation controls at the Municipality level for the calendar year before the
exam date. Attribute controls added to the regression include age (and its square), dummy variables for the maximum educational
attainment of the mother, dwelling stratum classification, whether the test taker is female, self-identification as part of an ethnic
minority, and whether the exam taker lives and studies in different Municipalities. Estimations allow for trend heterogeneity by
adding interactions between period (year-semester) dummies and quartiles of Municipality-level variables (namely Total High-
School Students, Total Number of School Teachers, and Total Number of School Administrative Personnel).

122

Figure 2B.10 Distribution of the exposure to Ser Pilo Paga first round at the Municipality level,
2014.

Notes — Information about SPP scholarship assignment was provided by the Colombian Ministry of Education (MEN). The graph
displays the distribution of the exposure measure, as well as the 90th, 95th, and 99th percentiles in red.

123

Table 2B.1 Correlation of student-level attributes and past-year weather variables: average
maximum temperature and average precipitation, 2014-II–2019-I.

Variables

(1)
Age

(2)
Female

(3)
Primary(a)

(4)
Secondary(a)

(5)

(6)

Tertiary(a) Minority(b)

Avg. maximum temperature

Mean precipitation

0.006
(0.007)
0.002
(0.002)

-0.000
(0.002)
-0.000
(0.001)

0.001
(0.003)
-0.000
(0.001)

-0.000
(0.003)
-0.000
(0.001)

-0.001
(0.002)
-0.000
(0.001)

0.001
(0.004)
-0.000
(0.001)

Observations
𝑅2
School FE
Time-State FE

2,202,869
0.119
Yes
Yes

2,202,869
0.077
Yes
Yes

2,202,869
0.090
Yes
Yes

2,202,869
0.064
Yes
Yes

2,202,869
0.263
Yes
Yes

2,202,869
0.557
Yes
Yes

Notes — Cluster (weather station level) robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. Avg. maximum
temperature captures the average daily maximum temperature (in °C) during the calendar year before the exam within each round.
(a) Variables
Mean precipitation is the average daily precipitation (mm/day), also for the calendar year before the exam date.
capturing the educational attainment of the test-taker’s mother, measured as whether the educational level was completed.
(b)
Minority is the exam-taker self-identification as part of an ethnic minority.

124

Table 2B.2 Effect of the past-year average maximum temperature on Saber 11 scores, 2014-II–2019-I.

Variables

Avg. maximum temperature

Mean precipitation

Observations
𝑅2
School FE
Time-State FE
Attribute Controls
Sample

(1)
Math

-0.575
(0.791)
-0.161
(0.270)

2,175,363
0.391
Yes
Yes
Yes
All

(2)
Math

(3)
Math

(4)
Spanish

(5)
Spanish

(6)
Spanish

(7)
Total

(8)
Total

(9)
Total

-2.896**
(1.158)
-0.852**
(0.385)

1,506,211
0.384
Yes
Yes
Yes
Urban

1.379*
(0.766)
0.0797
(0.191)

669,150
0.337
Yes
Yes
Yes
Rural

-0.403
(0.632)
-0.0971
(0.190)

2,175,363
0.334
Yes
Yes
Yes
All

-2.288***
(0.698)
-0.519
(0.345)

1,506,211
0.312
Yes
Yes
Yes
Urban

0.864
(0.609)
0.0627
(0.179)

669,150
0.284
Yes
Yes
Yes
Rural

-0.213
(0.656)
-0.135
(0.189)

2,175,363
0.442
Yes
Yes
Yes
All

-2.516***
(0.720)
-0.631**
(0.287)

1,506,211
0.430
Yes
Yes
Yes
Urban

1.435**
(0.665)
-0.0201
(0.154)

669,150
0.371
Yes
Yes
Yes
Rural

Notes — Cluster (weather station level) robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. The response
variable is 100𝑦, where 𝑦 is the standardized score of the specific subject (Math, Spanish) or total test score. Avg. maximum
temperature captures the average daily maximum temperature (in °C) during the calendar year before the exam within each round.
Mean precipitation is the average daily precipitation (mm/day) for the calendar year before the exam date. Attribute controls added
to the regression include age (and its square), dummy variables for the maximum educational attainment of the mother, dwelling
stratum classification, whether the test taker is female, self-identification as part of an ethnic minority, and whether the exam taker
lives and studies in different Municipalities.

125

Table 2B.3 Effect of the past-year average maximum temperature (alternate clustering) on Saber 11 scores, 2014-II–2019-I.

Variables

Avg. maximum temperature

Mean precipitation

Observations
𝑅2
School FE
Time-State FE
Attribute Controls
Sample

(1)
Math

-0.575
(0.787)
-0.161
(0.250)

2,175,363
0.391
Yes
Yes
Yes
All

(2)
Math

(3)
Math

(4)
Spanish

(5)
Spanish

(6)
Spanish

(7)
Total

(8)
Total

(9)
Total

-2.896**
(1.122)
-0.852**
(0.358)

1,506,211
0.384
Yes
Yes
Yes
Urban

1.379*
(0.706)
0.0797
(0.212)

669,150
0.337
Yes
Yes
Yes
Rural

-0.403
(0.690)
-0.0971
(0.175)

2,175,363
0.334
Yes
Yes
Yes
All

-2.288***
(0.600)
-0.519
(0.348)

1,506,211
0.312
Yes
Yes
Yes
Urban

0.864
(0.559)
0.0627
(0.203)

669,150
0.284
Yes
Yes
Yes
Rural

-0.213
(0.677)
-0.135
(0.177)

2,175,363
0.442
Yes
Yes
Yes
All

-2.516***
(0.665)
-0.631**
(0.274)

1,506,211
0.430
Yes
Yes
Yes
Urban

1.435**
(0.644)
-0.0201
(0.176)

669,150
0.371
Yes
Yes
Yes
Rural

Notes — Cluster (groups of weather station level) robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. As a
first stage, we cluster weather stations based on their spatial proximity (k-means clustering), then calculate standard errors at the
stations-cluster level. The response variable is 100𝑦, where 𝑦 is the standardized score of the specific subject (Math, Spanish) or
total test score. Avg. maximum temperature captures the average daily maximum temperature (in °C) during the calendar year
before the exam within each round. Mean precipitation is the average daily precipitation (mm/day) for the calendar year before the
exam date. Attribute controls added to the regression include age (and its square), dummy variables for the maximum educational
attainment of the mother, dwelling stratum classification, whether the test taker is female, self-identification as part of an ethnic
minority, and whether the exam taker lives and studies in different Municipalities.

126

Table 2B.4 Quality of weather-station data and the effect of the past-year average maximum temperature on Saber 11 scores,
2014-II–2019-I.

Variables

Avg. maximum temperature (AMT)

Mean precipitation

(1)
Math

-0.103
(0.973)
-0.182
(0.375)

Panel A: Dropping Municipalities matched to weather stations farther than 25 km
(8)
(5)
Total
Spanish

(4)
Spanish

(6)
Spanish

(3)
Math

(2)
Math

(7)
Total

(9)
Total

-2.517*
(1.489)
-0.915*
(0.487)

1.477
(0.928)
-0.0141
(0.315)

-0.434
(0.752)
-0.149
(0.280)

-2.434***
(0.759)
-0.900**
(0.386)

0.977
(0.767)
0.229
(0.239)

-0.0112
(0.787)
-0.112
(0.275)

-2.404***
(0.892)
-0.828**
(0.366)

1.439*
(0.816)
-0.0234
(0.231)

Observations
𝑅2

1,619,952
0.385

1,169,658
0.376

450,293
0.342

1,619,952
0.325

1,169,658
0.300

450,293
0.291

1,619,952
0.434

1,169,658
0.420

450,293
0.378

Variables

Avg. maximum temperature (AMT)

Mean precipitation

Observations
𝑅2
School FE
Time-State FE
Attribute Controls
Sample

(1)
Math

-0.178
(0.796)
-0.183
(0.287)

1,963,025
0.395
Yes
Yes
Yes
All

Panel B: Dropping weather stations with more than 25% of observations imputed
(8)
(5)
Total
Spanish

(4)
Spanish

(6)
Spanish

(3)
Math

(2)
Math

(7)
Total

-2.623**
(0.999)
-0.909**
(0.428)

1,353,948
0.387
Yes
Yes
Yes
Urban

1.739**
(0.819)
-0.0106
(0.199)

609,075
0.342
Yes
Yes
Yes
Rural

-0.0983
(0.710)
-0.0361
(0.191)

1,963,025
0.338
Yes
Yes
Yes
All

-2.368**
(0.887)
-0.411
(0.357)

1,353,948
0.316
Yes
Yes
Yes
Urban

1.049
(0.667)
0.0794
(0.189)

609,075
0.288
Yes
Yes
Yes
Rural

0.0246
(0.685)
-0.0886
(0.193)

1,963,025
0.446
Yes
Yes
Yes
All

-2.553***
(0.682)
-0.542*
(0.284)

1,353,948
0.434
Yes
Yes
Yes
Urban

(9)
Total

1.655**
(0.716)
-0.0287
(0.167)

609,075
0.377
Yes
Yes
Yes
Rural

Notes — Cluster (weather station level) robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. The response
variable is 100𝑦, where 𝑦 is the standardized score of the specific subject (Math, Spanish) or total test score. Avg. maximum
temperature captures the average daily maximum temperature (in °C) during the calendar year before the exam within each round.
Mean precipitation is the average daily precipitation (mm/day) for the calendar year before the exam date. Age-squared control
is demeaned, so the coefficient estimate of age is the partial correlation at the sample mean. Attribute controls added to the
regression include age (and its square), dummy variables for the maximum educational attainment of the mother, dwelling stratum
classification, whether the test taker is female, self-identification as part of an ethnic minority, and whether the exam taker lives and
studies in different Municipalities. Panel A drops municipalities matched to weather stations farther than 25 km. Panel B restricts
the group of weather stations to those where imputations were used in less than 25% of the days.

127

Table 2B.5 Effect of the past-year average maximum and exam date maximum temperatures on Saber 11 scores, 2014-II–2019-I.

Variables

(1)
Math

(2)
Math

(3)
Math

(4)
Spanish

(5)
Spanish

(6)
Spanish

(7)
Total

(8)
Total

(9)
Total

Avg. Maximum Temperature

Maximum temperature (exam days)

-0.571
(0.798)
0.00743
(0.172)

-2.709**
(1.143)
-0.362
(0.238)

1.370*
(0.770)
0.0705
(0.199)

-0.306
(0.612)
-0.154
(0.143)

-2.115***
(0.692)
-0.240
(0.184)

0.925
(0.591)
-0.101
(0.161)

-0.267
(0.662)
0.0724
(0.141)

-2.585***
(0.740)
-0.00796
(0.190)

1.467**
(0.666)
-0.0685
(0.155)

Mean precipitation

Precipitation (exam day)

Observations
𝑅2
School FE
Time-State FE
Attribute Controls
Sample

-0.163
(0.270)
0.00498
(0.0238)

2,158,356
0.392
Yes
Yes
Yes
All

-0.899**
(0.390)
0.0290
(0.0252)

1,492,021
0.385
Yes
Yes
Yes
Urban

0.0793
(0.193)
0.00181
(0.0293)

666,334
0.337
Yes
Yes
Yes
Rural

-0.130
(0.186)
-0.0327
(0.0205)

-0.565
(0.338)
-0.0298
(0.0193)

0.0337
(0.178)
-0.0464*
(0.0254)

-0.137
(0.185)
-0.00948
(0.0185)

-0.635**
(0.285)
0.0114
(0.0197)

2,158,356
0.335
Yes
Yes
Yes
All

1,492,021
0.313
Yes
Yes
Yes
Urban

666,334
0.284
Yes
Yes
Yes
Rural

2,158,356
0.443
Yes
Yes
Yes
All

1,492,021
0.431
Yes
Yes
Yes
Urban

-0.0357
(0.154)
-0.0228
(0.0217)

666,334
0.371
Yes
Yes
Yes
Rural

Notes — Cluster (weather station level) robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. The response
variable is 100𝑦, where 𝑦 is the standardized score of the specific subject (Math, Spanish) or total test score. Avg. maximum
temperature captures the average daily maximum temperature (in °C) during the calendar year before the exam within each round.
Mean precipitation is the average daily precipitation (mm/day) for the calendar year before the exam date within each round.
Attribute controls added to the regression include age (and its square), dummy variables for the maximum educational attainment
of the mother, dwelling stratum classification, whether the test taker is female, self-identification as part of an ethnic minority, and
whether the exam taker lives and studies in different Municipalities.

128

Table 2B.6 Effect of the exam-day maximum temperature on Saber 11 scores, 2014-II–2019-I.

Variables

(1)
Math

(2)
Math

(3)
Math

(4)
Spanish

(5)
Spanish

(6)
Spanish

(7)
Total

(8)
Total

(9)
Total

Maximum temperature (exam day)

Precipitation (exam day)

-0.0165
(0.174)
0.00607
(0.0238)

-0.511**
(0.239)
0.0246
(0.0244)

0.128
(0.203)
-0.000876
(0.0293)

-0.165
(0.151)
-0.0318
(0.0205)

-0.364
(0.221)
-0.0338
(0.0227)

-0.0609
(0.168)
-0.0481*
(0.0255)

0.0643
(0.145)
-0.00860
(0.0185)

-0.163
(0.214)
0.00637
(0.0193)

-0.000979
(0.158)
-0.0250
(0.0216)

Observations
𝑅2
School FE
Time-State FE
Attribute Controls
Sample

2,158,356
0.392
Yes
Yes
Yes
All

1,492,021
0.385
Yes
Yes
Yes
Urban

666,334
0.337
Yes
Yes
Yes
Rural

2,158,356
0.335
Yes
Yes
Yes
All

1,492,021
0.313
Yes
Yes
Yes
Urban

666,334
0.284
Yes
Yes
Yes
Rural

2,158,356
0.443
Yes
Yes
Yes
All

1,492,021
0.431
Yes
Yes
Yes
Urban

666,334
0.371
Yes
Yes
Yes
Rural

Notes — Cluster (weather station level) robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. The response variable
is 100𝑦, where 𝑦 is the standardized score of the specific subject (Math, Spanish) or total test score. Maximum temperature is the
highest recorded temperature (in °C) during the exam day within each round. Precipitation is the total precipitation (mm) during
the exam day within each round. Attribute controls added to the regression include age (and its square), dummy variables for the
maximum educational attainment of the mother, dwelling stratum classification, whether the test taker is female, self-identification
as part of an ethnic minority, and whether the exam taker lives and studies in different Municipalities.

129

Table 2B.7 Effect of the past-year number of days exposure to maximum temperatures on Saber 11 scores, 2014-II–2019-I.

Variables

DD < 15

DD 15 − 20

DD 25 − 30

DD 30 − 35

DD ≥ 35

(1)
Math

(2)
Math

(3)
Math

(4)
Spanish

(5)
Spanish

(6)
Spanish

(7)
Total

(8)
Total

(9)
Total

0.0519*
(0.0301)
0.0157
(0.0160)
0.00705
(0.0156)
0.00461
(0.0215)
-0.0190
(0.0283)

0.120***
(0.0241)
0.0447**
(0.0219)
-0.0251
(0.0166)
-0.0548**
(0.0264)
-0.117**
(0.0458)

-0.0212
(0.0388)
-0.0137
(0.0244)
0.0289*
(0.0161)
0.0542**
(0.0243)
0.0557*
(0.0312)

0.0389*
(0.0218)
0.00270
(0.0147)
0.00283
(0.0113)
-0.00992
(0.0180)
-0.0140
(0.0236)

0.0601**
(0.0226)
0.0205
(0.0214)
-0.0302**
(0.0133)
-0.0659***
(0.0172)
-0.0947***
(0.0311)

0.0155
(0.0369)
-0.0105
(0.0156)
0.0166
(0.0122)
0.0412**
(0.0193)
0.0370
(0.0247)

0.0384*
(0.0205)
0.0145
(0.0127)
0.00896
(0.0127)
0.00926
(0.0186)
-0.00382
(0.0242)

0.0794***
(0.0178)
0.0420***
(0.0154)
-0.0193
(0.0142)
-0.0477**
(0.0201)
-0.0987***
(0.0337)

-0.00117
(0.0342)
-0.0117
(0.0173)
0.0244*
(0.0133)
0.0558***
(0.0208)
0.0564**
(0.0271)

Mean precipitation

-0.157
(0.267)

-0.866**
(0.384)

0.0804
(0.183)

-0.0851
(0.190)

-0.525
(0.343)

0.0520
(0.180)

-0.132
(0.185)

-0.636**
(0.279)

-0.0362
(0.151)

Observations
𝑅2
School FE
Time-State FE
Attribute Controls
Sample

2,175,363
0.392
Yes
Yes
Yes
All

1,506,211
0.384
Yes
Yes
Yes
Urban

669,150
0.337
Yes
Yes
Yes
Rural

2,175,363
0.334
Yes
Yes
Yes
All

1,506,211
0.312
Yes
Yes
Yes
Urban

669,150
0.284
Yes
Yes
Yes
Rural

2,175,363
0.442
Yes
Yes
Yes
All

1,506,211
0.430
Yes
Yes
Yes
Urban

669,150
0.371
Yes
Yes
Yes
Rural

Notes — Cluster (weather station level) robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. The response
variable is 100𝑦, where 𝑦 is the standardized score of the specific subject (Math, Spanish) or total test score. DD counts the number
of days during the calendar year before the exam date for which the maximum temperature (in °C) was within the range. Mean
precipitation is the average daily precipitation (mm/day) for the calendar year before the exam date. Attribute controls added to
the regression include age (and its square), dummy variables for the maximum educational attainment of the mother, dwelling
stratum classification, self-identification as part of an ethnic minority, and whether the exam taker lives and studies in different
Municipalities.

130

Table 2B.8 Descriptive statistics of the data from ELCA.

Youths
Time-use variables (1 = One hour or more in a week)
Watch TV
Use computer
Spend time with parents
Play with siblings or friends
Help with household chores
Do exercise
Read for fun
Do homework
Play videogames
Participate in social groups
Independent variables
Age
Female (Yes=1)
Adults
Time-use variables (% of reported time)
Work on household farms
Work on other households’ farms
Commuting or looking for jobs
Domestic chores
Leisure and personal care
Other activities
Independent variables
Age
Female (1=Yes)

Mean Median

S.D.

Mean Median

S.D.

Rural (N=3,241)

Urban (N=3,162)

0.56
0.13
0.62
0.59
0.43
0.36
0.23
0.63
0.06
0.07

1.00
0.00
1.00
1.00
0.00
0.00
0.00
1.00
0.00
0.00

11.99
0.49

12.00
0.00

0.50
0.33
0.49
0.49
0.49
0.48
0.42
0.48
0.24
0.25

1.65
0.50

0.63
0.36
0.63
0.58
0.37
0.45
0.20
0.67
0.17
0.12

1.00
0.00
1.00
1.00
0.00
0.00
0.00
1.00
0.00
0.00

12.06
0.50

12.00
0.00

0.48
0.48
0.48
0.49
0.48
0.50
0.40
0.47
0.38
0.33

1.67
0.50

Household heads (N=12,155)

Spouses (N=9,158)

23.17
22.02
2.66
12.28
39.65
0.23

48.79
0.19

12.12
0.00
0.00
0.00
36.36
0.00

48.00
0.00

26.10
27.37
5.56
19.43
17.39
2.43

12.59
0.40

8.09
4.46
0.70
47.91
38.61
0.23

44.12
0.94

0.00
0.00
0.00
50.00
36.42
0.00

43.00
1.00

15.14
13.82
3.04
21.69
17.45
2.16

12.88
0.23

Notes — Individual-level information was accessed remotely using the computational infrastructure of the Data Center of the
Universidad de los Andes. Age is calculated as the exact number of years from the date of birth to the survey date. Female is a
binary variable based on self-identification from the respondents. The time-use variables for the sample of youths are dummies
equal to one if the respondent spent one hour or more in a typical day in the reference period in each of the activities mentioned in
the row. For the adults, they correspond to the percentage of time spent in the same reference period in each of the activities.

131

Table 2B.9 Effect of the past-year average maximum temperature on time use for youths (10 to 16 years of age), ELCA (2013 and 2016).

Variables

(1)
TV

(2)
Computer

(3)
Parents

Panel A: Rural
(4)
Play

(5)
Chores

(6)
Exercise

(7)
Read

(8)
Homework

Avg. maximum temperature

Mean precipitations

-0.171**
(0.073)
-0.012
(0.028)

0.100***
(0.026)
0.021**
(0.010)

0.012
(0.039)
0.014
(0.019)

-0.015
(0.052)
0.002
(0.029)

0.090*
(0.050)
0.003
(0.029)

0.173***
(0.032)
0.059***
(0.015)

0.050*
(0.028)
-0.002
(0.017)

0.094**
(0.043)
0.015
(0.025)

Observations
𝑅2

3,197
0.048

3,197
0.063

3,197
0.063

3,197
0.069

3,197
0.068

3,197
0.086

3,197
0.036

3,197
0.048

Variables

Avg. maximum temperature

Mean precipitations

Observations
𝑅2
Month-by-year FE
Municipality FE

(1)
TV

(2)
Computer

(3)
Parents

Panel B: Urban
(4)
Play

(5)
Chores

(6)
Exercise

(7)
Read

(8)
Homework

0.033
(0.044)
-0.011
(0.025)

3,068
0.066
Yes
Yes

-0.049
(0.053)
0.020
(0.022)

3,068
0.100
Yes
Yes

0.012
(0.061)
0.025
(0.031)

0.013
(0.037)
0.032
(0.023)

0.013
(0.051)
-0.011
(0.026)

3,068
0.059
Yes
Yes

3,068
0.074
Yes
Yes

3,068
0.083
Yes
Yes

0.074
(0.054)
0.050*
(0.030)

3,068
0.114
Yes
Yes

-0.036
(0.039)
-0.018
(0.021)

3,068
0.063
Yes
Yes

-0.004
(0.054)
-0.013
(0.028)

3,068
0.070
Yes
Yes

Notes — Cluster (Municipality level) robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. The dependent
variables are dummies equal to one if the respondent spent one hour or more in each of the activities mentioned in the column
header. Avg. maximum temperature captures the average of the daily maximum temperature (in °C) during the last year relative to
the week of the survey. Mean precipitation is the average daily precipitation (mm/day) also for the same reference period. Attribute
controls added to the regression include age dummies and a dummy variable for whether the individual is female.

132

Table 2B.10 Effect of the past-year average maximum temperature on time use for household
heads and spouses, ELCA (2010, 2013, and 2016).

Variables

Own-farm Off-farm Commuting Chores

(1)

Panel A: Household Heads
(2)

(3)

(4)

Avg. maximum temperature

Mean precipitations

Observations
𝑅2

-4.579*
(2.337)
0.653
(0.574)

12,084
0.128

2.376
(3.027)
-0.376
(0.867)

12,084
0.195

-0.331
(0.546)
-0.041
(0.108)

12,084
0.065

(5)
Leisure

2.054
(1.430)
-0.038
(0.397)

0.334
(1.001)
-0.246
(0.360)

12,084
0.575

12,084
0.133

Variables

Avg. maximum temperature

Mean precipitations

Observations
R-squared
Month-by-year FE
Municipality FE

Panel B: Spouses

(1)

(2)

(3)

(4)

Own-farm Off-farm Commuting Chores

0.436
(1.494)
-0.177
(0.450)

9,091
0.171
Yes
Yes

-1.277
(1.210)
0.185
(0.211)

9,091
0.135
Yes
Yes

0.017
(0.138)
-0.021
(0.054)

9,091
0.067
Yes
Yes

-0.277
(1.285)
-0.411
(0.498)

9,091
0.262
Yes
Yes

(5)
Leisure

1.064
(1.766)
0.371
(0.678)

9,091
0.118
Yes
Yes

Notes — Cluster (Municipality level) robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. The dependent
variables are the percentages of time spent on each activity mentioned in the column header. Avg. maximum temperature captures
the average daily maximum temperature (in °C) during the last year relative to the survey week. Mean precipitation is the average
daily precipitation (mm/day) for the same reference period. Attribute controls added to the regression include age and a dummy
variable for whether the individual is female.

133

Table 2B.11 Effect of the past-year average maximum temperature (per quarter) on Saber 11 scores, 2014-II–2019-I.

Variables

(1)
Math

(2)
Math

(3)
Math

(4)
Spanish

(5)
Spanish

(6)
Spanish

(7)
Total

(8)
Total

(9)
Total

Panel A: Only including AMT from Q1 of previous year

Avg. maximum temperature 𝑄1

𝑅2

0.362
(0.446)
0.391

-0.709
(0.851)
0.384

1.461***
(0.474)
0.337

0.110
(0.428)
0.334

-1.351**
(0.547)
0.312

1.062**
(0.422)
0.284

0.487
(0.415)
0.442

-1.044*
(0.608)
0.430

1.472***
(0.417)
0.371

Panel B: Only including AMT from Q2 of previous year

Avg. maximum temperature 𝑄2

𝑅2

0.0398
(0.674)
0.391

-1.571
(1.047)
0.384

1.460**
(0.646)
0.337

0.288
(0.480)
0.334

-1.188
(0.712)
0.312

1.071**
(0.459)
0.284

0.352
(0.539)
0.442

-1.178
(0.764)
0.430

1.400***
(0.519)
0.371

Panel C: Only including AMT from Q3 of previous year

Avg. maximum temperature 𝑄3

𝑅2

-1.146***
(0.410)
0.392

-1.612**
(0.647)
0.384

-0.195
(0.512)
0.337

-0.635*
(0.325)
0.334

-1.108**
(0.417)
0.312

-0.244
(0.373)
0.284

-0.787**
(0.318)
0.442

-1.261***
(0.426)
0.430

-0.110
(0.410)
0.371

Panel D: Only including AMT from Q4 of previous year

Avg. maximum temperature 𝑄4

𝑅2

Observations
Ref. Quarter Precipitation
School FE
Time-State FE
Attribute Controls
Sample

-0.381
(0.608)
0.391

2,175,363
Yes
Yes
Yes
Yes
All

-3.019***
(0.767)
0.384

1,506,211
Yes
Yes
Yes
Yes
Urban

0.335
(0.735)
0.337

669,150
Yes
Yes
Yes
Yes
Rural

-0.662
(0.502)
0.334

2,175,363
Yes
Yes
Yes
Yes
All

-1.560**
(0.664)
0.312

1,506,211
Yes
Yes
Yes
Yes
Urban

-0.0161
(0.576)
0.284

669,150
Yes
Yes
Yes
Yes
Rural

-0.471
(0.539)
0.442

2,175,363
Yes
Yes
Yes
Yes
All

-2.431***
(0.619)
0.430

1,506,211
Yes
Yes
Yes
Yes
Urban

0.437
(0.584)
0.371

669,150
Yes
Yes
Yes
Yes
Rural

Notes — Cluster (weather station level) robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. The response
variable is 100𝑦, where 𝑦 is the standardized score of the specific subject (Math, Spanish) or total test score. AMT Q(K) captures
the average daily maximum temperature (in °C) during the K-th quarter of the calendar year before the Saber 11 exam. Each panel
reports coefficients from a regression that includes only the AMT for the K-th quarter of reference as the temperature measure.
Attribute controls added to the regression include age (and its square), dummy variables for the maximum educational attainment
of the mother, dwelling stratum classification, whether the test taker is female, self-identification as part of an ethnic minority, and
whether the exam taker lives and studies in different Municipalities.

134

Table 2B.12 Heterogeneous effect of the past-year average maximum temperature on Saber 11 scores by exposure to Ser Pilo Paga,
2014-II–2019-I.

VARIABLES

Avg. maximum temperature (AMT)

Exposure to SPP × Post

AMT × Exposure to SPP × Post

Mean precipitation

Mean precipitation × Exposure to SPP × Post

Observations
𝑅2
School FE
Time-State FE
Attribute Controls
Heterogeneous Trends
Sample

(1)
Math

(2)
Math

(3)
Math

(4)
Spanish

(5)
Spanish

(6)
Spanish

(7)
Total

(8)
Total

(9)
Total

-0.446
(0.731)
3.484***
(1.103)
-0.00485
(0.0416)

-2.306**
(0.936)
6.056***
(1.786)
-0.0958*
(0.0523)

0.122
(0.258)
-0.156**
(0.0734)

2,175,363
0.392
Yes
Yes
Yes
No
All

-0.297
(0.413)
-0.230
(0.148)

1,506,211
0.384
Yes
Yes
Yes
No
Urban

1.454*
(0.778)
2.438**
(1.140)
-0.0163
(0.0415)

0.122
(0.206)
-0.0374
(0.0724)

669,150
0.337
Yes
Yes
Yes
No
Rural

-0.345
(0.625)
0.0940
(0.645)
-0.0195
(0.0242)

-2.142***
(0.729)
0.573
(1.170)
0.0134
(0.0397)

0.917
(0.622)
0.114
(0.714)
-0.0305
(0.0245)

-0.145
(0.628)
1.761**
(0.788)
-0.00708
(0.0284)

-2.293***
(0.585)
3.152**
(1.250)
-0.0297
(0.0398)

1.486**
(0.683)
1.013
(0.869)
-0.0176
(0.0294)

-0.139
(0.197)
0.0276
(0.0424)

-0.552
(0.404)
0.0751
(0.0981)

2,175,363
0.334
Yes
Yes
Yes
No
All

1,506,211
0.312
Yes
Yes
Yes
No
Urban

0.0925
(0.168)
-0.0267
(0.0525)

669,150
0.284
Yes
Yes
Yes
No
Rural

0.0146
(0.180)
-0.0876
(0.0532)

2,175,363
0.442
Yes
Yes
Yes
No
All

-0.266
(0.287)
-0.152
(0.105)

1,506,211
0.430
Yes
Yes
Yes
No
Urban

-0.00126
(0.156)
-0.0159
(0.0557)

669,150
0.371
Yes
Yes
Yes
No
Rural

Notes — Cluster (weather station level) robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. The response
variable is 100𝑦, where 𝑦 is the standardized score of the specific subject (Math, Spanish) or total test score. Avg. maximum
temperature captures the average of the daily maximum temperature (in °C) during the calendar year before the exam within each
round. Mean precipitation is the average daily precipitation (mm/day) for the calendar year before the exam date. Attribute controls
added to the regression include age (and its square), dummy variables for the maximum educational attainment of the mother,
dwelling stratum classification, whether the test taker is female, self-identification as part of an ethnic minority, and whether the
exam taker lives and studies in different Municipalities.

135

Table 2B.13 Heterogeneous effect of the past-year average maximum temperature on Saber 11 scores by exposure to Ser Pilo Paga,
Other subjects, 2014-II–2019-I.

Variables

Avg. maximum temperature (AMT)

Exposure to SPP × Post

AMT × Exposure to SPP × Post

Mean precipitation

Mean precipitation × Exposure to SPP × Post

Observations
𝑅2
School FE
Time-State FE
Attribute Controls
Heterogeneous Trends(a)
Sample

(1)
Nat-Sci

(2)
Nat-Sci

(3)
Nat-Sci

(4)
Soc-Sci

(5)
Soc-Sci

(6)
Soc-Sci

(7)

(8)
English SL English SL English SL

(9)

-0.163
(0.603)
-0.327
(0.746)
0.0292
(0.0271)

-2.092***
(0.567)
-0.244
(1.538)
0.0462
(0.0519)

1.708**
(0.738)
-0.526
(0.944)
0.0215
(0.0327)

0.0142
(0.644)
2.874***
(1.017)
-0.0499
(0.0349)

-2.297***
(0.630)
6.857***
(1.723)
-0.116**
(0.0530)

1.566**
(0.753)
1.147
(0.955)
-0.0323
(0.0315)

0.158
(0.905)
1.977**
(0.989)
-0.00242
(0.0397)

0.0677
(0.172)
-0.0449
(0.0622)

2,175,278
0.379
Yes
Yes
Yes
Yes
All

-0.390
(0.281)
-0.0805
(0.143)

1,506,211
0.373
Yes
Yes
Yes
Yes
Urban

0.0607
(0.196)
-0.0381
(0.0655)

669,066
0.322
Yes
Yes
Yes
Yes
Rural

-0.0488
(0.200)
-0.104
(0.0655)

0.156
(0.368)
-0.368***
(0.120)

-0.204
(0.175)
0.000701
(0.0632)

-0.141
(0.255)
0.0256
(0.0613)

2,175,278
0.321
Yes
Yes
Yes
Yes
All

1,506,211
0.312
Yes
Yes
Yes
Yes
Urban

669,066
0.260
Yes
Yes
Yes
Yes
Rural

2,175,275
0.461
Yes
Yes
Yes
Yes
All

-2.159**
(0.960)
2.534*
(1.270)
0.00583
(0.0519)

-0.0891
(0.592)
-0.276
(0.167)

1,506,210
0.458
Yes
Yes
Yes
Yes
Urban

2.276**
(1.079)
2.227**
(1.110)
-0.0467
(0.0404)

-0.196
(0.312)
0.118**
(0.0593)

669,064
0.308
Yes
Yes
Yes
Yes
Rural

Notes — Cluster (weather station level) robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. The response variable
is 100𝑦, where 𝑦 is the standardized score of the specific subject, namely Natural Sciences, Social Sciences, or English as a Second
Language. Avg. maximum temperature captures the average daily maximum temperature (in °C) during the calendar year before the
exam within each round. Mean precipitation is the average daily precipitation (mm/day) for the calendar year before the exam date.
Attribute controls added to the regression include age (and its square), dummy variables for the maximum educational attainment
of the mother, dwelling stratum classification, whether the test taker is female, self-identification as part of an ethnic minority,
and whether the exam taker lives and studies in different Municipalities. (a) Trend heterogeneity is allowed by adding interactions
between period (year-semester) dummies and quartiles of Municipality-level variables (namely Total High-School Students, Total
Number of School Teachers, and Total Number of School Administrative Personnel).

136

Table 2B.14 Heterogeneous effect of the past-year average maximum temperature on Saber 11
scores by exposure to Ser Pilo Paga under specific mother’s educational attainment levels,
2014-II–2019-I.

Variables

(1)

(2)

(3)

(4)

(5)

(6)

Math

Panel A: Overall
Spanish

Total Score

Avg. Maximum Temperature

Exposure to SPP × Post

AMT × Exposure to SPP × Post

Mother’s Education
Observations
𝑅2

0.556
(0.729)
3.330***
(0.923)
-0.0387
(0.0327)

-1.429
(0.874)
2.729**
(1.182)
-0.00731
(0.0406)

Low
965,202
0.281

High
1,209,602
0.388

0.117
(0.633)
0.284
(0.569)
-0.0148
(0.0193)

Low
965,202
0.236

-1.080
(0.740)
-0.239
(1.016)
0.00951
(0.0378)

High
1,209,602
0.313

0.711
(0.602)
1.833**
(0.711)
-0.0334
(0.0239)

Low
965,202
0.309

-1.135
(0.727)
0.879
(0.904)
0.0105
(0.0323)

High
1,209,602
0.432

Variables

(1)

(2)

(3)

(4)

(5)

(6)

Math

Panel B: Urban
Spanish

Total Score

Avg. Maximum Temperature

Exposure to SPP × Post

AMT × Exposure to SPP × Post

Mother’s Education
Observations
𝑅2

-1.570**
(0.756)
7.834***
(2.008)
-0.158***
(0.0545)

Low
550,045
0.256

-2.610*
(1.337)
4.233**
(2.000)
-0.0602
(0.0604)

High
955,949
0.386

-1.170
(1.057)
2.689*
(1.457)
-0.0408
(0.0427)

Low
550,045
0.205

-2.463**
(0.938)
-1.549
(1.602)
0.0749
(0.0550)

High
955,949
0.301

-0.917
(0.724)
5.713***
(1.474)
-0.108**
(0.0427)

Low
550,045
0.278

-2.855***
(0.878)
0.969
(1.671)
0.0251
(0.0520)

High
955,949
0.427

Variables

(1)

(2)

(3)

(4)

(5)

(6)

Math

Panel C: Rural
Spanish

Total Score

Avg. Maximum Temperature

Exposure to SPP × Post

AMT × Exposure to SPP × Post

Mother’s Education
Observations
𝑅2
School FE
Time-State FE
Heterogeneous Trends(a)
Additional Controls

2.081**
(0.797)
2.278**
(1.100)
-0.0303
(0.0421)

Low
415,155
0.290
Yes
Yes
Yes
Yes

0.554
(1.085)
1.823
(1.650)
0.00326
(0.0553)

High
253,649
0.343
Yes
Yes
Yes
Yes

1.208*
(0.724)
-0.254
(0.666)
-0.0192
(0.0244)

Low
415,155
0.234
Yes
Yes
Yes
Yes

0.612
(0.948)
0.108
(1.173)
-0.0265
(0.0409)

High
253,649
0.281
Yes
Yes
Yes
Yes

1.857***
(0.692)
0.856
(0.751)
-0.0259
(0.0271)

Low
415,155
0.314
Yes
Yes
Yes
Yes

0.988
(0.930)
0.417
(1.227)
0.00338
(0.0423)

High
253,649
0.373
Yes
Yes
Yes
Yes

Notes — Cluster (weather station level) robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. The response
variable is 100𝑦, where 𝑦 is the standardized score of the specific subject (Math, Spanish) or total test score. Avg. maximum
temperature captures the average daily maximum temperature (in °C) during the calendar year before the exam within each round.
Mean precipitation is the average daily precipitation (mm/day) for the calendar year before the exam date. Attribute controls
added to the regression include age (and its square), dummy variables for dwelling stratum classification, whether the test taker is
female, self-identification as part of an ethnic minority, and whether the exam taker lives and studies in different Municipalities. (a)
Trend heterogeneity is allowed by adding interactions between period (year-semester) dummies and quartiles of Municipality-level
variables (namely Total High-School Students, Total Number of School Teachers, and Total Number of School Administrative
Personnel).

137

Table 2B.15 Heterogeneous effect of the past-year average maximum temperature on Saber 11
scores by exposure to Ser Pilo Paga under specific House strata levels, 2014-II–2019-I.

Variables

Math

(1)

(2)

Panel A: Overall
Spanish

(3)

(4)

Total Score

(5)

(6)

Avg. Maximum Temperature (AMT)

Exposure to SPP × Post

AMT × Exposure to SPP × Post

-0.376
(0.712)
3.061***
(0.995)
-0.0109
(0.0360)

-0.909
(2.736)
1.189
(3.222)
-0.00880
(0.106)

-0.475
(0.565)
0.171
(0.594)
-0.00679
(0.0218)

-2.416
(1.963)
-0.250
(3.042)
0.0414
(0.117)

-0.139
(0.599)
1.534**
(0.736)
-0.0108
(0.0262)

-2.673
(2.048)
-1.803
(2.776)
0.117
(0.101)

House Strata
Observations
𝑅2

Variables

Avg. Maximum Temperature (AMT)

Exposure to SPP × Post

AMT × Exposure to SPP × Post

House Strata
Observations
𝑅2

Variables

Avg. Maximum Temperature (AMT)

Exposure to SPP × Post

AMT × Exposure to SPP × Post

House Strata
Observations
𝑅2
School FE
Time-State FE
Heterogeneous Trends(a)
Additional Controls

Low
2,011,055
0.337

High
162,315
0.537

Low
2,011,055
0.291

High
162,315
0.438

Low
2,011,055
0.379

High
162,315
0.592

Math

(1)

(2)

-2.319***
(0.789)
6.532***
(1.794)
-0.110**
(0.0494)

Low
1,354,424
0.320

0.144
(3.642)
0.566
(4.303)
-0.0496
(0.136)

High
150,960
0.485

Panel B: Urban
Spanish

(3)

(4)

Total Score

(5)

(6)

-2.062***
(0.707)
0.833
(1.050)
0.0114
(0.0360)

-1.525
(2.112)
-2.911
(3.001)
0.136
(0.125)

-2.102***
(0.529)
3.754***
(1.263)
-0.0489
(0.0386)

-2.234
(2.527)
-3.157
(2.878)
0.151
(0.109)

Low
1,354,424
0.263

High
150,960
0.370

Low
1,354,424
0.357

High
150,960
0.531

Math

(1)

(2)

Panel C: Rural
Spanish

(3)

(4)

Total Score

(5)

(6)

1.528*
(0.779)
2.142*
(1.102)
-0.0107
(0.0397)

Low
656,626
0.326
Yes
Yes
Yes
Yes

-7.397
(4.850)
21.49*
(12.17)
-0.765*
(0.414)

High
11,335
0.735
Yes
Yes
Yes
Yes

0.988
(0.624)
-0.148
(0.687)
-0.0208
(0.0237)

Low
656,626
0.275
Yes
Yes
Yes
Yes

-12.65**
(5.594)
17.62*
(9.758)
-0.571*
(0.293)

High
11,335
0.674
Yes
Yes
Yes
Yes

1.539**
(0.690)
0.757
(0.822)
-0.0128
(0.0281)

Low
656,626
0.356
Yes
Yes
Yes
Yes

-7.106*
(4.247)
13.46
(9.851)
-0.424
(0.344)

High
11,335
0.796
Yes
Yes
Yes
Yes

Notes — Cluster (weather station level) robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. The response
variable is 100𝑦, where 𝑦 is the standardized score of the specific subject (Math, Spanish) or total test score. Avg. maximum
temperature captures the average daily maximum temperature (in °C) during the calendar year before the exam within each round.
Mean precipitation is the average daily precipitation (mm/day) for the calendar year before the exam date. Attribute controls added
to the regression include age (and its square), dummy variables for the maximum educational attainment of the mother, whether
the test taker is female, self-identification as part of an ethnic minority, and whether the exam taker lives and studies in different
Municipalities. (a) Trend heterogeneity is allowed by adding interactions between period (year-semester) dummies and quartiles of
Municipality-level variables (namely Total High-School Students, Total Number of School Teachers, and Total Number of School
Administrative Personnel).

138

Table 2B.16 Heterogeneous effect of past-year average maximum temperature on Saber 11 scores by education of the mother (reporting
education-level dummies), 2014-II–2019-I.

Variables

Avg. maximum temperature (AMT)

AMT × Mother: Complete primary

AMT × Mother: Complete secondary

AMT × Mother: Complete tertiary

Mother: Complete primary

Mother: Complete secondary

Mother: Complete tertiary

Mean precipitation

Observations
𝑅2
School FE
Time-State FE
Attribute Controls
Sample

(1)
Math

(2)
Math

(3)
Math

(4)
Spanish

(5)
Spanish

(6)
Spanish

(7)
Total

(8)
Total

(9)
Total

-0.308
(0.782)
-0.154***
(0.0542)
-0.392***
(0.0604)
-0.526***
(0.143)
9.794***
(1.654)
24.33***
(1.653)
45.14***
(4.241)

-2.563**
(1.143)
-0.166***
(0.0596)
-0.350***
(0.0594)
-0.495***
(0.130)
10.66***
(1.716)
22.66***
(1.606)
42.96***
(3.498)

1.592**
(0.766)
-0.100*
(0.0563)
-0.498***
(0.0978)
-0.718***
(0.139)
7.375***
(1.659)
28.40***
(2.707)
54.04***
(3.732)

-0.145
(0.622)
-0.143***
(0.0455)
-0.384***
(0.0529)
-0.512***
(0.102)
8.503***
(1.398)
23.81***
(1.460)
43.89***
(3.017)

-1.946***
(0.701)
-0.183***
(0.0477)
-0.361***
(0.0551)
-0.499***
(0.0737)
9.757***
(1.428)
22.61***
(1.480)
42.17***
(1.861)

-0.161
(0.269)

-0.855**
(0.385)

0.0800
(0.191)

-0.0977
(0.190)

-0.521
(0.345)

2,175,363
0.392
Yes
Yes
Yes
All

1,506,211
0.384
Yes
Yes
Yes
Urban

669,150
0.337
Yes
Yes
Yes
Rural

2,175,363
0.334
Yes
Yes
Yes
All

1,506,211
0.312
Yes
Yes
Yes
Urban

1.050*
(0.611)
-0.0701
(0.0520)
-0.460***
(0.0885)
-0.671***
(0.124)
5.936***
(1.535)
26.87***
(2.543)
51.72***
(3.342)

0.0628
(0.179)

669,150
0.284
Yes
Yes
Yes
Rural

0.0722
(0.647)
-0.166***
(0.0518)
-0.412***
(0.0540)
-0.575***
(0.117)
8.787***
(1.624)
23.89***
(1.476)
46.22***
(3.450)

-2.138***
(0.714)
-0.200***
(0.0570)
-0.395***
(0.0611)
-0.560***
(0.109)
10.09***
(1.748)
23.03***
(1.665)
44.69***
(2.903)

1.640**
(0.665)
-0.0971**
(0.0478)
-0.471***
(0.0878)
-0.730***
(0.125)
6.257***
(1.418)
26.35***
(2.480)
53.51***
(3.367)

-0.135
(0.188)

-0.634**
(0.286)

-0.0200
(0.154)

2,175,363
0.442
Yes
Yes
Yes
All

1,506,211
0.430
Yes
Yes
Yes
Urban

669,150
0.371
Yes
Yes
Yes
Rural

Notes — Cluster (weather station level) robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. The response
variable is 100𝑦, where 𝑦 is the standardized score of the specific subject (Math, Spanish) or total test score. Avg. maximum
temperature captures the average of the daily maximum temperature (in °C) during the calendar year before the exam within each
round. Mean precipitation is the average daily precipitation (mm/day) for the calendar year before the exam date. Attribute controls
added to the regression include age (and its square), dummy variables for whether the test taker is female, self-identification as part
of an ethnic minority, and whether the exam taker lives and studies in different Municipalities.

139

Table 2B.17 Heterogeneous effect of past-year average maximum temperature on Saber 11 scores by house strata (reporting strata-level
dummies), 2014-II–2019-I.

Variables

Avg. maximum temperature (AMT)

AMT × House Strata: Level 2

AMT × House Strata: Level 3

AMT × House Strata: Level 4

AMT × House Strata: Level 5

AMT × House Strata: Level 6

House Strata: Level 2

House Strata: Level 3

House Strata: Level 4

House Strata: Level 5

House Strata: Level 6

Mean precipitation

Observations
𝑅2
School FE
Time-State FE
Attribute Controls
Sample

(1)
Math

(2)
Math

(3)
Math

(4)
Spanish

(5)
Spanish

(6)
Spanish

(7)
Total

(8)
Total

(9)
Total

-0.183
(0.766)
-0.564***
(0.0571)
-0.878***
(0.119)
-1.392***
(0.200)
-1.374***
(0.249)
-1.619***
(0.279)
15.15***
(1.590)
20.01***
(2.949)
28.24***
(4.830)
23.65***
(6.286)
26.52***
(7.803)

-2.311**
(1.146)
-0.554***
(0.0650)
-0.786***
(0.145)
-1.165***
(0.203)
-1.185***
(0.275)
-1.376***
(0.263)
15.78***
(1.795)
19.85***
(3.559)
26.59***
(4.889)
23.71***
(6.403)
25.54***
(7.113)

1.591**
(0.760)
-0.533***
(0.0826)
-0.986***
(0.182)
-1.559***
(0.347)
-0.619**
(0.312)
-0.990**
(0.477)
13.83***
(2.240)
17.22***
(4.462)
12.29
(10.84)
-18.00**
(8.730)
-8.810
(13.86)

-0.0721
(0.613)
-0.461***
(0.0571)
-0.794***
(0.110)
-1.165***
(0.163)
-1.211***
(0.184)
-1.338***
(0.209)
14.40***
(1.647)
22.68***
(2.779)
29.09***
(3.982)
27.02***
(4.680)
26.51***
(5.593)

-1.813**
(0.691)
-0.441***
(0.0651)
-0.669***
(0.136)
-0.913***
(0.157)
-0.941***
(0.181)
-1.108***
(0.208)
14.87***
(1.832)
21.47***
(3.335)
26.50***
(3.881)
24.76***
(4.689)
25.66***
(5.680)

-0.180
(0.263)

-0.862**
(0.380)

0.0635
(0.190)

-0.114
(0.184)

-0.527
(0.337)

2,175,363
0.392
Yes
Yes
Yes
All

1,506,211
0.384
Yes
Yes
Yes
Urban

669,150
0.337
Yes
Yes
Yes
Rural

2,175,363
0.334
Yes
Yes
Yes
All

1,506,211
0.312
Yes
Yes
Yes
Urban

1.045*
(0.592)
-0.439***
(0.0761)
-1.046***
(0.171)
-1.454***
(0.399)
-1.036***
(0.295)
-0.365
(0.370)
12.95***
(2.173)
24.75***
(4.403)
18.00
(12.49)
1.163
(8.318)
-17.97
(10.98)

0.0477
(0.179)

669,150
0.284
Yes
Yes
Yes
Rural

0.149
(0.632)
-0.504***
(0.0598)
-0.843***
(0.117)
-1.339***
(0.180)
-1.347***
(0.224)
-1.550***
(0.244)
14.49***
(1.704)
21.85***
(2.898)
30.81***
(4.421)
27.31***
(5.679)
28.91***
(6.760)

-1.964***
(0.711)
-0.512***
(0.0695)
-0.755***
(0.145)
-1.124***
(0.178)
-1.134***
(0.240)
-1.324***
(0.231)
15.85***
(1.896)
21.90***
(3.516)
29.64***
(4.355)
27.07***
(5.812)
28.86***
(6.250)

1.617**
(0.651)
-0.448***
(0.0735)
-0.995***
(0.169)
-1.442***
(0.351)
-0.796***
(0.292)
-0.627*
(0.327)
12.10***
(2.037)
20.57***
(4.233)
13.18
(11.18)
-10.36
(8.228)
-16.81*
(9.802)

-0.153
(0.181)

-0.641**
(0.276)

-0.0347
(0.153)

2,175,363
0.442
Yes
Yes
Yes
All

1,506,211
0.430
Yes
Yes
Yes
Urban

669,150
0.372
Yes
Yes
Yes
Rural

Notes — Cluster (weather station level) robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. The response
variable is 100𝑦, where 𝑦 is the standardized score of the specific subject (Math, Spanish) or total test score. Avg. maximum
temperature captures the average daily maximum temperature (in °C) during the calendar year before the exam within each round.
Mean precipitation is the average daily precipitation (mm/day) for the calendar year before the exam date. Attribute controls added
to the regression include age (and its square), dummy variables for the maximum educational attainment of the mother, whether
the test taker is female, self-identification as part of an ethnic minority, and whether the exam taker lives and studies in different
Municipalities.

140

Table 2B.18 Heterogeneous effect of the past-year average maximum temperature on Saber 11 scores by exposure to Ser Pilo Paga,
excluding Municipalities with top 1% of exposure, 2014-II–2019-I.

Variables

Avg. maximum temperature (AMT)

Exposure to SPP × Post

AMT × Exposure to SPP × Post

(1)
Math

(2)
Math

(3)
Math

(4)
Spanish

(5)
Spanish

(6)
Spanish

(7)
Total

(8)
Total

(9)
Total

-0.509
(0.712)
2.873**
(1.147)
-0.00435
(0.0397)

-2.267***
(0.843)
6.127***
(1.812)
-0.108**
(0.0504)

1.436*
(0.781)
1.748
(1.370)
0.00207
(0.0468)

-0.563
(0.582)
-0.110
(0.696)
-0.000992
(0.0253)

-2.033***
(0.727)
0.376
(1.102)
0.0231
(0.0379)

0.886
(0.629)
-0.516
(0.787)
-0.0125
(0.0272)

-0.268
(0.605)
1.327
(0.811)
-0.00390
(0.0281)

-2.182***
(0.546)
3.216**
(1.284)
-0.0356
(0.0393)

1.476**
(0.689)
0.455
(0.978)
-0.00454
(0.0325)

Mean precipitation

Mean precipitation × Exposure to SPP × Post

Observations
𝑅2
School FE
Time-State FE
Attribute Controls
Heterogeneous Trends(a)
Sample

0.0018
(0.235)
-0.0838
(0.0782)

2,171,492
0.392
Yes
Yes
Yes
Yes
All

-0.354
(0.381)
-0.235
(0.148)

1,506,211
0.384
Yes
Yes
Yes
Yes
Urban

0.123
(0.207)
-0.0108
(0.0842)

665,280
0.336
Yes
Yes
Yes
Yes
Rural

-0.0374
(0.203)
-0.00678
(0.0481)

2,171,492
0.334
Yes
Yes
Yes
Yes
All

-0.536
(0.407)
0.0525
(0.0978)

1,506,211
0.312
Yes
Yes
Yes
Yes
Urban

0.0910
(0.168)
-0.0245
(0.0636)

665,280
0.284
Yes
Yes
Yes
Yes
Rural

-0.0244
(0.172)
-0.0452
(0.0594)

2,171,492
0.442
Yes
Yes
Yes
Yes
All

-0.268
(0.277)
-0.166
(0.105)

1,506,211
0.430
Yes
Yes
Yes
Yes
Urban

-0.00933
(0.162)
0.00480
(0.0666)

665,280
0.371
Yes
Yes
Yes
Yes
Rural

Notes — Cluster (weather station level) robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. The response variable
is 100𝑦, where 𝑦 is the standardized score of the specific subject (Math, Spanish) or total test score. Avg. maximum temperature
captures the average daily maximum temperature (in °C) during the calendar year before the exam within each round. Mean
precipitation is the average daily precipitation (mm/day) for the calendar year before the exam date. Attribute controls added to the
regression include age (and its square), dummy variables for the maximum educational attainment of the mother, dwelling stratum
classification, whether the test taker is female, self-identification as part of an ethnic minority, and whether the exam taker lives
and studies in different Municipalities. (a) Trend heterogeneity is allowed by adding interactions between period (year-semester)
dummies and quartiles of Municipality-level variables (namely Total High-School Students, Total Number of School Teachers, and
Total Number of School Administrative Personnel).

141

Table 2B.19 Heterogeneous effect of the past-year average maximum temperature on Saber 11 scores by exposure to Ser Pilo Paga,
excluding Municipalities with top 5% of exposure, 2014-II–2019-I.

Variables

Avg. maximum temperature (AMT)

Exposure to SPP × Post

AMT × Exposure to SPP × Post

(1)
Math

(2)
Math

(3)
Math

(4)
Spanish

(5)
Spanish

(6)
Spanish

(7)
Total

(8)
Total

(9)
Total

-0.621
(0.727)
2.547*
(1.442)
0.0287
(0.0479)

-2.359***
(0.865)
5.727***
(1.952)
-0.0974*
(0.0541)

1.325
(0.811)
1.194
(1.910)
0.0507
(0.0698)

-0.581
(0.600)
-0.600
(0.945)
0.0147
(0.0340)

-2.167***
(0.730)
-0.291
(1.204)
0.0411
(0.0397)

0.960
(0.650)
-0.782
(1.076)
-0.00571
(0.0385)

-0.322
(0.617)
0.957
(1.026)
0.0236
(0.0351)

-2.270***
(0.552)
2.915**
(1.402)
-0.0267
(0.0420)

1.453**
(0.705)
-0.112
(1.321)
0.0326
(0.0488)

Mean precipitation

Mean precipitation × Exposure to SPP × Post

Observations
𝑅2
School FE
Time-State FE
Attribute Controls
Heterogeneous Trends(a)
Sample

0.0495
(0.234)
-0.114
(0.0859)

2,147,373
0.392
Yes
Yes
Yes
Yes
All

-0.345
(0.386)
-0.240
(0.149)

0.187
(0.211)
-0.0544
(0.0957)

-0.0436
(0.205)
-8.94e-05
(0.0540)

-0.537
(0.411)
0.0447
(0.0941)

1,504,785
0.384
Yes
Yes
Yes
Yes
Urban

642,587
0.331
Yes
Yes
Yes
Yes
Rural

2,147,373
0.335
Yes
Yes
Yes
Yes
All

1,504,785
0.312
Yes
Yes
Yes
Yes
Urban

0.0901
(0.168)
-0.0271
(0.0732)

642,587
0.280
Yes
Yes
Yes
Yes
Rural

0.00847
(0.169)
-0.0634
(0.0638)

2,147,373
0.443
Yes
Yes
Yes
Yes
All

-0.253
(0.279)
-0.176*
(0.103)

1,504,785
0.430
Yes
Yes
Yes
Yes
Urban

0.0294
(0.162)
-0.0217
(0.0746)

642,587
0.366
Yes
Yes
Yes
Yes
Rural

Notes — Cluster (weather station level) robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. The response variable
is 100𝑦, where 𝑦 is the standardized score of the specific subject (Math, Spanish) or total test score. Avg. maximum temperature
captures the average daily maximum temperature (in °C) during the calendar year before the exam within each round. Mean
precipitation is the average daily precipitation (mm/day) for the calendar year before the exam date. Attribute controls added to the
regression include age (and its square), dummy variables for the maximum educational attainment of the mother, dwelling stratum
classification, whether the test taker is female, self-identification as part of an ethnic minority, and whether the exam taker lives
and studies in different Municipalities. (a) Trend heterogeneity is allowed by adding interactions between period (year-semester)
dummies and quartiles of Municipality-level variables (namely Total High-School Students, Total Number of School Teachers, and
Total Number of School Administrative Personnel).

142

Table 2B.20 Heterogeneous effect of the past-year average maximum temperature on Saber 11 scores by exposure to Ser Pilo Paga,
excluding Municipalities with top 10% of exposure, 2014-II–2019-I.

Variables

Avg. maximum temperature (AMT)

Exposure to SPP × Post

AMT × Exposure to SPP × Post

(1)
Math

(2)
Math

(3)
Math

(4)
Spanish

(5)
Spanish

(6)
Spanish

(7)
Total

(8)
Total

(9)
Total

-0.548
(0.749)
3.700**
(1.693)
0.0228
(0.0562)

-2.084**
(0.914)
7.331***
(1.581)
-0.136***
(0.0483)

1.367*
(0.820)
3.037
(2.400)
0.0166
(0.0850)

-0.520
(0.619)
-0.242
(1.116)
0.0121
(0.0404)

-1.878**
(0.793)
0.158
(1.408)
0.00872
(0.0497)

0.936
(0.659)
-0.0889
(1.399)
0.00283
(0.0468)

-0.237
(0.639)
1.634
(1.174)
0.0244
(0.0415)

-1.930***
(0.596)
3.939***
(1.306)
-0.0482
(0.0410)

1.414*
(0.716)
0.818
(1.619)
0.0272
(0.0589)

Mean precipitation

Mean precipitation × Exposure to SPP × Post

Observations
𝑅2
School FE
Time-State FE
Attribute Controls
Heterogeneous Trends(a)
Sample

0.148
(0.240)
-0.188*
(0.107)

2,099,799
0.394
Yes
Yes
Yes
Yes
All

-0.257
(0.411)
-0.340**
(0.144)

1,480,860
0.385
Yes
Yes
Yes
Yes
Urban

0.277
(0.232)
-0.107
(0.147)

618,938
0.331
Yes
Yes
Yes
Yes
Rural

-0.0576
(0.204)
-0.0142
(0.0653)

2,099,799
0.337
Yes
Yes
Yes
Yes
All

-0.665
(0.422)
0.0449
(0.110)

1,480,860
0.313
Yes
Yes
Yes
Yes
Urban

0.161
(0.164)
-0.114
(0.102)

618,938
0.280
Yes
Yes
Yes
Yes
Rural

0.0629
(0.167)
-0.113
(0.0793)

2,099,799
0.445
Yes
Yes
Yes
Yes
All

-0.202
(0.302)
-0.242**
(0.109)

1,480,860
0.431
Yes
Yes
Yes
Yes
Urban

0.0969
(0.167)
-0.0812
(0.112)

618,938
0.366
Yes
Yes
Yes
Yes
Rural

Notes — Cluster (weather station level) robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. The response variable
is 100𝑦, where 𝑦 is the standardized score of the specific subject (Math, Spanish) or total test score. Avg. maximum temperature
captures the average daily maximum temperature (in °C) during the calendar year before the exam within each round. Mean
precipitation is the average daily precipitation (mm/day) for the calendar year before the exam date. Attribute controls added to the
regression include age (and its square), dummy variables for the maximum educational attainment of the mother, dwelling stratum
classification, whether the test taker is female, self-identification as part of an ethnic minority, and whether the exam taker lives
and studies in different Municipalities. (a) Trend heterogeneity is allowed by adding interactions between period (year-semester)
dummies and quartiles of Municipality-level variables (namely Total High-School Students, Total Number of School Teachers, and
Total Number of School Administrative Personnel).

143

Table 2B.21 Heterogeneous effect of the past-year average maximum temperature on Saber 11 scores by exposure to Ser Pilo Paga,
excluding Municipalities with no exposure, 2014-II–2019-I.

Variables

Avg. maximum temperature (AMT)

Exposure to SPP × Post

AMT × Exposure to SPP × Post

(1)
Math

(2)
Math

(3)
Math

(4)
Spanish

(5)
Spanish

(6)
Spanish

(7)
Total

(8)
Total

(9)
Total

-0.514
(0.731)
3.857***
(0.928)
-0.0311
(0.0338)

-2.321***
(0.811)
6.124***
(1.809)
-0.113**
(0.0507)

1.730**
(0.821)
2.936**
(1.189)
-0.0273
(0.0410)

-0.664
(0.606)
0.437
(0.636)
-0.0126
(0.0239)

-2.042***
(0.741)
0.427
(1.105)
0.0194
(0.0390)

1.058
(0.671)
0.289
(0.717)
-0.0325
(0.0256)

-0.344
(0.615)
2.116***
(0.660)
-0.0230
(0.0242)

-2.148***
(0.545)
3.289**
(1.286)
-0.0410
(0.0393)

1.641**
(0.730)
1.406*
(0.841)
-0.0271
(0.0284)

Mean precipitation

Mean precipitation × Exposure to SPP × Post

Observations
𝑅2
School FE
Time-State FE
Attribute Controls
Heterogeneous Trends(a)
Sample

0.0702
(0.299)
-0.121*
(0.0701)

2,047,351
0.385
Yes
Yes
Yes
Yes
All

-0.357
(0.384)
-0.213
(0.149)

1,499,740
0.383
Yes
Yes
Yes
Yes
Urban

0.232
(0.285)
-0.0563
(0.0737)

547,610
0.332
Yes
Yes
Yes
Yes
Rural

-0.0212
(0.263)
-0.0293
(0.0469)

-0.540
(0.404)
0.0659
(0.0954)

2,047,351
0.325
Yes
Yes
Yes
Yes
All

1,499,740
0.311
Yes
Yes
Yes
Yes
Urban

0.189
(0.239)
-0.0525
(0.0561)

547,610
0.279
Yes
Yes
Yes
Yes
Rural

0.0189
(0.222)
-0.0783
(0.0528)

2,047,351
0.434
Yes
Yes
Yes
Yes
All

-0.264
(0.274)
-0.153
(0.103)

1,499,740
0.429
Yes
Yes
Yes
Yes
Urban

0.00609
(0.224)
-0.0317
(0.0585)

547,610
0.364
Yes
Yes
Yes
Yes
Rural

Notes — Cluster (weather station level) robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. The response variable
is 100𝑦, where 𝑦 is the standardized score of the specific subject (Math, Spanish) or total test score. Avg. maximum temperature
captures the average daily maximum temperature (in °C) during the calendar year before the exam within each round. Mean
precipitation is the average daily precipitation (mm/day) for the calendar year before the exam date. Attribute controls added to the
regression include age (and its square), dummy variables for the maximum educational attainment of the mother, dwelling stratum
classification, whether the test taker is female, self-identification as part of an ethnic minority, and whether the exam taker lives
and studies in different Municipalities. (a) Trend heterogeneity is allowed by adding interactions between period (year-semester)
dummies and quartiles of Municipality-level variables (namely Total High-School Students, Total Number of School Teachers, and
Total Number of School Administrative Personnel).

144

Table 2B.22 Heterogeneous effect of the past-year average maximum temperature on Saber 11 scores by exposure to Ser Pilo Paga, with
SPP at the extensive margin, 2014-II–2019-I.

Variables

Avg. maximum temperature (AMT)

Exposed to SPP (1=Yes) × Post

AMT × Exposed to SPP (1=Yes) × Post

(1)
Math

(2)
Math

(3)
Math

(4)
Spanish

(5)
Spanish

(6)
Spanish

(7)
Total

(8)
Total

(9)
Total

-0.285
(0.716)
15.36***
(3.722)
-0.290**
(0.119)

-2.520**
(0.956)
22.29***
(7.968)
-0.287*
(0.170)

1.831**
(0.804)
15.60***
(4.733)
-0.420**
(0.161)

-0.514
(0.584)
0.407
(2.614)
-0.0431
(0.0860)

-2.281***
(0.705)
3.168
(6.392)
0.0604
(0.100)

0.955
(0.641)
2.104
(3.046)
-0.118
(0.0986)

-0.171
(0.607)
6.578**
(2.725)
-0.142
(0.0878)

-2.381***
(0.601)
12.21**
(5.814)
-0.0959
(0.117)

1.655**
(0.705)
6.601**
(3.222)
-0.216*
(0.112)

Mean precipitation

Mean precipitation × Exposed to SPP (1=Yes) × Post

Observations
𝑅2
School FE
Time-State FE
Attribute Controls
Heterogeneous Trends(a)
Sample

0.149
(0.275)
-0.456***
(0.163)

-0.239
(0.431)
-0.610***
(0.129)

2,175,278
0.392
Yes
Yes
Yes
Yes
All

1,506,211
0.384
Yes
Yes
Yes
Yes
Urban

0.213
(0.256)
-0.203
(0.272)

669,066
0.337
Yes
Yes
Yes
Yes
Rural

0.0227
(0.187)
-0.0849
(0.0999)

2,175,278
0.334
Yes
Yes
Yes
Yes
All

-0.425
(0.403)
-0.111
(0.113)

1,506,211
0.312
Yes
Yes
Yes
Yes
Urban

0.142
(0.168)
-0.155
(0.169)

669,150
0.284
Yes
Yes
Yes
Yes
Rural

0.0938
(0.177)
-0.287**
(0.140)

-0.170
(0.306)
-0.447***
(0.110)

2,175,363
0.442
Yes
Yes
Yes
Yes
All

1,506,211
0.430
Yes
Yes
Yes
Yes
Urban

0.0541
(0.178)
-0.118
(0.191)

669,150
0.371
Yes
Yes
Yes
Yes
Rural

Notes — Cluster (weather station level) robust standard errors in parentheses. *** p<0.01, ** p<0.05, * p<0.1. The response variable
is 100𝑦, where 𝑦 is the standardized score of the specific subject (Math, Spanish) or total test score. Avg. maximum temperature
captures the average daily maximum temperature (in °C) during the calendar year before the exam within each round. Mean
precipitation is the average daily precipitation (mm/day) for the calendar year before the exam date. Attribute controls added to the
regression include age (and its square), dummy variables for the maximum educational attainment of the mother, dwelling stratum
classification, whether the test taker is female, self-identification as part of an ethnic minority, and whether the exam taker lives
and studies in different Municipalities. (a) Trend heterogeneity is allowed by adding interactions between period (year-semester)
dummies and quartiles of Municipality-level variables (namely Total High-School Students, Total Number of School Teachers, and
Total Number of School Administrative Personnel).

145

CHAPTER 3

HEAT, YIELDS, AND INCIDENTAL TRUNCATION

3.1

Introduction

Global agricultural sustainability is at stake with projected increases in average temperatures,

as there are significant concerns about weather and climate impacts on the sector’s productivity

and its ability to adapt to a changing environment. The literature on the evaluation of the effects of

heat on staple crops is vast, highlighting that a warming environment will lead to decreases in the

productivity of corn, rice, soybean, sorghum, and wheat (Challinor et al., 2014; Lobell & Asseng,

2017; Miao, Khanna, & Huang, 2016; Miller, Tack, & Bergtold, 2021; Schlenker & Roberts, 2009;

Tack, Barkley, & Nalley, 2015), as well as their suitability and production risks (Akpoti, Groen,

Dossou-Yovo, Kabo-bah, & Zwart, 2022; Feng, Wang, Ma, Peng, & Shan, 2021; Li, Fleisher, &

Barnaby, 2024; Ramirez-Cabral, Kumar, & Shabani, 2017; Yue, Zhang, & Shang, 2019; Zhao

et al., 2021), thus meaning a continuous threat to global food security. These shocks further

imply a policy challenge as projected negative impacts translate to increasing needs in funding for

agricultural research and development in both developed and developing economies (Maredia &

Martinez, 2024; Ortiz-Bobea, Chambers, He, & Lobell, 2025). Therefore, properly measuring the

effects of warming temperatures remains a significant target of agricultural research.

The increasing availability of large-scale, farm-level data (e.g., yields, input use, timing of

production), along with more precise measures of weather conditions (e.g., more granular measures

of heat exposure and precipitation), has led to a significant share of recent research being focused on

identification issues. For instance, Mérel and Gammans (2021) discusses retrieving short- and long-

term climate impacts on yields, using fixed-effects (FE) estimations on large farm-level longitudinal

datasets. Moreover, more granular satellite crop data allows for assessing the emergence of

adaptation strategies like double-cropping as a response to larger-than-usual productive seasons in

traditionally seasonal climates (Gammans, Mérel, & Ortiz-Bobea, 2024). However, especially in

the former, most analyses rely on samples displaying considerable variability in farm-observability.

Namely, most of these studies leave aside the fact that some farms enter and exit the sample,

146

presumably considering it a completely random event, which could potentially lead to estimation

bias. While studies as Miller et al. (2021) acknowledge that sample selection cannot be completely

ruled out, they consider the possibility that the ultimate exit of a farm (i.e., no longer being observed

in the sample) resulting from crop failure–argued to be a rare outcome–is the only relevant factor

that could bias the estimation of weather-related effects. Nevertheless, this does not fully address

why a farm may choose to produce in some (but not all) periods before a total exit from the sample.

Are estimates of heat impact on crop yields sensitive to the inclusion of sample selection

corrections?

Is there a significant relationship between crop-specific production participation

decisions and weather and market variables? Using open-access information by Miller et al. (2021)

from the Kansas Farm Management Association (KFMA) for sorghum production during 1978-

2015, associated PRISM weather data, and yearly data on farmers’ received prices for sorghum

and corn from USDA-NASS during the period, we use (1) a random-effect Probit to estimate how

weather variables and prices affect the probability of farmer participation in sorghum production,

and (2) a Heckman-style approach in a fixed-effect framework to test (and potentially correct) for

sample selection in the estimation of heat impact on sorghum yields.

Under a working assumption of high accuracy in seasonal forecasts, such that farmers correctly

predict weather conditions, we find that additional exposure to temperatures favorable for sorghum

production (i.e., further degree days between 10 and 33°C) increase the probability of farmer partic-

ipation. Conversely, increasing exposure to temperatures unfavorable for production (degree days

below 10 or above 33°C) decrease the probability of participation. Meanwhile, our results further

suggest that increases in expected prices of sorghum increase the probability of participation, while

higher expected prices of corn lead to a lower likelihood of producing sorghum. Using contem-

porary or lagged weather variables yields statistically similar responses in production probability,

while using lags of observed prices leads to qualitatively similar (but statistically insignificant)

effects on the decision to produce sorghum. Overall, these findings suggest that production deci-

sions are climate-responsive, highlighting that period-specific farm observability in panel data is a

potential reflection of adaptation strategies.

147

Combining the results from various specifications of production participation, we implement

a Heckman-style test for sample selection. We find that when contemporary weather variables

are present in both participation and yield regressions (i.e., farmers correctly foresee weather

conditions), there is evidence of potential estimation bias. Conversely, whenever lags of weather

variables are used for the participation equation, I cannot reject the hypothesis that results from

fixed-effects estimates are consistent. While this finding is in line with recent literature on sample

selection in unbalanced panels (Baltagi, Jiménez-Martín, Labeaga, & Al Sadoon, 2023), I argue

that as weather predictions become increasingly precise and more widely available in agricultural

settings, the risk of bias in fixed-effects estimation should remain a concern to be addressed.

Our contribution is twofold. First, we contribute to the literature on productive participation,

highlighting that choosing to grow a crop of emerging importance is a decision responsive to

weather and price expectations. Moreover, we provide correlational evidence suggesting that

farmers may be following a forecasting strategy, potentially anticipating (at least partially) severe

shocks of heat, and thus adjusting their decision to produce accordingly. Second, we provide a

set of methodological recommendations to account for farm-observability in large longitudinal

datasets and test for potential sources of bias in estimating weather impacts on crop yields when

relying on traditional linear regression approaches. Some of our findings, for instance, indicate that

accounting for farm-observability due to productive participation decisions would lead to lower-

magnitude estimates–yet still significant–of high heat impacts on crop yields. In aggregate, our

findings suggest that in-sample patterns resembling climate adaptation are connected to attenuating

impacts from climate change.

The remainder of this chapter proceeds as follows: Section 3.2 briefly addresses our theoretical

background and the relevance of our crop of interest–sorghum. Section 3.3 presents our identifi-

cation strategy, while we present some summary statistics and stylized facts in Section 3.4. Then,

we present and discuss our main results and present some robustness checks in Section 3.5. A final

section presents a brief conclusion and discusses limitations and recommendations for future work.

148

3.2 Methods

3.2.1 Theoretical background and crop of interest

I consider a price-taking, profit-maximizing farmer 𝑖 in a perfectly competitive market, deciding

their crop productive allocation in a given agricultural season 𝑡, conditional on resource endow-

ments and technology constraints. Without loss of generality, assume that the farmer chooses to

allocate production between a crop of interest, 𝑆, and a single competing crop (𝐶). Under optimiza-

tion, the farmer sets production to levels at which the marginal rate of transformation (i.e., their

opportunity cost) equals the market price ratio (Nicholson & Snyder, 2012). From a comparative

statics perspective, increases in unfavorable weather conditions can impact both specific productive

allocations and total output by shifting the production possibility frontiers inward. Likewise, all

else equal, an increase in the price of a specific crop would lead to a larger production allocation

towards that crop.

The latter implies that in a dynamic setting, a rational farmer uses available information ahead

of a productive season to make a decision on productive allocation. For instance, if a farmer expects

harsh weather conditions, they would adjust their production allocation accordingly to maximize

their expected profits. Likewise, changes in expected prices would drive a decision to the extent of

production. Hence, this short-run response reflects an adaptation mechanism by affecting whether

a farm produces a given crop in a specific year. Particularly, the possibility that weather shocks

affect both the decision to produce a crop (via expectation) and final production outcomes, imply a

relevant connection between extensive and intensive margins that warrant further consideration in

estimating weather effects.

The relevance of sorghum stems from its (a) nutrient content comparability to that of corn, (b)

higher tolerance to heat and lower demand of water resources than other crops, and (c) increasing

global demand for biofuel, livestock feed, and processing for human consumption (Hariprasanna

& Rakshit, 2016; Wang, Upadhyaya, & Kole, 2014). Although its production has been suggested

as a productive alternative in light of harsher climate patterns in the US (Miller et al., 2021),

its suitability in global regions where it has been traditionally produced is being challenged by

149

climate change (Tack, Lingenfelser, & Jagadish, 2017). Moreover, its total acreage in the country

has steadily decreased for most of the past two decades, connected with reductions in private

investment and political impacts on international trade (Duff, Bice, Hoeffner, & Weinheimer, 2019;

Polansek, 2025). Therefore, evaluating how farmers make decisions about producing a crop that

could potentially serve as an adaptation to climate change, and whether said decisions could alter

the yield impacts of climate change, are questions of economic and policy relevance.

3.2.2 Empirical setting

In general, we are interested in estimating partial effects on the conditional mean of a variable 𝑦1,

namely 𝐸 [𝑦1|x1]. However, we face a case of partial non-response, leading to a scenario in which 𝑦1

is observed if and only if a condition occurs based on another variable 𝑦2.1 Therefore, we could only

approach the effects of x1 on 𝐸 [𝑦1|x1, 1[Cond(𝑦2)]], where 1[Cond(𝑦2)] is an index of whether a

specific condition on 𝑦2 is satisfied. In a linear case in which 𝐸 [𝑦1|x1, 1[Cond(𝑦2)]] = x1 𝜷+ 𝑓 (𝑦2),

and 𝐸 [x1𝑦2] ≠ 0, OLS estimates of 𝜷 that wave aside the nature of the relationship between x1 and

𝑦2 lead to biased and likely inconsistent estimates (Wooldridge, 2010), which is the standard result

of sample selection from Heckman (1979).2

As mentioned before, large longitudinal agricultural datasets at the farm level are becoming the

standard. If we consider a dataset with yearly information, 𝑡 = 1, ..., 𝑇, on crop yields 𝑦𝑖𝑡 for each

farm 𝑖 = 1, ..., 𝑛, during a common season (e.g., summer grains), a case of a perfectly balanced

panel, would allow us to observe a yield data point 𝑦𝑖𝑡 ≥ 0 for all 𝑖 and all 𝑡. Nevertheless, this

is not the norm. For most recent applications of yield panel data, we find that a given farm 𝑖 is

observed during a specific period going from

𝑇𝑖 to ¯𝑇𝑖, with
¯

𝑇𝑖 ≤ ¯𝑇𝑖 and
¯

𝑇𝑖, ¯𝑇𝑖 ∈ {1, ..., 𝑇 }. Moreover,
¯

we usually find several cases of missing yield data for some periods—i.e., there exists a farm 𝑖 such

1The problem of missing data resulting from non-response is most famously associated with that of consistently
estimating effects in wage equations vis-à-vis labor force participation (Gronau, 1974), where one only observes a wage
(𝑤 > 0) whenever an individual is effectively employed (𝑝 = 1). Otherwise, whenever an individual is not employed
(𝑝 = 0), the wage is not zero and is instead a missing value in the dataset. More specifically, Gronau (1974) defines
an individual as participating in the market whenever the wage offer is at least as large as the individual’s reservation
wage.

2Following the example of wage equations, it is natural to assume that increases in education will raise an
individual’s reservation wage. At the same time, this affects their decision to participate in the market and whether we
will observe a wage offer, thus rendering OLS estimates of the effects of education on a wage equation as potentially
inconsistent.

150

that 𝑦𝑖𝑡 is missing3 for some 𝑡 ∈ [

𝑇𝑖, ¯𝑇𝑖].
¯

Let 𝑝𝑖𝑡 be a binary variable taking the value of one if farm 𝑖 produces a crop of interest in period

𝑡, and zero otherwise. Hence, we face a scenario where the yield of our crop of interest (intensive

margin) follows:4

ln 𝑦𝑖𝑡 =

𝑓 (x𝑖𝑡; 𝜃) + 𝑢𝑖𝑡

if 𝑝𝑖𝑡 = 1

Unobserved

if 𝑝𝑖𝑡 = 0





where x𝑖𝑡 includes farm-level, and weather covariates, while 𝜃 is a vector of unknown parameters,

𝑢𝑖𝑡 is an error term, and 𝑓 is defined by the production technology. Meanwhile, the participation in

the crop of interest (extensive margin) is defined by a latent participation variable 𝑝∗

𝑖𝑡, such that

𝑝𝑖𝑡 =





1 if 𝑝∗

𝑖𝑡 = 𝑔(z𝑖𝑡; 𝜓) + 𝑒𝑖𝑡 ≥ 𝐺

0 if 𝑝∗

𝑖𝑡 = 𝑔(z𝑖𝑡; 𝜓) + 𝑒𝑖𝑡 < 𝐺

where z𝑖𝑡 includes farm-level, expected weather, and market (prices) covariates, 𝜓 is a vector of

unknown parameters, and 𝑒𝑖𝑡 is an error term. In this context, 𝑔 captures how weather and price

expectations affect the willingness to produce the crop of interest. Hence, there is a threshold of

indifference 𝐺, which defines whether a farmer chooses to produce.5

3.3

Identification strategy

3.3.1 A model of participation in sorghum production

We test whether a farmer’s decision to produce sorghum is influenced by weather and price

variables by defining our binary variable as 𝑝𝑖𝑡 = 1[Produces Sorghum]𝑖𝑡 taking the value of

one if farmer 𝑖 chooses to produce sorghum in year 𝑡, and zero otherwise. Specifically, we

assume that during a farm’s period of presence in the sample (i.e., all periods between

𝑇𝑖 and ¯𝑇𝑖)
¯

1[Produces Sorghum]𝑖𝑡 = 0 if yields are a missing data; otherwise, when yields are observed then

3It is worth noting that crop failure is a different case to that from missing values, as the former implies a yield of

zero.

4Note that in a case of crop failure (i.e., a significant or complete loss of harvestable yield) it could occur that 𝑦𝑖𝑡 = 0,
thus leading to an undefined value. However, unless in a case of extremely harsh conditions, this is rather a rare case.
In the open-access dataset, only 17 observations are seemingly reflecting a yield of zero (with an apparent imputation
of values as ln(𝑦𝑖𝑡 + 1) for cases where 𝑦𝑖𝑡 =0). Additional testing of our analyses excluding these observations (not
included here) does not change our inference.

5In a case with more disaggregated data on the cost structure of farms, one could consider the case of 𝑔 capturing
expected profits, in which case a farmer chooses to produce if and only if projected profits exceed a threshold 𝐺, which
is usually set to 𝐺 = 0 reflecting a decision to at least break even.

151

1[Produces Sorghum]𝑖𝑡 = 1. In addition, we assume that the probability of a farmer producing

sorghum in a given year follows a process such that:

Pr[ 𝑝𝑖𝑡 = 1|·] = Φ

(cid:20)

𝛼0 + 𝛼1𝐸 [Prec𝑖𝑡] + 𝜂1𝐸 [DLow𝑖𝑡] + 𝜂2𝐸 [DMed𝑖𝑡]+

𝜂3𝐸 [DHigh𝑖𝑡] + 𝜇1𝐸 [𝑃S

𝑡 ] + 𝜇2𝐸 [𝑃C

𝑡 ] + 𝜁𝑡

(cid:21)

,

(3.1)

where 𝐸 [X] is the expectation of X for the production period. The use of expectation follows

that the decision to produce occurs before the realization of weather or price variables. In this

specification, Prec𝑖𝑡 is total precipitation, while DLow𝑖𝑡, DMed𝑖𝑡, and DHigh𝑖𝑡 are, respectively,

the number of degree days during the productive period under 10°C, between 10-33°C, and above

33°C. Based on the analysis from Miller et al. (2021), an increasing exposure to degree days in

the lower- and upper-temperature bins (DLow𝑖𝑡 and DHigh𝑖𝑡) is detrimental to yields, while added

exposure to degree days in the bin DMed𝑖𝑡 captures a beneficial temperature range for the crop.

Finally, 𝑃𝑆

𝑡 and 𝑃𝐶

𝑡 are the prices of sorghum and an on-farm substitute (namely corn) received by

farmers in the State, respectively, while 𝑡 is also included as a trend variable.

From both an agronomic and an economic standpoint, it is a weak assumption having an

increasing number of degree days with unfavorable temperatures for the crop leading to a reduction

in the probability of producing (i.e., 𝜂1 and 𝜂3 are negative).6 Likewise, further exposure to degree

days in a range of favorable temperatures would increase the likelihood of producing (𝜂2 > 0). On

the other hand, higher prices of sorghum (corn) should potentially lead to an increased (decreased)

probability of producing sorghum, as it implies a case for likely higher (lower) revenues, and thus

𝜇1 > 0 (𝜇2 < 0). Meanwhile, the effect of precipitation is somewhat ambiguous. While additional

rain ultimately favors the vegetative development of the crop, sorghum has a yield gain advantage

over corn on lower levels of evapotranspiration and overall precipitation (Assefa et al., 2014).

Finally, we expect that the overall trend in participation is negative (𝜁 < 0), in line with national

trends of a reducing number of farms in the US according to the USDA-NASS.7

6For instance, they may prefer to focus on producing another crop that, while also potentially affected by the
Importantly, these ranges of
projected weather conditions, has a more favorable price and leads to higher profits.
degree days are also close to those found as optimal for modeling climate responsiveness in corn, soybean, and cotton
in the US (Schlenker & Roberts, 2009).

7https://www.ers.usda.gov/data-products/chart-gallery/gallery/chart-detail/?chartId=58268.

152

We consider various definitions of expected temperatures and prices, with two limiting cases

for weather variables. First, we take a case of high-accuracy predictions of weather conditions for

the production season, such that 𝐸𝑡 [X] = X𝑡. This scenario’s direct implication is that farmers have

perfect information, and any final exposure to unfavorable temperatures results from unavoidability

due to the necessary length of the productive period of sorghum. In our preferred specification, we

take contemporary values of weather variables and set up expected prices as the dynamic predictions

from a 2-lag vector autoregression.8 An alternate specification follows a naive estimation of prices,

taking the previous-year prices as reference values. Second, we consider a one-year lag scenario

for weather variables, assuming that farmers take the experienced previous year conditions as their

reference for making a decision on whether to produce, such that they set 𝐸𝑡 [X] = X𝑡−1. For this

case, we also include specifications based on dynamic price predictions and using the previous

year’s price information as a reference. Estimation of all specifications is done by maximum

likelihood.

3.3.2 Addressing sample selection: à la Heckman corrections

We argue that non-observability of yields (i.e., a farmer is not present in the crop yield data for

a given year) is potentially linked to climate (and price) responsiveness of production participation.

Under this assumption, estimating the unconditional effect of weather variables on yield, namely

𝜕𝐸 [𝑦𝑖𝑡]/𝜕𝑤 𝑗𝑖𝑡, is unfeasible. Yet, one can try to consistently estimate the effects among those who

choose to participate, namely 𝜕𝐸 [𝑦𝑖𝑡 | 𝑝𝑖𝑡 = 1]/𝜕𝑤 𝑗𝑖𝑡 (Wooldridge, 2010). Following the findings

of Baltagi et al. (2023), it would be expected that a problem of inconsistency would only arise

whenever contemporary weather variables appear in both the participation and yield models (i.e.,

under a high-precision prediction or perfect information scenario).

We consider an alternative specification of the log-yields equation from Miller et al. (2021),

which now follows

𝑦𝑖𝑡 = 𝛽0 + 𝛽1Prec𝑖𝑡 + 𝛽2Prec2

𝑖𝑡 + 𝛿1DLow𝑖𝑡 + 𝛿2DMed𝑖𝑡 + 𝛿3DHigh𝑖𝑡 + 𝜌 ˆ𝜆𝑖𝑡 + 𝛾𝑡 + 𝑐𝑖 + 𝑣𝑖𝑡,

(3.2)

8Given that using only a current year price to predict a next-year price would leave aside intertemporal adjustments of
price expectations (Nerlove, 1956), we define a vector autoregression 𝒀 𝑡 = 𝚷1𝒀 𝑡 −1+𝚷2𝒀 𝑡 −2+𝑼𝑡 , where 𝒀 ’
𝑡 ).
Therefore, the dynamic predictions for sorghum and corn prices for each year, are a function of both their prices in the
previous two years.

𝑡 = (𝑃S

𝑡 , 𝑃C

153

where 𝑦𝑖𝑡 is the sorghum log-yield for farmer 𝑖 in year 𝑡, the weather variables are defined as in the

previous subsection, 𝑣𝑖𝑡 is an idiosyncratic shock to yields, and 𝑐𝑖 is an unobserved, time-invariant,
individual heterogeneity. Under this specification, ˆ𝜆𝑖𝑡 is the inverse Mills ratio derived from the

participation model, such that

ˆ𝜆𝑖𝑡 =

𝜙 (cid:0)z𝑖𝑡; ˆ𝜶, ˆ𝜼, ˆ𝝁, ˆ𝜁 (cid:1)
Φ (cid:0)z𝑖𝑡; ˆ𝜶, ˆ𝜼, ˆ𝝁, ˆ𝜁 (cid:1)

,

(3.3)

where 𝜙 and Φ are the associated probability density and cumulative density functions from the

standard normal distribution, with coefficients retrieved from the participation analysis estimates.

Upon cluster-robust estimation, this specification approximates the fully robust test of sample

selection by variable addition suggested in Wooldridge (1995). Namely, the null hypothesis 𝜌 = 0

means no potential problems due to sample selection. Conversely, a failure to reject implies that

FE estimates of (3.2) that wave the role of ˆ𝜆𝑖𝑡 are likely inconsistent.9

3.4 Data

Our analysis is based on open-access information available in Miller et al. (2021) from the

database of the Kansas Farm Management Association (KFMA). The KFMA database is a lon-

gitudinal survey maintained by Kansas State University (KSU) and, while not representative of

farmers at the state level, it mainly represents large-scale operations (Kuethe, Briggeman, Paulson,

& Katchova, 2014). In addition to its application in Miller and coauthors’ research, the information

is used by Kansas State University researchers to derive production and financial recommendations

for farmers in the state. KFMA data has also been used to study the adoption of precision agriculture

(Griffin et al., 2017) and farm profitability (Stabel, Griffin, & Ibendahl, 2017).

The publicly available information contains a farm-level unique identifier, sorghum yield data,

observed weather variables during the farm- and year-specific production period, an anonymized

(yet unique) code that connects a farm to a specific PRISM data grid, and a code linking the farm

to a specific agricultural district. Since weather farm-specific data is only available for years in

which they produce sorghum, we input the sample-average weather variables10 from the grid (when

9In a case in which one fails to reject, it is necessary to adjust standard errors accordingly. However, in the case
of unbalanced panel data, this is no easy feat since the definition of a bootstrap is non-trivial. Hence, we take the
one-time estimate to provide some insights on potential problems of bias in usual FE estimations, but take the results
with caution since our calculated standard errors are likely inconsistent.

10Given that different farms choose different times for production, this imputation takes the simple average of degree

154

available) or the agricultural district for years when the farm does not produce sorghum. While the

available data is somewhat limited, as it does not include information about other crops’ production,

farm acreage, or input use, our setting of a partial equilibrium analysis of the extensive and intensive

margins of sorghum production in a competitive market only necessitates data on yields, climate,

and prices.

3.4.1 Descriptive statistics and stylized facts

We replicate the main descriptive statistics from Miller et al. (2021) in Table 3.1. The number

of farms producing sorghum in a given year varies from 449 to 1,676. The average farm is present

in the dataset producing sorghum for roughly 16 years,11 although a given farm may be present on

only one occasion or during the entire sample period (1978-2015). Average sorghum yields stand

at 63 hundredweight per hectare, with a maximum of 174 cwt/ha. In addition, farms experience

an average minimum (maximum) temperature of 16°C (29.8°C), along with an average cumulative

precipitation of 382 millimeters.

The KFMA data reveals a significant per-year decrease in the number of farmers producing

sorghum. In Figure 3.1, we present the number of farms throughout the 38-year period, along with

data on average temperatures experienced in Kansas per year, specifically for the June-September

period–the reference time for sorghum production. The average reduction in the number of farmers

was slight from 1978 to the late 1990s, but it became sharper in 1998. Given that farms in KFMA

usually represent large operations, the reduction is not necessarily trivial and could have important

implications at the sector level. Meanwhile, the State-level average temperatures have been slightly

increasing overall. Interestingly, there was an apparent downward trend in average temperatures

until the mid-1990s, followed by a sharp positive trend thereafter, in line with the global trends

reported by World Meteorological Organization (2023), which finds that average temperatures have

increased at an accelerated rate since the 1990s.12 From a pairwise perspective, this could point to

days experienced by farms within the grid.

11It is worth reemphasizing that this differs from the total range in which the farm is present in the sample, which

we define in the motivation as the length of [

𝑇𝑖, ¯𝑇𝑖].
¯

12I present an alternate version of this figure in Appendix, Figure 3A.1, allowing for a linear fit for the 1978-
1990 period, and another for 1990-2015. The marked differences in State-level temperature trends seem potentially
connected to farmers’ decisions to exit production.

155

the increases in temperature as a potential mechanism driving farmers out of sorghum production,

at least partially.13

Meanwhile, prices also reveal a significant trend to be considered. In Figure 3.2, we present

again the number of in-sample, KFMA sorghum-producing farms in the reference year, along with

farmers’ received price (in constant 2015 prices) for sorghum and corn, which are remarkably

coupled. Sorghum prices went from roughly 15 dollars per hundredweight ($/cwt) in the early

1980s to just above $5/cwt in the 2000s. It is precisely during the mid-2000s, when sorghum prices

were steadily at the lowest historic levels, that we observe the sharpest decline in the number of

farmers from the KMFA sample. With the price of corn (i.e., a leading competing crop with a

much broader market demand) following similar patterns, there are reasons to expect that structural

factors potentially affect the decision of farmers to stay out of sorghum production.

In aggregate, this preliminary data analysis hints that farmers’ decisions to participate in

sorghum production may be related to temperatures and prices in the market.14 These correlations

align with the extended literature on climate change impacts on agriculture, which points to

specialization and improvement in productive decisions as factors reducing the costs of a warmer

climate (Costinot, Donaldson, & Smith, 2016). Therefore, considering whether usual fixed-effects

estimates are potentially biased from a sample selection standpoint would suggest that estimations

may want to account for pre-existing adaptation strategies influencing the structure of observed

samples.

13While increasing temperatures can drive production away from sorghum, price increases could foster compensating
effects in observed acreage. For instance, while Kansas’ total acreage of sorghum decreased in 1997-2012 (Figure
3A.2)–correlated with increasing average temperatures and decreasing prices–the acreage increased significantly in
2017, likely connected to price increases in 2010-2016. However, the number of sorghum operations in Kansas
constantly decreased throughout 1997-2022 (Figure 3A.3), so acreage increases are concentrated across fewer farmers.
14Moreover, farmers could already be making decisions on their effective production period, potentially pulling
forward their planting time to experience more favorable weather. In line with the policy recommendation from Miller
et al. (2021), the farmer-observed temperatures display a decreasing trend, which contrasts with trends of increasing
average measures of temperature at the State level for the main productive season (Figure 3A.4).

156

3.5 Results and Discussion

3.5.1 Extensive margin: The case of production participation

We begin our testing of the responsiveness of production participation to weather and price

expectations in a case that assumes access to high-accuracy weather predictions, along with vari-

ations in definitions of price expectations, and summarize our findings in Table 3.2. Columns one

and three report Probit coefficients, while Columns two and four report average partial effects. In

line with the literature, we consistently detect a declining production participation rate through

time (Duff et al., 2019). Meanwhile, increases in precipitation and degree days in a range favorable

for the crop (DMed) are related to an increase in the participation rate.

In contrast, additional

degree days in ranges detrimental to the crop negatively relate to participation. Coefficients and

average partial effects of trend and weather variables are statistically equivalent regardless of the

type of price expectation variation.

In addition, price expectations operate as expected; using

dynamic predictions, we find that an additional dollar per cwt of sorghum (corn) correlated with

an increase (decrease) of 2.59 (2.66) percentage points in the participation in sorghum production.

These figures are roughly half their magnitude when relying on last-year prices as a reference for

expectations.15

While a scenario of access to high-accuracy forecasting is arguably a strong assumption, this

limiting case is informative as it points to a likely connection between farmers’ behavior and local

weather trends. Moreover, our findings suggest that both prices and expected shocks to production

possibilities likely affect the decision to produce a crop like sorghum, which aligns with our

theoretical framework. We then estimate the alternative scenario, in which weather predictions are

based on weather exposure levels from the past agricultural season, and summarize our findings in

Table 3.3. Our estimates for time-trend, precipitation, DLow, and DMed are qualitatively equivalent

and of similar magnitudes to those of the case of high-accuracy forecasting. Nevertheless, changes in

15These findings also indicate that while there are recommendations in switching to sorghum as an adaptation
strategy upon harsher climates, farmers still potentially adjust their decision to produce it to avoid yield impacts.
However, and perhaps more importantly, market demands (partially captured via price expectations) matter. While
sorghum could hypothetically serve as a substitute for a grain like corn, the market size for the latter is significantly
larger and could still lead to higher profits for farmers–even after accounting for the shocks of climate change.

157

price expectations are statistically insignificant regardless of their definition, and additional degree

days of high temperatures (DHigh) in the past season only weakly correlate with the decision to

participate in sorghum production. These findings could be partially explained by having high-

temperature degree days as events that are rarer and thus more difficult to predict, thus reducing the

impact of their realization on future behavior.16

3.5.2

Intensive margin: Yields and incidental truncation

Based on our extensive margin analysis findings, we now evaluate whether a climate-responsive

decision to participate can significantly alter (bias) the estimates of weather impacts on yields.

Starting again from a case of participation driven by access to high-accuracy weather forecasts,

we use a Heckman-style correction approach and summarize our results in Table 3.4. Column

one reports the benchmark result from Miller et al. (2021). On the other hand, Columns two and

three report the estimates adjusted by the added variable approach (inverse Mills ratio) derived

from a participation equation using contemporary weather variables and, respectively, dynamic

predictions and one-year lags of prices. We find that a case of high-accuracy weather forecasting,

along with price expectations based on dynamic predictions, suggests the presence of bias from FE

estimations that omit the role of farmers’ productive participation decisions. More specifically, this

adjustment leads to now statistically insignificant effects from degree days in low- and medium-

temperature bins, as well as to a reduction in the magnitude of the exposure to degree days above

33°C. Conversely, a correction based on a first-step estimation of participation based on price lags

suggests no risk of bias.

Next, we consider the alternate case in which the selection adjustment follows weather expec-

tations set as those realized during the previous year. We present these estimates in Table 3.5,

presenting again the benchmark estimation in Column one, while Heckman-adjusted regression

estimates are in Columns two and three. In this setting, we only detect a weakly significant effect

derived from the addition of the inverse Mills ratio. In practical terms, there is only a noticeable

16For example, in a setting like ours, where we argue that farmers use available information in preparation for the
season, past-year trends of cold- and medium-levels of temperature (which occur in the earlier part of the season) may
be more informative of the upcoming season.

158

change in the statistical significance of the trend covariate, thus acknowledging sustained growth

in sorghum productivity. Nevertheless, coefficient estimates for weather-related variables are sta-

tistically equivalent and are virtually unchanged. Ultimately, this alternate approach provides its

expected findings: past-season weather information has a seemingly limited (if any) effect on

productive decisions for upcoming productive seasons.

Nevertheless, the first set of results has important economic and policy implications regarding

the connection between adaptation strategies and estimating heat impacts on crop yields. For

instance, take a comparison between the benchmark and adjusted piecewise linear relationship

between temperature and yield (Figure 3.3). Our findings suggest that taking the role of produc-

tion participation into account as a climate-sensitive response can potentially lead to lower yield

responses to increased exposure to detrimentally high temperatures. Therefore, recent institutional

and research efforts that have been oriented towards improving seasonal forecasts for agricultural

producers respond to stated short-term demands from the sector and further contribute to its long-

term sustainability by allowing better adaptation strategies and fostering climate resilience (Klemm

& McPherson, 2017; Templeton, Shane Perkins, Aldridge, Bridges, & Lassiter, 2014).

3.5.3 Robustness checks

Our analysis rests on the hypothesis that farmers have access–to some degree–to precise infor-

mation about seasonal weather forecasts. However, our results so far exploit either previous-year

or contemporary weather variations. Thus, it could be argued that long-term trends in weather

patterns drive the qualitative similarity in estimated effects from lagged and contemporary weather

variables on participation, instead of access to forecasting in elaborating expectations. We pro-

vide correlational evidence in favor of our hypothesis in Table 3A.1, where we fit a participation

model based on contemporary and lagged weather variables, along with variations in the chosen

definition for price expectations. Under this specification, if it were the case that long-term trends

lead the results and forecasting played no role, we would expect that only lagged weather variables

have a statistically significant effect. Nevertheless, we find qualitatively equivalent and statistically

significant average partial effects for most past and contemporary weather variation. If anything,

159

the results would point to farmers potentially combining both sets of information to make decisions

regarding participation.17

In addition, we have argued that increasing access to high-accuracy forecasting is the most likely

driver of participation as an adaptation strategy, which could lead to potential bias in traditional

FE estimations. To approximate this, we consider a variation of our analysis that sets weather

expectations as a weighted average of past-season and contemporary weather variations. Let

𝐸𝑡 [𝑊] be the expectation of a weather variable 𝑊 for season 𝑡, such that

𝐸𝑡 [𝑊] = 𝜃𝑊𝑡 + (1 − 𝜃)𝑊𝑡−1,

where 𝜃 ∈ [0, 1] is a weight factor that reflects the precision of expectations, since whenever

𝜃 → 1 or 𝜃 → 0 we get our limiting scenarios. We reassess our analysis of participation in

scenarios that reflect increasing values of 𝜃 set at 0.25, 0.5, and 0.75 and report the results in Table

3A.2. The coefficients are qualitatively equivalent across all weight scenarios, in line with our

previous analyses, which suggest that we would mostly expect an impact from higher accuracy in

the intensive margin.

On the other hand, considering the role of participation based on increasingly accurate weather

expectations tells a different story. Using the previous results as a first-step analysis, we calculate

corresponding inverse Mills ratios based on different values of 𝜃 and re-estimate our main intensive

margin specification. We summarize these findings in Table 3A.3. Note that as 𝜃 increases, the

estimated magnitude of 𝜌 also increases–i.e., the odds of bias in traditional FE estimates increase.

Therefore, it could be argued that as seasonal weather forecasting increases its precision, farmers’

expectations become increasingly and more strongly correlated with effective weather patterns.

Thus, the role of production participation (extensive margin) will matter to determine the impacts

of weather on yields (intensive margin).

Additionally, we address three additional concerns when studying the case of production par-

ticipation. Namely, we evaluate whether climate responsiveness is sensitive to (a) considerations

of low yields in the immediately previous season, (b) farm-level invariant factors, and (c) alternate

17Noticeably, these specifications lead to insignificant effects from price variables. However, given that our price
data is at the State level, having two years of weather variation takes a significant share of year-to-year price variability.

160

definitions of production participation due to rotational considerations. We summarize these in

Table 3A.4. First, in columns 1-2, we add a binary covariate that takes the value of one if the

farm experienced production in the lowest decile of yields in the previous year.18 Our findings

suggest that farmers experiencing sorghum yields in the lowest decile in a year become less likely to

produce in the following year. Specifically, experiencing yield in the lowest decile reduces the odds

of producing the next year by roughly seven percentage points. Meanwhile, the remaining partial

effects align with our main inference. Therefore, farmers’ behavior regarding sorghum productive

decisions are seemingly responding to (i) weather expectations, (ii) expected prices, and (iii) their

assessment of their production efficiency.

Second, we consider the possibility of our random-effects first-stage results being driven by

unobserved heterogeneities. Therefore, we follow Wooldridge (1995) and implement a Mundlak

approach,19 by adding the farm-level averages of weather variables20 as control variables (see

Columns 3-4 in Table 3A.4. We find that while the magnitude of the average partial effects is

somewhat reduced, their statistical significance holds–and so does most of our inference. Interest-

ingly, we find that the between-effect from the Mundlak points to a higher probability of producing

sorghum for farms with larger averages in high temperatures. This likely suggests that while weather

foresight of high temperatures (short term) can hinder the decision to produce sorghum in a given

year, it also happens that farmers in warmer areas are more likely to remain in production.

Finally, we consider the potential sensitivity of estimates to assumptions regarding crop rotation.

In particular, we are interested in a case where some of the zeroes in the sorghum production decision

index may be misleading, as it is possible that sorghum remains in the farmer’s production plan but

18This addition implies a tradeoff in terms of the number of observations, as this limits the observations to cases in
which there exists production in the immediately previous period. Nevertheless, this specification allows us to consider
shorter-term dynamic production decisions based on farmers’ experienced outcomes.

19In brief, this approach assumes that individual-level, unobserved heterogeneities are a function of (a) the time
averages of time-varying covariates of interest, and (b) purely random individual-level variation. The former can be
directly captured as an added variable, leading to the standard Correlated Random Effects estimator.

20Following our framework, we consider the average of weather observed (or approximated) at each farm for each
𝑇𝑖, ¯𝑇𝑖]
¯

𝑡 ∈ [

161

is not produced in a given year due to a rotation system. We take a simple approach where

𝑝𝑖𝑡 =

1 if

(𝑖 produces sorgum in 𝑡) or (𝑖 produced sorgum in 𝑡 − 1)

0 if

𝑖 does not produce sorghum in 𝑡 or 𝑡 − 1,





which assumes a one-year rotation.21 Except for price variables, which result in statistically

insignificant estimates, we find qualitatively equivalent results to those of our main analysis. Thus,

even after accounting for potential strategies of crop rotation, the decision to produce sorghum is

climate responsive, which can potentially affect estimates of weather impacts on yields.

3.6 Conclusion

Using longitudinal farm-level data on yields and weather patterns, along with price time series,

we assess the potential role of farmers’ expectations in shaping productive participation and its

effect on yield-related weather impacts. With open-access data of sorghum production in Kansas

between 1978 and 2015, we use binary regression and Heckman-style linear regression adjust-

ments, respectively, to (1) evaluate how high-accuracy expectations affect the likelihood of farmers

producing sorghum in a given year, and (2) assess whether this behavior warrant a revision of heat

impacts on yields.

Our analysis reveals that a farmer’s decision to participate in sorghum production in a given

year is likely price- and climate-responsive. Moreover, we provide evidence that participation is

correlated with past and upcoming season weather variables, suggesting that farmers have some

degree of (or access to) high-accuracy forecasting and adapt their behavior accordingly. On the

other hand, corrections for potential bias from sample selection (emerging from participation

decisions) in yield equations, suggest that only when weather expectations are sufficiently close

(or correlated) to realized weather, traditional fixed-effect estimates are likely inconsistent.

In

particular, accounting for participation as a strategic adaptation choice leads to a lower magnitude

in impact estimates from crop exposure to high heat levels. As seasonal forecasting becomes

increasingly more precise, our findings signal the possibility of farmers properly using information

to adjust their production behavior, likely improving their resilience to climate change.

21While wheat-sorghum-fallow rotations are common in the region, and would point to longer lags to account for,

we do not find cases in our data pointing to potential rotations of more than one year.

162

Although our results provide relevant insights about the effective use of increasingly granular

data to estimate weather impacts on yields, we acknowledge two key limitations. First, future

studies could explore the role of expectations in the productive allocation of different crops at the

farm level. Our analysis focused on the extensive and intensive margins of sorghum production;

yet, it is still possible that farms are producing other in-farm substitute crops like corn, which we

cannot observe in the available data. Exploring how crop allocation shares change due to weather

and price expectations would provide a more comprehensive assessment of the potential impacts

of adaptation strategies on farm-level productive efficiency.

Second, future work could explore the role of farms’ and farmers’ attributes and financial

conditions as drivers of farm observability. Studies that include information about farmers’ age,

schooling, associativity, family background, and other farmer-level covariates could provide more

insights about the prioritization of specific crops in production, and of differentiated mechanisms of

adaptation based on changing rates of human and social capital. Likewise, detailed information on

the farm’s cost structures and financial conditions would allow for isolating the effects of the total

exit of agricultural production from those stemming from adaptation. Finally, any details about

the presence of contract farming would also allow us to draw important comparisons between a

scenario of flexibility towards adaptation vis-à-vis pre-existing commitments.

163

Figures and Tables

Figure 3.1 Number of farms and sorghum season average temperatures in Kansas, 1978-2015.

Notes — Figure displays the number of sorghum-producing farms registered in the KMFA (left axis) and the average temperature
in Kansas during the June-September period (right axis) for the reference year. Farm information is available from open-access
data provided by Miller et al. (2021), while Kansas-level temperature data comes from the National Centers for Environmental
Information. The dot-dashed line reports the linear fit of the average temperature as a function of the year variable for the entire
period.

164

Figure 3.2 Number of farms and farmers’ received prices for corn and sorghum, 1978-2015.

Notes — Figure displays the number of sorghum-producing farms registered in the KMFA (left axis) and farmers’ received prices
for sorghum and corn (right axis) for the reference year. Farm information is available from open-access data provided by Miller
et al. (2021), while Kansas-level temperature data comes from USDA NASS. Prices are adjusted by CPI from the U.S. Bureau of
Labor Statistics.

165

Figure 3.3 Comparison of marginal effects of degree-days exposure on sorghum yields, with and
without Heckman-style sample selection correction.

Notes — Figure displays the marginal effect of exposure to a particular 1°C temperature level on sorghum yields. The red line
reflects the baseline estimate from Miller et al. (2021), while the solid blue line reflects the alternate results upon a Heckman-style
correction. The dashed lines denote the 95% confidence intervals with standard errors clustered at the agricultural district-year
level.

166

Table 3.1 Descriptive statistics – Adapted from Miller et al. (2021).

Variable
Number of farms by year
Number of years by farm
Sorghum yield
Cumulative precipitation
Average daily min temp
Average daily max temp
DLow
DMed
DHigh

Mean
1,348.32
16.13
63.18
382.30
16.08
29.87
1,280.07
1,690.30
25.26

Std Dev
330.54
9.44
28.50
155.17
2.11
2.14
125.03
149.22
21.27

Min
449
1
0
68.07
7.36
20.27
989.97
1,134.01
0

Max
1,676
38
174.55
988.50
20.44
36.04
1,844.27
2,294.05
140.59

Notes — Total number of observations is 45,971 across 5,541 farms spanning 38 years (1978-2015). DLow is defined as degree
days above 0ºC minus degree days above 10ºC, DMed is defined as degree days above 10ºC minus degree days above 33ºC, and
DHigh is defined as degree days above 33ºC.

167

Table 3.2 Coefficient estimates of sorghum production participation model.

Variables

Trend

Precipitation

DLow

DMed

DHigh

ˆ𝑃sorghum

ˆ𝑃corn

Lag of 𝑃sorghum

Lag of 𝑃corn

Constant

(1)
Probit

(2)
APE

(3)
Probit

(4)
APE

-0.0102***
(0.000969)
0.000332***
(5.50e-05)
-0.000314***
(6.80e-05)
0.000437***
(5.79e-05)
-0.00218***
(0.000432)
0.0788**
(0.0343)
-0.0807**
(0.0324)

-0.00336***
(0.000318)
0.000109***
(1.81e-05)
-0.000104***
(2.24e-05)
0.000144***
(1.90e-05)
-0.000716***
(0.000142)
0.0259**
(0.0113)
-0.0266**
(0.0107)

0.248***
(0.0830)

-0.0112***
(0.00104)
0.000341***
(5.52e-05)
-0.000305***
(6.81e-05)
0.000430***
(5.81e-05)
-0.00211***
(0.000432)

-0.00369***
(0.000340)
0.000112***
(1.81e-05)
-0.000100***
(2.24e-05)
0.000142***
(1.91e-05)
-0.000695***
(0.000142)

0.0111***
(0.00397)
-0.0134***
(0.00402)

0.0337***
(0.0121)
-0.0407***
(0.0122)
0.293***
(0.0848)

Observations

66,959

66,959

66,959

66,959

Notes — Response variable is a binary index of whether the farm produces sorghum in the reference year. Reporting coefficients and
average partial effects (APE) from Random Effects Probit estimation. Robust standard errors in parentheses. *** p<0.01, ** p<0.05,
* p<0.1. Variables ˆ𝑃sorghum and ˆ𝑃corn are dynamic predictions of farmers’ received price from a 2-lag vector autoregression.

168

Table 3.3 Coefficient estimates of sorghum production participation model: Alternate results
using lags of weather variables.

Variables

Trend

Lag of Precipitation

Lag of DLow

Lag of DMed

Lag of DHigh

ˆ𝑃sorghum

ˆ𝑃corn

Lag of 𝑃sorghum

Lag of 𝑃corn

Constant

(1)
Probit

(2)
APE

(3)
Probit

(4)
APE

-0.00657***
(0.000993)
0.000361***
(5.63e-05)
-0.000855***
(6.95e-05)
0.000462***
(5.89e-05)
-0.000791*
(0.000476)
0.0141
(0.0337)
-0.0194
(0.0323)

-0.00216***
(0.000326)
0.000119***
(1.85e-05)
-0.000282***
(2.28e-05)
0.000152***
(1.94e-05)
-0.000261*
(0.000157)
0.00465
(0.0111)
-0.00639
(0.0106)

0.786***
(0.0893)

-0.00682***
(0.00105)
0.000353***
(5.67e-05)
-0.000836***
(7.02e-05)
0.000457***
(5.90e-05)
-0.000702
(0.000467)

-0.00225***
(0.000345)
0.000116***
(1.87e-05)
-0.000275***
(2.30e-05)
0.000150***
(1.94e-05)
-0.000231
(0.000154)

-0.000450
(0.00399)
-0.00194
(0.00403)

-0.00137
(0.0121)
-0.00588
(0.0123)
0.786***
(0.0896)

Observations

64,934

64,934

64,934

64,934

Notes — Response variable is a binary index of whether the farm produces sorghum in the reference year. Reporting coefficients and
average partial effects (APE) from Random Effects Probit estimation. Robust standard errors in parentheses. *** p<0.01, ** p<0.05,
* p<0.1. Variables ˆ𝑃sorghum and ˆ𝑃corn are dynamic predictions of farmers’ received price from a 2-lag vector autoregression.

169

Table 3.4 Coefficient estimates of sorghum yield model: Baseline and Heckman corrections.

Variables

Trend

Precipitation

Precipitation2

DLow

DMed

DHigh

ˆ𝜆𝑖𝑡

Constant

Weather Expectation
Price Expectation
Observations
R-squared

(1)
Log-Yield

(2)
Log-Yield

(3)
Log-Yield

0.00201
(0.00153)
0.00340***
(0.000461)
-3.12e-06***
(4.68e-07)
-0.000447**
(0.000180)
0.000588***
(0.000140)
-0.0124***
(0.00127)

3.100***
(0.210)

N/A
N/A
45,971
0.591

0.0137***
(0.00459)
0.00288***
(0.000507)
-3.09e-06***
(4.60e-07)
9.46e-05
(0.000270)
-8.49e-05
(0.000272)
-0.00903***
(0.00171)
-2.742***
(1.045)
4.950***
(0.712)

0.00572
(0.00385)
0.00324***
(0.000499)
-3.11e-06***
(4.68e-07)
-0.000278
(0.000252)
0.000377
(0.000241)
-0.0113***
(0.00157)
-0.867
(0.847)
3.685***
(0.590)

Contemporary Contemporary

Dynamic
45,971
0.593

Price Lag
45,971
0.591

Notes — Response variable is the log-yield of sorghum in the reference year. Reporting Fixed-Effect (farm-level) estimates, with
standard errors clustered at the agricultural district-year level in parentheses. *** p<0.01, ** p<0.05, * p<0.1. Variable ˆ𝜆𝑖𝑡 reflects
the inverse Mills ratio derived from the RE Probit estimates of a participation model, including contemporary values of weather
variables and either dynamic prediction of farmers’ received prices for sorghum and corn or lags of observed prices.

170

Table 3.5 Coefficient estimates of sorghum yield model: Baseline and Alternate Heckman
corrections.

Variables

Trend

Precipitation

Precipitation2

DLow

DMed

DHigh

ˆ𝜆𝑖𝑡

Constant

Weather Expectation
Price Expectation
Observations
R-squared

(1)
Log-Yield

(2)
Log-Yield

(3)
Log-Yield

0.00201
(0.00153)
0.00340***
(0.000461)
-3.12e-06***
(4.68e-07)
-0.000447**
(0.000180)
0.000588***
(0.000140)
-0.0124***
(0.00127)

3.100***
(0.210)

N/A
N/A
45,971
0.591

0.00441**
(0.00200)
0.00335***
(0.000471)
-3.04e-06***
(4.64e-07)
-0.000404**
(0.000180)
0.000562***
(0.000137)
-0.0121***
(0.00127)
-0.744*
(0.400)
3.452***
(0.238)

0.00415**
(0.00201)
0.00336***
(0.000473)
-3.05e-06***
(4.67e-07)
-0.000402**
(0.000182)
0.000560***
(0.000137)
-0.0121***
(0.00127)
-0.661*
(0.397)
3.413***
(0.239)

Lagged Weather Lagged Weather

Dynamic
44,400
0.587

Price Lag
44,400
0.587

Notes — Response variable is the log-yield of sorghum in the reference year. Reporting Fixed-Effect (farm-level) estimates, with
standard errors clustered at the agricultural district-year level in parentheses. *** p<0.01, ** p<0.05, * p<0.1. Variable ˆ𝜆𝑖𝑡 reflects
the inverse Mills ratio derived from the RE Probit estimates of a participation model, including lagged values of weather variables
and either dynamic prediction of farmers’ received prices for sorghum and corn or lags of observed prices.

171

REFERENCES

Akpoti, K., Groen, T., Dossou-Yovo, E., Kabo-bah, A. T., & Zwart, S. J. (2022). Climate change-
induced reduction in agricultural land suitability of West-Africa’s inland valley landscapes.
Agricultural Systems, 200, 103429. doi: 10.1016/j.agsy.2022.103429

Assefa, Y., Roozeboom, K., Thompson, C., Schlegel, A., Stone, L., & Lingenfelser, J. E. (2014).
Corn and Grain Sorghum Comparison: All Things Considered. Amsterdam, Netherlands:
Elsevier.

Baltagi, B. H., Jiménez-Martín, S., Labeaga, J. M., & Al Sadoon, M. (2023). Consistent estimation
of panel data sample selection models. Econometrics and Statistics, In press. doi: 10.1016/
j.ecosta.2023.11.003

Challinor, A. J., Watson, J., Lobell, D. B., Howden, S. M., Smith, D. R., & Chhetri, N. (2014). A
meta-analysis of crop yield under climate change and adaptation. Nature Climate Change,
4(4), 287–291. doi: 10.1038/nclimate2153

Costinot, A., Donaldson, D., & Smith, C. (2016). Evolving comparative advantage and the impact
of climate change in agricultural markets: Evidence from 1. 7 million fields around the world.
Journal of Political Economy, 124(1), 205–248. doi: 10.1086/684719

Duff, J., Bice, D., Hoeffner, I., & Weinheimer, J. (2019). The sorghum industry and its market
perspective. In I. Ciampitti & P. V. Prasad (Eds.), Sorghum (p. 503-514). John Wiley &
Sons, Ltd. doi: https://doi.org/10.2134/agronmonogr58.c23

Feng, L., Wang, H., Ma, X., Peng, H., & Shan, J. (2021). Modeling the current land suitability and
future dynamics of global soybean cultivation under climate change scenarios. Field Crops
Research, 263, 108069. doi: 10.1016/j.fcr.2021.108069

Gammans, M., Mérel, P., & Ortiz-Bobea, A. (2024). Double cropping as an adaptation to climate
change in the United States. American Journal of Agricultural Economics, Early view. doi:
10.1111/ajae.12491

Griffin, T. W., Miller, N. J., Bergtold, J., Shanoyan, A., Sharda, A., & Ciampitti, I. A. (2017).
Farm’s sequence of adoption of information-intensive precision agricultural technology.
Applied Engineering in Agriculture, 33(4), 521–527. doi: 10.13031/aea.12228

Gronau, R. (1974). Wage comparisons–a selectivity bias. Journal of Political Economy, 82(6),

1119–1143.

Hariprasanna, K., & Rakshit, S.

In S. Rakshit
& Y.-H. Wang (Eds.), The Sorghum Genome (pp. 1–25). Cham, Switzerland: Springer
International Publishing. doi: 10.1007/978-3-319-47789-3_1

(2016). Economic importance of sorghum.

172

Heckman, J. J.
153–161.

(1979). Sample selection bias as a specification error. Econometrica, 47(1),

Klemm, T., & McPherson, R. A. (2017). The development of seasonal climate forecasting for
agricultural producers. Agricultural and Forest Meteorology, 232, 384–399. doi: 10.1016/
j.agrformet.2016.09.005

Kuethe, T., Briggeman, B., Paulson, N., & Katchova, A. (2014). A comparison of data collected
through farm management associations and the Agricultural Resource Management Survey.
Agricultural Finance Review, 74(4), 492–500. doi: 10.1108/AFR-09-2014-0023

Li, S., Fleisher, D. H., & Barnaby, J. Y. (2024). Quantifying the impact of climate change and
extreme heat on rice in the United States. Agricultural and Forest Meteorology, 355, 110145.
doi: 10.1016/j.agrformet.2024.110145

Lobell, D. B., & Asseng, S. (2017). Comparing estimates of climate change impacts from process-
based and statistical crop models. Environmental Research Letters, 12(1), 015001. doi:
10.1088/1748-9326/aa518a

Maredia, M. K., & Martinez, J. M. (2024, November). Brewing Sustainability: The Need for
(Michigan State

Increased Agricultural R&D Investments in the Global Coffee Sector.
University. Department of Agricultural, Food, and Resource Economics. Working Paper)

Miao, R., Khanna, M., & Huang, H.

(2016). Responsiveness of crop yield and acreage to
prices and climate. American Journal of Agricultural Economics, 98(1), 191–211. doi:
10.1093/ajae/aav025

Miller, N., Tack, J., & Bergtold, J. (2021). The impacts of warming temperatures on us sorghum
yields and the potential for adaptation. American Journal of Agricultural Economics, 103(5),
1742–1758. doi: 10.1111/ajae.12223

Mérel, P., & Gammans, M. (2021). Climate econometrics: can the panel approach account for
long-run adaptation? American Journal of Agricultural Economics, 103(4), 1207–1238. doi:
10.1111/ajae.12200

Nerlove, M. (1956). Estimates of the elasticities of supply of selected agricultural commodities.
American Journal of Agricultural Economics, 38(2), 496-509. doi: 10.2307/1234389

Nicholson, W., & Snyder, C. (2012). Microeconomic Theory: Basic Principles and Extensions.

South-Western/Cengage Learning.

Ortiz-Bobea, A., Chambers, R. G., He, Y., & Lobell, D. B. (2025). Large increases in public R&D
investment are needed to avoid declines of US agricultural productivity. Proceedings of the
National Academy of Sciences, 122(11), e2411010122. doi: 10.1073/pnas.2411010122

173

Polansek, T. (2025, March). Trump trade war dries up sorghum sales to China but US farmers

plan to plant more.

Ramirez-Cabral, N. Y. Z., Kumar, L., & Shabani, F. (2017). Global alterations in areas of suitability
for maize production from climate change and using a mechanistic species distribution model
(Climex). Scientific Reports, 7(1), 5910. doi: 10.1038/s41598-017-05804-0

Schlenker, W., & Roberts, M. J. (2009). Nonlinear temperature effects indicate severe damages to
U.S. crop yields under climate change. Proceedings of the National Academy of Sciences,
106(37), 15594–15598. doi: 10.1073/pnas.0906865106

Stabel, J. S., Griffin, T., & Ibendahl, G. (2017). Do profitable farms remain profitable? Transition
probabilities using markov switching models applied to kansas farm data. Journal of Applied
Farm Economics, 2(1). doi: 10.7771/2331-9151.1009

Tack, J., Barkley, A., & Nalley, L. L. (2015). Effect of warming temperatures on US wheat yields.
Proceedings of the National Academy of Sciences, 112(22), 6931–6936. doi: 10.1073/
pnas.1415181112

Tack, J., Lingenfelser, J., & Jagadish, S. V. K.

(2017). Disaggregating sorghum yield re-
ductions under warming scenarios exposes narrow genetic diversity in US breeding pro-
grams. Proceedings of the National Academy of Sciences, 114(35), 9296–9301. doi:
10.1073/pnas.1706383114

Templeton, S. R., Shane Perkins, M., Aldridge, H. D., Bridges, W. C., & Lassiter, B. R. (2014).
Usefulness and uses of climate forecasts for agricultural extension in South Carolina, USA.
Regional Environmental Change, 14(2), 645–655. doi: 10.1007/s10113-013-0522-7

Wang, Y.-H., Upadhyaya, H. D., & Kole, C. (2014). Genetics, genomics and breeding of sorghum.

Boca Raton, FL: Taylor and Francis.

Wooldridge, J. M. (1995). Selection corrections for panel data models under conditional mean
independence assumptions. Journal of Econometrics, 68(1), 115-132. doi: 10.1016/0304
-4076(94)01645-G

Wooldridge, J. M.

(2010). Econometric Analysis of Cross Section and Panel Data (2nd ed.).

Cambridge, Mass: MIT Press.

World Meteorological Organization. (2023). The Global Climate 2011-2020: A decade of accel-

erating climate change. Geneva, Switzerland: United Nations.

Yue, Y., Zhang, P., & Shang, Y. (2019). The potential global distribution and dynamics of wheat
under multiple climate change scenarios. Science of The Total Environment, 688, 1308–1318.
doi: 10.1016/j.scitotenv.2019.06.153

174

Zhao, J., Wang, C., Shi, X., Bo, X., Li, S., Shang, M., . . . Chu, Q. (2021). Modeling climatically
suitable areas for soybean and their shifts across China. Agricultural Systems, 192, 103205.
doi: 10.1016/j.agsy.2021.103205

175

APPENDIX

Figure 3A.1 Number of farms and sorghum season average temperatures in Kansas with varying
linear fits, 1978-2015.

Notes — Figure displays the number of sorghum-producing farms registered in the KMFA (left axis) and the average temperature
in Kansas during the June-September period (right axis) for the reference year. Farm information is available from open-access
data provided by Miller et al. (2021), while Kansas-level temperature data comes from the National Centers for Environmental
Information. The dot-dashed line reports (a) the linear fit of the average temperature as a function of the year for the 1978-1990
period, and (b) a linear fit for the 1990-2015 period.

176

Figure 3A.2 Total sorghum area (harvested) in Kansas, 1997-2022.

Notes — Figure displays Kansas’s total sorghum harvested acreage (in thousands) between 1997 and 2022. The vertical red
line marks the final year of available data from KMFA, while the dashed gray line describes the linear fit of acreage and years.
Information is available from USDA-NASS Census Aggregates.

177

Figure 3A.3 Total sorghum operations (with area harvested) in Kansas, 1997-2022.

Notes — Figure displays Kansas’s total sorghum operations with harvested acreage between 1997 and 2022. The vertical red line
marks the final year of available data from KMFA, while the dashed gray line describes the linear fit of the number of operations
and years. Information is available from USDA-NASS Census Aggregates.

178

Figure 3A.4 Reference temperatures during sorghum season in Kansas and farmers’ observed
temperatures, 1978-2015.

Notes — Figure displays the average maximum, overall average, and average minimum temperatures for Kansas during the June-
September period, and the sorghum farmers’ observed values of these variables for the reference year. Kansas-level temperature
data comes from the National Centers for Environmental Information, while farm-level information is available from open-access
data provided by Miller et al. (2021).

179

Table 3A.1 Coefficient estimates of sorghum production participation model: Inclusion of both
contemporary and lagged values of weather variables.

Variables

Trend

Precipitation

DLow

DMed

DHigh

Lag of Precipitation

Lag of DLow

Lag of DMed

Lag of DHigh

ˆ𝑃sorghum

ˆ𝑃corn

Lag of 𝑃sorghum

Lag of 𝑃corn

Constant

(1)
Probit

(2)
APE

(3)
Probit

(4)
APE

-0.00809***
(0.00106)
0.000273***
(5.58e-05)
-4.28e-05
(7.29e-05)
0.000257***
(6.36e-05)
-0.00156***
(0.000465)
0.000223***
(6.13e-05)
-0.000811***
(7.28e-05)
0.000447***
(6.03e-05)
-0.00121**
(0.000490)
0.0463
(0.0347)
-0.0497
(0.0330)

-0.00266***
(0.000346)
8.95e-05***
(1.83e-05)
-1.41e-05
(2.39e-05)
8.45e-05***
(2.09e-05)
-0.000513***
(0.000153)
7.33e-05***
(2.01e-05)
-0.000266***
(2.38e-05)
0.000147***
(1.98e-05)
-0.000398**
(0.000161)
0.0152
(0.0114)
-0.0163
(0.0108)

0.397***
(0.116)

-0.00860***
(0.00111)
0.000274***
(5.60e-05)
-3.33e-05
(7.30e-05)
0.000249***
(6.37e-05)
-0.00150***
(0.000464)
0.000226***
(6.15e-05)
-0.000798***
(7.33e-05)
0.000443***
(6.04e-05)
-0.00117**
(0.000479)

-0.00282***
(0.000363)
9.01e-05***
(1.84e-05)
-1.09e-05
(2.40e-05)
8.16e-05***
(2.09e-05)
-0.000493***
(0.000152)
7.41e-05***
(2.02e-05)
-0.000262***
(2.40e-05)
0.000145***
(1.98e-05)
-0.000384**
(0.000157)

0.00476
(0.00406)
-0.00694*
(0.00409)

0.0145
(0.0124)
-0.0211*
(0.0125)
0.414***
(0.118)

Observations

64,934

64,934

64,934

64,934

Notes — Response variable is a binary index of whether the farm produces sorghum in the reference year. Reporting coefficients and
average partial effects (APE) from Random Effects Probit estimation. Robust standard errors in parentheses. *** p<0.01, ** p<0.05,
* p<0.1. Variables ˆ𝑃sorghum and ˆ𝑃corn are dynamic predictions of farmers’ received price from a 2-lag vector autoregression.

180

Table 3A.2 Coefficient estimates of sorghum production participation model: Weighted average of contemporary and lagged weather
variables.

Variables

Trend

Weighted Precipitation

Weighted DLow

Weighted DMed

Weighted DHigh

ˆ𝑃sorghum

ˆ𝑃corn

Constant

(1)
Probit

(2)
APE

(3)
Probit

(4)
APE

(5)
Probit

(6)
APE

-0.00727***
(0.00103)
0.000426***
(4.16e-05)
-0.000889***
(7.70e-05)
0.000677***
(7.05e-05)
-0.00298***
(0.000573)
0.0403
(0.0340)
-0.0426
(0.0324)
0.495***
(0.103)

-0.00239***
(0.000337)
0.000140***
(1.36e-05)
-0.000292***
(2.52e-05)
0.000222***
(2.31e-05)
-0.000981***
(0.000188)
0.0132
(0.0112)
-0.0140
(0.0107)

-0.00710***
(0.00103)
0.000438***
(4.77e-05)
-0.000879***
(8.58e-05)
0.000750***
(7.63e-05)
-0.00345***
(0.000657)
0.0393
(0.0340)
-0.0398
(0.0323)
0.347***
(0.112)

-0.00233***
(0.000339)
0.000144***
(1.56e-05)
-0.000289***
(2.81e-05)
0.000246***
(2.50e-05)
-0.00113***
(0.000216)
0.0129
(0.0112)
-0.0131
(0.0106)

-0.00811***
(0.00101)
0.000375***
(5.45e-05)
-0.000572***
(8.19e-05)
0.000623***
(7.04e-05)
-0.00315***
(0.000574)
0.0478
(0.0343)
-0.0490
(0.0324)
0.211**
(0.103)

-0.00267***
(0.000331)
0.000123***
(1.79e-05)
-0.000188***
(2.69e-05)
0.000205***
(2.31e-05)
-0.00104***
(0.000189)
0.0157
(0.0113)
-0.0161
(0.0107)

Weight Factor (𝜃)
Observations

0.25

0.5

0.75

64,934

64,934

64,934

64,934

64,934

64,934

Notes — Response variable is a binary index of whether the farm produces sorghum in the reference year. Reporting coefficients and
average partial effects (APE) from Random Effects Probit estimation. Robust standard errors in parentheses. *** p<0.01, ** p<0.05,
* p<0.1. Variables ˆ𝑃sorghum and ˆ𝑃corn are dynamic predictions of farmers’ received price from a 2-lag vector autoregression.

181

Table 3A.3 Coefficient estimates of sorghum yield model: Baseline and Heckman corrections
using weighted average of weather variables.

Variables

Trend

Precipitation

Precipitation2

DLow

DMed

DHigh

ˆ𝜆𝑖𝑡

Constant

(1)
Log-Yield

(2)
Log-Yield

(3)
Log-Yield

0.00423**
(0.00208)
0.00321***
(0.000498)
-3.03e-06***
(4.68e-07)
-0.000370*
(0.000196)
0.000547***
(0.000140)
-0.0122***
(0.00127)
-0.611
(0.455)
3.421***
(0.273)

0.00508**
(0.00252)
0.00314***
(0.000521)
-3.03e-06***
(4.70e-07)
-0.000239
(0.000261)
0.000428**
(0.000174)
-0.0116***
(0.00132)
-0.871
(0.656)
3.594***
(0.375)

0.0105**
(0.00421)
0.00286***
(0.000561)
-3.01e-06***
(4.68e-07)
0.000125
(0.000367)
7.47e-06
(0.000306)
-0.00945***
(0.00182)
-2.324**
(1.174)
4.595***
(0.730)

Weight Factor in Participation (𝜃)
Observations
R-squared

0.25
44,400
0.586

0.5
44,400
0.586

0.75
44,400
0.587

Notes — Response variable is the log-yield of sorghum in the reference year. Reporting Fixed-Effect (farm-level) estimates, with
standard errors clustered at the agricultural district-year level in parentheses. *** p<0.01, ** p<0.05, * p<0.1. Variable ˆ𝜆𝑖𝑡 reflects
the inverse Mills ratio derived from the RE Probit estimates of a participation model, including a weighted average of contemporary
and lagged values of weather variables and dynamic predictions of farmers’ received prices for sorghum and corn.

182

Table 3A.4 Coefficient estimates of sorghum production participation model: Response to low-yields, consideration of correlated
random effects, and potential case of productive rotation.

Variables

Trend

Precipitation

DLow

DMed

DHigh

ˆ𝑃sorghum

ˆ𝑃corn

Low yield in 𝑡 − 1

Prec

DLow

DMed

DHigh

Constant

(1)
Probit

(2)
APE

(3)
Probit

(4)
APE

(5)
Probit

(6)
APE

-0.0216***
(0.00133)
0.000466***
(7.22e-05)
-0.000480***
(8.83e-05)
0.000459***
(7.56e-05)
-0.00150***
(0.000563)
0.0902**
(0.0450)
-0.106**
(0.0426)
-0.236***
(0.0243)

-0.00649***
(0.000392)
0.000140***
(2.16e-05)
-0.000144***
(2.64e-05)
0.000138***
(2.26e-05)
-0.000451***
(0.000169)
0.0270**
(0.0135)
-0.0318**
(0.0128)
-0.0707***
(0.00726)

0.841***
(0.108)

-0.00363***
(0.000328)
3.24e-05
(1.98e-05)
8.57e-06
(2.60e-05)
6.10e-05***
(2.18e-05)
-0.000595***
(0.000149)
0.0297***
(0.0113)
-0.0310***
(0.0107)

0.000536***
(7.70e-05)
-0.000211***
(5.92e-05)
4.98e-05
(6.43e-05)
0.00296***
(0.000629)

-0.0111***
(0.00100)
9.88e-05
(6.02e-05)
2.61e-05
(7.92e-05)
0.000186***
(6.64e-05)
-0.00181***
(0.000454)
0.0904***
(0.0344)
-0.0944***
(0.0326)

0.00163***
(0.000235)
-0.000642***
(0.000181)
0.000152
(0.000196)
0.00901***
(0.00192)
0.0780
(0.284)

0.0184***
(0.00138)
0.000400***
(7.45e-05)
-0.000366***
(8.92e-05)
0.000394***
(7.47e-05)
-0.00207***
(0.000580)
-0.0221
(0.0540)
0.00784
(0.0508)

0.00212***
(0.000160)
4.59e-05***
(8.56e-06)
-4.20e-05***
(1.03e-05)
4.52e-05***
(8.59e-06)
-0.000238***
(6.67e-05)
-0.00254
(0.00620)
0.000900
(0.00583)

1.138***
(0.118)

Observations

45,466

45,466

66,959

66,959

66,959

66,959

Notes — Response variable is a binary index of whether the farm produces sorghum in the reference year. Reporting coefficients and
average partial effects (APE) from Random Effects Probit estimation. Robust standard errors in parentheses. *** p<0.01, ** p<0.05,
* p<0.1. Variables ˆ𝑃sorghum and ˆ𝑃corn are dynamic predictions of farmers’ received price from a 2-lag vector autoregression.

183