l LlBRARY

Michigan State
University

 

 

 

This is to certify that the
thesis entitled

The Effects of Land Quality and Natural Shocks
On Land and Labor Productivity
On Rice in Madagascar Highland

presented by

Pierre Jean Claude Randrianarisoa

has been accepted towards fulfillment
of the requirements for

M.S. Agricultural

degree in
Economics

 

 

 

 

Major professor

 

0-7639 MS U is an Affirmative Action/Equal Opportunity Institution

PLACE IN RETURN Box to remove this checkout from your record.
TO AVOID FINES return on or before date due.
MAY BE RECALLED with earlier due date if requested.

 

DATE DUE DATE DUE DATE DUE

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

6/01 cJCIHC/DatoDmsz-sz

THE EFFECTS OF LAND QUALITY AND NATURAL SHOCKS
ON LAND AND LABOR PRODUCTIVITY
ON RICE IN MADAGASCAR HIGHLAND
by

Pierre Jean Claude Randrianarisoa

A THESIS

Submitted to
Michigan State University
In partial fulfillment of the requirements

For the degree of

MASTER OF SCIENCE

Department of Agricultural Economics
2003

ABSTRACT

THE EFFECTS OF LAND QUALITY AND NATURAL SHOCKS
ON LAND AND LABOR PRODUCTIVITY
ON RICE IN MADAGASCAR HIGHLAND
By

Pierre Jean Claude Randrianarisoa

Despite economic reforms towards a more liberalized market, rice land and labor
productivity remain low in Madagascar Highland. The purpose of this paper is to address
the importance of the effects of land quality and natural shocks on land and labor
productivity, and to explore the implication for agricultural policy actions.

Using data from 563 rice plots, the study finds that: (a) land quality and natural
shocks do influence rice production in Madagascar and their omission in modeling efforts
leads to a bias of the marginal effects of land, labor, and other factors; (b) return to land
and labor varies across smallholder farms of different size. Smallholder farms of all Size
have low family labor productivity, and only medium and larger smallholder farms find
interesting profit in using hired labor; and (c) improvement in Rice Seedling
Transplantation seems to be a way to overcome bad land quality and to reduce production
vulnerability from exogenous natural shocks. However, its negative effect on the return to
labor seems to explain farmer’s reluctance to adopt this technique.

Major implications to draw from this study the necessity are to include land
quality and exogenous natural shocks in agricultural production analysis in developing
countries. Also, one should look at the way to improve the return to labor as land
productivity increases, for example, the use of mechanical small tools or the use of

animal traction or a better allocation of family labor in order to increase its return.

To Tojo, Rojo, and 01y,
To my beloved father,

To my mother,

“Ny hazo no vanon-ko lakana, ny tany naniriany no tsara

iii

ACKNOWLEDGEMENTS

I would like to thank the member of my guidance committee: Pr. Michael Weber,
Pr. Scott Swinton, Pr. Thomas Reardon, all three from the Agricultural Economics
Department at Michigan State University, Pr. Jack Meyer from the Economics
Department at Michigan State University, and Pr. Christopher Barrett from Cornell
University for their wise advice and support during the process of the thesis writing.

This thesis was not possible without the financial support provided by USAID
ATLAS, the Cornell ILO project, and FOFIF A. Special thanks to Mrs. Michelle Roberts
at the ATLAS program, Mary Norris, Francois Vezina, and the training office at the
USAID/Antananarivo, Francois Rasolo and his staff at FOFIFA. The success of this work
is inevitably associated with the involvement of the ILO project in Madagascar, with Dr.
Bart Minten and the team in Madagascar.

And lastly, I thank my family for their sacrifice, their love and support during

these last years.

TABLE OF CONTENTS

LIST OF TABLES
INTRODUCTION

1. RICE IN MADAGASCAR
1.1. Country Background
1.2. Prior Study on Rice Production In Madagascar

2. ANALYTICAL FRAMEWORK, ESTIMATION METHOD,
AND RESEARCH DESIGN
2.1. Analytical Framework
2.2. Estimation Method
2.2.1. Regression Estimation
2.2.2. Choice of Functional Form
2.3. Research Design
2.3.1 Description of Variables
2.3.2. Data Patterns

3. REGRESSION RESULTS AND INTERPRETATIONS
3.1. Effect of Land Quality and Natural Shocks
on Agricultural Production
3.2. Changes in Land, Labor, and Fertilizer Marginal Return Due to
the Omission of Land Quality and Natural Shocks
3.3. Effect of Farm Size on MVP
3.4. Incremental Land and Labor MVP of Productivity Interventions

CONCLUSIONS

APPENDIX 1 Madagascar - Mapping of Survey Sites
APPENDIX 2 Mackinnon, White, and Davidson Tests for
Non-Nested Models
APPENDIX 3 Correlation between Choice, Conditioners,
Land Quality — natural shocks variables
APPENDIX 4 Wald Tests Results
APPENDIX 5 Bootstrap Method and t-test Results Between two Means

REFERENCES

vi

GA-b

14
14
17
20
20
27

35
35
39
43
49

53

56
57
58
59
60

61

Table 1

Table 2

Table 3

Table 4

Table 5

LIST OF TABLES

Descriptive Statistics of the Variables Used

in The Production Function

Madagascar Highland Rice Production Function
Without and With Land Quality and Natural Shocks
Marginal Input Productivity With and Without

Land Quality and Natural Shocks, Estimation

At the Sample Means

Input MVP and Switching Value of Binary Variables
By Farm Size

Incremental Land and Labor MVP from Adoption

of Improved Rice Seedling Transplantation Technique

vi

28

36

4O

45

50

Introduction

During the last decades, the use of modern inputs and adoption of different new
technologies were thought to be available to boost agricultural production in Sub-Saharan
African countries. Experience from the Green Revolution in Eastern Asia was usually
taken as an example of the successful effects of this kind of intensification on agricultural
productivity. However, many African countries failed to attain the expected results. The
case of rice in Madagascar falls under this situation. Agricultural land and labor
productivity remain low, and adoption of new technologies is disappointing despite
economic reforms towards a more liberalized market and a reduction of institutional
constraints facing the agricultural sector (Minten et al, 2000; UPDR, 2000; USDA, 1999)

A better understanding of the return to inputs from rice production at the farm
level is crucial to address agricultural policy adjustments. One needs to identify what
factors of production offer the highest marginal returns to land and labor, on which
agricultural development policy might thus focus.

Usually, researchers use inputs and conditioner variables to assess the
determinants of agricultural production. Land quality and natural shocks are sometimes
overlooked, with the assumption that farmers are facing homogenous land characteristics
and the same production risks. Thus, farmer’s decisions are independent from these
exogenous factors (Kelly et al, 1996; Frisvold, 1994; Bemier and Dorosh, 1993;
Tshibaka, 1989; Hossain, 1988).

In reality, farmer’s choice variables are likely to be correlated with land quality,

hence a mutation of the parameter estimates of the variable inputs. Consequently the

effects of variable inputs on agricultural production partially depend on the
characteristics of land quality.

Moreover, the magnitude of natural shocks during the agricultural season affects
also output level (Sherlund et al, 2001; Reardon et al, 1996). It was shown for example
that annual change in rainfall, which caused drought or flood, leads to a different level of
production. This paper will then examine whether omitting land quality and natural
shocks in the estimation of the production function will lead to a bias in the importance of
the effects of common farmer’s choice variables on agricultural production.

It is also important to know to what extent there are different returns among
different farm size of smallholders. Can such results help to better understand the low
rate of adoption of chemical fertilizer or adoption of improved production technology on
rice production in Madagascar? Consequently, the objective of this paper is to give
decision-maker additional information in order to better identify policy interventions that
might foster greater uptake.

These kinds of research can be addressed by the analysis of the production
function relating the use of physical inputs with output. We will use rice data from
households in Madagascar highland for the analysis. Our analytical method is based on
econometric models that explain observed production, as a function of biophysical
characteristics of land units, exogenous natural shocks, farmer’s choice, and some
conditioner variables.

We have opted to use the primal approach1 for the analysis because of the

following reasons. First, there is the difficulty to assess the relationship between market

 

I The primal approach here involves a maximization of an objective function (production) subject to
physical constraints such as land, labor, capital, and other exogenous variables.

prices and smallholder’s production decision especially in the case of developing
countries. A second point is that with a cross-section data, prices have less variation than
the volume of inputs, and this would provide less variability in the sample. Lastly, as
farmers make decisions on input uses prior to realization of either output or output prices,
these latter variables are almost surely not the ones over which farmers made their input
decision.

For the organization of the report, we first present the rice situation in
Madagascar, the analytical framework, the estimation method, and data explanation. Next
we will demonstrate that the hypothesis of non-neutral impact of land quality and natural
shocks holds for rice production in Madagascar highland. By the analysis of marginal
products, we will show that omitting land quality and exogenous shocks in the analysis of
agricultural production can mislead agricultural policy decisions. Then we will look at
the differences in marginal value product by smallholder farm size, and draw some policy
implications. And last, we will identify the farmer’s decisions that result in highest land

and labor marginal return under different land quality and natural shocks conditions.

1. Rice in Madagascar

1.1. Country Background

Agriculture represents 30 percent of Madagascar’s Gross Domestic Production
(GDP) while almost 78 percent of the total active population gets a livelihood out of it.
Rice is the staple food and it occupies more than the half of total cultivated land. Its
contribution to agricultural GDP is around 43 percent (UPDR, 2000, World Bank, 1998).
Because rice is such an important staple food crop, it has long been central to any
development strategy. Over the years, every government has implemented agricultural
policies designed to encourage output increases.

The majority of rice production and thus marketed rice comes from smallholders.
Large farms of more than 50 hectares represents less than 0.05 percent of total farms
(SMTIS, 1989). Growing rice is part of the “cultural identity” of the rural community, as
92 percent of farmers are involved in rice cultivation. However, the average cultivated
rice area per household among smallholders is only 0.8 hectare with important inter- and
intra-regional differences. For example, in the North West and in the Lac Alaotra region,
the average rice area per household exceeds 150 ares while in the Vakinankaratra and
F ianarantsoa highland regionsz, it is only around 50 ares per household (SMTIS, 1989).
The same study reports also a high intra-regional variability, with a coefficient of
variation more than 100 percent on cultivated area of rice.

With only an average of 1.2 percent increase per year during the last 25 years, rice

production is experiencing very little improvement, and is lagging far behind population

 

2 The Vakinankaratra and Fianarantsoa highland regions are our study area.

4

growth3 (Minten et al, 2000; Roubaud, 1997; IF PRI, 1997). A liberalized economy and
privatized market and production systems have not yet caused substantial changes to this
trend. Findings from an IFPRI (1997) study showed that market reforms were still not
enough to cause increases in rice production, and it was estimated that land and labor
productivity in rice production remain low (UPDR, 2000; USDA, 1999).

Traditionally, rice is grown on narrow depressions ("bas-fonds") in lowland
locations or on fields in terraces on the slope of hills. Besides these traditional rice fields,
there are also some modern irrigated perimeters in flat areas. Natural exogenous shocks
are frequent as illustrated by the incidences of numerous cyclones every year. These
affect rice production seriously through the risk of long inundations".

Rice is generally transplanted in Madagascar highland where the present study is
focused. Direct seeding and "slash and burn" are rare. In the early 90’s, a new technology
“Systéme de Riziculture Intensive” (SRI) was promoted by extension services and various
NGOs. This is a very labor intensive technology, which requires an almost permanent
management of the rice field. It is based, among others, on (i) the use of young plants of
8 to 15 days for transplanting, (ii) several weeding cycles, i.e. from 3 to 4 times, (iii) a
square transplantation allowing the use of small tools for weeding; and (iv) intensive
water management practices by farmers. The SR1 technology does not require the use of
specific or improved rice varieties nor any fertilizer application. Expected yield are very
high, reaching 12 to 18 tons per hectares. Given farmer’s difficulties for meeting all the

required conditions for a complete SR] technology package, extension services started to

 

3 Madagascar demographic growth rate is 2.8 percent per year.
’ Inundation is defined as an entirely flooded plot during a few days that will affect negatively rice
production.

This is an extrapolated yield from small plots. Moreover, in most of the cases, the average area cultivated
with the SRI technology is below 10 ares.

recommend a less labor-intensive technology called “Syste‘me de Riziculture Améliorée”
(SRA). This consists of using one or two parts of SRI, such as rice seedling
transplantation along more traditional use of organic or mineral fertilizers. In this study,
we will use the "Improved Rice Seedling T ransplantation" (IRST) as the improved

technology. This is the innovation related to SR1 and SRA that farmers mostly retained.

1.2. Prior Studies on Rice Productivity in Madagascar

Using farm-level data, Bemier and Dorosh (1993) observed an average yield of
17.7 kg per are in the highland among smallholders". They found that chemical fertilizer
affects positively rice production in Madagascar with a marginal physical return of 6.2 kg
of paddy for one additional kg of fertilizer at 90 kg per hectare average rate of
application. However, the limitation of this study is that the region boundary does not
follow the agro-ecological zones in Madagascar, leading to a higher variability in the
sample, specifically for the highland region. Actually, the highland agro—ecological zone
is a mountainous area situated above 1,000 meters of altitude. Average farm size is small.
The Lac Alaotra area is part of the Middle East agro-ecological zone, situated in a lower
altitude of less than 800 meters, and characterized by the dominance of larger farms
established in some large irrigated perimeter. While the highland area is deficient in rice,
the Lac Alaotra is surplus and is the primary rice supplier for the country. Another
important drawback is the neglect of land quality and natural Shocks in the analysis that

might lead to a bias on the effect of input variables on productivity.

 

6 This yield corresponds to the category small farms in Plateau Central area in the Bemier and Dorosh
Study.

A second set of analyses, particularly focused on the impact of market reform on
agricultural productivity, showing that farmers who have the opportunity to increase their
landholding are likely to have less incentive to intensify rice production (Zeller et al,
2000). Their yield is lower than the average farm in the area. First, farmer might apply
less input per land tmit when cultivated area increases, thus resulting in lower
productivity. Second is that land quality may differ between existing plots and new
additional plots, leading to a productivity difference. Without considering land quality
and natural Shocks, one can attribute the change in productivity to only the change in
input use, which might be misleading. In their study in Brazil, Nerlove and Vosti (1996)
found similar results. Application of the "Green Revolution" technical packages in
marginal land resulted in lower increases associated with higher variances in the yield
compared to the same technology applied on high-potential land.

Using a recursive model7, Minten et al (2000) found that input expenditures affect
positively rice production. Irrigation, which is one variable used to control for land
quality, had a significant and positive effect on production, but its magnitude was
relatively low. Irrigation only contributed up to a 7 percent increase in yield.

Also, the effect of extension service and higher education were found to be
insignificant. The advantage of this study is the introduction of land quality and natural
shocks in the analysis, although they acknowledged that there were not enough variables
that would control for these factors. Land quality was represented by irrigation and land
prices while two indexes of climatic and disease risks were used to control for natural

shocks.

 

7 The recursive model consists in considering cultivated land, input expenditures, and yield as endogenous
in a system of regression equations.

However, the limitation of the Minten et al (2000) study is that the use of
aggregation in input expenditures does not permit identifying the individual effect of each
input on rice production. For example, one cannot observe the difference between the
return from the use of chemical fertilizer from hired labor use. Also for the econometric
estimation, the functional form used (Cobb-Douglas) precludes the analysis of the
interaction terms between pair of variables separately.

The current study tries to improve the estimation of the effect of inputs on rice
productivity by considering land quality and natural shocks variables, combined with the

use of flexible functional form.

2. Analytical Framework, Estimation Method, and Research Design

2.1. Analytical Framework

The production function represents the physical relationship between inputs and
output. Farmers are assumed to maximize production subject to constraints on land,
labor, capital availability, and technology.

The level of production y might be affected:

(1) by farmer’s decision -- choice variables -- in determining the quantity

of available labor, land, and capital allocated to each plot;

(2) by fixed and quasi-fixed factors at the farm or regional level that act as
conditioner variables;
(3) by land quality and natural shocks variables.

We have then the following model for our production function:

y=g(x,z,a;l3)+£ (1)
where y is rice output per plot in kg;

x is a vector of the farmer’s choice variables;

z is a vector of conditioning variables;

a is a vector representing land quality and natural shocks;

D are parameters and 8 is the stochastic disturbance term.

The z and a variables are exogenous factors and cannot be changed by farmers at
least in the medium term. Researchers usually account for the z factors in agricultural

productivity analysis. However, the a factors are sometimes overlooked either because of

missing data or because of insufficient observations precluding the estimation of large
number of parameters.

If land quality and exogenous Shocks do not affect agricultural production, the
entire coefficient estimates of the (1 variables will be equal to zero, so:

y = 8(x. 2’ a) =f(x. Z)- '

However, if the a variables affect the level of y, which means Cov(y, a) at 0, and
the x and the a variables are not correlated at all i.e Cov(x, a) = 0, we would deduce that
y = g(x, z, a) =f(x, z) + h(a) under assumption of strong separability. This is indeed too
strong an assumption to be true in the real world. Our hypothesis is that the a variables do
have significant effect on the production level, and are correlated to x, which means
Cov(y, a) ¢ 0 and Cov(x, a) at 0. Therefore, the coefficient estimates of x and the marginal
productivity of the input variables will be affected by the omission of the a variables in
the analysis. In order to test this hypothesis, we will use a model integrating proxy
variables that control for land quality and natural shocks variables.

To assess the potential biases introduced by omitting land quality and exogenous
shocks, we will estimate two models. The first one includes variables that control for land
quality and exogenous shocks, called "full specification model" and the second one, the
"short specification model" excludes this set of variables.

The differences will be estimated by the relative change between the MPPS from
the two models, and computed such that:

Mpg?” - MPPg’OH

%Change = * 100 (2a)
MP P gull

 

10

If the percent change is statistically equal to zero, there is no bias in using the
short model specification. Otherwise, the short specification model gives biased estimates
of the marginal values of different factors of production.

Suppose that the farmer is seeking to maximize his profit without constraints and
with the assumption of a competitive market where the farmer is a price taker in both
inputs and outputs.

If w is a vector of prices of inputs and p a vector of prices of output, y = f(x) is the
quantity of outputs, and if we assume that (l) farmer would maximize his profit, and (2)
for the sake of simplification, farmer produces only one output, then:

”at. w) = "mp/(x) - wx

The first order condition for the single output profit maximization is:

——af(x*)=w- i=12
8x,- I ,

which mean that the marginal value product (MVP) of each factor must be equal
to its marginal factor price (MVP = MFP) for a profit maximization. Marginal Value
Product is computed from the MPP and the average output price. It represents the
expected monetary return from the use of one more unit of a given input x,- .

The benefit of allocating more or less input, say x,- depends on the value of MVP

related to MFP, with the assumption that the production function is concave to the origin.

0 If MVP > MF P: There is an under-use of input xi. Farmers can still apply
more inputs and having higher MPP ofxi;

D If MVP = MFP: This is the point of profit maximum that farmer Should seek;

11

D If MVP < MFP: There is an over-use of input x,- resulting in a lower value of

the MVP ofx; compared to the price ofx,-.

However, changes in land and labor productivity can be thought of as a
combination of two phenomena:

1. The direct change due to the difference in the use of the inputs and;

2. The indirect effect of land quality and natural shocks.

To clarify the idea, we will present a couple of examples. We assume that the
changes in the input use are related to the adoption or not of a given new technology, e.g.
in the use of improved rice seedling transplantation.

If a farmer decides to adopt a given new technology on the same plot, it will affect
land and labor Marginal Physical Product (MPP)8, by the direct effect of the impact of the

technology use.

D If a farmer applies the same technology in two plots with different land
quality, land or labor MPP would differ from the difference in land
quality, with the hypothesis that land quality affects agricultural
production.

D If a farmer opts to adopt the new technology on good quality land and not
to use on bad land quality, the total MPP change between the two plots
would be the sum of the change from the land quality difference plus the

change from the new technology adoption (1) + (2).

 

8 MPP is the physical amount of production given by one additional unit of an input. The advantage of its
use is that it permits the analysis of the allocation efficiency of each input, thus allowing farmers to stop or
to continue using the given input.

12

Hence, if one wants to know the effect of the adoption of a new technology on the
productivity, it might be useful to look at the characteristics of land and eventual natural
shocks.

To see the effect of land quality and exogenous shocks on the use of factor x,-,

the Obj ective is then to look at the magnitude A:

(iltzmk =1,;j,§)_(EY_I,=o,ak =0,rj,2)
13‘ 3x; axi

 

3 (2b)
— — — . —

Where t is the technology to be analyzed, t is equal to 1 if farmer adopt the
technology and is equal to zero if there is no adoption;

xj is the specific farmer’s land or labor variable,

xj are other choice variables influencing productivity,

the al, are different categorical variables representing land quality and natural
shocks. The value 1 represents a good state of land and a value 0 a bad state, for example
ak=1 is for irrigated land while ak=0 is for rainfed;

and z other conditioner variables such as dummy villages, localization etc.

a If A > O: The adoption of technology t gives farmer greater profit in using the

factor of production x,-;

13

a If A = 0: Farmers will be indifferent in adopting or not t because it does not
have impact on the MPP of the factor of production x,-. The cost minimizing
theory should then influence farmer's decision;

0 If A < 0: The adoption of technology 1 does not pay off farmer's expenditures

on adopting technology t.

Therefore, besides the presentation of the marginal return for each input, we will
also Show the changes in land and labor marginal returns when both the crop growing

conditions change, for example from rainfed to irrigated land.

2.2. Estimation Method
2.2.1. Regression estimm

The analysis is conducted by using econometric estimation. However with data
from a cross-section survey, heteroskedasticity issues are likely to be present because of
the similarity between farmer’s behavior and input use per village (Deaton, 1997). Hence,
the choice of the econometric estimation method becomes crucial because Ordinary Least
Square (OLS) does not anymore give efficient estimators. With heteroskedasticity, the
variance a) is not constant over the explanatory variables, then obviously the magnitude

of which y will depend on 03,2 should not be weighted equally. Therefore, the points for

which 0}) is comparatively large should be down-weighted.

14

Using the test for heteroskedasticity suggested by Wooldridge (2000), we found
that our sample exhibits heteroskedastic propertiesg. We will then do the estimation by
allowing the conditional variance, as well as the conditional mean, to vary with the
explanatory variables by using the Feasible Generalized Least Square (GLS) method.

Feasible GLS weights each variable by the inverse of the conditional variance of
the residual from an original equation. As we do not know the value of the variances, we
have to model the heteroskedasticity form from an original equation.

Suppose we have the original model:

k
yi = a0 + Xaixi + 8i (3)
i =1
where a’s are the coefficient parameters to be estimated, xi’s are the explanatory
variables, and 8 the error terms. With OLS, we assume that 8 is i. i.d(0, 0'2). However, if 8
exhibits heteroskedasticity i.e. 8 is N~(0, 03-2), the approach is to estimate the model
allowing the variance of 8 to change.

The method we are using relies heavily on the work of Maddala (1971) and
Wooldridge (2000).

First, we estimate equation (3), disregarding heteroskedasticity. The objective is
to get the estimated disturbances é which represent the heteroskedastic error terms of the

model.

Suppose that Var(£/x) = 03-2 = 0” h(x)

 

9 Wooldridge (2000), page 276 presents a specific case of White test for heteroskedasticity using predicted
fitted value of the dependent variables and their squared terms. We got a r squared of 0.015 and a
Lagrangean Multiplier (LM) value of 8.45, which indicates heteroskedasticity at 5 percent of significance
level.

15

where h(x) is some functions of the explanatory variables that determine the

heteroskedasticity. In our case, we use an exponential model for h(x) in order to get

positive expected values which represent the conditional variances").

k
(50 + 251%)
Var(£|x)=0’2e i=1 (4)
where x; are the RHS in equation (3), 5,. ’s are unknown parameters.
From equation (4), we have:
k
32 = O'ze I =1 v (5)
where v is the error term of the conditional variances.

Taking the natural logarithm of (5):

k
1n(82)=70+ Z5iXi+fl ~ (6)
i=1

where ,u = In(v) has a zero mean and independent of x, thus homoskedastic;

and ya =1n(o’) + 50

Replacing 8 by I? , we run the conditional variance regressionll in order to get the

effect of each input on the variability of the outcomes y. The fitted values, let say g from

equation (4) after few modifications, are the weights to be applied to the original model

in order to have an unbiased estimator of the coefficient parameters and efficient standard

errors. Because we used the natural logarithm in equation (4), the modification consists in

taking the exponential of the expected value from equation (6) in order to restore the

 

'0 Linear model for the heteroskedastic error term gave us negative values for some observations. The issue

of the sign of the expected value is that we cannot use a negative weight with WLS method.
” OLS can be used in this case with the assumption that It is i.i.d(0, 0’2)

16

original magnitude of the numbers. Then following the Weighted Least Square estimation

method, we use the inverse of eg as weight in the final equation.

2.2.2. Choice of functioned forms

 

Several types of functional forms have been used to analyze agricultural
productivity. Each of them has their advantage and drawback related to the simplicity of
use and the behavior toward assumptions of microeconomic theory. For example, the
linear form is relatively simple and easy to manipulate but violated some economic
theories such as diminishing marginal return. The Cobb Douglas is also nicely behaved
and easy to compute functional form but it does not allow looking at the pairwise
interaction terms between two variables and leads to a zero output level when one of the
variables input is zero. We have a high likelihood of having such case for example in
fertilizer use. Therefore, we eliminate these two functional forms.

During the last three decades, the use of Quadratic, Translog, and generalized
Leontief (GL) models has become more familiar in agricultural production analysis in
developing countries (Sherlund et al, 2001; Byiringiro, 1995; Clay et al, 1995; Savadogo
et a1, 1994). They were chosen because of their flexibility, which allows an ability to
provide second order approximations to any arbitrary function. They share the common
characteristics of linearity in parameters and permit looking at the interaction terms
between two independent variables. They also would fit most of the assumptions of
economic theories on the analysis of agricultural production.

The Mackinnon White Davidson (MWD) test for non-linear non-nested models is

used to identify the “best among the tested” fimctional form for our data. Pairwise tests

17

between three functional forms (quadratic, Translog, and square root generalized
Leontief) indicated a better fit for square root generalized Leontief. Details of the tests
are given in appendix 2.

Specifically, with the three sets of vectors representing farmer’s variable choices,
conditioners, and land quality — natural shocks variables presented in equation (1), square

root GL model is of the form:

n m I n n
JE=0£0+ ZaiJJTi—‘i 25ka + 25th + Z priixz' +
i=1 k=1 t=1 i=1 i=1

i i [3ng EEIkafiIwLE imam: (7)
i=1 j=l i=1 k=1 i=1t=1
where x is a vector of variable inputs;
z a vector of conditioner variables;
a a vector of land quality and natural shocks variables;
0!. B (p, 5, yare parameters to be estimated;
and [.1 is i.i.d error term.
It is the number of input variables,
m the number of conditioner variables,

I the number of land quality and natural shocks variables.

Square root GL exhibits diminishing marginal returns to factors of production. Its

marginal productivity does have unrestricted Sign, allowing it to represent all stages in the

18

production process. GL is linear in parameters, so it can be estimated with linear
regression method. .

A drawback of the squared root GL is that we cannot normalize the variables to
have zero means”. We then work with non-normalized variables; therefore we might not
have exact second order approximation anywhere.

If land quality and natural shocks are not considered, we have the short
specification model:

r: “0+ Elair+12 Zfiij‘jxix} + kZlfika+ 21 kaiksz—i‘fli (8)
t- i=1j=l t-

From the equations 7 and 8, we have computed the marginal return of an input x;
and the elasticity ofx; to y by taking the partial derivative of y with respect to input x; for
the marginal return, and multiplying the result by (x; y") for the elasticity.

For the full specification, we have then:
a n —l m I x.
MPPxi = al, = (OH + Max/73+ 2 fiijW/xj + 251'ka + Zritat)><,/—' (9)
I: j = l k=1 1:1 y

3_yfy_

X.
_1
ex)“: ax =MPle. y (10)

The MPP and elasticity for the short specification are Obtained from equation (9)

and (10) by dropping the term with the a’s variables.

 

'2 Negative values are not allowed by the square root function of the CL model. With normalization, one
would expect to have exact approximation at any point of the curve.

19

2.3. Research Design

Theoretically, land quality and natural shocks variables differ in the sense that
farmers have knowledge of their own land quality while natural shocks are truly
stochastic. Farmer’s input use decision might have then higher probability to be affected
by land quality rather than natural shocks. An obvious example is that the quantity of
fertilizer applied in a given plot can be thought as independent of the existence of drought
that might occur during the agricultural cycle”. However, different irrigation quality may
induce different fertilizer use decisions by farmers.

Therefore, to investigate the individual effect of these types of factors, two
separate F—tests based (1) only on land quality -- irrigation, topography location, and soil
texture -- and (2) only on natural shocks -- inundation and drought -- were performed.
The null hypothesis is that all the parameter estimates for land quality or natural shocks

are equal to zero.

2.3.1. Description of variables
a) Dependent variables
y, the dependent variable, represents the total quantity of paddy rice produced per
plot. It is measured in kilograms.
b) Explanatory variables.
Plot area
Plot area is obvious in having a strong positive relationship with the level of

production. Area is quantified in are (100 square meters). Data are obtained from

 

’3 Notice that Malagasy farmer’s generally applies fertilizers only at the beginning of the agricultural
season.

20

farmer’s estimation of their plot areas, cross-checked by various traditional area
measurements.

Labor

There is an ongoing discussion of the homogeneity or not of family labor and
hired labor. Results in the literature diverge but findings seem to indicate that they are not
perfect substitute (Feder, 1985; Deolalikar and Vivj erberg, 1982; Rao and Chotigeat,
1981; Brown and Salkin, 1974). Frisvold (1988) attributed the difference between family
labor and hired labor to the higher incentive of family labor to better execute agricultural
tasks. However, he also demonstrated that hired labor efficiency depends on the level of
supervision by family labor. Other findings showed that hired labor are more
professional, more homogeneous, thus have the expectation to give higher output.
Following these findings, we will separate family and hired labor in our analysis.
Moreover, we will treat child labor as an independent and separate variable, avoiding
weighting and aggregating issues14 and in order to get more homogeneous adult family
labor composition.

Chemical Fertilizer

In general, we expect chemical fertilizer to have a positive relationship with
agricultural production. In practice, some farmers use chemical fertilizer to accelerate the
growth of the plants in the nursery plot. This constitutes an indirect effect of fertilizer use
in agricultural production, providing healthier plants. The current analysis does not

differentiate such a practice.

 

” For female and male labor, we assume that they perform different agricultural tasks, usually non-
substitutable; therefore they can be considered as equal. For example, transplantation, weeding tasks are for
women while plowing task is for men.

21

Other variables imuts

Use of chemical (pesticide, herbicide) is very low, so we decided to ignore it.
Seed is difficult to assess. Its level of use depends heavily on the technology adopted by
farmers. With the SRI technologys, farmers use a very low quantity of seeds, up to 15 kg
to transplant to one hectare. On the opposite, traditional technologys require three to six
times more seeds. Thus, by controlling the cultivated area and the technology used by
farmers, we control for seed.

For rice varieties, Goletti et al (1997) showed that there was no significant
productivity difference between “new varieties” and “traditional varieties”. Actually,
with the exception of specific varieties for high altitude rice produced by F OF IFA in the
late 1990's, the agronomic research did not proposed improved new rice variety for the
areas of study during the last five years, leading to a misunderstanding on the term “new”
and “traditional” varieties by farmers.

Managerial capital

Another important variable in the production function estimation is the level of
education at the household. Usually, there are many alternatives to measure this variable
but for our study; we will use the household's head number of years of education. This is
justified by the role and the place of the household head in agricultural process decision.
Its expected positive effect on agricultural production comes from the ability of educated
person to a better management and a more efficient information processing.

Animal Traction

We hypothesize that the use of animal traction is a technology choice, and it

depends on the availability of drafi power on the farm or the existence of an animal

22

traction market at the village level. Its omission can be misleading because it affects total
required labor (Savadogo et al, 1994). However its effect on agricultural productivity is
not very clear. First, the effect is hypothesized to be positive for land productivity in the
sense that it permits farmers to have land prepared on time. However, possession of
animal traction is also associated with the practice of agricultural extensification strategy
because it would allow farmers cultivating more land. Nonetheless the aggregated
quantity of animal traction collected during the survey is considered to be endogenous to
production15 , hence the number of draught oxen present at the farm should be used to
control for this variable. However, there is a high degree of correlation between draught
oxen and total oxen possessed by farms. As we use this latter to control for the use of
manure, we will drop the draught oxen variable from the regression.

Improved Rice Seedling Transplantation

Improved Rice seedling transplantation is used to control for the adoption of SRA
technologys. It is measured by the age of transplanted plants. A lower age of the plants at
the transplantation would result in higher production, but in turn will require higher need
for labor. There is also a high correlation between the age of transplanted plants and the
number of tillers per hill and water control. For example the lower the age, the fewer the
number of tillers and the better the irrigation control. Thus, the variable age of

transplanted plants will control for the adoption of an improved rice seedling technology.

 

'5 Animal traction was the sum of all tasks performed by animal traction for the plot from plowing,
leveling, to transportation. If plowing and leveling, might be assumed as exogenous, transportation is
endogenous because the number of days needed to move the paddy-rice from the plot to the farm-storage is
a linear function of the quantity produced.

23

Extension Service

This is a qualitative variable, which shows if farmers have had contact with
extension service on the rice production at least once during the last five years. Several
technologys besides SRI and SRA were diffused by the extension service. One can adopt
the recommendation on seeding date, number of weeding, mulching, use of chemical
fertilizer etc. This variable would control for all these technologies, and expected to have
positive effect on productivity.

gem

In Madagascar highland, there are two agricultural seasons for rice: the “vary
aloha” and the “vakiambiaty”. The “vakiambiaty ” rice is sown during the rainy season
and harvested in April — May and is the most important in term of cultivated area, while
the “vary aloha” is sown in July — August and harvested in December — January. We use
a dummy variable representing these two seasons. Sometimes, the “vary aloha” crop will
give higher production because of the relative good land quality that allows farmers to
cultivate rice before the rainy season. However, it is penalized by the low temperature
during the first stage of plant development, and the negative effect of cyclones during the
maturation stage.

RagiO—n

We have here two regions: Vakinankaratra and Fianarantsoa plateau, both part of
the Highland agro-ecological zone (See appendix 1). Unfortunately, the limited number
of zones —- only 2 -- precludes the use of other climatic measurement such as rainfall in

the model, because of the perfect collinearity issue.

24

Land tenure

There are two major land rental market systems that can be observed in
Madagascar: leasing and sharecropping. The literature is ambiguous on the effect of land
tenure on productivity. The common wisdom is that temporary property rights might
have a negative effect on agricultural productivity, as there is less incentive for
investments by the temporary or insecure owner. However, empirical results on land
titling are mixed. Studies for different countries in Central Africa show neutral effects of
legal land rights on agricultural productivity (Place and Hazell, 1993; Platteau, 1996).
Other studies have documented that insecure land rights lead to less input use,
investments, and therefore, lower productivity (Anim, 1999; Lutz et al, 1994; Reardon et
al, 1999; Feder and F eeny, 1991). We expect the argument for land tenure, which is

controlled by a categorical variable for land rental to have negative effect on agricultural

productivity.
Land Qlalig

Land quality is very difficult to define and to measure. Most of the studies on
agricultural productivity used categorical variables to develop a proxy for land quality.
Soil color, soil texture, and soil depth were used by Bhalla and Roy (1988). IFPRI (1997)
used irrigation and prices of land. In his study on Indian agriculture, Bhalla (1988)
reported that there is a big difference in intra-zonal land quality, so land characteristics
cannot be included in the categorical region variables. An individual estimate of land
quality is needed for the estimation of the production function.

Land quality can be assessed as a combination of two factors: the inherent soil

quality and human-made adjustments. Inherent soil quality is determined by the physical

25

and chemical characteristics of the soil. By physical, we mean the slope, the location of
the plot, and the soil texture and structure. Chemical characteristics regroup the
composition of the soil in different nutrients: nitrogen (N), phosphorus (P), potassium
(K), iron, minor and trace elements etc. Physical human made adjustments are changes in
the land environment such as irrigation, drainage surface, and soil conservation
investments, which affect land quality. For example, two plots with the same inherent soil
quality are likely to produce different outcomes if the irrigation system is different.

For our study, we will use three categorical variables to control for land quality:
the topography location of the plot, the soil texture that is function of the ability of the
soil to retain water, and the irrigation system”. The location of the plot differs from
lowland rice field vs. terraced rice field. Texture is between sandy and clay/silty soil.
Clay and silty soils are expected to have the ability to retain water longer than sandy soil,
resulting in better weeds and water control. Irrigation distinguishes between (i) human-
made such as pond, dam, well and (ii) natural sources of irrigation such as rainfall or
spring water source. Farmers with controlled irrigation scheme are expected to use more
inputs giving the reduction of production risks.

Lastly, many studies introduced the number of plots cultivated by farm as an
important variable that would affect farmer's technical efficiency. A high number of plots
is hypothesized to have negative influence on the productivity because of a more

complex field management.

 

'6 A drawback of this analysis is that we do not have the soil initial endowment in organic matter and
chemical elements. The levels of these components are assumed to be controlled by the use of fertilizer and
manure by farmers and by the land quality proxy variables (texture, location).

26

Natural Shocks

This set of factors that would influence agricultural production is represented by
two categorical variables of natural constraints during the agricultural season: inundation
and drought problems. They are both expected to have negative relationship with the

level of rice production and assumed to be stochastic over the years.

2.3.2. Data patterns

Our study uses data from smallholders in Madagascar highland. This is a
mountainous area, with a lot of narrow depressions (bas-fond) where farmers usually
grow lowland rice. The average altitude is around 1,200 meters.

For the climatic condition, temperature varies from an average of 26°C in the
summer to 13°C in the winter, which restricts rice cultivation during the dry and cold
season. Average rainfall is between 1300 and 1600 mm per year, and is concentrated
from November to April. This is also a populous area with a demographic density of

more than triple the country’s average.

27

Table 1 — Descriptive Statistics on rice production variables by farm size tercile,

563 fields, Madagascar highlands, 2000.

 

 

 

 

 

 

 

 

Overall samples Tercile

Variables Units Sample Standard Small Med. Large

means Deviation
Dependent variable
Production quantity Kg of rice 1,009 997 352 820 2,012
Area and yield
Total cultivated rice area Ares per household 44.82 47.64 8.60 30.33 95.55
Yield Kg / are (w) 22.52 15.16 45.99 25.28 19.40
Choice variables
Family labor Man-days / are (w) 0.99 1.38 3.39 1.20 .70
Hired labor Man-days / are (w) 1.04 1.12 2.66 1.27 .80
Child labor Child-days / are (w) 0.14 0.32 .34 .13 .12
Number of oxen Number 2.53 2.99 1.15 2.21 3.77
Fertilizer use Dummy (1 = use) 0.10 0.30 .12 .10 .09
Fertilizer application Kg of fertilizer/are (w) 0.08 0.34 .12 .09 .07
Improved RST Dummy (l = yes) 0.06 0.24 .07 .06 .06
Conditioner variables ~
(Education level Number of years 4.91 2.89 4.54 5.39 4.70
Number of plots Number of plots 3.00 1.38 2.16 2.89 3.66
Extension service contact Dummy (1 = yes) 0.17 0.37 .12 .21 .15
Land tenure Dummy (1 = renting) 0.11 0.31 .06 .16 .10
Season vary aloha Dummy (l = yes) 0.15 0.36 .19 .16 .l 1
Region Vakinankaratra Dummy (1 = yes) 0.56 0.50 .72 .54 .l 1
Land quality and natural shocks variables
Irrigation Dummy (1 = yes) 0.40 0.49 .49 .32 .42
Lowland location Dummy (1 = yes) 0.39 0. 49 .42 .37 .38
Texture clay Dummy (1 = yes) 0.73 0. 44 .60 .78 .78
Flood problem Dummy (1 = yes) 0.44 0. 50 .37 .44 .48
Drought problem Dummy (l = yes) 0.68 0.47 .63 .65 .72
Other interesting variables
Household size Number 6.7 2.9 6.02 6.43 7.42
Family labor for weeding Day/are 4.4 7.6 .68 .27 .18
Hired labor for weeding Day/are 3.6 6.4 .42 .23 .15
Family labor for transplantation Day/are 2.1 2.7 .37 .13 .08
Hired labor for transplantation Day/are 3 .7 4. 7 .48 .21 .16
Water management Day/are 1.2 2.5 .21 .06 .05

 

(w) Average weighted by plot area

Source: Agricultural Production Survey, Madagascar Highland June 2000

28

The data come from a farm survey conducted by the Cornell ILO/FOFIF A project
in Madagascar in July - August 2000. It includes rice production during the dry season
1999 (vary aloha) and the wet season 1999-2000 (vakiambiaty). The sample is
constituted by 237 rice growers, which correspond to 563 rice plots afier data cleaning
process. The map in appendix (1) shows the area of study.

The questionnaire has data on the characteristics of each plot cultivated by
farmers: plot area, texture, irrigation system, and topography location. It also has
information about the cultivated area, the quantity of labor used (family, hired, male,
female, children), the number of days of animal traction used (family, hired), the quantity
of chemical fertilizer, the level of production, the type of seedling transplantation
methods adopted, and the natural conditions during the production process including
flood and drought.

Table 1 shows the descriptive statistics of the variables used in oUr study. In
general, there is relative high variability in the sample. For example, the total rice
cultivated area, and all the labor use, have coefficients of variation higher than 100
percent. Chemical fertilizer use and improved rice seedling transplantation (RST)
adoption exhibit also a very high variation more than 400 percent.

For our sample, the average yield, weighted by plot area, is around 22.5 kg per
are, with a standard deviation of 15.2 kg. This average yield is higher than those reported
by Bemier and Dorosh (1993), Roubaud (1997), and IFPRI (1997). These previous

studies reported average yield between 13.0 and 20.1 kg per are for aquatic rice. We have

29

however to notice that these studies used different yield computation method compared to
the current analysis”, thus we cannot compare the yield or the change over time.

Use of labor is quite high with averages of 0.99 and 1.04 man-day per are
respectively for family and hired labor. First, there is the fact that rice technology
requires a lot of manual works from the nursery field, the planning, the transplantation,
the weeding, through the harvesting. As some of these tasks should be done at a limited
period, intensive need of hired labor is quite normal. Also there is the existence of a kind
of fixed tasks such as water management, guarding against crop-thief, where labor
quantities do not depend on the size of the plots. These “fixed” labor increase the per-are
labor used18 inversely proportional to plot area.

One can observe the usual low percentage of chemical fertilizer adopters, with
only 10% of the sample. When we observe the rate of application, for all plots (those
which received and not received chemical fertilizer), we have a figure of 8 kg per hectare,
while for plots that received fertilizer; farmers apply 81 kg per hectare”. This
corresponds approximately to one fourth of the extension service recommendation”. In
most of the cases, farmers apply chemical fertilizer on cash crop under a contracting
system such as tobacco, sugarcane, cotton, and barley rather than on rice (IFPRI, 1997).
The key reason for this difference in input use is that the contracting firms supply

fertilizers to farmers. Also, farmers have access to more information especially price

 

'7 The current analysis use average yield computed directly from plot data weighted by the plot area while
the other two studies seem to use average yield computed at the farm level, which is already an average
from the plot level. Bemier and Dorosh (1993) includes low land productivity slash and burn technologys
from the Middle East zone (excluded from our sample). Also, we did not use any weight to get
representative samples.

'8 We will come back to this issue later on the analysis of the labor use by farm size.

'9 Previous studies in Madagascar (Bemier and al, 1993; IFPRI, 1997) presented similar results.

2° For rice in the Highland, the extension service from the Ministry of agriculture recommends 300 kg of
NPK 11-22-16 plus 66 kg of Urea per hectare (Bemier and Dorosh, 1993).

30

information that allows them to make more optimal decisions on input use and input
allocation. Similarly, improved RST adopters are also very few with only 6 percent of
total sample. We have classified farmers as adopters of improved RST if they transplant
young plants less than 25 days.

Although the two regions of study are part of the highland region, which is the
most densely infrastructured region in Madagascar, it is surprising that only 17 percent of
farmers have had a contact with the extension service.

The percentage of plots with physical human-made irrigation with 40 percent of
all plots is higher than the national average of 24 percent (FAO, 2000). This contributes
significantly to drop the inundation to affect only 44 percent of the total plots21 (Table 1).

Most of the rice lands are clay or silty textured. This is a common morphological
characteristic of the narrow lowland in the Highland region. These plots are frequently
flooded, thus subject to sedimentation from the watershed.

Table 1 also shows descriptive statistics by farm size. Farms were classified into
tercile according to the total rice cultivated area”. We found that for our sample, the
average rice area cultivated by household is 8.60; 30.33; and 95.55 ares for respectively
tercile 1, 2 and 3, which represent small, medium and large smallholder farms. Each
tercile contains around 79 households23 .

Like the findings from many studies in developing countries, average land
productivity indicates that small farms have higher land productivity than large farms

(Byiringiro, 1995; Barrett, 1994; Feder, 1985, Deolalikar et al, 1981). From tercile 1 to

 

2‘ In the coastal area where modern irrigation is almost inexistent, inundation risk may reach 90 percent of
cultivated land during the cyclone season.

22 We use total rice cultivated area as a proxy for farm size. This constitutes a good proxy for farm size.
The correlation coefficient between these two variables is greater than 0.70 (SMTIS, 1989).

23 Use of per capita cultivated area shows similar results (see Appendix 4)

31

tercile 3, we respectively have significant different average yields of 45.99; 25.28; 19.40
kg per are. i

To eliminate the doubt that the inverse relationship between riceland holding and
rice productivity may be caused by some artifacts in regrouping data by tercile, we will
use the method suggested by Deolalikar (1981). It consists in running a single regression
fiinction of the form:

Y, = a + 13X, + e,
where Y I is the yield in kg; X I is the size of the farm in are, and 8, is the i.i.d. error term.

There is inverse relationship if the parameter [3 exhibits a negative sign.

For our sample, the result of this regression presents a negative and significant
coefficient of -49.0 for X, with a t-value —10.0 and an R squared of 0. 16. This
corroborates the hypothesis that there is indeed an inverse relationship between farm size
and rice productivity. We found also that the pairwise correlation coefficient between
plot area and land productivity is —0.41, statistically significant at 1 percent level.

There are significant differences on the quantity of input uses by tercile of farm
size. For example, family labor per are decreases from 3.4 for small farms to .7 for large
farms. Hired labor, and use of organic and chemical fertilizers follow similar patterns.
Inversely, the number of oxen is more important for large farms than for small and
medium farms. Logically, a decrease in family and hired labor should correspond to an
increase in the use of animal traction because of the possible substitution effect between
these two categories of inputs. We can observe this tendency in Table 1.

These high quantities per-are of family and hired labor confirm the hypothesis

that small farms would allocate more labor to the available land than large farms, and this

32

partially explains their higher land productivity. The per-are average quantity of labor
allocated to weeding and transplanting (Table 1) illustrate this situation. For example,
either for weeding or transplanting, family labor use per-are for large farms represents
only one fourth of the same variable for small farms. As these two tasks (weeding and
transplanting) are key-factors for rice production, it is not surprising if small farms have
higher land productivity than large farms.

Small farms are also facing lower opportunity costs of labor than large farms. It is
then in their logic to allocate large quantity of labor on the available land.

As there are some fixed tasks to be performed independently of the size of the
plots, farmers are sometimes obliged to spend the same amount of time whatever the plot
size is. This is the case of water management, which occupies .21 man-day per are for
small farms against only .05 man-day per are for large farms, a ratio of 4 to 1. The overall
result is an increase of the relative per are use of labor for small farm. Time spent in
guarding against harvest theft falls also under this category. Theft problem is crucial for
certain villages and can make big differences in labor productivity (IFPRI, 1997).

For the conditioner variables, the existence of extension contact presents no
significant differences between the tercile of farm size, while monotonic changes are
noticed for seasonal and regional variables.

Irrigation quality does not vary significantly across tercile of farms size. In
general, one can say that large farms are facing higher production risks than small and
medium farms. For example, in 1999-2000, table 1 shows that inundation and drought
occurrences were statistically higher as tercile goes from 1 to 3 (small to large farms).

Lands of large farms seem to be located in worse area than those of small farms. Thus

33

they are more subject to bad natural conditions such as inundation or drought. On the
other hand, large farm have more plots, and if all plots have the same, independent
probabilities of being flooded, then large farms will have higher probability of
inundation.

Following the same pattern, small farms have higher chance to grow rice during

the dry season, indicating a relative good irrigation scheme.

3. Regression Results and Interpretation

3.1. Effect of Land Quality and Natural Shocks on Agricultural Production

The results from estimating equation (8) are shown in Table 2. Overall, the model

fits the data extremely well. It is evident that land quality and natural shocks do influence

rice production in Madagascar. Some coefficient parameters of the variables controlling

for land quality and natural shocks or their interaction terms with other variables were

found to be statistically different from zero.

a.

Jointly tested at 1 percent level, the F-test on variables
controlling for land quality and natural shocks from the full
specification model leads to reject the null hypothesis of “all 6 's
and 7‘s are equal to zero” 24 with a F(30, 470) = 3.36.

Tested separately, at the 1 percent level, we alsoreject the null
hypothesis (See. Paragraph 3.3) that all land quality variables do
not have any influence on productivity and conclude that at least
one of the land quality coefficients is different from zero with a
F(18, 470) = 2.85.

The same test applied to natural shocks variables presents a
F(12, 470) = 3.39, which supports also the null hypothesis

rejection.

 

2‘ Respectively 8 and 7 represent the parameters of land quality and natural shocks variables and their
interaction terms— See equation 7.

35

Table 2 — Madagascar Highland Rice Production Function
With and Without Land Quality and Natural Shocks

Plot level analysis, 563 fields, Madagascar Highland 2000

 

 

 

 

Full Specification Short Specification
WITH land quality and WITHOUT land quality
natural shocks and natural shocks
Dependent variable = Rice production in kg Coeff. Sig. Std. Err. Coeff. Sig. Std. Err.
Plot area in are (AREA) 0.991 "‘ 0.596 1.459 " 0.593
Family labor in day (LABF) 2.440 "* 0.683 1.431 " 0.656
Hired labor in day (LABH) 1.167 "‘* 0.407 0.352 0.415
Child labor in day (LABC) -0.134 0.693 0.853 0.696
Number of oxen owned by farmer (OXEN) 3.306 *“ 0.931 -0. 193 0.880
Chemical fertilizer use - YES = 1 (FERT) 4.442 " 2.223 0.851 2.332
Improved RST - YES = 1 (IRST) 3.988 3.074 1.539 3.160
Region Vakinankaratra — YES = 1 (DVAK) 1.414 1.375 -0.660 1.373
Extension service contact - YES = 1 (VULG) 1.851 1.281 0.985 1.406
Land tenure - Renting land = 1 (RENT) 5.621 "" 1.676 2.332 1.675
Head education level in year (EDUC) -0.266 0.238 -0.185 0.242
Season vary aloha - YES = l (DSEA) -0.909 1.360 0.157 1.488
Number of plots per farm (PLOT) -0.662 0.477 -0.3 1.9 0.501
Human made irrigation - YES = l (IRRI) 2.632 " 1.106
Topography location - low = 1 (TOPO) 2.371 "" 1.017
Soil texture - Clay/silty = 1 (TEXT) 4.237 "" 1.178
Inundation in 1999- YES = l (FLOO) 0.719 1.099
Drought in 1999 - YES = 1 (DROU) 1.360 1.104
Interaction Terms
AREA‘AREA -0.193 """" 0.054 -0. 150 " 0.062
LABF*LABF -0.193 "* 0.051 -0. 150 *" 0.052
LABH‘LABH 0.036 0.031 0.018 0.035
LABC*LABC 0.210 ” 0.098 0.157 0.116
OXEN*OXEN 0.228 0.209 0.489 " 0.212
AREA‘LABF 0.347 *” 0.081 0.250 "" 0.089
AREA’LABH 0.168 "* 0.065 0.163 *"' 0.077
AREA‘LABC -0.091 0.089 0.017 0.103
AREA‘OXEN 0.062 0.132 0.024 0.141
AREA‘FERT 0.833 " 0.434 0.857 " 0.488
AREA‘IRST -0.956 ‘ 0.551 -1 .294 *“ 0.607
AREA‘DVAK 1.883 ""‘ 0.280 1.655 "* 0.298
AREA*VULG -0.553 * 0.295 -0.261 0.334
AREA‘TENU 0.164 0.284 0.108 0.320
AREA’PLOT 0.008 0.092 0.136 0.100

36

 

 

 

 

Full Specification Short Specification
WITH land quality and WITHOUT land quality
natural shocks and natural shocks
Dependent variable = Rice production in kg Coeff. Sig. Std. Err. Coeff. Sig. Std. Err.
AREA‘EDUC -0.094 ‘“" 0.042 -0.064 0.045
AREA‘DSEA -1.300 "* 0.280 -1.463 "* 0.325
LABF'LABH -0.193 *" 0.060 -0. 106 0.067
LABF’LABC -0.157 0.100 -0.268 ** 0.106
LABF‘OXEN -0.010 0.137 0.189 0.133
LABF*FERT -0.739 *“ 0.361 -0.469 0.400
LABF’IRST 0.475 0.528 0.914 * 0.514
LABF‘DVAK -1.110"”'“" 0.342 0878*" 0.334
LABF‘VULG 0.097 0.259 0.239 0.255
LABF‘TENU -0.553 0.375 0.097 0.378
LABF‘PLOT 0.013 0.109 -0.054 0.1 14
LABF*EDUC 0.130 " 0.054 0.074 0.054
LABF'DSEA 0.678 "" 0.261 0.425 0.297
LABH‘LABC 0.076 0.062 -0.045 0.074
LABH‘OXEN -0.114 0.102 -0.119 0.1 13
LABH‘FERT 0534 0.328 -0.200 0.390
LABH'IRST 0.061 0.296 0.128 0.315
LABH‘DVAK -0.053 0.201 0.341 0.217
LABH*VULG 0.129 0.198 0.038 0.229
LABH‘TENU -0.639 "”" 0.251 -0.530 " 0.252
LABH‘PLOT 0.109 0.074 -0.030 0.082
LABH‘EDUC 0.047 0.032 0.059 0.036
LABH‘DSEA 0.719 ""' 0.237 0.826 *" 0.269
LABC‘OXEN -0.053 0.164 -0. 143 0.171
LABC‘FERT -0.390 0.441 -0.019 0.501
LABC*IRST -0.712 0.714 -0.605 0.809
LABC*DVAK «0.196 0.316 0.024 0.341
LABC*VULG -0.296 0.347 -0.1 12 0.412
LABC‘TENU -0.191 0.413 -0.021 0.428
LABC‘PLOT 0.250 " 0.120 0.138 0.120
LABC‘EDUC -0.026 0.060 -0.049 0.058
LABC*DSEA 0.686 0.464 0.298 0.507
OXEN‘FERT -1. 182 "' 0.623 -0.447 0.668
OXEN‘IRST -1.070 0.660 -0.047 0.713
OXEN‘DVAK -0.780 " 0.364 -0.393 0.387
OXEN‘VULG 0.336 0.489 0.281 0.493
OXEN‘TENU -0.625 0.550 -0.432 0.556
OXEN’PLOT -0. 153 0.130 -0. 132 0.132
OXEN‘EDUC -0.009 0.066 0.026 0.065
OXEN‘DSEA -0.394 0.502 0.390 0.548
CONSTANT TERM -4.496 ** 2.246 2.852 2.025
Land characteristics and Natural shocks
AREA‘IRRI -0.281 0.209

37

 

 

 

 

Full Specification Short Specification
WITH land quality and WITHOUT land quality
natural shocks and natural shocks
Dependent variable = Rice production in kg Coeff. Sig. Std. Err. Coeff. Sig. Std. Err.
AREA‘TOPO 0.752 *" 0.208
AREA*TEXT 0.532 ** 0.228
AREA‘FLOO -0.308 0.240
AREA*DROU 0.605 ** 0.279
LABF‘IRRI -0.038 0.244
LABF*TOPO ~0.564 " 0.233
LABF*TEXT -0.620 ** 0.265
LABF*FLOO 0.143 0.247
LABF‘DROU -0.312 0.240
LABH‘IRRI -0.1 16 0.152
LABH‘TOPO —0.441 *** 0.168
LABH*TEXT -0.389 ""’ 0.191
LABH‘FLOO -0.058 0.187
LABH*DROU -0.452 " 0.197
LABC'IRRI -0.503 " 0.234
LABC*TOPO -0. 162 0.288
LABC*TEXT -0.505 * 0.276
LABC*FLOO -0. 122 0.314
LABC’DROU 0.836 "* 0.318
OXEN‘IRRI 0.200 0.291
OXEN‘TOPO -0.310 0.321
OXEN*TEXT -1.020 ‘" 0.379
OXEN*FLOO -0.626 0.409
OXEN*DROU -1.316 ""' 0.386
R squared .91 .88
Adjusted R squared .89 .86
Overall F test F(98, 464)= 47.5 F(68, 494): 51.5
Residual Sum of Squares 3510 5376

 

*, “, "* Respectively significant at 10, 5, and 1% level

The F test between the two models specification (full and short) showed a F(30, 563) = 141 which implies
that at 10 percent level, we reject the null hypothesis and conclude that the two models do not have the
same explanatory power.

The very high R squared is an artifact resulting from the econometric method used. With a linear functional
form, the R squared is .62 for the full specification.

Source: Agricultural Production Survey, Madagascar Highland June 2000

 

38

Moreover, the Wald25 tests performed on each variable confirm the hypothesis
that land quality and natural shocks affect significantly rice production in Madagascar
(Appendix 4). It shows that topography location, irrigation, and drought are significant at
the 1 percent level. Soil texture is significant at the 10 percent level while inundation risk

is significant at the 5 percent level.

3.2. Changes in Land, Labor, and Fertilizer Marginal Return Due to the Omission

of Land Quality and Natural Shocks in the Model Specification

Table 3 presents the changes in MPP between a full specification model with land
quality and natural shocks variables and a short specification model without these two
sets of variables. Marginal productivity is obtained from equation (9), itself obtained
from the first derivative of the production function presented in equation (8). The overall
results show that the omission of land quality and natural shocks variables in the model
specification affects significantly the estimation of the marginal productivity of some
factors of production.

Although there are no significant changes on land and hired labor between the
two models, one can observe a small variation on the return to family labor, with a 9%
change. The short specification tends to overestimate the return to this factor. For both
models, hired labor shows a higher return than family labor. For example, the full
specification model shows a hired labor MPP of 4.7 kg per man-day and a family MPP of

2.7 kg per man-day, therefore indicating a net advantage of using hired labor.

 

25 See Appendix 4 for detail of Wald test.

39

Table 3 — Marginal Input Productivity With and Without Land Quality

and Natural Shocks, Estimation at the Sample Means

Plot level analysis, 563 fields, Madagascar Highland 2000

 

Full Specification Short Specification

 

 

 

 

MPP Output MPP Output Relative Sig.
ElasticitL Elasticity chflge ‘
MPP and Elasticity of some choice variables
Land K8 nee per are 13.5 .60 13.5 .60 0%
Family labor Kg rice per mannay 2.7 .12 2.9 .13 9% *
Hired labor K8 rice per manner 4.7 .22 4.8 .22 2%
Child labor Kg rice per childnay 4.3 -.03 -1.2 -.01 -71% ***
Number of plots K8 rice per number -6.5 -5% -7.8 -5% 19%
Number of oxen K8 nee per ox 6.7 .04 12.3 .07 84% "*
Switchingeffect of some categorical variables
Unit of quantity Quantity Percent Quantity Percent Relative Sig.
chilnge change change change change '
Education of head Kg rice per year 1.7 2% 3.4 4% 102% "
Chemical Fertilizer Kg rice if using fertilizer 396 10% 414 10% 4%
Extension service Kg rice if access 207 8% 360 14% 74% "*
Rice seedling Kg rice if IRST 391 6% 109 2% -72% *
Season Kg rice if season 2 -111 4% 26 1% -124% "*
Region Kg rice if Vakinankaratra. 231 30% 241 31% 4%
Land tenure Kg rice if rented 291 8% 241 6% -17%

Switching effect of Land quality and Natural shocks variables

 

 

Unit of quantity Quantity Percent

change change

Irrigation Kg rice if irrigated 37 8%
Topography location Kg rice iflowland 97 9%
Texture Kg rice if clay 4] 7%
Inundation Kg rice if no flood -100 _] 0%
Drought Kg rice if no drought -10 -20/0

 

a = Change relative to the full specification model, computation from equation (2a)

‘, ", ”* are respectively significant at 10, 5, and 1% level, See Appendix 5 for the Bootstrap method and

the t-test results.

Source: Agricultural Production Survey, Madagascar Highland June 2000

Several explanations for this exist. First there is the effect of “supervision” by

family labor that would result in better performance of hired labor (F risvold, 1994). A

40

second hypothesis is the nature of tasks attributed to hired labor, such as transplantation
and weeding. These are among the key constraints to get higher rice productivity as any
delay in transplantation and weeding will result in lower production. Third, there is the
issue of heterogeneity in family labor, i.e. quality of family labor is on average lower
even if we take out child labor. It should be noted that family labor might have low
retums but its presence is required in order to get higher returns from hired labor.

Child labor MPP from the short specification is estimated to be 71 percent less for
the short specification compared to the value from the full specification model. This
indicates that child labor is also correlated with some land quality and natural shocks
variables. The coefficients are both negative for the two model specifications.
Nonetheless, low production might not be the effect of child labor per se, but is related to
the task attributed to children. For example, it is expected that child labor is used to keep
birds away, but children’s presence might not reduce the loss to zero. The low production
would then be the effect of the damage done by birds but not child labor on itself. In
some cases, there might also be a negative direct effect of child labor, i.e. for example, if
they perform poorly some difficult tasks such as transplantation or harvesting. Moreover,
children can have a negative MPP if they require supervision from adults that reduce the
work the adult can do.

Other variables that have substantial positive effects are ownership of oxen, the
level of education of the head of household, and access to extension services. Ignoring
land quality and exogenous shocks would inflate the effect of these factors on agricultural
production. For example, in the case of access to extension services, instead of having a

production increase of 201 kg from the full specification, one would expect to get 360 kg

41

from the short specification, i.e. a 74 percent difference. The expected bias as a result of
agricultural policy interventions is clear. A policy focused on‘ extension services will
have a smaller impact on production when land quality and natural shocks are taken into
account.26

The seasonal effect and improved rice seedling transplantation are also affected
negatively by the omission of land quality and natural shocks as the short specification
shows a reduction of the magnitude of the switching effect from the adoption of IRST
and the practice of "vary aloha".

In table 3, we also have additional information on the effect of land quality and
natural shocks variables on other determinants of rice production. The long specification
shows a marginal return of 8 percent for irrigation. This value is very close to the
findings from IF PR1 (1997), which reported an irrigation effect between 4 and 7 percent
on land productivity. Lowland plots are expected to produce 9 percent more than rice
fields in terraces. Clay or silty soil texture contributes to an increased total production of
9 percent. With respect to natural shocks, inundation is expected to reduce rice
production by 10 percent. This is clearly an average value as in certain cases, flood can
completely destroy the crop”.

Lastly, the land rental dummy shows a significant and positive effect on
agricultural productivity. It indicates that instead of reducing productivity, short-term
land markets improve it. In theory, land rental arrangement might have two opposite

effects on productivity: creating more insecurity to the cultivator or leading to

 

2° The Ministry of agriculture reports an average increase of 4.68 kg of rice per kg of fertilizer used from 0
to 366 kg per hectare. Our fertilizer marginal physical product at an average rate of application of 81 kg per
hectare is here 3.24 kg with the full specification.

27 Some plots in the sample did not report production because of flooding.

42

reallocation of the land to more efficient tenants. The results of the regression indicate
that the latter effect is more important as land under rental agreement slightly increase
productivity. This supports the hypothesis that in our area of study, land was rented out to
a group of farmers who could cultivate it more efficiently (see also Dorosh et al, 1998).
However, we also can see in the descriptive statistics that access to rental lands seems to

be difficult for poorer households, as they use only 5.5% of the rental land.

3.3. Effect of Farm Size on the MVP of Land, Labor, and other factors
M_a_rgin_al Land Productivity

The patterns of marginal land productivity presented in Table 4 support the
existence of the inverse relationship between land productivity and farm size for rice
production in Madagascar highland. The Marginal Value Product (MVP) is computed
from MPP and an exogenous average price of one kilo of paddy-rice, collected at the
village level. From our sample, this average output price is 274 ariaryzs. The ratio is that
of MVP divided by the factor price”.

Table 4 also shows that for any tercile, land MVPs lay above the market price of
land. We have estimated the market price of land from the cost of renting one “are” of
land during one agricultural season. As such, we obtained from our sample a value of
around 2,600 ariary. The fact that these land MVP is greater than the factor price seem to
indicate that land markets do not function perfectly and that poor people are ofien

constraint to rent out land because of liquidity constraints.

 

28 In 1999, the average exchange rate for $ 1 US is 1,300 ariary.
29 A ratio less than 1 indicates an over-use of the factor while inversely, a ratio greater than 1 is a sign of an
under-use of the factor of production.

43

One can also notice the relative low elasticity output estimates of land. In table 4
we see an average output elasticity varying from .49 to .54, which is comparable to the
.50 reported by Minten et a1 (2000) but definitely lower than the land output elasticity
found by Sherlund et a1 (2001) on rice in Ivory Coast. Does this result imply that
agricultural policy should focus on land redistribution in order to increase land
productivity? The answer is not straightforward as if the behavior of farmers is a
consequence of their landholding, an increase in the cultivated area for small farms would
result in a reduction of their technical efficiency. On the other hand, reducing the land
area for a large farm would not automatically result in an increase of their technical
efficiency. Other factors might inhibit this improvement such as the reduction of the
return to labor as land productivity increases.

Marginagabor Productivity

For labor MVP, both family and hired labors present a significant difference inter-

tercile (Table 4). We see the same pattern observed for the whole sample. Independently

of farm size, family labor MVPs are lower than the hired labor MVP.

Table 4 — Inputs MVP and Switching Value of Binary Variables

by Farm Size Tercile, 563 fields, Madagascar Highland 2000

 

Full specification model

Tercile of farm size

 

 

 

 

 

 

 

 

Average F actor prices Small Medium Large
Land MVP (ariary per are) 3693 2,600 (a) 6138 3725 2873
From small to (C01) -3 9% -53%
From medium to (C01) -23%
Ratio MVP to Factor prices 2.4 1.4 1.1
Elasticity of land .49 .54 .54
Family Labor MVP (ariary per day) 741 760 (b) 184 861 1088
From small to (C01) 369% 492%
From medium to (Co!) 26%
Ratio MVP to Factor prices .2 1.1 1.4
Elasticity of family labor .05 .15 .14
Hired labor MVP (ariary per day) 1236 7 60 (b) 845 1217 1650
From small to (C01) 44% 95%
From medium to (C01) 36%
Ratio MVP to Factor prices 1.1 1.6 2.2
Elasticity of hired labor . 18 .22 .25
Child labor MVP (ariary per day) -1185 - -l803 -1573 -494
From small to -13% -73%
From medium to -69"o
Elasticity of child labor -.05 -.03 -.01
Number of oxen MVP (ariary per ox) 1824 - 2230 1629 1858
From small to -2 7% -1 7%
From medium to 14%
Elasticity of oxen .05 .04 .04

Percent Change on M VP from Switchirg Value of Binary Variables
Chemical fertilizer use 10% 13% 1 1% 7%
Extension access 8% 24% 1 1% 1%
Irrigation 3% 22% 1 1 % 1%
Topography 9% 4% 9% 1 0%
Texture 7% 15% 9% 3%
Land tenure 8% 25% 10% - 1 %
Inundation 40% -1% -10% - l 4%
Second season 4% 17% -3% -l 7%
SRA 6% 43% 10% -1 1%
Drought 4% -4% -4% 0%
Education (d) 2% 9% 2% -1%
Number of plots ((1) -5°/o -8% -6% -1%

 

(a) This is the average price in ariary for renting one are of land for one season.
(b) This is the average agricultural wage rate at the village level

(col) = medium or large

(d) These are the change in MVP from one unit of change in the dependent variable
Source: Agricultural Production Survey, Madagascar Highland June 2000

45

The ratio for family labor for small farms is less than one, indicating an over-use
of this factor. This corresponds to a MVP of 184 ariary per day, which at first, appears to
be irrational. In theory, within a competitive labor market environment, the optimal
solution for profit maximizing farmers is to equate labor MVP from rice to the
agricultural wage rate, i.e. assuming that renting out labor constitutes the next best
alternative. If labor markets are imperfect - which might to be the case in rural
Madagascar highland - farmers might implicitly be forced to use family labor on the
available land as long as the marginal return is not negative.

A failure in the output rice marketing system will also alter the optimal solution
previously mentioned. If there are high uncertainties for rice supply during the lean
period, risk-averse farmers would maximize rice production in order to minimize the risk
of rice shortage. Real prices of rice would be higher because of transaction and
transportation costs, resulting in higher MVP that farmers would use as benchmark for
their input allocation decision. Over-using family labor on rice production would then be
rationally justified.

Historical changes in the labor market may explain the adoption of the
controversial SR1 technology. In the late 1980’s, Madagascar just came out of the
centrally planned economic system. Thus labor mobility was limited. Moreover, the rice
market was not entirely liberalized; rice distribution remained under the control of the
government and did often not reach rural and remote area. Given these two constraints,
many small farmers were over-using family labor for rice production, resulting in an
increase in the numbers of SRI adopters. As the country is now moving towards a more

liberalized economic system with an improvement in rice distribution, farmers who can

46

find better opportunities by selling their labor would reduce labor use on their own rice
production, thus leading to a SRI disadoption. This situation would affect small farms
more than medium or large farms because using hired labor is not a rational option for
them, giving that their ratio of hired labor MVP over wage rate is already less than unity.
A second important finding concerning labor MVP is that hired labor presents a
beneficial marginal return for medium (MVP = 1217 ariary per day) and large farms
(MVP = 1650 ariary per day), but not for small farms. Approximately, the ratio of MVP
over wage rates are 1.6 and 2.2 respectively for medium and large farms. Hence, they
should be better of by using more hired labor for rice production. However, they might be
constrained by the non-existence of a well functioning credit market. As hired labor
requires cash in a period where resources are scarce (IF PRI, 1997; SECALINE, 1997),
the existence of a well-functioning credit market constitutes a necessary condition for
farmers. We estimate that an interest rate of 21 percent30 applied on the current wage rate
will raise agricultural wage rate to 920 ariary later in the year. This results in reducing the
relative ratio of marginal return over wage rate for medium and large farms to
respectively 1.2 and 1.8. The break-even for medium farms will be an interest rate of 60
percent during the agricultural season. Large farm would continue to obtain positive net
return (marginal return — factor price) so long the interest rate stays below 117 percent.
At this level of interest rate, farmers would no longer have an incentive to use additional

hired labor.

 

3° Currently, bank and micro-finance organization practice an interest rate varying from 2.5 to 3 percent per
month. This corresponds approximately to a 21 percent interest during the rice season.

47

Also, labor may be simply unavailable when needed. In many agricultural
settings, the MVP becomes very high in peak labor demand period such as for weeding

and transplanting. In that case, they clearly exceed prevailing wages.

Other factors MVP

Chemical fertilizer is shown to have a higher marginal return for small farms than
for large farms. This is consistent with the findings from Bemier and Dorosh (1993). This
might indicate that small farms have better agronomic practices such as timely fertilizer
application, more homogenous and timely weeding, which reduces the competition
between weeds and rice. These conditions improve the effect of chemical fertilizer on
rice production. This result also supports the idea that small farms have lower
vulnerability from natural risks, which allows fertilizer to have higher MVP. For
example, Table 1 shows that small and medium farms have lower inundation risk than
large farms. Consequently, they would have a higher expected marginal return for
fertilizer. Fertilizer has less effect on flooded plots because of the bigger problem with
leaching.

Other factors show a difference in return between small and large farms. Such
results are useful for poverty reduction strategies. For example, access to extension
services and to education offer higher absolute and relative returns for small compared to
large farms.

The differential effects of natural conditions and land quality on different farm
size indicate that large farms face higher risks than smaller farms. For example,

inundation decrease rice production by 1 percent for small farms while the losses may

48

reach 14 percent for large farms. As this variable should be stochastic, the sole
explanation seem to come from the difference in the plot location. Plots located in the
lowlands are flooded longer than plots in higher areas.

Lastly, RST would give a higher return to small than large farms. It is expected to
increase rice production by 43 percent for small farms. This might be the result of timely

and homogeneous agricultural practices combined with better land quality.

3.4. Incremental Land and Labor MVP of Productivity Intervention

In this section, we will analyze the change in land and labor MVP when farmers
choose to move from a given level of technology to another. Improved Rice Seedling
Transplantation (RST) is a good example of an alternative technology for increasing land
productivity, using local production factors. It also constitutes a feasible alternative to the
labor-intensive SR1 technology.

The percentage of improved RST adopters in Table 5 indicates clearly that
farmers adopt different technologies on different plots. It seems that farmers take soil
characteristics into consideration and act accordingly. For example, there are fewer
farmers who adopt IRST on flooded soil, which will have a high risk of submerging

young plants.

49

Table 5 — Incremental Land and Labor MVP from
Adoption of Improved Rice Seedling Transplantation Technology (RST)

563 fields, Madagascar Highland 2000

 

 

 

 

 

State of land Percent of Improved Ratio MVP Traditional Ratio MVP Incremental
quality and natural adopters RST over factor RST over factor change from
shocks price price without to
with RST
Land
Rainfed 6% 4639 1-8 3786 1-5 23%
Irrigated 7% 4593 1-3 3438 1.3 34%
Low 3% 3938 1-5 3336 1.3 18%
Terrace 4% 6799 2.6 4148 1.6 64%
Sandy 5% 3627 l-4 3455 1.3 5%
Clay 9% 5212 2.0 3715 1.4 40%
Non-flooded 9% 4753 1.8 3930 1.5 21%
Flooded 3% 4143 1.6 3337 1.3 24%
Family labor
Rainfed 6% 762 1-0 702 -9 9%
Irrigated 7% 570 -3 81 l 1-1 -30%
Low 8% 1050 1.4 969 1.3 8%
Terrace 4% -419 --6 391 -5 -207%
Sandy 5% 1321 1-7 981 1-3 35%
Clay 9% 224 .3 643 -8 -65%
Non-flooded 9% 763 1-0 714 .9 7%
Flooded 3% 509 -7 776 1-0 -34%
Hired labor
Rainfed 6% 13 74 1.8 1296 1.7 6%
Irrigated 7% 1150 1‘5 1275 1.7 ~10%
Low 8% 1583 2.1 1407 1-9 12%
Terrace 4% 561 ~7 1075 1.4 -48%
Sandy 5% 2008 2-6 1526 2-0 32%
Clay 9% 888 1.2 1208 1.6 -26%
Non-flooded 9% 1366 1.8 1489 2.0 -8%
Flooded 3% 798 1.1 1065 1.4 -25%

 

Source: Agricultural Production Survey, Madagascar Highland June 2000

 

Table 5 presents the results of the change in land and labor MVP due to farmer’s

choice to adopt or not the improved rice seedling transplantation technology. All land

50

MVPs show a significant increase ranging from 5 to 64 percent. Obviously, if the
increase of land MVP is the main objective, the conclusion is that all farmers should
adopt improved rice seedling transplantation. The overall result indicates that this
technology seems to overcome bad land quality and high risk from natural shocks. To
achieve national self-sufficiency of its staple food, this technology might be a feasible
alternative. It is therefore not surprising that in the mid-1990's, the Government
propagated this technology as a national strategy to increase rice production.

However, adoption of this RST technology seems to affect other production
factors. Family labor MVP is affected negatively. We are also facing a significant
decrease in the MVPs for certain situations such as on terraced, irrigated, clay textured,
and flooded land. These cases are also associated with low marginal productivity of
family labor, showing often a ratio MVP over factor price of labor less than unity. For
plots in terraces, the MVP is even negative, which might be interpreted as follow: the
values of the increase in the production do not offset the costs of the increase in the
number of family labor. More generally, the results show that the return to family labor
from an RST adoption is not beneficial for farmers who have irrigated land. However,
these farmers are the target population for this technology as it requires better water
management control. The RST technology requires not as much a large quantity of
irrigation water but more so a very high quality of the irrigation system, as it should
allow the plot be watered or drained at any period.

The results for hired labor follow a similar pattern as for family labor. However,
the magnitude of changes is smaller. The sole exception is that IRST increases

significantly the family and hired labor MVP on sandy soil. This might be the effect of

51

the combination of the special water management system for IRST and the relative ease
of working on sandy soil. On top of IRST, certain farmers apply the SR1 technology,
which is based on the assumption that rice is not an aquatic plant, thus there is no need to
maintain a film of water above the rice field. The SRI technology requires that the best
management is to treat rice like other plants, and alternate dry and humid soil without
flooding the root. In this case where permanent water is not needed, sandy soil that
requires less labor force because of its texture will give higher output and might value
labor more.

In general, the positive but low marginal return and the decrease in the marginal
return to labor might explain the relative disappointing level of adoption of improved rice
seedling transplantation. However, this is not the only reason. The constraint of available
cash during the beginning of the agricultural season is also important. Improved RST
requires the need of large amount of labor during a peak period, often situated in the lean
season. Thus labor and credit market imperfections would play an important role in the
relative disappointing level of adoption of IRST.

Nonetheless, the results show that using hired labor is a good choice in many
cases as the marginal returns are staying above the factor price. Table 5 indicates that
with the exception of terraced and flooded soil, one can have a ratio MVP over factor

price of hired labor greater than 1.

52

Conclusions

This thesis analyzes the effect of land quality and natural shocks on rice
productivity in Madagascar. It is shown that these variables do have non-neutral impacts
on rice productivity and that they are correlated with some of farmer’s choice variables. It
is shown that the omission of land quality and natural shocks in analytical models leads
to a bias in the marginal physical product of land, labor, and other factor of production,
which would result in misleading agricultural policy-makers about the productivity of
these factors.

As findings in other developing countries studies where many agricultural tasks
are still manually performed have shown, small farms allocate more labor per unit of
land. This allows them to reach higher land productivity. However, this is usually
associated with lower labor productivity, especially for family labor. The consequence
might be less interest to adopt labor-saving new technologies, unless imperfections in the
local labor market and in rice markets are solved, as these might put them in an insecure
food situation.

Medium and large smallholders show a high return from hired labor. This makes
them a potential target for labor-intensive technology, under the condition that there is a
well-fimctioning and financially accessible credit market. In general, results from the
comparison between tercile of farm size appear to be important in targeting different
households who might be interested or not in a given technology. If farmers were
considered as customers for various technologies, then a market segmentation approach

would be a better strategy for extension service delivery.

53

More generally, all farms are experiencing a low family labor marginal value
product. Increasing labor productivity could involve two types of related policies. First is
the improvement policy, which consists in improving MVP by reducing labor input. This
would lead to a higher labor MVP, but might also lead to lower land MVP because of the
reduction of the input use. One can imagine a movement along the production curve,
toward the origin. For a profit maximizing manager viewpoint, the optimal point, where
the average product maximizes the return to labor, would be the one where labor MVP is
equal to the implicit wage rate.

This might be solved practically by identifying and promoting small tools that
allow farmers to reduce the amount of labor needed per land unit, without banning the
quality of the labor while maintaining the level of land productivity. A good illustration
for this is the use of mechanical tools for weeding given that this task occupies almost 21
percent of total labor needed by farmers. In practice, farmers are aware of this situation of
low productivity of family labor, thus they already modified the SRI technology to
optimize their labor input use. Various other practices by farmers illustrate this behavior.
For example, they use older plants instead of a 8-day old plants in order to reduce labor
for the transplantation. One also notes a reduction of the number of weeding from 5 to 3;
or a more flexible water management, all of this in order to reduce the quantity of labor
needed.

A second policy option is the transformation approach. This consists of a more
structural change, led by agricultural research. The objective here is to improve the
responsiveness of rice from the use of different inputs, in order to have higher

productivity and result in an increase in labor MPP. One can imagine here a non-parallel

upward shift of the production curve, corresponding to higher output from the same level
or even reduced inputs. This approach involves substantial changes and relies more on
the efficiency of the agricultural research system. Of course, a combination of these two
approaches would solve more efficiently the low labor productivity issues.

Lastly, credit policy should not be targeted only in the supply of chemical inputs
or seeds, usually seen in contracting financial scheme. We demonstrated that the use of
hired labor, which means a need for cash, constitutes a good investment for medium and
large farms. The use of improved rice seedling transplantation technology appears to be
promising in many ways. It allows an increase of land MVP under any state of land
quality and natural shocks. However, the effect on labor productivity, both family and
hired, seems to be negative, resulting in a moderate level of adoption of this technology.
This is however a good alternative for farmers having low opportunity costs of labor and
facing high rice supply insecurity.

A potential extension of this thesis might be the analysis of the variance effects of
different factors of production. This research would contribute to the reason why farmers
do not adopt a given technology. Also, to better assess initial land quality, it is suggested
to have some variables that control for the initial endowment of the plot. It is indeed
interesting to know what decision rule farmers apply in allocating fertilizer on their plots.
Do they apply fertilizer on good quality plot or is the reasoning vice-versa? The results
will change the estimated effect of fertilizer application on the production level because

of the effect of previous fertilizer application.

55

Appendix 1 — Madagascar - Mapping of Surveys Sites

Highland defined in the Current study

IFPRI (1997) study

 
   
   
   
   
   
 
   
   

Plateau Central defined in
Bemier and Dorosh (1993) study

 

Lac Alaotra

Antananarivo

Vakinankaratra

N
Fianarantsoa T
Mozambique .
I d
Channel See]:

56

Appendix 2 - Mackinnon White Davidson (MWD) Test For Non-Nested Models

The MWD test consists in identifying the “best among the tested” functional form
for the data. It is used for non-nested models that have different dependent variables
measurement units. The test involves the following steps (Gujarati, 1995):

1. Estimate model 1 and compute the estimated value mhatl
Estimate model 2 and compute the estimated value mhat2
Convert mhatl in model 2 dependent variable units, let say we have mhatlb
Take the difference between mhatlb — mhat2, let say mhat
Estimate model 2 with adding mhat in the RHS
Using asymptotic t-test, test mhat

9‘99?!"

a. If mhat estimates = 0 -) Reject model 1
b. If mhat estimates not equal to zero -) cannot reject model 1

7. Repeat in the opposite way to test for model 2

1 - MWD Test between Square root generalized Leontief (GL) and Translog

 

 

Coefficient t-value p-value
GL/Translog 0.64 3.54 0.000 GL is a valid functional form
Significant at 1% level
Translog/GL 0.63 3.10 0.003 Translog is a less valid functional form

Even significant at 1% level

 

2 — MWD Test between Square root genegrlized Leontief (GL) and Ouflratic

 

 

Coefficient Stand. Err. p-value
GL/Quadratic 1.33 11.88 0.000 GL is a valid functional form
Significant at 1% level
Quadratic/GL 0.31 3.25 0.001 Quadratic is a less valid functional
form

Even significant at 1% level

 

57

Appendix 3 - Correlation between choice, conditioners and land quality — natural

shocks variables

 

 

 

Location
Irrigation topography Soil texture Inundation Drought

Choice variables

Plot area in ares .040 .029 .067 .068* .085*
Family labor in days -.097"‘ .061 -.030 .056 -.007
Hired labor in days .006 .024 .082* .013 .046
Child labor in days .014 .141“ .001 -.018 .021
Chemical fertilizer in kg .030 .020 .046 -.062 -.090
Number of oxen .023 .030 .037 -. 181 "‘ .098“
Rice seedling transplantation in days .006 -.072* -.072"‘ -.1 13* -.099*
Conditioner variables

Extension contact = 1 .032 .023 .052 -.031 .155*
Region Vakinankaratra = l .047 -.074" .087" -.117* .158“
Season Vary aloha = 1 .030 -.006 .071 "‘ -.037 -.068
Education of household head .034 -.086* .013 -.123* .1 16*
Number of plots .040 -.029 .013 .450* .063

 

"‘ Significant at least at 10 percent level

Source: Agricultural Production Survey, Madagascar Highland June 2000

58

Appendix 4 - Wald Test results

The F statistic for testing a set of linear restrictions is no longer distributed as F

because neither the enumerator nor the denominator has the necessary chi-square

distribution. However, the Wald statistic JF [J , n — k] does have a chi-square distribution

asymptotically and can be used instead (Greene, 1993). ‘Wald test’ tests linear hypothesis

about the estimated parameters. The significance means that the regressors jointly explain

a significant amount of variation in the production level.

 

 

 

 

Variables t-value p-value Significance
level
Plot area F(18, 470) 35.07 .000 ""'
Family labor F(18, 470) 8.02 .000 ”*
Hired labor F(18, 470) l 1.58 .000 ""'
Child labor F(18, 470) 1.94 .012 "”"
Fertilizer F(18, 470) 3.88 .000 "*
Draught oxen use F (6, 470) 1.04 .396
Dummy Vakinankaratra F(6, 470) 20.52 .000 **“
Acces to extension service F(6, 470) 1.20 .306
Dummy season F(6, 470) 5.10 .000 “*
Head education F(6, 470) 2.38 .028 "
Land tenure F(6, 470) 1.51 .170
Rice seedling transplantation F(6, 470) l 1.85 .000 ""'
Irrigation F(6, 470) 3.25 .004 ***
Topography location F(6, 470) 3.93 .000 ""‘
Soil texture F(6, 470) 1.93 .074 "'
Inundation F(6, 470) 2.53 .020 ”
Drought F(6, 470) 3.69 .001 "“

 

”', ", "" are respectively significant at 10, 5, and 1% level

59

Appendix 5 - Bootstrap Method and t-test between two means

Difference significance is based on a t-test between the two means of MPP from
the full and the short specification. Because we computed the MPP from the sample mean
on each regression result, there is no variation on the values of MPP, thus the variances
would be estimated by bootstrapping the distribution of MPP.

This method consists in drawing randomly lots of replicates (50 in our cases) of
the original data -- with replacement -- and then computing the MPP for each of those
random samples. We can take the variances of MPP from the hundreds resulting MPPs
and use t-test to look at the significant difference between the two means.

This is easily done with STATA using the following syntax:

bs "regress y x1 x2 x3" "MPP = ay/axi", reps(50)

Results of t-test between the MPP from full and short specification

 

 

 

Full specification Short specification

Mean Std. Err. Mean Std. Err. I t-value p-value
Plot area 13.49 .89 13.45 .80 .236 .814
Family labor 2.70 .81 2.95 .64 -l.712 .090
Hired labor 4.69 .47 4.76 .70 -.587 .558
Child labor -4.32 6.13 -1.24 5.50 -2.644 .009
Number of oxen 6.66 3.68 12.25 3.40 -7.889 .000
Land tenure 291 250 241 292 .918 .360
Extension access 207 163 360 138 -5.050 .000
Season "aloha" -1 1 1 227 26 198 -3.221 .002
Region Vakinankaratra 231 37 241 39 -1.246 .216
Improved RST 391 851 109 723 1.76 .077
Chemical fertilizer use 396 327 414 453 -.224 .823
Education ofhead 1.39 3.59 2.79 2.85 -2. 159 .033
Number of plots -5.43 7.33 -6.46 5.83 .777 .489

 

60

REFERENCES

Anim, F ., "A Note on the Adoption of Soil Conservation Measures in the Northern
Province of South Africa", Journal of Agricultural economics 50(1999), 2, 336-345.

Antle J. M., Capalbo S.M. “Chapter 2 - Agricultural productivity measurement and
explanation.”, Measuring Agricultural Productivity, 1988.

Bemier R., Dorosh P. "Constraint on Rice Production in Madagascar: The Farmer’s
Perspective", Cornell Food and Nutrition Policy Program, 1993.

Bhalla S. S. “Does land quality matter? Theory and measurement” Journal of
Development Economics: 29 (1988) : 43-62.

Bhalla S.S., Roy P. “Mis-specification in Farm Productivity Analysis: The Role of Land
Quality”. Oxford Economic Papers 40(1988), 55 — 73.

Brown J ., Salkin J ., “Underemployment in Rural South Vietnam: A Comment and
Discussion of Family Labor”, Economic Development and Cultural Change 23 N01, 571
- 574.

Byiringiro F. E. Determinants of F arm Productivity and the Size-Productivity
Relationship under Land Constraints - The Case of Rwanda. Ms. Thesis, MSU, 1995.

Christensen L. R.. “Concepts and Measurement of agricultural Productivity,” American
Journal of agricultural Economics, 57(1975)5: 910 — 915.

Clay D., Byiringiro F ., Kangasnierni J ., Reardon T., Sibomana B., Uwamariya L.,
Promoting food security in Rwanda through sustainable agricultural productivity:
Meeting the challenges of population pressure, land degradation and poverty. Staff paper
No 95-08, Department of Agricultural Economics, Michigan State University, 1995.

Coelli T., Prasada D.S, Battese G. An Introduction to efliciency and Productivity
Analysis, Kluwer Academic Publishers, 1998

Debertin D. L. Agricultural Production Economics. Macmillan Publishing Company :
1992

Deolalikar AB. “The inverse relationship between productivity and farm size: a test
using regional data from India.” American Journal of Agricultural Economics, May
1981: 275-279

Deolalikar, A. and Vij verberg W. “The Heterogeneity of Family and Hired Labor in
Agricultural Production: A Test Using District-Level Data from India.” Journal of
Economic Development: 8(1983) 45 — 69.

Deolalikar, A. and Vijverberg W. “A test of Heterogeneity of Family and Hired Labor in
Asian Agriculture.” Oxford Bulletin of Economics and Statistics: 49(1987) 291 - 305.

Dorosh P., Haggblade S., Rajemison S., Ralantoarilolona H., Sirnler K., Structure et
F acteurs Déterminants de la Pauvreté ti Madagascar, IN STAT, 1998.

FAO. World Agriculture: Towards 2010, An FAO Study. Alexandratos Eds, 2000

61

Feder G., “The Relation Between Farm Size and Farm Productivity. The role of family

labor, supervision, and credit constraints”. Journal of Development Economics, 18(1985)
297 - 313. -

Feder G., F eeny D., "Land Tenure and Property Rights: Theory ans Implications for
Development Policy. "World Bank Economic Review, 5(1991) 135-153.

Frisvold G.B. “Does Supervision Matter? Some Hypothesis Tests using Indian F arm-
level Data.” Journal of Development Economics: 43 (1994) 217-238.

Goletti F ., Randrianarisoa C., Rich K. How Good Is Rice Seed in Madagascar? The
Structure and Performance of the Seed Sector, Part 2, Final Report, International Food
Policy Research Institute - F OFIFA project, 1997, 15 - 71.

Greene, W. H. — “Chapter 7: Hypothesis Tests with the Multiple regression Model”,
Econometric Analysis, New York, NY: Macmillan Publishing Company, 1993

Griffin R.C, Montgomery J.M, Rister M.E. “Selecting Functional Form in Production
function Analysis.” Western Journal of Agriculture Economics 12(Dec 1987): 216-227.

Gujarati D.,. Basic Econometrics, Third Edition, McGraw-Hill, Inc, Chapter 8., 1995.

Hossain M., Nature and Impact of the Green Revolution in Bangladesh, Research Report
No 67, Washington DC, International Food policy Research Institute, 1988

IFPRI - FOFIF A. Structure and Conduct of Major Agricultural Input and Output Markets
and Response to Reforms by Rural Households in Madagascar, Part 1, Final Report,
International Food Policy Research Institute - FOFIFA project, 1997.

Just R., Pope R., “Stochastic Specification of Production Functions and Economic
Implications”, Journal of Econometrics, 7(1978) 67 — 86.

Lau L.J. “Profit Functions of Technologies with Multiple Inputs and Outputs,”. Review of
Economic Statistics 54(1972)3: 281 — 289.

Lau L., J. “Functional Forms in econometric Model Building”, Handbook of
Econometrics, Volume 111. Z. Griliches and MD. Intriligator Eds, 1986.

Lutz E., Pagiola S., Reiche C., "The Costs and Benefits of Soil Conservation, the
Farmer's Viewpoint". The World Bank Research Observer, 9(1994) 2, 273 - 295.

Maddala G.S., “Generalized Least Squares with an Estimated Variance Covariance
Matrix”, Econometrica, Volume 39, Issue 1(1971), 23 — 33.

Minten B., Zeller M., Randrianarisoa C. Determinants of A gricultural Production, Farm
Income, and Marketed Surplus in Madagascar. Part 4 — Final Report, International Food
Policy Research Institute - FOFIFA project, 1997.

Minten B., Zeller M., Randrianarisoa C., “Chapter 6 - Factor Use and Agricultural
Productivity”, Beyond Market Liberalization: Welfare, Income Generation and
Environmental sustainability in Rural Madagascar, Bart Minten and Manfred Zeller Eds,
Ashgate, 2000.

62

Nerlove M, Vosti S., Basel W. "Role of F arm-Level Diversification in the Adoption of
Modern Technology in Brazil". Research Report 104, Washington D. C. International
Food Policy Research Institute, 1996.

Place E, Hazell P., "Productivity Effects of Indigenous Tenure Systems in Sub Saharan
Afiica", American Journal of Agricultural Economics, 75(1993), 10 - 19.

Platteau J. P., "The Evolutionary Theory of Land Rights as Applied to Sub Saharan
Africa: A Critical Assessment", Development and Change, 27(1996), 29 - 86.

Rao V., Chotigrat T., “The Inverse Relationship Between Size of Land Holding and
Agricultural Productivity”, American Journal of Agricultural Economics 63, 3(1981) 571
-574

Reardon T., Kelly V., Crawford E., Jayne T., Savadogo K., Clay D., Determinants of
Farm Productivity in Africa: A synthesis of F our Case Studies. MSU International
Development Paper No 22, Michigan State University, 1996

Roubaud F., “La Question Rizicole a Madagascar: Les Résultats d’une Décennie de
Libéralisation”, Economic de Madagascar, vol.2 (1997) 37 - 62.

Savadogo K. T., Reardon T., Pietola K. Farm productivity, Technology Choice, and Non-
F arm Income in the Sahel. Staff Paper No 94-54, Michigan State University, Department
of Agricultural Economics, 1994.

Secaline, Projet Sécurité Alimentaire et Nutrition Elargie, La Situation Alimentaire et
Nutritionelle d Madagascar, Stratégie Nationale de Sécurité A limentaire et de Nutrition,
Gouvemement Malagasy IDA, MAG, No 2474, 1997.

Sherlund S.M., Barrett C. B., Adesina A.A. Smallholder Technical Efliciency with
Stochastic Exogenous Production Conditions, Forthcoming, 2001.

Shively G.E. “Modeling Impacts of Soil Conservation on Productivity and Yield
Variability: evidence from a Heteroskedastic Switching Regression”, American Journal
of Agricultural Economics :Selected paper at annual meeting, 2-5 August 1998.

SMTIS, Résultats du Recensement Agricole, République de Madagascar, Ministére de
l’Agriculture, Service de Methodologies et Traitements des Informations Statistiques,
1989

Tshibaka T., Food Production in a Land-Surplus, Labor-Scarce economy: The Zairian
Basin, Research Report No 74, Washington DC. International Food Policy Research
Institute, 1999.

UPDR, Diagnostic et Perspectives de la F iliére Riz ti Madagascar, République de
Madagascar, Ministere de l’Agriculture, Draft version, octobre 2000

Upton M. The Economics of Tropical Farming Systems, Cambridge University Press,
1996.

USDA, Food Security Assessment, lntemational Agriculture And Trade Reports,
December 1999.

63

Valerie K., Diagana B., Reardon T., Gaye M., Crawford E., Cash Crop and F oodgrain
Productivity in Senegal: Historical View, New Survey Evidence, and policy Implications,
MSU lntemational Development Paper No 20, 1996.

Wooldridge J. M., Introductory Econometrics, A Modern Approach, South-Westem
College Publishing, 2000

Zeller M., Lapenu C. The critical Triangle of Rural Development — The Case of
Madagascar. Part 6 - Final Report, lntemational Food Policy Research Institute -
FOFIFA project, 1997.

64

Will 1 11111111111

93 024 3394

1