INSTITUTIONS AND INCENTIVES FOR THE EFFICIENT USE OF ENERGY

By

Dylan Brewer

A DISSERTATION

Michigan State University

in partial fulfillment of the requirements

Submitted to

for the degree of

Economics – Doctor of Philosophy

Environmental Science and Policy – Dual Major

2019

ABSTRACT

INSTITUTIONS AND INCENTIVES FOR THE EFFICIENT USE OF ENERGY

By

Dylan Brewer

Chapter 1: Moral Hazard and Equilibrium Sorting in Residential Energy Markets
This chapter studies tenant behavior in rental housing when the landlord pays for heating. I develop
a model in which renters have heterogeneous preferences for home size and indoor temperature.
When energy is costly, renters choose smaller apartments and turn down the heat—or sort into
apartments with landlord-pay energy bills. I estimate the model using exogenous variation in energy
prices and use a machine-learning algorithm to explore preference heterogeneity. Surprisingly, I
find that renters who prefer hotter temperatures do not systematically choose landlord-pay units,
though I am unable to rule out sorting on preferences for unobserved home attributes. Eliminating
moral hazard by forcing all renters to pay their own bill reduces energy consumption by 25% due to
renters turning down the heat (22%) and choosing smaller units (3%). Moral hazard in residential
energy contracts cost the United States $839 million per year in welfare losses including $238
million from carbon emissions.

Chapter 2: The Remarkably Inelastic Demand for Home Heating Services: A Choice Exper-
iment
I conduct a stated-choice experiment that presents research subjects with hypothetical costs to adjust
their thermostats. I estimate responses to the cost of heating and analyze the causes for heterogeneity
in household demand for energy services, using the experimental results as a complete-information
baseline. I find that even at the highest price level, half of the participants exhibit zero response
to price. Further, I find that participants’ experimental behavior with complete information can
explain the full range of observed real-world temperature settings, suggesting a limited role for
informational barriers or price salience issues in energy-service demand heterogeneity. Individuals
with higher temperature preferences are more price responsive, suggesting that increasing block

pricing for energy may reduce energy consumption while minimizing the regressivity of the energy
pricing program.

Chapter 3: Negawatts v. Megawatts: Demand Response in Wholesale Electricity Markets
(with Katherine Wagner)
We study the Federal Energy Regulatory Commission’s (FERC) wholesale demand-response pro-
gram.
In this program, participants sell a reduction in electricity use from a baseline on the
wholesale electricity market as if it was electricity generation. We show that demand-response par-
ticipants have different incentives to consume electricity based on whether they participate directly,
through a third-party aggregator, or through a utility aggregator. We argue that the current whole-
sale demand-response program is inefficient and should be replaced by retail demand-response
programs run by utilities.

ACKNOWLEDGEMENTS

I thank Soren Anderson for his careful mentorship, friendship, and genuine interest in my well-
being. I would not be the same economist without Soren’s advice and guidance. In addition, I
thank Joseph Herriges, Kyoo il Kim, and Scott Swinton for serving on my committee, reading my
work, and passing on their wisdom. I would also like to thank Steven Haider for his professional
mentorship, Cody Orr for many brilliant insights, and Michael Conlin for getting me to go the extra
mile. Special thanks to Lee Coppock for introducing me to economics and inspiring this journey.
My time spent at the Property and Environment Research Center in Bozeman, Montana was
critical to the development of this thesis. I would like to thank the many individuals that provided
feedback on my research while I was there: Randal Rucker, V. Kerry Smith, Timothy Fitzgerald,
Maximilian Auffhammer, Matthew Kahn, Roger Meiners, and Ryan Kellogg. I also want to thank
the office staff at PERC, in particular Sharie Rucker and Dianna Reinhart, for their excellent support.
The Environmental Science and Policy Program at Michigan State University provided funding
for the choice experiment in the thesis. I would like to thank Joseph Hamm for his help in the
experimental design. I am grateful for fellowship support from the Department of Economics at
Michigan State University, the Graduate School at Michigan State University, the Environmental
Science and Policy Program at Michigan State University, the Institute for Humane Studies, and
the Property and Environment Research Center while earning my PhD. I appreciate the office staff
of the Department of Economics, and in particular Lori Jean Nichols, for their excellent support of
my endeavors while at Michigan State.

Thank you to Katherine Wagner for her important work as co-author of Chapter 3. I am grateful
for comments by seminar participants at Michigan State University, the CU Environmental & Re-
source Economics Workshop, Midwest Energy Fest (Northwestern), Energy Camp at Haas, Camp
Resources (NC State), the Midwest Economics Association Annual Meeting, the Heartland Work-
shop (University of Illinois-Urbana Champaign), Davidson College, James Madison University,
Wake Forest University, the University of Georgia, and Georgia Tech.

iv

Finally, thank you to my family for their support, advice, and love. I especially want to thank my
future wife Amanda for being my best friend, supporting my dreams, reading my drafts, becoming
an honorary economist, and moving with me to Georgia. We did it!

TABLE OF CONTENTS

LIST OF TABLES .
LIST OF FIGURES .

.
.

.
.

.
.

.
.

.
.

.
.

.
.

.
.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viii
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

ix

CHAPTER 1 MORAL HAZARD AND EQUILIBRIUM SORTING IN RESIDENTIAL

.

.

.

.

.

.

. .

Introduction .

ENERGY MARKETS . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1
1
1.1
. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
5
1.2 Data and descriptive evidence of moral hazard . . . . . . . . . . . . . . . . . . .
.
7
1.3 Modeling and empirical strategy . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7
1.3.1 General model
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.3.2 Empirical model
1.3.3
Parameter identification . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.3.4 Empirical model strengths and limitations . . . . . . . . . . . . . . . . . . 15
. . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.4.1 Heat costs .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.4.2 Temperature disutility and bliss points . . . . . . . . . . . . . . . . . . . . 20
Exploring heterogeneity with LASSO . . . . . . . . . . . . . . . 22
. . . . . . . . . . . . . . . . 23
. 25
. 26

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

1.4.3 Hedonic prices and preferences for attributes

1.5 Simulation results .
1.6 Conclusions .
.

1.4 Estimation of model parameters

1.4.2.1

. .

. .

.
.

.
.

.
.

.
.

.

CHAPTER 2 THE REMARKABLY INELASTIC DEMAND FOR HOME ENERGY

.

.

.

. .

Introduction .

SERVICES: A CHOICE EXPERIMENT . . . . . . . . . . . . . . . . . . . 29
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.1
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.2 Experimental procedure .
2.3 Data .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
.
2.4 Estimating demand for heating . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.5 Heterogeneity analysis
2.6 Conclusion .
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

. .

. .

. .

. .

.
.

.
.

.

.

.

.

.

CHAPTER 3 NEGAWATTS VS. MEGAWATTS: DEMAND RESPONSE IN WHOLE-

.

.

.

.
.

. .

Introduction .
.
Institutional background .

3.1
3.2
3.3 A model of demand response incentives

SALE ELECTRICITY MARKETS (WITH KATHERINE WAGNER) . . . . 39
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
. . . . . . . . . . . . . . . . . . . . . . . 44
3.3.1 Direct retail consumer participation . . . . . . . . . . . . . . . . . . . . . 47
. . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.3.2 Third-party aggregator
3.3.3 Utility aggregator .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.3.4 Comparing total surplus in the linear case . . . . . . . . . . . . . . . . . . 50
3.4 Can FERC improve demand response? . . . . . . . . . . . . . . . . . . . . . . . . 54
3.5 Conclusion .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

. .

.

.

.

.

.

vi

APPENDICES .

.

.

.

.

.

.

.

.

.

.

.

.

.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
APPENDIX A FIGURES AND TABLES FOR CHAPTER 1 . . . . . . . . . . . . . . 57
APPENDIX B SUPPLEMENTAL INFORMATION FOR CHAPTER 1 . . . . . . . . 78
APPENDIX C FIGURES AND TABLES FOR CHAPTER 2 . . . . . . . . . . . . . . 81
APPENDIX D FIGURES AND TABLES FOR CHAPTER 3 . . . . . . . . . . . . . . 87
. 92

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

.

.

.

.

.

BIBLIOGRAPHY .

vii

LIST OF TABLES

Table A.1: Residential Energy Consumption Survey sample means . . . . . . . . . . . . . . 70

Table A.2: American Housing Survey sample means . . . . . . . . . . . . . . . . . . . . . 71

Table A.3: Hedonic price estimates using AHS data . . . . . . . . . . . . . . . . . . . . . . 72

Table A.4: Fuel use estimates using RECS data . . . . . . . . . . . . . . . . . . . . . . . . 73

Table A.5: Average marginal effects on monthly heating costs using RECS data . . . . . . . 74

Table A.6: Temperature setting estimation using RECS data . . . . . . . . . . . . . . . . . 75

Table A.7: Temperature setting estimation using choice experiment

. . . . . . . . . . . . . 76

Table A.8: LASSO heterogeneity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76

Table A.9: Simulation results: Effect of changing all landlord-pay contracts to tenant-pay . . 77

Table C.1: Participant sample mean and standard deviation . . . . . . . . . . . . . . . . . . 86

Table C.2: Estimated semi-elasticities. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86

viii

LIST OF FIGURES

Figure A.1: Average temperature setting by year compared between landlord-pay and
tenant-pay regimes. The tenant-pay temperature settings statistically vary
between sample years, possibly due to changes in energy prices and resulting
behaviors. This figure does not include households that turn off the heat
completely: 28 (3.1%) landlord-pay households and 244 (6.8%) tenant-pay
households.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

. .

. .

.

Figure A.2: Average temperature setting when out of the house between landlord-pay
and tenant-pay regimes. This figure does not include households that turn
off the heat completely: 28 (3.1%) landlord-pay housholds and 244 (6.8%)
tenant-pay households.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58

Figure A.3: Time series variation in the average ratio of electricity to natural gas price per

BTU. .

.

.

.

.

.

.

.

.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

Figure A.4: Plots of estimated average markup for having the landlord-pay heating ∆p
∆Rj
by census division from the preferred correlated-random-effects regression
in table A.3. 95 percent confidence intervals are derived using cluster-robust
standard errors.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60

.

.

.

Figure A.5: Plots of estimated average hedonic price for 100 square feet ∂ ˆp
∂xj

by census
division from the preferred correlated-random-effects regression in table A.3.
95 percent confidence intervals are derived using cluster-robust standard errors.

61

Figure A.6: Density of bliss point temperature settings. Does not include 11 (1.2%)
households that did not use heat in the winter. Plotted against a normal
distribution.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

. .

. .

Figure A.7: The LASSO-predicted temperature settings when price is zero (bliss point)
for landlord-pay and tenant-pay households. Landlord-pay households are
predicted as having lower bliss point temperature preferences. . . . . . . . . . . 63

Figure A.8: The implied temperature disutility parameters β2,i estimated from the LASSO

regression.

.

.

. .

. .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64

Figure A.9: The LASSO-predicted effect on temperature setting of requiring landlord-pay

households pay their own heating bills.

. . . . . . . . . . . . . . . . . . . . . . 65

Figure A.10:The LASSO-predicted effect on temperature setting of moving tenant-pay

households to a landlord-pay regime.

. . . . . . . . . . . . . . . . . . . . . . . 66

ix

Figure A.11:The predicted temperature setting using LASSO versus the observed temper-

ature setting used to train the algorithm.

. . . . . . . . . . . . . . . . . . . . . 67

Figure A.12:Estimated bliss points using the preferred β2 = −0.79 by regime.
Figure A.13:Kernel density of the distribution of estimated preferences for square feet by

. . . . . . . . 68

regime. .

.

.

.

.

.

.

.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69

Figure C.1: Kernel density of participants’ bliss point temperature preferences and real
temperature settings. A Kolmogorov-Smirnov test of equivalence of distri-
butions easily rejects the null hypothesis that the distributions are equal.

. . . . 81

Figure C.2: Scatterplot of bliss point temperature preferences and actual temperature
settings with a 45 degree line for reference. 54 percent of respondents set the
thermostat equal to the bliss point. 6.5 percent reported setting the thermostat
greater than the bliss point. Random noise has been added to the data to show
clustering on common temperature choices such as 70 degrees Fahrenheit.

. . . 82

Figure C.3: Histogram of estimated arc semi-elasticities (η).

. . . . . . . . . . . . . . . . . 83

Figure C.4: Histogram of semi-elasticities (η) estimated using OLS.

. . . . . . . . . . . . . 84

Figure C.5: The marginal effects from a Tobit estimation of the estimated arc semielas-
ticities on average price and participant demographics. The marginal effects
from this estimation can be interpreted as the change in arc semi-elasticity for
a one-standard-deviation change in the predictor variable holding the other
predictor variables constant. 95 percent confidence intervals are bootstrapped
using 1000 replications with sampling at the participant level. . . . . . . . . . . 85

Figure D.1: Aggregate supply and demand curves for electricity consumption. Area 1
corresponds to the surplus loss in the no-demand-response baseline and area
2 corresponds to the surplus loss from a demand response compensation level
greater than pw − r.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87

.

.

Figure D.2: The equilibrium outcome and surplus loss under direct participation. Area
1 corresponds to the surplus loss under the no-demand-response baseline.
When electricity consumers particpate directly in demand response, the op-
portunity cost already includes the retail rate, so the demand curve shifts
downward to D(pw + r). The intersection of D(pw + r) = S(pw) deter-
mines the market-clearing wholesale price under direct participation, which
decreases. Area 2 corresponds to the surplus loss under direct participation.

. . 88

Figure D.3: Simulated deadweight loss by regime type plotted against the demand elas-

ticity parameter.

. .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89

x

Figure D.4: Simulated deadweight loss by regime type plotted against the retail rate.

. . .

. 90

Figure D.5: Simulated deadweight loss for each regime type versus the retail rate and
the elasticity parameter. Darker colors indicate larger simulated deadweight
losses. The gray regions indicate where the wholesale price is lower than the
retail rate.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91

.

.

.

.

.

.

xi

CHAPTER 1

MORAL HAZARD AND EQUILIBRIUM SORTING IN RESIDENTIAL ENERGY

MARKETS

1.1 Introduction

Economic wisdom suggests that individuals make socially efficient decisions only when they
face the full costs and benefits of those decisions. In landlord-tenant rental regimes, the energy-
efficiency and energy-use decisions are fundamentally divided in what is called the “split-incentive
problem." When the landlord pays for utilities, tenants face no marginal cost for energy use and
so are expected to use more energy than if they were facing the full marginal cost—a form of
moral hazard.1 Meanwhile, household sorting exacerbates incentive problems through the choice
of payment regime and home attributes that affect energy efficiency. First, a household that expects
to consume more energy services will have a higher willingness to pay for landlord-pay utilities, all
else equal. Second, conditional on choosing a landlord-pay unit a household may choose a larger
home because it knows it will not pay the full cost to heat the larger home.2 Thus, the relevant causal
effect of landlord-pay utilities on energy use includes both the direct effect via temperature setting
as well as the indirect effect via the choice of payment regime and housing attributes. An estimated
17 percent of US rental housing has landlord-pay heating contracts, suggesting that eliminating
moral hazard may provide large private energy savings and reduced external costs from fossil fuel
combustion (US Census, 2013).

In this paper, I develop and estimate a model of renters making joint home-efficiency and energy-
use decisions that characterize moral hazard and equilibrium sorting in the residential energy market.
1Conversely, when the tenant pays for utilities, landlords do not directly benefit from energy
efficiency and so are expected to under-invest in energy efficiency—also a form of moral hazard.
2Recent work on energy use suggests that individuals are cognizant of energy costs and savings.
Myers (2017) provides convincing evidence that when purchasing a home, households discount
future energy costs at 8-10 percent. In commercial real estate, properties built under more stringent
energy-efficiency standards rent and sell at a premium relative to those built under less stringent
energy-efficiency standards (Papineau, 2017).

1

I find that landlord-pay households spend 25 percent more on energy for heating than they would if
all rental contracts required the household to pay for energy.3 The moral hazard effect dominates,
contributing to 22 percent of extra heating expenditures per household. Meanwhile, landlord-pay
households choose housing units that are 140 square feet larger, contributing an additional 3 percent
of extra heating expenditures. In addition, I find that for a fixed cost of heating, households that
choose tenant-pay contracts prefer higher thermostat settings than households that choose landlord-
pay contracts as predicted by observables. This counter-intuitive result suggests that selection into
the landlord-pay regime is driven by preferences for unobserved characteristics of landlord-pay
units. Finally, I provide evidence of energy-efficiency economies of scale in multi-unit apartment
buildings, suggesting a possible explanation for the persistence of landlord-pay utility regimes.

In the model, renters have heterogeneous preferences for home size and temperature setting.
When energy is costly, renters either choose smaller apartments and turn down the heat to conserve
energy, or they sort into apartments with landlord-pay energy bills. Using data on housing rents,
temperature settings, and exogenous variation in energy prices, I estimate equilibrium hedonic
prices for home size and landlord-pay heating, the engineering energy cost of home size, and the
engineering energy cost of winter temperature settings. Given these estimates, I then estimate
how an average household’s indoor temperature setting responds to the estimated marginal cost
of heating the home.
I corroborate the results using a choice experiment that I designed and
implemented on a nationally representative sample of US households. These parameter estimates
allow me to construct unique household utility functions from the model that I use to simulate
counterfactuals. I simulate the effects of eliminating moral hazard by requiring that all tenants
pay for their own energy bills and allow tenants to respond by re-sorting into housing units of
different sizes. With this counterfactual, I quantify the energy impacts of moral hazard, selection
into housing regime, and selection of housing attributes in the landlord-pay utility regime.

This paper makes several contributions to the economics literature studying energy efficiency.
This is the first paper to use a structural model to analyze the split-incentive problem for residential
3This calculation is relative to the average bill for tenant-pay units because the actual heating

bill is not observed for landlord-pay households.

2

energy contracts. Experimental interventions requiring tenants to pay the bills find average energy
savings as high as 20 and 25 percent (Dewees and Tombe, 2011; Elinder et al., 2017), while quasi-
experimental approaches measure energy savings of less than 3 percent (Levinson and Niemann,
2004; Jessoe et al., 2018).4 I measure an average impact consistent with the literature, but my
structural approach provides new insights into the relative sizes of moral hazard and equilibrium
sorting. This paper complements theoretical and empirical work studying underinvestment in
energy efficiency due to split-incentive problems (e.g. Jaffe and Stavins, 1994; Harding et al., 2000;
Davis, 2012; Gillingham et al., 2012). While recent reviews conclude that the “energy-efficiency
gap" is small in most sectors (Allcott and Greenstone, 2012; Gillingham and Palmer, 2014), recent
work on energy use in rental housing suggests that asymmetric information between landlords
and tenants significantly distorts energy-efficiency investment incentives (Myers, 2018). Further,
I add to recent work investigating energy-efficiency scale economies in urban space (Borck and
Brueckner, 2017). The modeling and estimation strategy used here can be adapted to other contexts
with an up-front energy-efficiency choice followed by a downstream energy-use choice.

In addition, this paper connects the literature on equilibrium sorting in housing markets (Kumi-
noff et al., 2013) to the literature estimating moral hazard and adverse selection under asymmetric
information.5 I show how to estimate two-stage partial-equilibrium models of selection and moral
4Estimates of conservation behavior by tenants who pay for utilities vary by context. Levinson
and Niemann (2004) estimate that paying for heat reduces energy expenditures by less than 1
percent on average. On the other extreme, Elinder et al. (2017) find 25 percent average reductions
in electricity use when Swedish renters were experimentally switched to tenant-pay. The top 20
percent of energy users completely drive this result; the majority of households in the study did not
change behavior. Furthermore, these households did not use electricity for heating, and heat use is
highly inelastic. An experiment based on a single condominium complex found that submetering
decreased electricity consumption by 20 percent (Dewees and Tombe, 2011). Jessoe et al. (2018)
find that switching to tenant-pay results in a 2.9 percent reduction in average daily electricity use
among the top decile of commercial users, but that the effect for the remaining 90 percent of users
is negligible. The difference between single-site experimental and multi-site quasi-experimental
treatment effects may reflect site-selection bias on the experimental side (Allcott, 2015).

5Most recent papers in this literature focus on asymmetric information in health insurance
markets; e.g., Brot-Goldberg et al. (2017); Finkelstein et al. (2016); Baicker et al. (2015); Autor
et al. (2014); Einav et al. (2013). Other applications study lending markets (e.g. Veiga and Weyl,
2016; Crawford et al., 2018), crop insurance markets (He et al., 2017a,b), online marketplaces (Hui

3

hazard using an assumption on the parametric form of the utility function as in Bajari and Benkard
(2005). One key insight from this paper is that the first-stage choice of home attributes affects
the second-stage heating costs. Households’ choices of housing and subsequent heating use reveal
relative preferences for housing attributes and energy services when housing and fuel marginal
costs vary over time and space. The main appeal of this approach is that it allows for the estima-
tion of completely heterogeneous utility functions without making assumptions on the distribution
of the error term and with minimal supply-side assumptions. Another benefit is that estimation
is computationally simple and transparent. The parametric assumption on the utility function
provides closed-form solutions for the unknown parameters of each household’s utility function.
These parameters can be evaluated using household consumption decisions, hedonic prices, and
downstream energy costs estimated from the data.

Finally, I introduce a novel machine-learning approach to estimate energy-preference parameter
heterogeneity in the structural model. Households that pay their own energy bills do not reveal the
counterfactual winter temperature settings they would choose if the landlord was to pay for heat.
Households in landlord-pay units do reveal their “bliss point" temperature preferences (i.e., the
temperature chosen at zero marginal cost). This bliss point temperature setting is likely a function
of physiological characteristics such as age and sex.6
I train a machine-learning algorithm to
predict underlying bliss-point temperature settings using demographic covariates and the revealed
bliss points of households in landlord-pay units. These predictions characterize household heating
demand when there is zero marginal cost for temperature setting. I use these predictions to explore
heterogeneity in household energy demand, contributing to work analyzing household energy
demand heterogeneity (e.g. Reiss and White, 2005; Auffhammer and Rubin, 2018).
et al., 2016), worker contracts (Jackson and Schneider, 2011, 2015), and vehicle leasing (Weisburd
et al., 2018).

6Appendix B.1 discusses the physiological determinants of temperature setting preferences with

a thorough review of science and engineering studies.

4

1.2 Data and descriptive evidence of moral hazard

In this section, I provide descriptive evidence of moral hazard in rental energy use. In 2013, an
estimated 17 percent of US renters had landlord-pay heating (US Census, 2013).7 Space heating
accounts for approximately 42 percent of residential energy use by American households and is the
single largest non-transportation residential energy item (EIA, 2013). For this reason, I focus on
space heating and winter thermostat setting to analyze heterogeneous demand for energy services
and moral hazard in residential energy contracts.

Moral hazard arises in the landlord-pay regime because households do not face the marginal
cost of energy use. A housing unit’s utility payment regime has a larger effect on the household’s
marginal heating cost than home size or any other housing attribute. If utility payments are bundled
into the rent, the household has an effective marginal cost of zero for heating services regardless
of how large the home is. When the marginal cost of heating is zero, the household will heat
their home to a “bliss point," regardless of energy price or outside temperature. Thus, for a given
household, heating use will be weakly higher in a landlord-pay regime than in a tenant-pay regime.
Landlord-pay households do not respond to market signals of scarcity or to the increased cost of
heating due to low outside temperatures.

Sorting into landlord-pay regimes on temperature preferences may also occur. Given heteroge-
neous preferences for heat, households will differ in willingness to pay for the zero-marginal-cost
heating offered in a landlord-pay regime. Households with higher bliss point temperature will be
willing to pay more to heat without marginal cost all else equal. Those with high energy demand
are thus likely to sort into landlord-pay units. This sorting results in higher energy use than would
occur under randomly assigned regimes.

To test for moral hazard, I turn to data in the Residential Energy Consumption Survey (RECS).
The RECS asks landlord-pay and tenant-pay renters about their winter temperature settings. The
7This paper focuses on the rental case. While some owner-occupied housing units have the
utility payment bundled into the homeowners’ association fee, this case is relatively rare. Two
percent of homeowners had utility payments bundled into the homeowners’ association fee.

5

RECS is a repeated cross-sectional survey of nationally representative US households collected
every four years. I examine 3,993 electric-heating and gas-heating renters from the 2001, 2005,
and 2009 surveys—3,164 tenant-pay households and 829 landlord-pay households.8 In addition to
thermostat settings, the survey collects heating bills for tenant-pay units, housing characteristics,
heating and cooling degree days, location at the census division level, and household demographics.
Figure A.1 shows that landlord-pay households set their thermostats higher during the winter
than tenant-pay households during the winter. This difference by regime in average thermostat
setting can be explained either by moral hazard or by selection into landlord-pay regime on
temperature preference. Table A.1 shows differences in the demographics of tenant-pay and
landlord-pay households. Landlord-pay renters are older and poorer than tenant-pay renters—and
thus may also be expected to differ in temperature preferences. However, figure A.2 shows a larger
difference in temperature setting between the two regimes when the household is out of the home.
This evidence suggests that some of the difference in temperature setting between regimes is driven
exclusively by moral hazard rather than sorting.

Households also endogenously select housing attributes that influence energy cost. For example,
perhaps an additional 100 square feet increases the rent by five dollars each month but also increases
energy cost by a dollar each month. Then the true cost of 100 additional square feet is six dollars
per month rather than five. Even if households take energy prices as given, a household knows that
it can influence the cost of energy services by selecting different housing attributes. Likewise, it is
more expensive to increase the thermostat setting in a larger housing unit than a smaller unit.

I investigate the relationship between housing attributes, rent price, and utility payment regime
using data from the American Housing Survey (AHS). The AHS is a panel of housing units labeled
at the Metropolitan Statistical Area (MSA) level and includes information on rental price and unit
characteristics for 106,062 rental units from 1997 to 2013.9 Table A.2 displays sample means
for landlord-pay and tenant-pay units. On average, landlord-pay units are cheaper, smaller, older,

8See the appendix for notes on how the final sample is chosen.
9An MSA is a statistical region often centered around a city and the surrounding population

centers.

6

and in larger buildings than tenant-pay units. In addition, landlord-pay units are less likely than
tenant-pay units to have large energy-using appliances such as in-unit laundry and air conditioning.
Thus, landlord-pay units are older but otherwise have attributes associated with lower energy use.
It is difficult to draw conclusions about the impacts of equilibrium sorting in this context, because
if more landlord-pay units are in high-cost rental areas we may be mistaking a movement along the
demand curve of an attribute for a systematic difference in regime type. The role of home attribute
prices becomes clear in the formal model—where I turn to next.

1.3 Modeling and empirical strategy

In this section, I develop a model that describes how renters make joint choices of housing and
heating. Renters with heterogeneous preferences select housing units with differentiated attributes
and utility payment regimes that influence the cost of heating. Renters subsequently choose
temperature setting, trading off heating with other consumption.
I begin with a general model
that demonstrates the connection between housing attributes, heating cost, and temperature setting.
Next, I specify parametric forms of the utility function, heating cost function, and hedonic price
function that allow me to isolate and estimate the welfare impacts of moral hazard and equilibrium
sorting.

1.3.1 General model
Consider households i ∈ {1, ..., I} choosing to live in housing units j ∈ {1, ..., J} located in a
housing market. Housing units differ in basic housing characteristics xj (e.g., square feet), and
the utility payment regime Rj (a binary variable equal to one if the tenant pays the marginal cost
of utilities and zero if the landlord pays). The equilibrium per-month rental price of each housing
unit is a function of attributes and is given by pj = p(xj, Rj, ξj), where xj are observed housing
characteristics and ξj are researcher-unobserved housing characteristics. In this model, the choice
of which market to live in is made before considering housing and energy use so that a rental
unit in one housing market is not a substitute for a rental unit in another housing market. For

7

example, households choose jobs that determine which housing market is available to them before
considering housing choice.10

Given incomes yi, households derive utility from the housing characteristics xj and ξj, con-
sumption of an outside good zi, and the temperature setting within the housing unit si. I represent
household preferences with the utility function ui(xj, si, ξj, zi), which is concave in xj, si, and ξj.
Furthermore, there exists a bliss point temperature preference sb

i such that

= 0, and

< 0.

(1.1)

(1.2)

(cid:12)(cid:12)(cid:12)(cid:12)s=sb
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)s=sb

i

i

∂ui
∂s
∂2ui
∂s2

Intuitively, the bliss point sb
is an individual’s preferred heating temperature when facing a zero
i
marginal cost for heating. In addition, I adopt the assumptions of Bajari and Benkard (2005) to
guarantee the existence of a differentiable hedonic price surface in equilibrium.11

The amount the household spends on heating is a function of temperature setting si, housing
characteristics xj, the ambient temperature in the heating season Tj, and the market price of en-
ergy Pe. Let H(xj, si, Tj, Pe, ξj) be the expected per-month cost of heating unit j to temperature
setting s given outdoor temperature Tj and energy prices Pe. H(·) can be understood as a heating
cost function. Households sign leases without exact knowledge of the future outside temperature.
Households choose among units j with different housing attributes xj and ξj, taking into considera-
tion future heating costs. Once the households have moved in, they observe the outside temperature
and choose the temperature setting si. At this point, the households cannot change their optimally
chosen housing units j∗(i), so the housing attributes are fixed and denoted xj∗.
10This assumption is realistic in modeling the short run where a household’s employment is fixed
and therefore tied to a housing market. In the long run, a household may move to take advantage of
differential housing and energy prices across markets. The short run view is more realistic given
large moving costs.
11Specifically, the assumptions on the utility function are that (1) ui is continuously differentiable
in zi and strictly increasing in zi with ∂ui/∂zi > 
 for some 
 > 0 and zi > 0, (2) ui is Lipschitz
continuous in the housing attributes, and (3) ui is strictly increasing in ξj. Given discrete housing
attributes, condition (2) is not satisfied, but preferences over continuous attributes are identified as
the choice set becomes large (Bajari and Benkard, 2005).

8

In this framework, each household maximizes utility by first choosing a housing unit and
then choosing a temperature setting. Each household solves this problem by reverse induction:
each decides how much heat to use in a housing unit with given characteristics xj and ξj and
housing regime Rj and then uses this planned behavior to choose the preferred housing unit j∗(i).
The subsequent heating choice is a budget-constrained trade-off between consumption of energy
services and consumption of an outside good zi with price equal to one.

Heating choice: max
{si,zi}

ui(xj∗, si, ξj∗, zi)
s.t. yi ≥ p(xj∗, Rj∗, ξj∗) + Rj∗ · H(xj∗, si, Tj∗, Pe, ξj∗) + zi.

(1.3)

Note that the household cannot change the housing characteristics xj∗ because it has already chosen
its housing unit j∗(i). This problem yields the first-order condition for maximization:

(cid:30) ∂ui

∂zi

=⇒ [si] :

∂ui
∂si

= Rj∗ · ∂H
∂si

.

(1.4)

Equation 1.4 describes the household’s temperature setting behavior. The left side of the equation
is the marginal rate of substitution between energy services and consumption of the outside good.
The right side is the price ratio with the price of the outside good normalized to one.
If the
household is in a landlord-pay regime (Rj∗ = 0), the household sets temperature to the bliss point
sb
i regardless of cost. Intuitively, the bliss point is the temperature at which the utility function
peaks—beyond this point, a person feels too hot and before this point a person feels too cold. Denote
si = s∗
i (xj∗, Rj∗, Tj∗, Pe, yi) as the optimal “home heating rule" as determined by the first-order
condition in equation 1.4.

When sorting into a housing unit j, households consider the rent price of the unit and future

heating costs:

Housing choice: max

j

ui(xj, s∗

i , ξj, zi)
yi ≥ p(xj, Rj, ξj) + Rj · H(xj, si, Tj, Pe, ξj) + zi.

s.t.

(1.5)

I treat future energy prices as known by households, which is a realistic assumption given regulated
residential energy prices. I further assume that utility is linearly separable in the outside good to

9

allow me to interpret the opportunity cost of housing and heating in dollar terms.12 If there is a
continuum of available housing attributes such that there is continuity in attributes x and ξ, the
first-order conditions to this problem characterize sorting behavior:

=

∂p
∂xj
i ) ≥ ui(x−r

+ Rj · ∂H
∂xj
, s−r
i

j

−r
j z−r

i

, ξ

) for Rj = r ∈ {0, 1}, −r (cid:44) r.

(1.6)

(1.7)

(cid:30) ∂ui

=⇒ [xj] :

∂ui
∂xj
[Rj] : ui(xr

∂zi
i , ξr
j, sr

j , zr

These first-order conditions are instructive. The left side of equation 1.6 is the household’s
marginal rate of substitution between consumption of the housing attribute and consumption of the
outside good, while the right side is the price ratio with the cost of the outside good normalized
to one. Note that the implicit price to which each household responds depends on whether the
household sorts into a landlord-pay or tenant-pay regime. If in a tenant-pay regime, then Rj = 1, so
choices of xj (such as square feet) will affect the marginal cost of energy use. If in a landlord-pay
regime, then Rj = 0, so choices of xj no longer affect the marginal cost of energy use, eliminating
the price incentive to conserve energy.13 Equation 1.7 says that the household sorts into the payment
regime that brings it the highest utility when it considers the optimal housing characteristics for
each. The dependence of the first-order conditions on Rj confirms the intuition that households
have different incentives for energy consumption under each payment regime. These conditions
also illustrate the simultaneous choice of heating contract and other housing characteristics that
necessitates studying housing and heating decisions in a joint framework.

How much do the energy-use incentive distortions matter? To answer this question, the differ-
ence in energy use under the current regime and under an alternative regime in which all households
have tenant-pay contracts must be calculated, relying fundamentally on counterfactual housing and
12Linear separability abstracts away from risk aversion. Given regulated energy prices, price
uncertainty plays a small role in this market. Empirical studies of farm tenancy (Allen and Lueck,
1992, 1999), timber contracts (Leffler and Rucker, 1991), groundwater contracts (Aggarwal, 2007),
and even health insurance markets (Einav et al., 2013) reject risk aversion as a determinant of
contract choice. Allen and Lueck (1995) review the literature of contract choice across several
settings and conclude generally that risk aversion does not play a large role in contract choice.

13A similar condition exists describing the choice of the unobserved attribute ξj but is not the

focus of the analysis.

10

heating scenarios that cannot be observed. While the housing and heating choices of households
can be observed, household choices in a world with only landlord-pay or tenant-pay options cannot
be observed. To identify the counterfactual energy outcomes and welfare implications, I estimate
household utility functions with preferences revealed by household sorting and temperature choices.
These utility functions can be used to estimate households’ behavioral responses to policy changes
and to calculate the welfare effects of moral hazard and choice of housing attributes.

1.3.2 Empirical model

Building upon Bajari and Kahn (2005) and Bajari and Benkard (2005), I specify a flexible utility
function for households:

ui(xj, si, ξj, zi) = β1,iln(xj) − 1
2β2

(si − sb

i )2 + β3,iln(ξj) + zi.

(1.8)

Households have heterogeneous preferences for how warm they like to keep their homes as well
as for consumption of housing attributes. Note that the utility parameters β1,i and β3,i for housing
attributes in equation 1.8 vary by household. The term sb
is household i’s heterogeneous bliss
i
point temperature setting. Deviation from the bliss point temperature causes disutility at a rate that
depends on parameter β2 and the bliss point for household i.14

Monthly average heating cost for household i is the price of fuel Pe multiplied by the quantity
of fuel used Qe,i and the binary regime indicator Rj. The quantity of fuel used to heat to a
given indoor temperature setting is the difference in the temperature setting si from the outside
temperature during winter Tj, scaled by the housing unit’s efficiency level.15 I assume that home
attributes act as energy-efficiency complements (e.g., lower square footage and low temperature
14The distance norm is chosen for differentiability and simplicity. This parametric form is similar
to that adopted in Einav et al. (2013). A more general functional form could be used that allows
for stronger disutility from being too cold than too hot or vice versa, but this is not necessary given
the focus on heating.
15To reflect heating-season temperatures, the outdoor temperature during winter Tj is interpreted
as mean outdoor temperature conditional on being colder than 65◦F. This is to reflect heating-season
temperatures.

11

setting could have a multiplicative effect):

Qe,i = (si − Tj)σ · xγ

∂H
∂xj

= Rj

γ · Pe · Qe,i
xj

j

j · ξ κ
(cid:19)
(cid:19)

,

(cid:18)
(cid:18)

(1.9)

(1.10)

(1.11)

where σ, γ, and κ are fuel-intensity parameters. From this specification, the marginal heating cost
of attribute xj is

while the marginal heating cost of the temperature setting si is
σ · Pe · Qe,i
si − Tj

∂H
∂si

= Rj

.

In the model, I make the computational assumption that households respond to marginal heating
costs averaged by year, census division, regime, and fuel type.16

In addition, I assume rents are a function of housing attributes xj, fuel prices Pe, regime type Rj,
and potentially unobserved housing or market attributes νdm j, according to the rent price function:

rentdm j = µ0,m + α1,d xj + α2,dPe,d + α3,d(1 − Rj)Pe,d + α4,d(1 − Rj) + µ1wj + νdm j,

(1.12)

where rentdm j is the price of unit j in census division d, and MSA m.17 I allow coefficients on
attributes to vary by census division and allow mean rents to vary by MSA according to µ0,m. The
term wj are controls in estimation used to account for unobserved unit heterogeneity.18 Thus, α1,d
represents the marginal rent cost of attribute xj in division d. The markup for having the landlord
pay for utilities is the difference in energy price pass-through, α3,dPe,d. A higher fuel price should
be capitalized into the implicit rental markup for a landlord-pay housing unit. Different fuel prices
16While my specification provides estimates of completely heterogeneous marginal costs due to
the functional form of heating costs, any given estimate for i is likely a poor estimate (Wooldridge,
2010). Taken on average over division, year, regime, and fuel type, the average marginal heating
cost estimates are unbiased and consistent while preserving the variation due to fuel prices and
average housing stock efficiency differences to identify heterogeneous effects.

17Census divisions are regional groups of states: New England, Mid Atlantic, East North Central,
West North Central, South Atlantic/East South Central, West South Central, and Mountain/Pacific.
18For example, wj includes indicators by unit in the fixed-effects specification or individual

means in the correlated random effects specification.

12

should induce changes in markups for landlord-pay units with different fuel types over time. If the
fuel price was equal to zero, it would not matter whether a housing unit’s heating bill was paid by
the landlord or the tenant. Thus, the interaction between fuel price and landlord-pay regime traces
out the markup, while the landlord-pay term controls for unobserved differences correlated with
regime type.

Just as before, households set an optimal heating rule and choose housing attributes. Households
have unlimited wants and limited resources such that total spending on housing, energy services,
and other goods must be less than or equal to income yi. Under my funtional-form assumptions,
the first-order conditions in equations 1.4, 1.6, and 1.7 yield

[si] :

−1
β2

(s∗
i − sb

i ) = Rj∗

xj∗ = α1,d + Rj∗

β1,i

[xj] :
[Rj] : ui(xr

(cid:33)

(cid:32)

(cid:32)
σ · Pe,d · Qe,i
(cid:33)
si − Tj∗
γ · Pe,d · Qe,i
−r
j

xj∗
, s−r
i

, ξ

j, sr

i , ξr

j , zr

i ) ≥ ui(x−r

j

, z−r
i

(1.13)

(1.14)

(1.15)

) for Rj = r ∈ {0, 1}, −r (cid:44) r.

These equations characterize the choices of home heat setting si, home attributes xj, and energy
contract regime Rj.

1.3.3 Parameter identification

Identification of parameter β1,i is straightforward. Solving for β1,i in equation 1.14 reveals that for
the optimal choice j∗(i):

(cid:32)

(cid:33)(cid:33)

(cid:32)

γ · Pe,d · Qe,i

xj∗

β1,i = xj∗

α1,d + Rj∗

.

(1.16)

A household’s choice of housing characteristic xj∗, with estimates of the hedonic price α1,m and
marginal cost of heating reveal an estimate of parameter β1,i. Estimation is discussed in more detail
later, but these are relatively simple applications of a hedonic price regression and a “hedonic heat
cost" regression. Note that the first-order conditions and the estimated parameters can be solved
for the counterfactual choices of xj under the two payment regimes.

13

Temperature preferences β2 and sb

i are more difficult to identify. Using the home heating rule

in equation 1.13 for individuals in tenant-pay regimes (Rj = 1),:
σ · Pe,d · Qe,i
si − Tj∗

s∗
i = sb

i − Rj∗

β2.

(cid:32)

(cid:33)

(1.17)

(cid:44) sb

Thus, temperature settings by landlord-pay households (Rj = 0) reveal bliss-point temperature
settings, so sb
i is directly observed for landlord-pay households. Likewise, temperature settings by
tenant-pay households reveal s∗
i for tenant-pay households. With an estimate of the disutility
i
parameter β2 and the marginal cost of temperature setting, I can use equation 1.17 to back out the
unobserved bliss point for tenant-pay households as well as the counterfactual optimal temperature
setting under positive energy prices for current landlord-pay households. I use two complementary
approaches to identify the key parameter β2, both based on equation 1.17: I estimate β2 using a
revealed-preference approach in the RECS as well as design and implement a choice experiment via
a nationally-representative survey to elicit changes in temperature setting in response to randomized
variation in marginal heating costs.

First, I regress observed temperature setting on marginal costs, controlling for being in landlord-
pay and observed demographics. The bliss point is a physiological parameter that depends on
= S(Di) + δ(1 − Rj∗) + hi, where S(Di)
demographic characteristics Di.
is a function of demographics, δ(1 − Rj∗) represents unobserved differences in bliss points for
households which sort into landlord-pay units, and hi represents household taste heterogeneity.19
Using these assumptions and equation 1.13, the optimal heating choice can be rewritten as

I therefore model sb
i

i =(cid:2)S(Di) + δ(1 − Rj) + h1,i

s∗

(cid:3) −(cid:104)

where MC(·) is the marginal cost of temperature setting:

MC(Pe, Qe,i, si, Tj∗, Rj∗) = Rj∗ ·

β2MC(Pe, Qe,i, si, Tj∗, Rj∗)(cid:105)
(cid:32)

(cid:33)

σ · Pe · Qe,i
si − Tj∗

.

,

(1.18)

(1.19)

19I later discuss the possibility that the bliss point temperature setting may be different in different
housing units; e.g., a higher temperature may be preferred in an old and drafty unit relative to a
new unit.

14

Equation 1.18 traces out each household’s home heating rule. The first bracketed term represents the
bliss point as a function of demographics, while the second bracketed term represents the response
to the cost of home heating. Thus, a regression of temperature setting on household demographics
and the marginal cost of temperature setting (estimated in the heating cost regression) provides
an estimate of the mean disutility parameter β2. As a side benefit, the coefficients on (1 − Rj)
and D have economically meaningful interpretations. The δ coefficient on (1 − Rj) represents the
mean difference in bliss point preference for households in landlord-pay regimes not explained by
demographics. An estimate of S(Di) provides an estimate of each household’s bliss point that is
explained by observed demographic coefficients.

Second, to supplement this revealed-preference approach, I conduct a choice experiment that
allows me to estimate equation 1.17 directly using exogenous variation in energy costs. To do so,
I first elicit a subject’s bliss point by asking at what temperature they would set their thermostats
if heating was free. Next, I draw a low, medium, and high cost of changing the thermostat that is
calibrated from the estimated heating costs. Finally, I regress temperature setting on marginal cost
to get an estimate of β2 that is free of endogeneity.

Thus, the model provides closed-form expressions for households’ heterogeneous preferences
β1,i and sb
i . I use the estimated utility function parameters and heating costs to calculate partial-
equilibrium counterfactual energy use scenarios for each household and characterize the relative
effects of moral hazard, choice of home size, and sorting based on temperature preference. The
model allows me to estimate the welfare costs of moral hazard and equilibrium sorting through
deadweight loss and the external cost of emissions from fossil fuels.

1.3.4 Empirical model strengths and limitations

The functional-form assumptions on the utility function eliminate the need for several other as-
sumptions common in the literature. First, the functional form allows the utility parameters to
vary arbitrarily by household. Though I specify the parametric form of the utility function, the
parameters are extremely flexible and general. Second, the random coefficients do not require

15

a researcher-imposed error term as is common in most random coefficient utility models (Bajari
and Kahn, 2005). Finally, some functional form assumptions are necessary to simulate realistic
counterfactuals in any setting.

Another attractive feature of the estimation strategy is that it does not require any assumptions
about supply-side market structure (Bajari and Kahn, 2005). Households react to prices set outside
of their influence. These prices could arise in a perfectly competitive or monopolistic setting. The
downside to this aspect of flexibility is that I cannot model general-equilibrium supply-side effects.
For example, in a counterfactual simulation of removing the landlord-pay regime, landlords would
likely respond by changing the level of efficiency in the existing housing stock. While housing
attributes such as square footage are difficult to change, changes to appliances are a realistic response
to a policy change. In addition, the hedonic prices of energy-efficiency features of housing units
will adjust in general equilibrium. Computed counterfactuals therefore only reflect the response of
households in a partial equilibrium.

(cid:16)

(cid:17)

In addition to these relaxed constraints, the estimated parameters have an intuitive reduced-form
interpretation that lends credibility to the functional-form assumption. Equation 1.16 that identifies
β1,i states that a household’s preference for a housing attribute is revealed by the amount of a
housing attribute xj purchased as well as the true cost,
, of the attribute.
This makes sense: if household A purchases more square feet than household B at the same price,
all else equal, then household A’s preference for square feet β1,A is revealed to be greater than
B’s preference, β1,B. Similarly, if household A faces higher cost per square foot than household
B but chooses the same amount of square feet, it is revealed that β1,A > β1,B. Thus, the utility
parameters incorporate information from variation in quantity chosen, the market price, and the
cost to downstream heating.

∂p/∂xj + Rj · ∂H/∂xj

The empirical identification of the utility parameters depends on the flexibility of the estimated
gradients. To see this, consider estimating an inflexible price gradient by restricting the coefficient on
square feet to be the same for all households in the United States. Call this estimate ˆps f , the marginal
effect of square feet from a regression of price on housing attributes. Then consider two households:

16

household A chooses a 1,000-square-foot apartment with landlord-pay utilities in Los Angeles,
California, and household B chooses a 1,000-square-foot-apartment with landlord-pay utilities in
Raleigh, North Carolina. Using the framework developed above, ˆβs f ,A = ˆβs f ,B = 1, 000· ˆps f . The
issue is that the price of an additional square foot is not the same in Los Angeles and Raleigh—the
household in Los Angeles has likely paid much more per square foot and has a stronger preference.
The utility differences between the two households are not identified due to unobserved price
differences between markets.
In fact, the estimation of a single marginal price for square feet
adds no more information beyond the level choice of square footage. In this example, accurate
estimates of the marginal prices of square footage in each market identifies the difference between
two households’ utility parameters in different markets. The difference in quantities purchased
identifies within-market variation in a household’s utility parameters.

I proceed by estimating the model parameters used to construct the utility parameters above.

1.4 Estimation of model parameters

The key empirical objects needed to identify the model’s parameters are the marginal impacts
of thermostat setting and home size on heating cost as well as the hedonic price of landlord-pay
utilities and home size.
I estimate these objects using a heat cost regression and hedonic rent
price regression. Using these estimates and individual home choices, I construct estimates of the
preference for square feet β1,i. I then use a regression-based approach to estimate the temperature
disutility parameter β2 and selection into landlord pay based on bliss point. I corroborate these
estimates using data from a choice experiment. Armed with estimates of β2, I am able to back out
unobserved bliss points for current tenant-pay households. Finally, I use a LASSO regression to
explore heterogeneity in bliss-point temperature.

1.4.1 Heat costs

This step estimates the causal effect of home attributes and temperature-setting behavior on home
heating cost. How much more money does it cost to heat an additional square foot of home space,

17

(cid:16)
σ · (Pe · Qe,i)/(si − Tj)(cid:17)

ceteris paribus? Given that fuel use is governed by the relationship in equation 1.9, the marginal
heating cost of attribute xj is
and the marginal heating cost of temperature
setting is

γ · (Pe · Qe,i)/xj
. I estimate γ and σ by log-linearizing equation 1.9:20

(cid:16)

(cid:17)

ln(Qe) = σln(si − Tj) + γln(xj) + κln(ξj),

(1.20)

where si − Tj is the difference between the winter thermostat setting si and the average outdoor
temperature conditional on being less than 65◦F.21 Of course, the chosen heating intensity si −Tj is
an endogenous function q(·) of the fuel price Pe, home efficiency due to attributes ln(xj), heating
degree days HDDj, demographics Di, and unobserved tastes hi:22

ln(si − Tj) = q(Pe, ln(xj), HDDj, Di, hi).

(1.21)

Equations 1.20 and 1.21 are a classic demand system that I estimate using two-stage least squares.
Equation 1.20 is the structural equation of interest, in which si − Tj is endogenous. Equation 1.21
points to a first-stage regression in which energy prices affect temperature setting but satisfy the
standard exclusion restriction and can be used as an instrumental variable for temperature setting.
The estimation sample is all tenant-pay renters in the RECS whose fuel use is obtained directly from
20For attributes with zero or negative values, I use the inverse hyperbolic sine transformation as

an approximation to the natural log.

j,d

be (1/365)
365
temperature when it is less than 65◦F outside (i.e., HDDj =
365

21The ideal but infeasible approach would be to calculate an individual measure of heating
intensity for each household i that measures the cumulative daily difference between the indoor
temperature setting and the outside temperature when the temperature outside is lower than the
thermostat setting. Thus if Tideal
is the daily outdoor temperature, the ideal measure would
}. Because of data-privacy concerns, I do not observe each
d=1 max{0, si − Tideal
household’s exact daily outdoor temperature over the year. Instead, I observe “65◦F heating-degree
days" for each household (HDDj), which is the cumulative difference between 65◦F and the outdoor
j,ddeal|Tideal
65)}). Using the provided HDDj, I back out the average outdoor temperature conditional on the
outdoor temperature being less than 65◦F and label this Tj. The constructed measure of heating
intensity is the difference between the temperature setting si and the average outdoor temperature
conditional on the outdoor temperature being less than 65◦F.
22See footnote 21 for a definition of heating degree days, a commonly-used measurement of

d=1 max{0, 65−(Ti

j,d

<

j,d

frequency and intensity of cold days per year.

18

the utility company.23 Nearly all renters heat with either electricity or gas; therefore, I estimate heat
regressions for electricity and gas only. The final sample of electricity-heated households includes
a repeated cross section of 1,511 renters surveyed in 2001, 2005, and 2009. The final sample of
gas-heated households includes a repeated cross section of 1,653 renters surveyed in 2001, 2005,
and 2009.

Table A.4 displays the ordinary least squares and two-stage least squares estimates of equation
1.20. The two-stage least squares estimates of the energy-impacts of temperature setting σ and
unit characteristics γ are precise. For gas-heated homes (with an average monthly bill of $59), a
one percent increase in thermostat setting increases the heating bill by 0.49 percent, all else equal.
A one percent increase in square feet increases the heating bill by 0.21 percent, all else equal.
Similarly, for electricity-heated homes (with an average monthly bill of $91), a one percent increase
in thermostat setting increases the heating bill by 0.2 percent, all else equal. A one percent increase
in square feet increases the heating bill by 0.26 percent, all else equal.

Table A.5 displays the estimated marginal costs using the two-stage least squares estimates.
Increasing the thermostat setting by one degree increases the heating bill by $3.57 on average for
gas-heated homes. It costs an average of $1.01 per month to heat an additional 100 square feet for
gas-heated homes. Increasing the thermostat setting by one degree increases the heating bill by
$2.89 on average for electricity-heated homes. It costs $2.42 per month to heat an additional 100
square feet for electricity-heated homes. I use estimated marginal heating costs averaged by year,
census division, and by heating fuel type to construct preferences and estimate welfare effects in
the simulation model.

I also display the estimated savings from living in a larger apartment building. For gas- and
electricity-heated homes, heating bills are lower in buildings with a larger number of units. This
relationship between building size and energy savings is not due to the number of floors as argued
by Borck and Brueckner (2017). These results suggest economies of scale in heating for large
buildings with many units, independent of the number of floors.

23I exclude households whose fuel use is estimated. Note that fuel consumption and billing is

not available for landlord-pay households.

19

1.4.2 Temperature disutility and bliss points

I first estimate the temperature disutility parameter (β2) in a regression of temperature setting on
estimated marginal cost using revealed-preference data from the RECS. My ordinary least squares
estimates are precise but inconsistent, while my two-stage least squares estimates are consistent
but imprecise. Thus, I use data from a choice experiment that I conducted to get a precise causal
estimate of β2 to corroborate the cross-sectional estimates. Finally, I explore potential heterogeneity
in the disutility parameter using a LASSO approach. The point estimates of my revealed-preference
approach are corroborated by the choice experiment data.

First, I directly estimate the temperature disutility parameter β2 from the RECS using a specifica-
tion suggested by the heating rule in equation 1.18. I regress temperature settings on demographics

Di, a dummy for being on landlord-pay heating (1 − Rj), and the estimated marginal cost(cid:100)MC:

si = ψ0 + ψ1Di + ψ2(1 − Rj) + ψ3(cid:100)MCi + hi,

(1.22)

where hi is mean-zero heterogeneity. The marginal cost is estimated for household i in a first
stage as described in the preceding section.24 These predicted marginal cost terms are endogenous
because they depend mechanically on temperature setting, so I use fuel prices as an instrument
ˆψ3 is an estimate of the mean
for temperature setting. Given the heating rule in equation 1.18,
disutility parameter β2, while ˆψ2 is the estimated difference in bliss-point temperature preferences
for households in landlord and tenant pay due to factors not in Di.

Table A.6 displays these estimates. The coefficient on the marginal-cost term is imprecise
and sensitive to controls in the two-stage least squares specifications. Meanwhile, the ordinary
least squares estimates are tightly estimated around -0.55. Surprisingly, I find that landlord-pay
households have bliss points that are 1-3◦F lower than tenant-pay households on average (67◦F vs
69-70◦F).

Next, to supplement imprecise estimates from the cross-sectional approaches, I conduct a
24Standard errors are bootstrapped and allow for sampling error in the estimation of the marginal

cost term.

20

separate choice experiment to corroborate the estimates.25 The steps are as follows: First, I elicit
bliss point by asking subjects what temperature setting they would choose if there was no cost to
setting the thermostat. Next, I draw a low, medium, and high marginal cost for temperature setting
from three independent uniform distributions. The temperature marginal costs are calibrated from
the estimates obtained from the RECS and range from $1 to $8 per month to change the thermostat
by five degrees Fahrenheit. I then reveal each price to the respondents and ask at what temperature
they would set their thermostat, tracing out each respondent’s demand for energy services.

I conducted the choice experiment in March 2018 using the Qualtrics Survey Panel. The
final sample includes 414 individuals drawn from a nationally representative sample who each
completed three temperature setting choices; combined, these choices make 1,242 individual-
choice observations.26 Using the responses, I directly estimate equation 1.17 using the following
specification:

si,c − sb

i = a + bMCi,c + εi,

(1.23)

where si,c is respondent i’s temperature setting for choice c and MCi,c is the randomly drawn
marginal cost. Given this estimating equation, ˆb is an estimate of β2.

Table A.7 displays the ordinary least squares estimates, which indicate a β2 of -0.8. The
estimates are precise and are robust to controlling for individual fixed effects. These results are
similar to those obtained from the RECS approach, lending confidence to the estimates above.

Across the well-powered estimation methods, the disutility parameter β2 ranges from -0.5 to -1.
For the final simulation, I use the fixed-effects estimate of ˆβ2 = −0.79 from the choice experiment
and explore the sensitivity of the results to alternative parameter values of β2 = -0.5 and -1. Figure
A.12 displays the estimated bliss-point temperature settings for landlord and tenant-pay households
assuming β2=-0.79. Surprisingly, landlord-pay households have lower bliss points than tenant-pay
households. Thus, I do not find strong evidence that households select into landlord-pay regimes

25As described in chapter 2.
26See the appendix for a more detailed description of sampling methodology, a description of

how the final sample was chosen, and exact question text.

21

based on bliss-point temperature preference.27

1.4.2.1 Exploring heterogeneity with LASSO

Finally, I use a reduced-form double LASSO approach to flexibly explore heterogeneity in the
bliss point while allowing for heterogeneity in the temperature disutility parameter β2.28 Here, I
estimate an equation for temperature setting as a function of demographics Di, housing attributes xi
including Rj, and heating fuel prices faced by tenant-payers (Rj Pe).29 I also include the interactions
and squared terms of all these variables, resulting in 1,590 covariates. I estimate the equation

si =Υ0 + Υ1xj + Υ2x2

+ Υ7Rj Pe + Υ8Rj P2

j + Υ3Di + Υ4D2

i + Υ5xj Di + Υ6(1 − Rj)
e + Υ9(Rj Pe)xj + Υ10(Rj Pe)Di + ηi

(1.24)

(1.25)

where ηi is unobserved mean-zero heterogeneity in preferences using the L1-norm LASSO penal-
ization.30 This approach flexibly traces out a predictive equation for temperature setting. I use
this equation to predict bliss-point temperatures for all households when prices are equal to zero
ˆsi,lasso|Pe = 0 (landlord pay), and temperature settings for all households when prices are positive
ˆsi,lasso|Pe > 0 (tenant pay).

Figure A.7 displays the distribution of predicted bliss points using the LASSO algorithm. The
figure implies that landlord- and tenant-pay households have nearly indistinguishable bliss points
as predicted by observables. The distribution of estimated temperature disutility parameters β2,i
in figure A.8 shows similar temperature disutilities by regime. Landlord-pay households are less
responsive to marginal heating costs than tenant-pay households on average, though the distribution
of parameters is very similar overall. Finally, figures A.9 and A.10 display the distributions of

27I explore alternative explanations to this finding in the appendix.
28The double LASSO uses a first-stage LASSO algorithm to select variables then performs
ordinary least squares regression on the selected variables. This approach eliminates the parameter
bias introduced by LASSO. See Belloni and Chernozhukov (2013).

29By using fuel prices rather than marginal cost in this regression I avoid the endogenous marginal

cost terms.

30I choose the penalization weight λ using 10-fold cross validation.

22

predicted temperature settings under moral hazard and fully internalized prices for landlord-pay
and tenant-pay households. Both groups of households display significant predicted response to
paying for heating.

To demonstrate the improved performance of the LASSO over OLS, I include cross-validation
performance metrics in table A.8 and plot the predicted temperature settings versus the true
distribution in figure A.11. I compare the LASSO to ordinary least squares and find that LASSO
significantly reduces mean-squared error and bias in cross validation. Despite these performance
gains and the wide variety of included predictors, the root-mean-squared error of the best performing
LASSO is still on the order of 5◦F. I conclude that a large degree of heterogeneity in temperature
preferences is idiosyncratic and difficult to predict.

1.4.3 Hedonic prices and preferences for attributes

The goal of this step is to recover the marginal prices of apartment unit attributes. For example,
how much does it cost to rent an apartment with more square feet, ceteris paribus? Using the
AHS data, I regress rental rate on unit characteristics while controlling for housing market and year
effects. I allow for different marginal price coefficients by census division, which is the finest level
of geography that can be used to link prices between the AHS and RECS.

I estimate the hedonic gradient, allowing coefficients to vary by census division:

rentdm j = µ0,m + α1,d xj + α2,dPe,d + α3,d(1 − Rj)Pe,d + α4,d(1 − Rj) + µ1wj + νdm j,

(1.26)

where rentdm j is the price of unit j in census division d, and MSA m. I allow coefficients on
attributes to vary by census division and allow mean rents to vary by MSA according to µ0,m. The
term xj represents attributes of unit j, (1− Rj) is a binary variable equal to one if the landlord pays
for heat, Pe,d is the heating fuel price, wj are controls in estimation used to account for unobserved
unit heterogeneity, and νdm j are unobserved attributes or market characteristics affecting rent.31
Pass-through of regional electricity and natural gas prices into the rent identifies the landlord-pay
31I control for unit square feet, bedrooms, bathrooms, air conditioning type, in-unit laundry,
number of units in the building, frequency of the rent payment, heating equipment type (e.g.,

23

markup ∆p/∆R—similar to a strategy used in two recent papers by Myers (2017, 2018). If the
fuel price was equal to zero, it would not matter whether a housing unit’s heating bill was paid by
the landlord or the tenant. However, a positive fuel price should be capitalized into the implicit
rental markup for a landlord-pay housing unit. Further, variation in relative fuel prices should
induce differential changes in markups for landlord-pay vs tenant-pay units relying on different
fuel types over time. This intuition allows for the inclusion of a rich set of controls such as unit
fixed effects, indicators for fuel type, and flexible vintage by year indicators. Because of the
relative variation in fuel prices over time (depicted in figure A.3), the landlord-pay markup is still
identified when including unit fixed effects. Intuitively, the identifying variation for the markup
comes from exogenous changes of fuel prices that are capitalized into higher rents for landlord pay
units. This strategy requires the assumption that rentals do not vary over time and space between
payment regimes in unobserved ways that are correlated with changes in relative fuel prices over
time. Figure A.3 displays the time-series variation in the relative fuel prices (in $/MMBTU) used
to identify the markup.

Table A.3 reports the estimates of the hedonic price of square feet and the estimated markups
by census division. The specifications include a “vanilla" ordinary least squares estimation, a
housing unit random-effects estimation, a housing unit fixed-effects estimation, and a housing
unit correlated-random-effects estimation. The preferred correlated-random-effects specification
uses the within-unit variation on fuel prices over time to trace out the markup for having the
landlord pay for heating but uses across-unit variation in square feet to estimate the hedonic price
of square feet. Figures A.5 and A.4 display the results from the preferred correlated-random-effects
specification. I find that 100 additional square feet costs anywhere from $5-$20 extra per month to
rent. These estimates are consistent across all specifications. The high estimates in New England
furnace, electric room heaters, etc), non-heating utilities that the landlord pays, whether the unit
was too cold last year, a subjective rating of neighborhood quality, an indicator for gas heat, division
interacted with fuel type, division indicators, vintage by year interactions, landlord-pay interacted
with division, MSA, and year dummies interacted with the landlord-pay term. The wj term includes
indicators by unit int he fixed-effects specification or individual means in the correlated random
effects specification.

24

and the Mountain/Pacific divisions are likely driven by expensive housing in New York City, San
Francisco, and other such cities. The confidence intervals for the estimated markup for having
heat included in the rent are wide and include zero for most census divisions. The point estimates
suggest a markup as low as zero and as high as $100 per month.

I construct household preferences for square feet β1,i by substituting the estimated values for
the hedonic price and average marginal heating cost into equation 1.16. Figure A.13 displays the
distribution of estimated preferences for square feet. Landlord-pay households have significantly
lower preferences for square feet than tenant-pay households on average and their choices are more
sensitive to changes in the price per square foot.

1.5 Simulation results

Consider requiring all tenants to pay for their own heating bills. Landlord-pay households would
respond by re-sorting into different-sized housing units and changing their temperature settings.
I use the estimated utility parameters for temperature setting and housing unit size to estimate
household responses to such a change. Table A.9 contains the results of the simulation. Column
(1) uses the preferred estimate of the temperature disutility parameter β2 = −0.79 and columns (2)
and (3) display results for alternative values of β2. On average, landlord-pay households re-sort
into housing units that are 140 square feet smaller and reduce their thermostat settings by 4 degrees
Fahrenheit. Overall, these changes translate into lower energy expenditures of $218 per household
per year or a 25 percent decrease relative to observed energy bills.32 While 140 square feet seems
like a large decrease in home size, it only accounts for $30 per household per year. The moral
hazard effect dominates, accounting for a full $188 of lower expenditures per household per year.
What do these results imply for the welfare losses associated with the landlord pay regime? I
calculate the deadweight loss from moral hazard and the sorting effect by considering not just ad-
ditional expenditures but also each household’s benefit from consuming additional energy services
and square feet when the landlord pays. In addition, I calculate the additional carbon dioxide emis-
32This change is relative to the average bill for tenant-pay units because the actual heating bill is

not observed for landlord-pay households.

25

sions due to increased demand for fossil fuels.33 The total calculated welfare loss is the deadweight
loss plus the external cost of the carbon dioxide emissions under landlord pay relative to a scenario
in which all tenants pay for their own heating and sort into different-sized housing units.34

Table A.9 breaks down the yearly welfare cost per household and for the entire United States. I
calculate an average welfare loss of $154 dollars per household per year; external emissions costs
account for $43.68 of this total. Using the RECS survey weights, I calculate the annual welfare
loss across all landlord-pay households in the United States. I find that the total welfare loss is
$839 million per year. Of this amount, $238 million is due to the external cost of carbon dioxide
and $601 million is due to the deadweight loss. I also calculate the welare costs based on the high
and low range of disutility parameter (β2) estimates and find that the total welfare loss ranges from
$682 million to $1.26 billion per year.

The model estimates significant regional heterogeneity in welfare impacts. Both the East North-
Central and West North-Central regions have the lowest total deadweight loss impacts at $117 and
$70 per household per year. The Mountain/Pacific and South Atlantic/East South Central regions
have the highest total deadweight loss impacts at $203 per household per year. Regional differences
in fuel prices and the number of carbon-intensive electricity-heated units drive this heterogeneity.

1.6 Conclusions

When the landlord pays for heating, the renter does not face any price incentive for energy
conservation—a clear form of moral hazard. Landlord-pay households thus choose higher thermo-
33I assume a $40/ton external cost of carbon dioxide. For natural gas, I assume that natural
gas combustion releases 53.12 kilograms of carbon dioxide per 1000 cubic feet (EIA, 2016). For
electricity, I calculate a simple carbon emission intensity for each census division by dividing total
electricity production by total carbon emissions from electricity production.

(cid:17)
pay household i isÞ sb
society from sorting on attribute xj isÞ xj|Rj =0

34In terms of the model, the total monthly welfare loss to society from moral hazard in landlord
+ Cs(e, m)ds where Cs(e) is the marginal carbon cost
of temperature setting for fuel type e in census division m. The total monthly welfare loss to
x + Cx(e, m)dx where Cs(e) is

i
si|Rj =1

∂H
∂si

− β1,i

+ β2

s − sb
i

xj|Rj =1

∂H
∂xj

+ ∂P
∂xj

(cid:16)

the marginal carbon cost of square feet for fuel type e in census division m.

(cid:19)

(cid:18)

26

stat settings than if they paid for heating. In addition, when the landlord pays for heating, the renter
does not pay the implicit energy cost of of square feet and other home attributes. Landlord-pay
households thus choose larger housing units than if they paid for heating. Landlord-pay residential
energy contracts create perverse energy-use incentives for moral hazard and equilibrium sorting
that if eliminated could provide private savings and reductions in fossil-fuel emissions.

In this paper, I build a model in which households jointly choose home attributes and temperature
setting. Households sort into housing units by selecting the heating contract and the size of the
unit, considering the difference in heating cost. After selecting a home, the household chooses the
indoor temperature setting. If in a tenant-pay unit, the household faces a tradeoff between energy
services and other consumption.
If in a landlord-pay unit, the household ignores heating costs
and heats to its temperature bliss point. The model allows me to estimate the welfare impacts of
choosing a housing unit that is too large and setting the thermostat too high.

Using observed rent prices, home attributes, energy bills, temperature settings, and a choice
experiment, I estimate key parameters in the model. I find that landlord-pay households heat 4◦F
higher than if they were made to pay their own heating bills. In addition, these households choose
housing units 140 square feet larger than if they were made to pay their own bills. These distortions
result in welfare losses of about $154 per household per year. Aggregated across all US landlord-
pay households, these results imply $839 million in welfare losses per year from landlord-pay
contracts, including $238 million due to the external cost of carbon emissions.

I find suggestive evidence for economies of scale in home heating:

The landlord’s role in contract choice is still not well-understood. Future work should explore
the landlord’s role in selecting housing attributes and choosing to bundle utility payments into
the rent.
for gas-heated
units, units in larger apartment buildings are significantly less expensive to heat. Units in larger
apartment buildings are also significantly more likely to have landlord-pay utilities. Landlord-pay
units may have more energy-efficient appliances and insulation than tenant-pay units (Myers, 2018;
Gillingham et al., 2012). The results here and in the literature suggest that landlords choose to pay
for heating when it is profitable. The landlord’s joint decision has not received much attention in

27

the current literature and would shed light on the other side of the market.

28

CHAPTER 2

THE REMARKABLY INELASTIC DEMAND FOR HOME ENERGY SERVICES: A

CHOICE EXPERIMENT

2.1 Introduction

The law of demand states that as the price of a good increases, the quantity of the good a
consumer demands will decrease. When applied to environmentally damaging goods, the law of
demand suggests that prices are a powerful tool for conservation. Since the seminal work of Pigou
(1920), environmental economists have studied how prices of polluting goods can be increased to
reduce consumption and account for the external costs of pollution. Energy is the quintessential
polluting good. Energy production typically requires burning fossil-fuel inputs such as oil, natural
gas, or coal.1 Combustion of fossil fuels produces costly local pollutants (e.g., particulate matter,
nitrogen dioxide, and sulfur dioxide) and global pollutants (e.g., carbon dioxide). At the same time,
the energy services derived from fossil fuels provide profound benefits to human well-being (e.g.
lighting, heating, and other electronic technologies). Understanding demand for energy is important
for creating institutions that properly incorporate environmental costs into decision making.2

Central to this story is the assumption that individuals respond to changing energy prices
and trade off the benefit from energy consumption with the cost of energy consumption. Recent
empirical work challenges this assumption. One study of electricity-use data finds that 44 percent
of studied households did not respond to prices at all (Reiss and White, 2005). An analysis of
natural gas billing data finds that both low-income and high-income households do not respond to
prices (Auffhammer and Rubin, 2018). An experiment in Sweden where renters were switched from
landlord-pay to tenant-pay electricity shows that while average electricity consumption decreased by
1In 2017, of all useful thermal output in the United States, 37 percent came from non-fossil
fuels. In the electric power sector, only 9.6 percent of thermal output came from non-fossil fuels
(EIA, 2018).

2Energy demand is also a key concern for electricity and natural gas utilities. For example, if the
supply and demand of electricity is not balanced at all times during a day there will be a blackout.

29

24 percent, two-thirds of the reduction came from just 20 percent of the studied households (Elinder
et al., 2017). Price responsiveness is low for most energy users but high for a select group. Why do
some individuals fail to respond to prices while other individuals cut energy consumption drastically
when prices are high? There are two main potential explanations for low price responsiveness:
behavioral heterogeneity and preference heterogeneity. Either non-price-responsive individuals
are uninformed or face high costs to monitor energy prices, or these individuals know prices and
rationally choose not to respond because of high valuation of energy services.3

This paper tests whether individual heating-choice behavior is consistent with households
having full information about the cost of energy. I conduct and analyze a stated-choice experiment
with a nationally representative sample of individuals in which participants make choices about
how high to set their thermostat during the winter when told the hypothetical cost of doing so. In
this experimental setting, energy costs are easy to understand, costless to monitor, and salient. The
experimental environment is clean of any potential confounding factors such as unobserved energy
efficiency, thermostat or meter placement, and attrition bias which makes it difficult to interpret
results from field experiments. The results from the choice experiment serve as a fully informed
benchmark to compare to real temperature-setting behavior. If hypothetical temperature-setting
behavior and real temperature-setting behavior match, this provides evidence for heterogeneous
preferences as the primary driver for energy-use heterogeneity. If experimental temperature-setting
behavior differs from real temperature-setting behavior, this is evidence pointing to price salience,
informational barriers, and adjustment costs as determinants of energy-use heterogeneity.

From the experiment, I find that under perfect experimental conditions, 50 percent of individuals
do not change the thermostat at any treatment cost. On average, a 100 percent increase in the cost of
heating the home one degree Fahrenheit warmer induces a 0.31-0.97 degree Fahrenheit reduction
3The literature suggests a number of behavioral responses or informational barriers for energy
use. For example, Ito (2014) find evidence that energy users respond to average rather than marginal
prices. Jessoe and Rapson (2014) argue that consumers do not know prices or face a high cost of
determining energy prices. Allcott and Rogers (2014) find that social comparisons impact energy
use and observe behavior consistent with short attention spans. Finally, Allcott and Taubinsky
(2015) argue that consumers do not pay attention to energy prices when choosing light bulbs.

30

in the winter heating level. In addition, I find that reported actual household temperature-setting
behavior is consistent with realistic beliefs about the cost of heating. This evidence suggests
that preference heterogeneity plays a large role in driving empirical observations of inelastic and
heterogeneous energy demand. I analyze the heterogeneous price responses and find that individuals
with higher temperature preferences are more price responsive. This is evidence that increasing
block pricing of emissions can reduce emissions while minimizing the regressive properties of
emissions fees.

2.2 Experimental procedure

Experiment participants comprise a nationally representative sample of US individuals drawn
from the online Qualtrics Panel. I eliminated respondents if they failed Qualtrics speeding checks,
failed attention check questions placed in the survey (e.g. “Agree or disagree: I breathe more than
once per day."), if they do not use heat at home in the winter, or if they provided poor-quality
responses (e.g. gibberish in free response boxes). The final sample includes 414 individuals. The
experiment took place in early March 2018, the end of winter for most of the United States; thus,
respondents completed the survey after making heating decisions for several months.

First, I elicit each individual’s temperature preference baseline by asking what temperature they

would choose if there was no cost to heat:4

Imagine that you do not have to pay for heating your home during the winter. In this
situation, what temperature setting (degrees F) would you choose when you are at
home?

4The science and engineering literatures argue that temperature preference is determined by
physiological characteristics such as age (Taylor et al., 1995; Schellen et al., 2010), sex (Kingma
and van Marken Lichtenbelt, 2015; Karjalainen, 2012, 2007; Fanger, 1970; Parsons, 2002; Cena
and de Dear, 2001; Muzi et al., 1998; Pellerin and Candas, 2003; Griefahn and Künemund, 2001;
Nakano et al., 2002; Nagashima et al., 2002), diet (Ringsdorrf Jr. and Cheraskin, 1982), and
previous exposure (Young, 2010). There is some evidence that temperature preferences of men and
women differ by country (Beshir and Ramsey, 1981; Karjalainen, 2007; Indraganti and Rao, 2010)
and that individuals may be able to consciously alter the body’s internal response to temperature
(Kox et al., 2014).

31

This baseline temperature preference with no price can be thought of as a bliss point temperature
preference for heating. Next, respondents see an example of how much money would be saved for
a variety of thermostat settings.

I draw a low, medium, and high marginal cost from three indepdendent uniform distributions
spanning $1 to $8 per month for a five degree Fahrenheit change when they are home.5 Respondents
see a price and are asked to input their chosen temperature setting. For example,

Choice #3: Imagine increasing your thermostat by one degree Fahrenheit will increase
your monthly heating bill by $1.60 (or changing your thermostat by five degrees
Fahrenheit will increase your heating bill by $8).

When a one degree change in temperature costs $1.60 per month, what temperature
setting would you choose?

Remember that you said you would set your thermostat to 70 degrees Fahrenheit if you
weren’t paying for heating.

Respondents input their chosen temperature into a text-response box. After completing the experi-
ment, respondents supply their demographic information. Qualtrics compensates each respondent a
small sum after participating successfully. A copy of the survey instrument is available for viewing
in the supplementary dissertation files.

2.3 Data

Table C.1 displays summary statistics from the experimental sample after cleaning the data.
The sample is higher income, older, more female, and more white than the nationally representative
quotas used to recruit subjects. Figure C.1 displays kernel density plots of participants’ bliss
point temperature preferences and actual temperature settings. The distribution of bliss point
temperatures appears to have a higher mean and similar variance to the distribution of actual
5The first price is a random draw from a U(1,2.67) distribution, the second price is a random
draw from a U(2.67,5.33) distribution, and the third price is a random draw from a U(5.33,8)
distribution.

32

temperature settings. A Kolmogorov-Smirnov test of equivalence of distributions easily rejects the
null hypothesis that the distributions are the same.

Half (54 percent) of participants report that they set their actual thermostats equal to their bliss
point temperature preference. Figure C.2 shows a plot of bliss point vs actual temperature setting.
6.5 percent of participants report that the thermostat is higher than the bliss point, perhaps because
they did not understand the question or because they are not in control of the thermostat. For
experimental temperature settings, 50 percent of participants continue to choose their bliss point
temperature setting at the highest cost level (including 73 percent of individuals who set their actual
home temperature equal to their bliss point).

2.4 Estimating demand for heating

The experiment provides points on each respondent’s temperature demand curve. The most
intuitive measure of an individual’s temperature response to a change in the price of an additional
degree is the semi-elasticity, or the degree change in the thermostat for a percent change in price.
For each experiment choice c ∈ {1, 2, 3}, an individual i sees a price pricei,c to change the
thermostat and chooses a temperature setting tempi,c. Thus, for a household i, I model the choice
of temperature setting as some function f(·) of the price:

tempi,c = f(pricei,c) + εi,

where εi is individual heterogeneity that is orthogonal to pricei,c. The semi-elasticity is

(∂temp/∂price) · pricei,c.

(2.1)

(2.2)

First, I pool the sample and estimate the mean semi-elasticity using ordinary-least-squares and

fixed-effects estimation. I estimate the following equation on the pooled temperature choices:

tempi,c = α + βln(pricei,c) + 
i.

Taking the derivative with respect to the price variable and solving for β reveals that β =
(∂temp/∂price) · pricei,c. Thus with this functional form, the estimate ˆβ serves as an estimate of
the average semi-elasticity.

33

%∆c,c−1pricei =(cid:169)(cid:173)(cid:171) pricei,c − pricei,c−1

pricei,c+pricei,c+1

2

(cid:170)(cid:174)(cid:172) ,

(2.4)

Next, I estimate each individual’s unique semi-elasticity for temperature setting by calculating
the arc semi-elasticity directly and using a regression adjustment method. Following the intuition
developed in Allen and Lerner (1934), I calculate the arc semi-elasticity between each pair of
points on each participant’s temperature demand curve. The arc semi-elasticity between two
choices tracing out the demand curve c and c − 1 is

Arc semi-elasticityc,c−1 =

(2.3)
where ∆c,c−1tempi = tempi,c − tempi,c−1, the difference in temperature settings chosen by the
participant, and

,

∆c,c−1tempi
%∆c,c−1pricei

the percentage difference in researcher-assigned energy price. One benefit of this approach is that
the arc semi-elasticity can utilize information from the bliss point choice (i.e., when price is zero),
while the regression-based approaches cannot due to the log of zero being undefined. In addition,
this approach provides a heterogeneous and non-parametric measure of price responsiveness.

Finally, I measure each participant’s semi-elasticity using a regression-adjustment approach.

For each participant i, I estimate the following equation separately with ordinary least squares:

tempi,c = αi + βiln(pricei,c) + 
i.

(2.5)

The estimate ˆβi is an estimate of each individual’s mean semi-elasticity over individual i’s exper-
imental choices c. This method provides heterogeneous semi-elasticities but does not incorporate
information provided from the bliss point choice.

Table C.2 displays the estimated average semi-elasticities using all four methods. I bootstrap
the 95 percent confidence intervals of the averages using 1,000 replications and re-sampling at the
participant level. The estimates imply that for a 100 percent increase in the cost of heating, an indi-
vidual reduces the thermostat setting by 0.31-0.97 degrees Fahrenheit. This small average response
is due to the large number of price-insensitive participants and hides significant heterogeneity,
which I analyze in the following section.

34

2.5 Heterogeneity analysis

Figure C.3 displays a histogram of participants’ arc semi-elasticities, and figure C.4 displays
a histogram of participants’ regression-adjustment semi-elasticities. The distributions display a
similar bunching of individuals completely unresponsive to prices with a significant portion of
more price-responsive individuals in the tail.

The distribution of elasticities is characteristic of those found in other energy settings. Reiss and
White (2005) estimate a similarly skewed distribution of annual elasticities for electricity use with
a mass of relatively price-insensitive households and a fat tail of more elastic households. They
also find that low-income households are more elastic and conclude that households with space
heating are significantly more elastic than other households. The experiment here shows that the
skewed distribution can be generated without the energy-efficiency responses included in a yearly
elasticity.

I explore what drives heterogeneity in temperature response by regressing the arc semi-
elasticities on standardized vectors of the average price on the arc pricei,c,c−1 and participant
demographics demographicsi. I use a Tobit maximum-likelihood estimation to account for the
clustering at zero in the dependent variable. Thus, denoting Z(·) as the function that transforms a
sample draw of a random variable into its z-score, I estimate the equation

Arc semi-elasticityi,c,c−1 = a + bZ(pricei,c,c−1) + dZ(demographicsi) + ei,c,c−1

(2.6)

using maximum likelihood, treating all non-negative arc semi-elasticities as a corner solution.
Standardization allows the marginal effects of the regression to be easily compared. The marginal
effects from this estimation are interpreted as the change in arc semi-elasticity for a one-standard-
deviation change in the predictor variable holding the other predictor variables constant.6

I explain heterogeneity as a function of bliss point temperature preference, average monthly
heating bill, income, age, household size, number of children living at home, gender, race, ur-
effect
where x is a matrix of predictor variables and xj is a single predictor

∂E(Arc semi-elasticity|Z(x))

6The marginal

“unconditional"

I

estimate

is

average

partial

variable using the results provided in (Wooldridge, 2010).

effect

∂z(xj)

the

35

ban/rural status, education, and political party. Figure C.5 plots the estimated marginal effects with
the 95 percent confidence intervals bootstrapped using 1,000 replications with repeated sampling
at the participant level. Most strikingly, individuals with a one-standard-deviation-higher bliss
point temperature have on average a -.42 higher arc semi-elasticty (i.e. are more elastic), all else
equal. Higher-income and higher-education respondents are less elastic on average, all else equal,
although the confidence intervals for the education marginal effects include zero. Participants
living in urban areas are more responsive to price changes. Older participants are less elastic, with
a one-standard-deviation increase in age corresponding with a .23 lower arc semi-elasticity, all else
equal. The marginal effects estimates of average heating bill, participant gender, race, number of
children, and household size have confidence intervals that contain zero.

Political party is not a strong determinant of elasticity, with Republicans, Democrats, and
Independents having statistically indistinguishable elasticity measures when controlling for other
covariates. In two papers, Costa and Kahn estimate heterogeneous energy use patterns by political
ideology. First, Costa and Kahn (2013a) show that total household electricity use is lower for polit-
ically progressive households. Second, Costa and Kahn (2013b) finds that politically progressive
homeowners are more responsive to non-price nudges. The experiment in this paper measures a
different dimension of energy use, but nonetheless the lack of heterogeneity by political group is
surprising. It is possible that in the literature, total energy use and ownership of energy-efficient
appliances are correlated with progressive energy-efficiency policies and thus reflect these policies
rather than individual behavior. In this estimation, I include many controls that are correlated with
ideology and whose influence may be spuriously attributed to ideology (e.g., urban or rural).

In this experiment, age plays a large role in determining elasticity whereas sex does not. While
the science and engineering literatures focus on measuring differences in temperature preference
and sensitivity, differences in behavior are often ignored. While a group of people may on average
be able to detect a difference in temperature in a laboratory more readily, this does not translate
necessarily to differences in thermostat setting behavior. Indeed, I find here that men and women
do not respond to prices differently after other characteristics have been controlled for despite

36

numerous findings that women prefer higher temperatures than men.7 The findings in the literature
may reflect how temperature decisions are made in settings that affect multiple individuals with
heterogeneous temperature preferences or other barriers to adjusting the thermostat.8

2.6 Conclusion

The experiment reproduces energy-use heterogeneity distributions comparable to those seen in
actual energy-use data. Half of participants report that they set their actual thermostats equal to their
bliss point temperature preference when at home. Of these individuals, 70 percent were similarly
unresponsive to the cost of heating in the experiment. This is evidence that for these 70 percent of
individuals, there is some perceived negative preference for deviating from their temperature bliss
point in excess of the savings that they could have made in the experiment. These participants’
behavior is consistent with a rational zero response to the cost of heating at the relevant price level.
Under perfect-information conditions, energy-use behavior features significant heterogeneity and
unresponsiveness.

There are two main empirical findings in the paper. First, a majority of individuals set their
thermostats consistent with having perfect information. It is not likely that every individual knows
the exact cost-per-degree change on the thermostat, but over time most people have adjusted their
behavior based on feedback from energy bills. Second, more than half of all individuals are
completely unresponsive to prices. People simply do not like to be cold. The cost of heating is low
enough to take heating for granted, but it is likely that even if the cost of heating was to dramatically
increase (perhaps due to a pollution fee), behavior would respond very little. Inelastic demand for
energy services does not mean that prices are ineffective; instead, it means that the benefits from
energy services are high. As long as the inelasticity does not arise from an artificial barrier such as
false information about the energy cost savings, individuals will make the proper tradeoff between

7See Karjalainen (2012) for a review of this literature.
8For example, Kingma and van Marken Lichtenbelt (2015) discuss temperature demand in
shared office buildings and Karjalainen (2007) finds that women are less likely to change the
thermostat settings than men are.

37

costs and benefits from energy use when facing prices that reflect the full external costs of energy
use.

These results imply that the gains from home energy efficiency are likely high.

If energy-
use behavior is fixed for many individuals, energy efficiency savings are large and will not be
cannibalized by a rebound effect.
It is not clear whether households optimally adopt energy-
efficiency upgrades (i.e., whether there is an “energy-efficiency gap"), but a recent review of the
literature did not find much evidence that individuals systematically fail to adopt energy efficiency
(Gillingham and Palmer, 2014).

One important implication of these findings is that increasing block pricing can be used to reduce
energy use without large incidence for a bulk of inelastic energy users.9 The largest determinant
of elasticity in the experiment was having a high bliss point temperature preference, implying that
large energy users are more price responsive. By increasing the price of energy for large energy
users who are most price-responsive, a regulator or regulated energy provider can reduce load
(and corresponding emissions) without increasing payments from inelastic users. For example, a
carbon tax with a zero-price carbon allowance may not sacrifice efficiency gains but may reduce
the regressivity of the policy.

9Increasing block pricing charges a higher marginal cost per unit of energy for consumption of
units of energy over a threshold. It essentially provides users with an allowance of cheap energy
each billing period before having to spend more on additional energy consumption.

38

CHAPTER 3

NEGAWATTS VS. MEGAWATTS: DEMAND RESPONSE IN WHOLESALE

ELECTRICITY MARKETS (WITH KATHERINE WAGNER)

3.1 Introduction

Between 2007 and 2017 in the United States, retail consumers spent $377 billion per year on
electricity—about 2.3% of annual GDP (EIA, 2018; BEA, 2019). Before a utility sells electricity to
a retail consumer, the utility purchases electricity from generators on real-time wholesale markets.
Despite electricity market liberalization in the 1990s, the typical retail rate is regulated and does not
vary with the real-time marginal cost of electricity; thus, consumers do not receive price signals of
scarcity when electricity is expensive during peak hours. During peak hours, the wholesale price of
electricity often exceeds the retail price, and utility companies operate at a loss during these hours
to keep the lights on. To address variability, regulators and utilities use a number of “demand-side
management" tools to induce reductions in electricity demand when the price is high. One of the
most controversial of these tools is wholesale demand response.

Wholesale demand response (hereafter “demand response") is a Federal Energy Regulatory
Commission (FERC) program that allows an end user or middleman to sell a reduction in electricity
use on the wholesale electricity market as if that reduction was generation. In a typical wholesale
electricity market, renewable and fossil-fuel generators make bids to supply megawatts of electricity
in an hour as long as the wholesale price exceeds the generator’s reserve price. The market operator
then accepts the cheapest generation required to meet demand in the hour and pays all suppliers the
bid of the marginal generator. With demand response, an end user or middleman can participate
in the wholesale market by submitting a bid to supply “negawatts" or reductions in electricity
use in an hour.
If the demand response bidder’s reserve price is less than the market-clearing
wholesale price, the bid is accepted, and the bidder must curb electricity use by the accepted
amount or pay a fine. The utility company pays the bidder the wholesale price multiplied by the

39

amount of reductions supplied. The logic is that it is cheaper to pay for reduced electricity demand
than it is to produce these marginal units during peak periods when energy production would be
inefficient and expensive. The demand response compensation scheme was the subject of a 2016
Supreme Court case in which the Court ultimately ruled it did not have the expertise to question
FERC’s demand response compensation rate, and that FERC was justified in maintaining a demand
response program “that no one...disputes will curb prices and enhance reliability in the wholesale
market" (Christiansen, 2016). Certainly no one disputes the benefits of a program that properly
passes wholesale prices to electricity consumers; however, institutional design details dictate the
incentives and the ultimate costs and benefits of the program.

In this paper, we analyze electricity consumers’ incentives to contract with middlemen and
reduce electricity use through the demand response program. By mandating that utilities com-
pensate reductions at the wholesale price, FERC allows real-time prices to pass through to retail
consumers during peak hours; however, the economic incentives created by the program are not
well understood. The Supreme Court’s ruling upheld FERC Order 745, which mandates that utili-
ties compensate reductions in electricity at the wholesale electricity price. Thus, for an electricity
consumer participating in demand response, the opportunity cost of consuming electricity includes
both the retail rate and the wholesale price of electricity. However, most consumers do not di-
rectly participate in demand response. Most demand response is contracted through a middleman
aggregator called a “curtailment service provider."1 The aggregator contracts with any number
of electricity consumers and can be any entity participating in the wholesale electricity market,
including utilities or third-party firms. Previous work on demand response ignores the role of the
aggregator and relies on informal conceptions of how demand response operates.2

1For example, in the first 9 months of 2018 over 80 percent of demand response came from four

curtailment service providers (Monitoring Analytics, 2018).

2Eryilmaz et al. (2017) provide some suggestive evidence that wholesalers may be more price-
responsive during demand response periods but they ignore that wholesalers who do not reduce
their demand during demand response periods pay a higher effective price because they forgo
compensation for each unit of electricity purchased. Walawalkar et al. (2008) build a model of
demand response that does not include the retail rate and ignores the effect of aggregators. O’Connel
et al. (2014) review the state of demand response and note a possible effect of aggregators, but do

40

To understand the incentives created by demand response, we build a theoretical model of
the PJM electricity market with demand response. Under minimal assumptions, we show that
incentives for electricity use are not surplus-maximizing under the current demand-response rules.
In addition, we show that direct participation, third-party aggregation, and utility aggregation create
different (and inefficient) electricity-use outcomes. When end users participate directly in demand
response, the combined opportunity cost of the retail rate and wholesale price creates an incentive
to provide too many reductions. When there is an aggregator, only some of the incentive to reduce
is passed through to the end user. The degree of pass through depends on the demand elasticity and
which entity is the aggregator. Because the utility loses the retail rate for each unit of electricity
not sold, the utility has a lower incentive to seek demand-response reductions than a third-party
aggregator who does not forgo retail revenue.

Next, we compare the relative size of the economic surplus losses under different regimes.
The effect on surplus depends on the slope of the electricity demand curve and the difference
between the retail electricity rate and the wholesale electricity rate. Using a linear supply and
demand framework, we show that when demand is less elastic and the retail rate is low, third-
party aggregators perform better than the other second-best alternatives. When demand is more
elastic and the retail rate is higher, utility aggregators perform better than the other second-best
alternatives. We also show that direct participation can reduce surplus relative to the no-demand-
response baseline.

Finally, we suggest that to maximize surplus, only utilities should participate in demand re-
sponse. Because the utility is the residual claimant of surplus from load reductions, the utility is the
best steward of demand response. In addition, regulators should recognize that demand response
is a channel through which to exercise market power. A profit-maximizing utility does not have an
incentive to pass on the surplus-maximizing compensation for reductions.
not dwell deeply on the subject.

41

3.2 Institutional background

Wholesale demand response is a load reduction program that is connected to the wholesale
market and receives a price formally tied to the current wholesale price of electricity.
In the
PJM service area, wholesale demand response is designated as either “emergency" or “economic."
Emergency demand response is actively called upon by the market operator (PJM) during a peak
demand event. Economic demand response allows electricity users to submit reductions from
baselines in electricity use on the wholesale electricity market as generation (PJM, 2014a). Retail
demand response is a load reduction program operated by a utility that is not formally connected to
the wholesale market and is similar to the interruptible load studied by Caves et al. (1988, 1992).
In this paper, we focus on wholesale economic demand response as implemented in PJM. Further
discussion of demand response refers to the wholesale “economic" variety.

Without demand response, utilities buy electricity from generators on a regional real-time
wholesale market and re-sell it to electricity consumers at a regulated rate. On the wholesale
side of the market, renewable- and fossil-fueled generators submit bids to generate megawatts of
electricity at a minimum reserve price in each hour. PJM orders these bids from lowest to highest
cost, creating an electricity wholesale market supply curve. On the retail side of the market, end
users pay a regulated price for electricity determined by state regulatory agencies. These retail rates
are typically disconnected from the real-time marginal cost of generating electricity and instead
reflect regulatory desire for low and stable retail prices with a reliable grid. The utility is compelled
by regulation to purchase electricity on the wholesale market sufficient to supply any level of
demand at the fixed retail rate. The wholesale price of electricity the utility pays is equal to the bid
of the last unit of generation required to meet demand. The utility is required to meet this demand
even when the wholesale price of electricity is higher than the retail rate. Thus, end users do not
have any extra incentive to conserve electricity during peak hours and subsequently demand more
electricity than the surplus-maximizing level.

Demand response is designed to pass on real-time wholesale price fluctuations to end users.
The idea is to subsidize reductions in electricity use by allowing end users to participate directly in

42

the wholesale market. This raises the end user’s per-unit opportunity cost of consuming electricity
without raising retail rates—raising retail rates is outside of FERC’s jurisdiction (Christiansen,
2016). Via demand response, end users participate in the wholesale market by submitting reductions
in electricity use (“negawatts") as bids on the wholesale market. PJM treats a demand response bid
exactly like an energy supplier’s bid to generate electricity. The demand-response bid consists of an
hourly marginal cost curve of reductions. If the market-clearing price exceeds the price of the bid
for the demand-response reduction, PJM accepts the bid, and the bidder must provide the reductions
within 20 percent of the amount accepted. The utility then must compensate accepted bids at the
locational marginal price of electricity (the wholesale price plus delivery cost) plus shutdown costs.
If the reduction is outside of the 20 percent window, the bidder forfeits any shutdown compensation,
is charged a “balancing operating reserve charge," and receives the locational marginal price of
electricity multiplied by the actual reductions. The utility pays for the demand-response reductions
and no longer purchases actual generation on the wholesale market (PJM, 2013).

PJM determines what qualifies as a reduction by calculating the baseline electricity demand
ahead of time. There is an initial certification process during which the demand response provider
submits hourly meter data from 30 of the past 60 days. Once certified, the baseline load adjusts for
each demand response event based on the electricity use within the past five days and on the day
of the event (PJM, 2014b). Baseline manipulation is possible both during the certification process
and during the same day adjustment (Kema, 2011; Chen and Kleit, 2016).

The market operator only accepts demand response bids when the locational marginal price of
electricity exceeds a “net benefits threshold" set by PJM. The net benefits threshold is determined
monthly by PJM and is meant to represent the price above which reductions in electricity use is
worth soliciting (Rich, 2017). Between April 2012 and January 2013, the wholesale price exceeded
the net benefits threshold for 75 percent of all hours (PJM, 2013).

Reductions must come from electricity end-users with an interval meter on the PJM market;
however, any individual or entity is eligible to submit bids of reduction on the PJM wholesale market
(PJM, 2014b). PJM refers to bidders as “curtailment service providers." In theory, curtailment

43

service providers can be the electricity end-users themselves, utilities, or any third party entity.
In practice, many third party firms have emerged as aggregators of demand response reductions
and mainly handle establishing baseline load and submitting bids (McAnany, 2018). While PJM
compensates the curtailment service providers at the locational marginal price, contracts between
the curtailment service provider and the electricity consumer can take any form (PJM, 2017). We
examine these contracts formally.

3.3 A model of demand response incentives

Here we develop an analytical framework to study the differing incentives of demand response

aggregators.

Consider an electricity utility purchasing electricity generation on a competitive wholesale
electricity market and selling electricity to consumers at a retail opportunity cost of pr. In the
wholesale electricity market, generators supply electricity at marginal cost and get paid the cost
of the last accepted bid to generate electricity. Denote the wholesale market-clearing price as pw
and the wholesale market-clearing price without demand response as p0
w. In the retail electricity
market, the electricity price is constrained to a fixed retail rate r by regulation, so without demand
response, the opportunity cost of electricity is the retail rate: pr = r. During peak hours without
demand response, the wholesale price is higher than the retail price p0
w > r, so the utility operates
at a loss during those hours.3

We denote D(p) the aggregate electricity demand at opportunity cost p in an hour and S(p) the
electricity supplied at opportunity cost p in an hour.4 Let MB(q) be the corresponding aggregate
marginal benefit from consumption of q units of electricity and MC(q) be the corresponding
aggregate marginal cost of generating q units of electricity. The retail price cannot fluctuate and
thus the inelastic aggregate hourly electricity demand is I = D(r), but the wholesale price paid
by the utility increases to induce supply that meets demand. Regulation compels the utility to

3In practice, state regulators allow the utility to recover costs through a flat fee per customer.
4Aggregate demand is downward sloping in p: ∂D

∂p > 0. Supply is upward sloping in p: ∂S

∂p > 0.

44

supply any level of quantity demanded at the regulated price. Thus, the equilibrium wholesale
price equates supply with demand D(pr) = S(pw), and the market-clearing wholesale price without
demand response p0

w is found by solving D(r) = S(p0

w) for p0
w.

The total surplus in the electricity market is the total benefit to consumers minus the total
cost of generation at the quantity demanded at the retail opportunity cost D(pr). Define the total
surplus function W(D(pr)) in the electricity market as total benefits minus total generation costs of
electricity generation:

W(D(pr)) =

MB(φ) − MC(φ)dφ.

(3.1)

Þ D(pr)

0

Total surplus is maximized when the marginal benefit of electricity consumption equals the marginal
cost, which is achieved when retail consumers face the same opportunity cost as wholesale suppliers.
Thus, the surplus-maximizing electricity consumption is at D(pw), and the retail opportunity cost
equals the market-clearing wholesale price of electricity p∗
r = pw. When the retail opportunity cost
is set equal to the wholesale price, the market-clearing wholesale price equals the retail opportunity
cost.

We examine demand response incentives for electricity consumers and aggregators during peak

hours.

First consider an electricity consumer i consuming electricity in a peak hour.5

In the hour,
customer i has an underlying demand curve Di(p) with corresponding marginal benefit MBi(q).6
Because the retail price is fixed, inelastic demand for electricity is Ii, which is determined by the
underlying demand at the fixed retail price: Ii = Di(r). When the wholesale price exceeds the retail
price, the consumer uses too much electricity relative to the surplus-maximizing level q∗
= Di(p∗
w).
i
Now consider a demand-response program designed to incentivize reductions in electricity
consumption. The demand response program consists of a compensation payment to the consumer
Ii − qi.7 The cost
c > 0 per unit of reductions from baseline electricity consumption, i.e.

5The electricity consumer could be an industrial, commercial, or residential customer.
6Aggregate demand is the summation of individual demand curves in the market: D(p) =


i Di(p).

7We describe how c is set and who pays later.

45

of consuming q units of electricity in the hour is then rqi − c(Ii − qi).8 The corresponding
marginal opportunity cost of electricity consumption is then the retail rate plus the demand-
response payment, i.e. r + c. When choosing how much electricity to consume, the electricity
consumer sets the marginal benefit from electricity consumption equal to the marginal opportunity
= Di(r + c). Thus, a demand-response incentive equal to c∗ = pw − r
cost and consumes qDR
leads the consumer to respond to the wholesale price and demand the surplus-maximizing amount
of electricity qDR

= Di(p∗
w).

(c∗) = q∗
i

i

i

The demand-response program lowers the equilibrium wholesale price of electricity. With
demand response, the aggregate demand for electricity is D(r + c). The equilibrium wholesale
price of electricity with demand response pDR
w clears the market so that all demand is supplied:
w ). Because demand is downward sloping and supply is upward sloping, the
D(r + c) = S(pDR
equilibrium wholesale price with demand response is lower than without, i.e., pDR
w. With the
optimal demand-response compensation rate, c∗ = pw −r, the equilibrium wholesale price satisfies
D(p∗

w) = S(p∗
Figure D.1 plots the aggregate electricity supply and demand curves. Maximum surplus is
at point q∗. Area 1 corresponds to the surplus loss in the no-demand-response baseline where
consumers demand I units of electricity and the market-clearing wholesale cost is p0
w. Area 2
corresponds to the surplus loss for a demand response compensation level c > pw − r that reduces
demand past the surplus-maximizing level.

w) and maximizes total surplus.

w < p0

In practice, FERC requires that the utility pays pw per unit of accepted demand-response
reductions. Further, the compensation rate passed through to the final electricity consumer depends
on whether the consumer participates directly in the demand response market or through an
aggregator. In the next sections, we analyze how aggregators set the compensation rate c and the
differing incentives for electricity use when consumers participate directly, through a third-party
aggregator, or through a utility aggregator. We analyze the impact of each case on the equilibrium
wholesale price of electricity. Finally, we compare the ultimate effect on total surplus in each case.
8This only holds for qi ≤ Ii, but we omit this condition from the remaining notation for

simplicity.

46

3.3.1 Direct retail consumer participation

Given demand response, a retail consumer i can sell reductions in electricity use from baseline
consumption Ii and directly receive compensation c1 = pw per unit of reductions. Thus for a direct
participator, the marginal opportunity cost of electricity use is the retail rate plus the wholesale
price pr = r + pw. The consumer sets marginal cost equal to marginal benefit and consumes
q1 = Di(r + pw) < Di(pw). The consumer reduces electricity consumption too much because the
opportunity cost for consumption is artificially too high.

If all consumers participate directly, the aggregate electricity demand is downward sloping in
) =
). With direct participation, the equilibrium wholesale price of electricity drops too far

the wholesale price: D(r + pw). The equilibrium wholesale price then satisfies D(r + pdirect
S(pdirect
relative to the optimal wholesale price and too little electricity is consumed.

w

w

Figure D.2 is a graph with the wholesale price on the vertical axis and quantity of electricity on
the horizontal axis. The optimal amount of electricity consumption is at q∗ when D(pw) = S(pw).
As before, in the baseline without demand response, the surplus loss is equal to area 1. Under
direct participation in demand response, consumers always have an opportunity cost of the retail
rate r. Thus, the demand curve is shifted downward to D(pw + r).9 The intersection where
D(pw + r) = S(pw) is the equilibrium outcome. The equilibrium wholesale price decreases, but
the total demand-response compensation increases. Area 2 corresponds to the surplus loss under
direct participation in demand response.

The relative effect on total surplus depends on the slope of the electricity supply and aggregate
demand curves. To generate useful comparisons between the cases, we calculate total surplus using
linear specifications of supply and demand.10 In particular, let aggregate demand be D(p) = α− βp
and let the marginal cost of electricity supply be MC = γq.11 With the linear demand and supply

9When the wholesale price is zero, D(pw + r) = D(r) = I, which can be seen on the graph.
10This assumption is meant to generate comparisons. To the extent that electricity supply and
demand are locally linear this may be a good approximation to reality, but one should not read into
this assumption as empirically motivated.

11One may also specify MC = δ + γq, but for ease of explanation we normalize δ = 0 (or

equivalently the demand intercept includes the supply intercept).

47

specifications, total surplus W is

W =

(α − βpr)

1
2

(cid:19)

,

+ pr − pw

(cid:18) α

β

(3.2)

where pw = γ(α − βpr) and pr = r + c. Taking the first order conditions for maximization
and solving for c shows that the surplus-maximizing demand response compensation is c∗
(αγ)/(1 + βγ) − r with corresponding equilibrium wholesale price p∗
= (αγ)/(1 + βγ). In
other words, the optimum compensation eliminates the gap between the efficient wholesale price
and retail opportunity cost. When there is no demand response, the equilibrium wholesale price is
p0
w,linear

= γ(α − βr).

w,linear

=

linear

We return to the case of direct participation by end consumers. With linear aggregate demand,
= (γα − βγr)/(1 + βγ). Thus, the equilibrium demand-
the equilibrium wholesale price is pdirect
response compensation is c1,linear = (γα− βγr)/(1 + βγ). The equilibrium compensation is larger
than the optimum by r/(βγ +1). Thus, the distortion is increasing in the retail rate r and decreasing
in β and γ. We compare the relative surplus effects in section 3.3.4.

w

3.3.2 Third-party aggregator

Assume that a third-party aggregator (called a curtailment service provider) contracts with retail
consumers for demand-response reductions. The third-party aggregator acts as a middleman,
receiving pw per unit of reductions sold. The third-party aggregator offers consumers a payment
of c2 per unit of electricity reduced from the baseline and each consumer uses q2,i = Di(c2 + r)
units of electricity.12 We assume that a monopolist aggregator takes the wholesale price as fixed
but can exert market power because it does not compete with other aggregators when setting the
demand-response compensation.13 The aggregator sets c2 to maximize the arbitrage profits:

c2 = arg max
c

(pw − c)(I − D(r + c)).

(3.3)

12We assume that there is a single aggregator for this exercise and therefore has market power in

the arbitrage.

13If there is perfect competition in the aggregation arbitrage market and zero cost of aggregation

services, then the third party aggregator case is equivalent to the direct participation case.

48

The arbitrage profits are less than or equal to zero when c ≥ pw and when c ≤ 0. Thus,
0 < c2 < pw and electricity demanded is q2 < D(r + c2) < I. The first-order condition for equation
3.3 determines the profit-maximizing c2:

−pwD(cid:48)(r + c) = I − D(r + c) − cD(cid:48)(r + c).

(3.4)

The left-hand side of equation 3.4 is the marginal arbitrage revenue and the right-hand side is the
marginal arbitrage cost. Thus, the aggregator sets marginal revenue equal to marginal cost when
w satisfies the condition D(r + c2) =
choosing c2. Finally, the market-clearing wholesale price p3rd
w ). This is a downward shift in the demand curve, so the equilibrium wholesale price decreases
S(p3rd
relative to the zero-demand response case.

The relative effect on total surplus depends on the slope of the electricity supply and aggregate
demand curves. We again consider a linear aggregate demand of D(p) = α − βp and linear
marginal cost of electricity supply of MC = γq in order to compare the size of the distortion.
Directly applying the result from the first-order condition in equation 3.3, the aggregator sets
c2 = pw
2 . Combining the first-order condition with the equilibrium condition, the equilibrium
= (γα − βγr)/((βγ)/2 + 1). Thus, the equilibrium demand response
wholesale price is p3rd
compensation is c2,linear = 0.5(γα − βγr)/((βγ)/2 + 1). We compare the relative surplus effects
in section 3.3.4.

w,linear

3.3.3 Utility aggregator

Assume that the utility contracts with consumers for demand response reductions. The utility saves
pw − r for each unit of electricity reduced from the baseline (because it no longer has to purchase
it on the wholesale market but also no longer sells it on the retail market) and offers a payment to
electricity consumers of c3 per unit of reductions. The utility sets c3 to maximize savings from

49

participating:14

c3 = arg max
c

(pw − r − c)(I − D(r + c))

s.t. pw = MC(D(r + c)).

(3.5)

(3.6)

Note that the utility is a large enough player to know how the reductions will influence the
marginal cost of electricity generation. The savings function in equation 3.5 makes it clear that
0 < c3 < pw−r. If c3 ≤ 0, electricity consumption will not fall and the utility does not save money.
If c3 = pw − r, the utility passes all the savings on to the electricity consumer, and if c3 > pw − r,
the utility pays out more than the marginal benefit for reductions. Thus, the utility pays too little
for reductions relative to the surplus maximizing level. The first-order condition for optimization
determines the profit maximizing c3:

(r − pw)D(cid:48)(r + c) + (I − D(r + c3)) MC(cid:48)(D(r + c3))D(cid:48)(r + c3)
= (I − D(r + c3)) − cD(cid:48)(r + c)

(3.7)

The left-hand side of equation 3.7 is the marginal revenue and the right-hand side is the marginal
cost. Thus, the utility sets marginal revenue equal to marginal cost when choosing c3.

The relative effect on total surplus depends on the slope of the electricity supply and demand
curves. We again consider a linear aggregate demand of D(p) = α − βp and linear marginal cost of
electricity supply of MC = γq in order to compare the size of the distortion. Applying the result
from the first-order condition in 3.7, the utility sets c3 = 0.5((γα)/(1+ βγ)−r). Combining this with
= 0.5γ(α−βr +(α)/(1+βγ)).
the equilibrium condition, the equilibrium wholesale price is p3rd
In the next section, we compare the relative surplus effects.

w,linear

3.3.4 Comparing total surplus in the linear case

We now compare the total surplus from each regime in the linear case. The equilibrium deadweight
loss from retail opportunity cost of electricity pr = r + c is

DW L =

|D(pr) − D(p∗

w)||pw − pr|,

1
2

(3.8)

14This is equivalent to minimizing costs.

50

where pw = MC(D(pr)). Substituting the equilibrium compensation levels and wholesale prices
from each of the cases, we find that the deadweight losses are:

(cid:18)
(cid:18)
(cid:18)
(cid:18)

β

2 + 2βγ

β

2 + 2βγ

β

2 + 2βγ

β

(cid:19)
(cid:19)
(cid:19)(cid:18)1
(cid:19)(cid:18)1

4

(cid:19)(cid:32)
(cid:19)

2 + 2βγ

4

DW LNoDR
linear

=

DW Ldirect
linear

=

DW L3rd

linear

=

DW LUtilit y
linear

=

(αγ − r − βγr)2 ,
(r)2 ,

αγ − 2r − 2βγr

1 + βγ
2

(αγ − r − βγr)2 .

(cid:33)2

,

(3.9)

(3.10)

(3.11)

(3.12)

These deadweight losses are difficult to compare analytically, but are easy to compare compu-
tationally. First, we consider 100 retail rates r ∈ [40, 100], the equivalent to a typical PJM retail
rate between 40 and 100 dollars per megawatt hour. We set α = 120, 000 and use 100 values of
β ∈ [100, 350]. With these parameters, the demand elasticity at the retail price ranges from -0.43 to
-0.03, covering a wide range of electricity demand elasticities estimated in the literature.15 Without
demand response, total load ranges from 85,000 to 116,000 megawatt hours, roughly matching
the above-average (or peak) real-time loads experienced in PJM from 2010-2017 (Bowring, 2018).
For supply, we set γ = 0.0011, leading to wholesale prices ranging from 77 to 104 dollars per
megawatt hour which matches the above-average (or peak) wholesale electricity prices seen in PJM
(Bowring, 2018).

First, we plot the deadweight losses by regime type as we vary the retail rate for a fixed level of
the slope of demand, β. Figure D.3 shows how the deadweight loss varies with r. As the retail rate
increases, the gap between the wholesale price and the retail price of electricity shrinks. Thus, the
no-demand-response case becomes less damaging as the retail rate increases. In the case of direct
participation by electricity consumers, as the retail rate increases, the total opportunity cost of
electricity overshoots the optimum level by a larger amount and increases losses. Similarly, when
15The most commonly estimated elasticities are yearly and typically range from -0.1 to -0.5 on
average (Deryugina et al., 2019; Reiss and White, 2005, e.g.). Long-run elasticities are typically
larger. Hourly elasticities are smaller, for example around -0.16 in Jessoe and Rapson (2014). The
chosen values are meant to span the feasible set of elasticities.

51

a third-party aggregator sets demand-response contracts, the total opportunity cost of electricity
for the consumer is too high and increases too quickly. The deadweight loss increases at a slower
rate than in the direct-participation case because the third-party aggregator reduces the demand-
response compensation as the retail price increases. Finally, when the utility is the aggregator, the
deadweight loss decreases as the retail rate increases. Similar to the third-party-aggregator case,
the utility reduces the demand-response compensation as the retail rate increases.

Next, we plot the deadweight losses by regime type as we vary the demand elasticity parameter
β. Figure D.4 shows how the deadweight loss varies with β. In the no-demand-response baseline,
deadweight losses are increasing and then decreasing in elasticity. When demand is perfectly
inelastic, there is no deadweight loss. When electricity consumers directly participate in demand
response, the losses are increasing in elasticity because the consumers respond more strongly to
the artificially-inflated opportunity cost of electricity consumption. This is much the same story
when there is a third-party aggregator—increased price responsiveness increases the distortion.
When the utility is the aggregator, the deadweight loss exhibits a quadratic shape similar to the
no-demand-response baseline.

Finally, we examine the relative sizes of the deadweight losses from the regimes over all the
parameter values tested. Figure D.5 is a set of heat maps of the deadweight losses from the regimes
over the tested values of the retail rate and demand parameter. Each horizontal axis is the range of
retail rates and each vertical axis is the range of demand parameters β. The darker colors are higher
deadweight losses and the lighter colors are smaller deadweight losses. The gray area on the heat
maps is the set of elasticity/retail price combinations at which the wholesale price is less than the
retail rate.16

No regime dominates. Over the range of tested values, the third-party and utility aggregators
outperform the no-demand-response and direct-participation cases. The utility-aggregator case has
the lowest average deadweight loss overall, though when elasticity is low and retail rates are low,
the utility aggregator performs better.

16This occurs when the retail rate is high and demand is more elastic. When demand is more

elastic, the wholesale price is lower and thus closer to the retail price.

52

The no-demand-response baseline heat map illustrates that without demand response, the dead-
weight loss is larger when the retail rate is low and elasticity is high. When retail rates are low,
electricity consumers do not have a strong incentive to conserve electricity, and thus surplus is lost
from too much electricity consumed. However, the surplus loss from a too-low retail rate is smaller
if the consumer is less price elastic because behavior under the surplus-maximizing price is not
that different.

When consumers participate directly in demand response, the deadweight loss is larger when
the retail rate is high and elasticity is high. Because directly-participating consumers are overcom-
pensated for reducing consumption by the retail rate, the size of this distortion directly increases in
the retail rate. When consumers are more elastic, they respond more to the improper incentive and
reduce their electricity below the surplus-maximizing level even more.

The third-party aggregator’s heat map shows that the deadweight loss is larger when the retail
rate is high and elasticity is high. The third-party aggregator’s pricing incentives face two partially-
offsetting distortions. First, the third-party aggregator’s marginal revenue per unit of reduction
includes the full wholesale price rather than the difference between the wholesale price and the
retail price, creating an incentive to over-provide reductions. Thus when the retail price is higher,
the marginal revenue distortion is larger. Second, the third-party aggregator’s market power creates
an incentive to reduce reductions by not passing through the full demand-response incentive. This is
easily seen from the aggregator’s profit function in equation 3.3. If the firm compensates consumers
at a full pw per unit of reductions, profit goes to zero.

Finally, the utility aggregator’s heat map shows that the deadweight loss is larger when the
retail rate is small and customers are more elastic. In this case, the utility’s marginal revenue is
pw − r and the remaining distortion is because the savings from demand response are zero if the
utility passes on the full compensation to consumers. The utility as aggregator is the only case
in the current demand-response formulation in which surplus is guaranteed to increase from the
zero-demand-response baseline (easily seen by noting the deadweight loss in this case is always
one-fourth the deadweight loss without demand response in equations 3.9 and 3.12).

53

3.4 Can FERC improve demand response?

The analysis above shows that current wholesale demand-response institutions fail to incen-
tivize surplus-maximizing electricity consumption.
In fact, unless the utility is the aggregator,
offering demand response is not even guaranteed to increase electricity market surplus. The
FERC-mandated demand-response program requires utilities to compensate third-party and direct-
participating consumers for reductions in load. Thus, while wholesale demand response passes
real-time wholesale price variation on to electricity consumers, the institutional design creates poor
incentives for electricity consumption.

Even without a FERC mandate, utilities have an incentive to enroll customers in programs such
as retail demand response or other time-varying price programs. Utilities supply demand at fixed
retail prices, so the utility is the residual claimant for any programs that increase the opportunity
cost of electricity consumption during peak hours. The only remaining inefficiency is due to
market power, which is currently dealt with through rate-of-return regulation or retail choice on a
state-by-state basis.

This begs the question: why even have wholesale demand response? Proponents of wholesale
demand response often argue that the program provides a path for behind-the-meter renewable
generation to participate in wholesale markets (O’Connel et al., 2014). While perhaps true,
allowing negawatts to be sold as megawatts at the same time reduces efficiency. If FERC wants to
provide a path for distributed generation, it should pursue this goal directly through the wholesale
market rather than through a convoluted demand-response program.

3.5 Conclusion

In this paper, we demonstrate that the aggregator is a potential source of inefficiency in FERC’s
wholesale demand response program. Using a theoretical model, we show that direct participants,
third-party aggregators, and utility aggregators have different incentives to provide reductions
in electricity use.
In the current structure of demand response, direct particpants over-provide
reductions because the opportunity cost includes the retail rate plus the wholesale-price incentive.

54

Third-party aggregators similarly have an incentive to over-provide reductions because of the
compensation scheme but also have an incentive to under-provide reductions by exercising market
power. Utility aggregators have an incentive to under-provide reductions if the utility can exert
market power.

We argue that the wholesale demand-response institution is inefficient and superfluous to retail
demand-response programs. The compensation scheme does pass on some real-time-wholesale-
price variation to electricity consumers, but the pass-through does not create efficient electricity-use
incentives and can even reduce economic surplus. The primary beneficiary of demand-response
pricing is the utility, which can adopt retail demand-response or time-varying pricing programs
separately from the wholesale market. Thus, we recommend that FERC end its wholesale demand-
response program in favor of retail demand-response programs.

55

APPENDICES

56

APPENDIX A

FIGURES AND TABLES FOR CHAPTER 1

Figure A.1: Average temperature setting by year compared between landlord-pay and tenant-pay
regimes. The tenant-pay temperature settings statistically vary between sample years, possibly
due to changes in energy prices and resulting behaviors. This figure does not include households
that turn off the heat completely: 28 (3.1%) landlord-pay households and 244 (6.8%) tenant-pay
households.

57

Figure A.2: Average temperature setting when out of the house between landlord-pay and tenant-
pay regimes. This figure does not include households that turn off the heat completely: 28 (3.1%)
landlord-pay housholds and 244 (6.8%) tenant-pay households.

58

Figure A.3: Time series variation in the average ratio of electricity to natural gas price per BTU.

59

Figure A.4: Plots of estimated average markup for having the landlord-pay heating ∆p
by census
∆Rj
division from the preferred correlated-random-effects regression in table A.3. 95 percent confidence
intervals are derived using cluster-robust standard errors.

60

Figure A.5: Plots of estimated average hedonic price for 100 square feet ∂ ˆp
by census division from
∂xj
the preferred correlated-random-effects regression in table A.3. 95 percent confidence intervals are
derived using cluster-robust standard errors.

61

Figure A.6: Density of bliss point temperature settings. Does not include 11 (1.2%) households
that did not use heat in the winter. Plotted against a normal distribution.

62

Figure A.7: The LASSO-predicted temperature settings when price is zero (bliss point) for landlord-
pay and tenant-pay households. Landlord-pay households are predicted as having lower bliss point
temperature preferences.

63

0.02.04.06.08.1kdensity predictedbliss406080100xLL payers, price = 0Tenant payers, price = 0LASSO predicted bliss point by regimeFigure A.8: The implied temperature disutility parameters β2,i estimated from the LASSO regres-
sion.

64

0.2.4.6.8kdensity beta2_2−2−1.5−1−.50xLL payersTenant payersLASSO estimated disutility parameterFigure A.9: The LASSO-predicted effect on temperature setting of requiring landlord-pay house-
holds pay their own heating bills.

65

0.02.04.06.08.150607080xLL payers, price = 0LL payers, price > 0Landlord−pay household moral hazard (LASSO)Figure A.10: The LASSO-predicted effect on temperature setting of moving tenant-pay households
to a landlord-pay regime.

66

0.02.04.06.08.15060708090100xTenant payers, price = 0Tenant payers, price > 0Tenant−pay household moral hazard (LASSO)Figure A.11: The predicted temperature setting using LASSO versus the observed temperature
setting used to train the algorithm.

67

0.02.04.06.08.1405060708090xTemperature settingPredicted temperature setting (LASSO)LASSO goodness of fitFigure A.12: Estimated bliss points using the preferred β2 = −0.79 by regime.

68

Landlord payTenant pay0.02.04.06.08.1Density405060708090Temperature setting for no energy costLL payersTenant payersEstimated bliss point temperature preferenceFigure A.13: Kernel density of the distribution of estimated preferences for square feet by regime.

69

Landlord payTenant pay0.005.01.015Density0100200300400500Beta_1LL payersTenant payersEstimated preferences for square feetTable A.1: Residential Energy Consumption Survey sample means

Mean winter heat temp
Temp when home
Temp when gone
Temp at night
Household size
Age
Female
White
Black
Native American
Asian
Pacific Islander
Other
Multi-racial
Unemployed
Part-time
Income $0 to $4,999
Income $5,000 to $9,999
Income $10,000 to $14,999
Income $15,000 to $19,999
Income $20,000 to $29,999
Income $30,000 to $39,999
Income $40,000 to $49,999
Income $50,000 to $74,999
Income $75,000 to $99,000
Income $100,000 or more
New England
Middle Atlantic
East North Central
West North Central
South Atlantic
East South Central
West South Central
Mountain
Pacific

(1)

(2)

Tenant-pay renters Landlord-pay renters
Mean
67.36
70.12
66.38
68.59
2.65
40.56
0.43
0.68
0.20
0.02
0.04
0.00
0.03
0.02
0.29
0.17
0.05
0.08
0.09
0.09
0.17
0.15
0.12
0.14
0.06
0.05
0.05
0.07
0.09
0.10
0.16
0.06
0.14
0.09
0.24

Std dev Mean
67.40
(5.88)
(4.91)
70.34
68.02
(7.02)
68.84
(5.83)
(1.57)
2.09
46.59
(15.97)
0.40
(0.49)
0.70
(0.47)
(0.40)
0.19
0.01
(0.13)
0.05
(0.20)
0.00
(0.07)
(0.17)
0.04
0.01
(0.15)
0.39
(0.45)
(0.38)
0.19
0.09
(0.22)
0.16
(0.28)
0.16
(0.29)
(0.28)
0.10
0.16
(0.38)
0.11
(0.35)
0.08
(0.32)
(0.35)
0.09
0.03
(0.24)
0.03
(0.21)
(0.23)
0.14
0.21
(0.26)
0.17
(0.29)
0.09
(0.30)
(0.37)
0.10
0.03
(0.23)
0.10
(0.34)
0.07
(0.28)
(0.43)
0.11

Std dev
(7.07)
(4.67)
(5.94)
(5.69)
(1.38)
(20.16)
(0.49)
(0.46)
(0.39)
(0.08)
(0.22)
(0.05)
(0.19)
(0.12)
(0.49)
(0.39)
(0.28)
(0.37)
(0.36)
(0.29)
(0.36)
(0.32)
(0.27)
(0.29)
(0.16)
(0.17)
(0.34)
(0.41)
(0.38)
(0.28)
(0.30)
(0.17)
(0.30)
(0.25)
(0.31)

Observations
Sample means for tenant- and landlord-pay units using electric or
gas heating in the Residential Energy Consumption Survey. Standard
deviations in parentheses.

3,164

829

70

Table A.2: American Housing Survey sample means

(1)

Tenant-pay units
Mean
803.01
23.07
11.99
1098.00 (1133.66)

1098.00 (1133.66)

Monthly rent
Energy price per MMBTU
Freq rent payment
Sq ft
# Bedrooms
# Baths
# Rooms
Sq ft
Neighborhood rating
Unit age (years)
# units in building
Urban
Laundry
Central air
Furnace heat
Gas heat
Unit too cold last year

1.98
1.29
4.42

7.48
40.76
16.79
0.58
0.52
0.60
0.61
0.44
0.09

(2)

(0.84)
(0.35)
(1.22)

(0.89)
(0.49)
(1.32)

Landlord-pay units
Std dev
(401.02)
(10.18)
(3.42)

Std dev Mean
(388.74)
701.65
19.29
(10.45)
12.25
(1.62)
876.79 (1139.78)
1.50
1.09
3.70
876.79 (1139.78)
7.45
47.48
48.28
0.62
0.19
0.43
0.53
0.69
0.09

(2.02)
(23.14)
(49.95)
(0.49)
(0.50)
(0.49)
(0.49)
(0.50)
(0.28)

(2.16)
(21.73)
(89.50)
(0.49)
(0.39)
(0.50)
(0.50)
(0.46)
(0.29)

Observations
Sample means for tenant- and landlord-pay units using electric or
gas heating in the American Housing Survey. Standard errors in
parentheses.

17,371

88,700

71

Table A.3: Hedonic price estimates using AHS data

(1)
OLS

(2)
RE

(3)
FE

(4)
CRE

Y = rentj
Price, 100 sq ft
New England

Mid Atlantic

East North Central

West North Central

West South Central

South Atlantic/
East South Central
Mountain/Pacific

Landlord-pay markup
New England

Mid Atlantic

East North Central

West North Central

West South Central

South Atlantic/
East South Central
Mountain/Pacific

13.15

6.874

7.416

8.223

9.148

7.888

19.43

-79.09

-83.20

5.812

18.12

39.63

92.19

-159.8

[10.86,15.43]

[10.12,17.61]

[5.466,8.282]

[3.208,7.638]

[6.283,8.549]

[6.649,9.271]

[6.457,9.989]

[6.039,10.02]

[8.010,10.29]

[8.720,11.38]

[7.049,8.727]

[7.883,10.04]

[18.59,20.27]

[19.10,21.70]

13.87

5.423

7.960

8.030

10.05

8.961

20.40

14.78

-54.54

-11.72

12.97

81.98

115.0

-102.6

[9.044,16.33]

[3.025,7.483]

[5.690,8.223]

[5.499,9.353]

[7.065,9.616]

12.69

5.254

6.957

7.426

8.340

7.740

18.42

[6.671,8.809]

[17.08,19.76]

128.4

-19.69

-24.39

29.03

141.1

141.5

-41.19

[-192.2,34.05]

[-112.3,141.9]

[-69.39,316.4]

[-66.60,323.4]

[-178.4,12.02]

[-152.2,43.08]

[-160.0,106.1]

[-156.9,117.5]

[-56.59,68.22]

[-68.37,44.93]

[-92.44,50.18]

[-96.58,47.81]

[-63.72,99.96]

[-55.75,81.69]

[-58.80,121.0]

[-61.33,119.4]

[-53.41,132.7]

[-7.716,171.7]

[19.40,277.4]

[12.11,270.1]

[24.66,159.7]

[37.85,192.2]

[32.83,260.9]

[27.16,255.8]

[-209.0,-110.7]

[-162.4,-42.77]

[-118.2,42.79]

[-121.7,39.33]

123.5

-26.90

-21.13

31.08

148.4

146.9

-37.71

Y
Y

Y

Y

Y
Y

Y
Y

Y

Unit characteristics
Vintage × year
Unit FE/means
LL pay × year
Observations
95% confidence intervals in brackets constructed with cluster-robust standard errors.
Interpreta-
tion is the additional monthly rent for 100 additional square feet and the additional monthly rent
to have landlord-pay utilities (i.e., ˆα3,d ¯Pe,d), ceteris paribus. Unit characteristics controlled for
are bedrooms, bathrooms, air conditioning type, in-unit laundry, number of units in the building,
frequency of the rent payment, heating equipment type (e.g., furnace, electric room heaters, etc),
non-heating utilities that the landlord pays, whether the unit was too cold last year, a subjective rating
of neighborhood quality, an indicator for gas heat, MSA, and division interacted with fuel type.

107,725

107,725

107,725

107,725

Y
Y
Y
Y

72

Table A.4: Fuel use estimates using RECS data

Natural gas

Electric

Y = ln(Fuelquantit y)
IHS(Tempindoor − Tempoutdoor)

(1)
OLS
0.100

[0.071,0.13]

(2)
2SLS
0.49

(3)
OLS
0.019

(4)
2SLS
0.20

[0.39,0.59]

[-0.0029,0.042]

[0.079,0.33]

ln(Square feet)

ln(Units in building)

ln(Floors in building)

0.21

[0.14,0.28]

0.21

[0.14,0.29]

0.28

[0.21,0.36]

0.26

[0.14,0.39]

-0.14

[-0.18,-0.10]

-0.15

[-0.19,-0.10]

-0.057

[-0.081,-0.033]

-0.038

[-0.078,0.0015]

0.042

[0.0041,0.080]

0.027

[-0.018,0.072]

-0.0035

[-0.028,0.021]

-0.021

[-0.056,0.015]

1511
Observations
91.0
Mean heat bill ($/month)
Yes
Unit characteristics
Yes
Appliances
VintageXYear
Yes
Interpretation: For a 1% increase in x, there is a ˆb% increase in monthly heating bill for a fixed energy price, ceteris
paribus. Fuel prices at the division level and yearly heating and cooling degree days are instruments. IHS is the inverse
hyperbolic sine function, which approximates the natural log, but is defined for non-positive values. 95% Confidence
intervals calculated using heteroskedasticity-robust standard errors.

1511
91.0
Yes
Yes
Yes

1653
58.8
Yes
Yes
Yes

1653
58.8
Yes
Yes
Yes

73

Table A.5: Average marginal effects on monthly heating costs using RECS data

Heat setting - one degree change

Square feet - 100 sq ft change

Units in building - one unit change

Floors in building - one floor change

Natural gas

(1)

3.57

[2.76,4.37]

(2)

Electric

2.89

[1.06,4.71]

1.01

[0.63,1.39]

-5.86

[-7.71,-4.01]

1.11

[-0.80,3.02]

2.42

[1.31,3.53]

-1.74

[-3.61,0.13]

-1.18

[-3.25,0.90]

Observations
Unit characteristics
Appliances
VintageXYear
Interpretations: For a unit change in x, there is a ˆb dollar increase in monthly heating bill, ceteris
paribus. Estimates from the two-stage-least-squares specification in table A.4. 95% confidence intervals
are bootstrapped using 1,000 replications.

1511
Yes
Yes
Yes

1653
Yes
Yes
Yes

74

Table A.6: Temperature setting estimation using RECS data

Y = temp setting
ˆβ2
ˆδ
i |Rj∗ = 0
ˆsb
i |Rj∗ = 1
ˆsb

(1)
-0.56

-1.76

OLS

(3)
-0.56

-1.29

(2)
-0.57

-1.65

(4)
-0.53

-1.15

(5)
-0.98

-3.14

2SLS

(6)
-1.06

-3.16

(7)
-0.36

-0.69

(8)
1.44

4.92

[-0.73,-0.38]

[-0.75,-0.39]

[-0.73,-0.39]

[-0.69,-0.37]

[-9.43,7.46]

[-13.0,10.9]

[-2.86,2.14]

[-8.14,11.0]

[-2.36,-1.16]

[-2.24,-1.06]

[-1.89,-0.70]

[-1.73,-0.58]

[-32.1,25.8]

[-42.0,35.7]

[-8.94,7.56]

[-24.6,34.4]

67.4

67.4

67.4

67.4

67.4

67.4

67.4

67.4

[66.9,67.9]

[66.9,67.9]

[66.9,67.9]

[66.9,67.9]

[66.9,67.9]

[66.9,67.9]

[66.9,67.9]

[66.9,67.9]

69.2

69.2

69.2

69.1

70.5

70.8

68.5

62.7

[68.7,69.6]

[68.8,69.6]

[68.7,69.6]

[68.6,69.5]

[41.7,99.4]

[30.2,111.4]

[59.3,77.8]

[31.6,93.8]

3993

3993

3993

Observations
First stage F
HH characteristics
Division
Year
Estimation of equation 1.18. ˆβ2 is an estimate of the mean disutility parameter for deviation from the bliss point. Estimated using a two-step
estimator for marginal cost. Instrument for marginal cost in 2SLS estimates is division fuel price. 95% confidence intervals including the
sampling error for marginal cost bootstrapped with 1,000 replications. Sample of all renters from the RECS.

3993
35.3
Yes
Yes

3993
21.0
Yes

Yes
Yes
Yes

3993
29.6

Yes
Yes

3993

Yes

3993
4.96
Yes
Yes
Yes

75

Table A.7: Temperature setting estimation using choice experiment

ˆβ2

Constant

(1)
OLS
-0.76

[-1.08,-0.43]

-0.45

[-0.77,-0.14]

(2)
FE
-0.79

[-0.97,-0.61]

-0.42

[-0.60,-0.24]

1242

Observations
FE
Estimation of equation 1.23. 95 percent confidence intervals in brackets. Estimated on a sample of
choice experiment respondents for which hypothetical marginal cost is varied randomly. I conducted the
choice experiment in March 2018 using the Qualtrics Panel and a nationally representative sample of US
individuals.

1242
Yes

Table A.8: LASSO heterogeneity

OLS
72.01
Mean predicted bliss temp
Std dev.
5.56
Mean predicted tenant pay temp 67.31
Std dev.
4.44
-1.54
Mean implied β2
Std dev.
10.57

LASSO
67.82
4.39
67.31
4.24
-9.31
59.29

RMSPE
MSPE
Bias
Covariates
FinalCovariates
Observations
Mean-squared error, root-mean-squared error, and bias statistics calculated using ten-fold cross validation.

5.71
32.55
0.00
1590
1209
3993

8.73
76.15
-0.01
1590
1590
3993

76

Table A.9: Simulation results:

Effect of changing all landlord-pay contracts to tenant-pay

Reduction in energy expenditures per household, per year
Temperature reduction
Square feet reduction

6.54
139.57

(1)

4.15
139.57

(2)

Implied welfare loss from landlord pay per household, per year
Moral hazard DWL
Sorting DWL
Carbon damages

$12.33
$1.34
$5.46

$7.83
$1.34
$3.64

(3)

3.27
139.57

$6.17
$1.34
$2.97

Total US, per year average
US yearly welfare loss

$839,000,000 $1,260,000,000

$682,000,000

β2
Average household responses to eliminating moral hazard by requiring tenant-pay heating in all
units. Total welfare losses include deadweight losses from moral hazard, equilibrium sorting,
and external cost of carbon ($40/ton CO2) over all landlord-pay units using survey weights.

-1.00

-0.79

-0.5

77

APPENDIX B

SUPPLEMENTAL INFORMATION FOR CHAPTER 1

B.1 Science and engineering findings on temperature preferences

The science and engineering literatures characterize temperature preferences along three dimen-
sions: temperature sensitivity, temperature discomfort, and mean temperature preference (Kingma
and van Marken Lichtenbelt, 2015; Karjalainen, 2012; Indraganti and Rao, 2010; Schellen et al.,
2010). Temperature sensitivity is the ability to detect changes in temperature from a baseline.
Temperature discomfort is a subjective rating of how uncomfortable deviations from a bliss point
temperature are. Finally, mean temperature preference refers to the temperature bliss point itself. In
terms of the model in this paper, temperature sensitivity and discomfort are factors in the disutility
parameter (β2), while mean temperature preference is a measure of bliss point sb
i (where the indoor
temperature’s contribution to utility is (−1/β2)(si − sb

i )2).

Human temperature preferences depend on several physiological factors. Diets rich in Vitamin
C improve extreme heat tolerance in many studies (Ringsdorrf Jr. and Cheraskin, 1982). Older
individuals prefer higher temperatures than younger individuals (Schellen et al., 2010), though
the elderly find colder temperatures less unpleasant perhaps due to decreased sensitivity (Taylor
et al., 1995). The majority of scientific temperature studies conclude that women prefer higher
temperatures and are more sensitive to temperature extremes (Kingma and van Marken Lichtenbelt,
2015; Karjalainen, 2012).1

Culture may play a role in driving temperature preference differences between sexes. A study
of Finnish men and women found that while women report more discomfort than men in both low
and high temperatures, women are less likely to adjust the thermostat and are more likely to prefer
warmer settings in general (Karjalainen, 2007). In another study, Indian men and women exhibited
1See also Karjalainen (2007); Fanger (1970); Parsons (2002); Cena and de Dear (2001); Muzi
et al. (1998); Pellerin and Candas (2003); Griefahn and Künemund (2001); Nakano et al. (2002);
Nagashima et al. (2002)

78

a slightly different effect: while women were more sensitive to temperature changes than men, they
were less likely to report discomfort (Indraganti and Rao, 2010). This cross-country difference is
suggestive of a cultural role in temperature preference. In the United States, women express greater
temperature discomfort than men (Beshir and Ramsey, 1981). Another study examining more
general aspects of workplace environment such as air quality, social conditions, and noise levels in
addition to temperature finds that women report a greater number of work-related health symptoms
than men in general (Reynolds et al., 2001). Surprisingly, temperature was a statistically significant
determinant of workplace health problems for men and not women. The authors speculate that
these findings are likely driven by cultural attitudes of “stoic" behavior among men. In any case,
sex and culture are both important determinants of heat preference.

When exposed to heat and cold over time, the human body adapts physiologically to become
more tolerant of hot and cold temperatures (Young, 2010). Households living in colder climates will
likely be more tolerant of cool indoor temperatures. Through this channel, temperature preference
may also be related to income if low-income households develop a resilience to cold weather when
conserving on heating costs, or if low-income households tend to work or commute outdoors in cold
weather. Indraganti and Rao (2010) demonstrate an income-temperature effect for heat resilience
in India with less wealthy individuals reporting a higher comfort level in extreme summer weather
conditions. Thus, local climate and income will likely influence temperature preference.2

Households reveal bliss-point temperature preferences when heating without price constraints.
The science and engineering literature has found that physiological factors such as age, sex, and
previous temperature exposure influence temperature preferences. Other factors such as culture,
behavioral or psychological adaptation (de Dear and Brager, 1998), and idiosyncratic preferences
2In theory, this acclimatization process could be used as a “defensive" behavior to mitigate energy
expenditures. There is recent experimental evidence that suggests individuals can consciously alter
the body’s internal temperature response with training (see Kox et al.’s (2014) study of the “Wim
Hof method" in the Proceedings of the National Academy of Sciences). This relatively unknown
method involves bathing in temperatures near 32◦F over weeks to become acclimatized to low
temperatures. While this form of defensive behavior might be used in theory to reduce heating
costs, I do not think it is likely.

79

probably play a large role. If temperature preference is largely determined physiologically, then
preferences can be predicted using reported demographics and revealed heating choices.

B.2 Data cleaning notes

Vacation homes, vacant units, timeshares, tents, mobile homes, and hotel rooms are dropped
from the sample. This essentially leaves only traditional apartment units and rental homes in
the sample.
In addition, units where no rent is paid, the rent is reported to be adjusted due to
relationship with the owner, the frequency of the rent payment is not reported, and units where
the square footage is not reported are dropped from the sample. The American Housing Survey is
topcoded in rent price, so units with the top 3% of rents in the sample in the National sample are
dropped, and the top 3% of rents in each geographic area in the Metropolitan sample are dropped.

B.3 Alternative explanations

I provide descriptive evidence showing that landlord-pay households are in control of their
temperature settings. One alternative explanation for the lack of evidence that households select
into landlord pay regimes based on bliss point temperature preference is that thermostat settings
by landlord-payers may not reflect the true bliss point.
If the landlord controls the thermostat
in these regimes, interpreting thermostat settings as revealed bliss point is improper; however,
there is only limited evidence to support this hypothesis. In the 2015 version of the Residential
Energy Consumption Survey, respondents are asked whether they control their thermostat. 86%
of landlord-pay households report that they control the temperature setting in their home. The
difference in temperature settings between those in control and not in control is not statistically
different from zero. Given the small number of households and the lack of perceivable differences
in temperature settings, I conclude that it is not likely that this effect is driving the results.

80

APPENDIX C

FIGURES AND TABLES FOR CHAPTER 2

Figure C.1: Kernel density of participants’ bliss point temperature preferences and real temper-
ature settings. A Kolmogorov-Smirnov test of equivalence of distributions easily rejects the null
hypothesis that the distributions are equal.

81

Figure C.2: Scatterplot of bliss point temperature preferences and actual temperature settings with
a 45 degree line for reference. 54 percent of respondents set the thermostat equal to the bliss point.
6.5 percent reported setting the thermostat greater than the bliss point. Random noise has been
added to the data to show clustering on common temperature choices such as 70 degrees Fahrenheit.

82

Figure C.3: Histogram of estimated arc semi-elasticities (η).

83

Figure C.4: Histogram of semi-elasticities (η) estimated using OLS.

84

Figure C.5: The marginal effects from a Tobit estimation of the estimated arc semielasticities
on average price and participant demographics. The marginal effects from this estimation can
be interpreted as the change in arc semi-elasticity for a one-standard-deviation change in the
predictor variable holding the other predictor variables constant. 95 percent confidence intervals
are bootstrapped using 1000 replications with sampling at the participant level.

85

Table C.1: Participant sample mean and standard deviation

(1)
Mean

(2)

Std dev

Income
Monthly heat bill
Bliss point
Temperature at home
Household size
Age
Children
Female
Non-white
Urban
High school
Some college
College
Graduate degree
Republican
Democrat
Respondents
Observations
Sample means presented with sample standard deviations in parenthe-
ses.

(71251.68)
(145.81)
(3.59)
(3.71)
(2.85)
(16.60)
(1.11)
(.49)
(.45)
(.43)
(.48)
(.45)
(.39)
(.30)
(.45)
(.47)

79925.13
121.29
70.77
69.70
2.91
50.22
0.44
0.60
0.29
0.76
0.37
0.28
0.19
0.10
0.28
0.33
414
1242

Table C.2: Estimated semi-elasticities.

(2)

(3)

Elasticity Lower bound Upper bound

(4)
N
1242
OLS
1242
FE
1242
Arc semi-elasticity
Individual OLS
1242
The interpretation is for a 100 percent increase in price, the average participant will
reduce the thermostat by η degrees. 95 percent confidence intervals bootstrapped
using 1000 replications with sampling at the participant level.

-.73
-.69
-.97
-.86

-.31
-.43
-.43
-.46

(1)

-.52
-.56
-.69
-.66

86

APPENDIX D

FIGURES AND TABLES FOR CHAPTER 3

Figure D.1: Aggregate supply and demand curves for electricity consumption. Area 1 corresponds
to the surplus loss in the no-demand-response baseline and area 2 corresponds to the surplus loss
from a demand response compensation level greater than pw − r.

87

Figure D.2: The equilibrium outcome and surplus loss under direct participation. Area 1 corre-
sponds to the surplus loss under the no-demand-response baseline. When electricity consumers
particpate directly in demand response, the opportunity cost already includes the retail rate, so the
demand curve shifts downward to D(pw +r). The intersection of D(pw +r) = S(pw) determines the
market-clearing wholesale price under direct participation, which decreases. Area 2 corresponds
to the surplus loss under direct participation.

88

Figure D.3: Simulated deadweight loss by regime type plotted against the demand elasticity
parameter.

89

4045505560657075808590Retail rate012345678Deadweight loss105Deadweight loss as a function of retail rate No DR Direct Third LSEFigure D.4: Simulated deadweight loss by regime type plotted against the retail rate.

90

100150200250300350(Less elastic)   Elasticity parameter   (More elastic)0123456Deadweight loss105Deadweight loss as a function of elasticity No DR Direct Third LSEFigure D.5: Simulated deadweight loss for each regime type versus the retail rate and the elasticity
parameter. Darker colors indicate larger simulated deadweight losses. The gray regions indicate
where the wholesale price is lower than the retail rate.

91

Deadweight Loss ComparisonsNo DRDirectThird-partyLSELowHighLowHighRetail rateHighLowHighLowElasticitySmall DWLLarge DWLBIBLIOGRAPHY

92

BIBLIOGRAPHY

Aggarwal, R. M. (2007). Role of risk sharing and transaction costs in contract choice: Theory
and evidence from groundwater contracts. Journal of Economic Behavior and Organization,
63(3):475 – 496.

Allcott, H. (2015). Site selection bias in program evaluation. The Quarterly Journal of Economics,

130(3):1117–1165.

Allcott, H. and Greenstone, M. (2012). Is there an energy efficiency gap? Journal of Economic

Perspectives, 26(1):3–28.

Allcott, H. and Rogers, T. (2014). The short-run and long-run effects of behavioral interventions:
Experimental evidence from energy conservation. American Economic Review, 104(10):3003–
37.

Allcott, H. and Taubinsky, D. (2015). Evaluating behaviorally motivated policy: Experimental

evidence from the lightbulb market. American Economic Review, 105(8):2501–38.

Allen, D. and Lueck, D. (1992). Contract choice in modern agriculture: Cash rent versus cropshare.

The Journal of Law and Economics, 35(2):397–426.

Allen, D. W. and Lueck, D. (1995). Risk preferences and the economics of contracts. The American

Economic Review, 85(2):447–451.

Allen, D. W. and Lueck, D. (1999). The role of risk in contract choice. The Journal of Law,

Economics, and Organization, 15(3):704–736.

Allen, R. G. D. and Lerner, A. P. (1934). The concept of arc elasticity of demand. The Review of

Economic Studies, 1(3):226–230.

Auffhammer, M. and Rubin, E. (2018). Natural gas price elasticities and optimal cost recovery
under consumer heterogeneity: Evidence from 300 million natural gas bills. Technical report,
Energy Institute at Haas.

Autor, D., Duggan, M., and Gruber, J. (2014). Moral hazard and claims deterrence in private

disability insurance. American Economic Journal: Applied Economics, 6(4):110–41.

Baicker, K., Mullainathan, S., and Schwartzstein, J. (2015). Behavioral hazard in health insurance.

The Quarterly Journal of Economics, 130(4):1623–1667.

Bajari, P. and Benkard, C. L. (2005). Demand estimation with heterogeneous consumers and unob-
served product characteristics: A hedonic approach. Journal of Political Economy, 113(6):1239–
1276.

Bajari, P. and Kahn, M. E. (2005). Estimating housing demand with an application to explaining

racial segregation in cities. Journal of Business and Economic Statistics, 23(1):20–33.

93

BEA (2019). Gross domestic product [gdpa]. Technical report, United States Bureau of Economic

Analysis.

Belloni, A. and Chernozhukov, V. (2013). Least squares after model selection in high-dimensional

sparse models. Bernoulli, 19(2):521–547. arXiv: 1001.0188.

Beshir, M. and Ramsey, J. (1981). Comparison between male and female subjective estimates of

thermal effects and sensations. Applied Ergonomics, 12(1):29 – 33.

Borck, R. and Brueckner, J. K. (2017). Optimal energy taxation in cities. Journal of the Association

of Environmental and Resource Economists, 5(2):481–516.

Bowring, J. (2018). 2017 State of the market report for PJM. Technical report, Monitoring

Analytics, LLC presentation.

Brot-Goldberg, Z. C., Chandra, A., Handel, B. R., and Kolstad, J. T. (2017). What does a deductible
do? The impact of cost-sharing on health care prices, quantities, and spending dynamics. The
Quarterly Journal of Economics, 132(3):1261–1318.

Caves, D. W., Hanser, P., Herriges, J. A., and Windle, R. J. (1988). Load impact of interruptible
and curtailable rate programs: Evidence from ten utilities. IEEE Power Engineering Society
Transactions, 3(4):1757 – 1763.

Caves, D. W., Herriges, J. A., and Windle, R. J. (1992). The cost of electric power interruptions in
the industrial sector: Estimates derived from interruptible service programs. Land Economics,
68(1):42–61.

Cena, K. and de Dear, R. (2001). Thermal comfort and behavioural strategies in office buildings
International

located in a hot-arid climate. Journal of Thermal Biology, 26(4):409 – 414.
Thermal Physiology Symposium.

Chen, X. and Kleit, A. (2016). Money for nothing? Why FERC order 745 should have died. Energy

Journal, 37(2):201 – 222.

Christiansen, M. R. (2016). FERC v. EPSA: Functionalism and the Electricity Industry of the

Future. Stanford Law Review Online 100, 68.

Costa, D. L. and Kahn, M. E. (2013a). Do liberal home owners consume less electricity? A test of

the voluntary restraint hypothesis. Economics Letters, 119(2):210 – 212.

Costa, D. L. and Kahn, M. E. (2013b). Energy conservation “nudges" and environmentalist
ideology: Evidence from a randomized residential electricity field experiment. Journal of the
European Economic Association, 11(3):680–702.

Crawford, G. S., Pavanini, N., and Schivardi, F. (2018). Asymmetric information and imperfect

competition in lending markets. The American Economic Review. (forthcoming).

Davis, L. (2012). Evaluating the slow adoption of energy efficient investments: Are renters less
likely to have energy efficient appliances?, chapter 19. National Bureau of Economic Research
Conference Report.

94

de Dear, R. and Brager, G. S. (1998). Developing an adaptive model of thermal comfort and

preference. ASHRAE Transactions, 104:145–167.

Deryugina, T., MacKay, A., and Reif, J. (2019). The long-run elasticity of electricity demand: Ev-
idence from municipal electric aggregation. American Economic Journal: Applied Economics,
forthcoming.

Dewees, D. and Tombe, T. (2011). The impact of sub-metering on condominium electricity demand.

Canadian Public Policy / Analyse de Politiques, 37(4):435–457.

EIA (2013). Heating and cooling no longer majority of U.S. home energy use. Technical report,

United States Energy Information Administration (EIA).

EIA (2016). Carbon dioxide emissions coefficients by fuel. Technical report, United States Energy

Information Administration (EIA).

EIA (2018). Electric power annual 2017. Technical report, United States Energy Information

Administration (EIA).

Einav, L., Amy Finkelstein, S. R., Schrimpf, P., and Cullen, M. (2013). Selection on moral hazard

in health insurance. American Economic Review, 103(1):178–219.

Elinder, M., Escobar, S., and Petre, I. (2017). Consequences of a price incentive on free riding and
electric energy consumption. Proceedings of the National Academy of Sciences, 114(4):3091–
3096.

Eryilmaz, D., Smith, T. M., and Homans, F. R. (2017). Price responsiveness in electricity markets:

Implications for demand response in the midwest. Energy Journal, 38(1):23–49.

Fanger, P. (1970). Thermal comfort: Analysis and applications in environmental engineering.

Danish Technical Press.

Finkelstein, A., Gentzkow, M., and Williams, H. (2016). Sources of geographic variation in health
care: Evidence from patient migration. The Quarterly Journal of Economics, 131(4):1681–1726.

Gillingham, K., Harding, M., and Rapson, D. (2012). Split incentives in residential energy

consumption. The Energy Journal, 33(2).

Gillingham, K. and Palmer, K. (2014). Bridging the energy efficiency gap: Policy insights from
economic theory and empirical analysis. Review of Environmental Economics and Policy,
8(1):18–38.

Griefahn, B. and Künemund, C. (2001). The effects of gender, age, and fatigue on susceptibility
International Thermal

to draft discomfort. Journal of Thermal Biology, 26(4):395 – 400.
Physiology Symposium.

Harding, J., Miceli, T. J., and Sirmans, C. (2000). Do owners take better care of their housing than

renters? Real Estate Economics, 28(4):663–681.

95

He, J., Rejesus, R. M., Zheng, X., and Yorobe Jr, J. (2017a). Advantageous selection in crop

insurance: Evidence from the philippines. Journal of Agricultural Economics.

He, J., Rejesus, R. M., Zheng, X., and Yorobe Jr, J. (2017b). Moral hazard and adverse selection

effects of cost-of-production crop insurance: Evidence from the philippines. Working paper.

Hui, X., Saeedi, M., Shen, Z., and Sundaresan, N. (2016). Reputation and regulations: Evidence

from eBay. Management Science, 62(12):3604–3616.

Indraganti, M. and Rao, K. D. (2010). Effect of age, gender, economic group and tenure on thermal
comfort: A field study in residential buildings in hot and dry climate with seasonal variations.
Energy and Buildings, 42(3):273 – 281.

Ito, K. (2014). Do consumers respond to marginal or average price? Evidence from nonlinear

electricity pricing. American Economic Review, 104(2):537–63.

Jackson, C. K. and Schneider, H. S. (2011). Do social connections reduce moral hazard? Evidence
from the New York City taxi industry. American Economic Journal: Applied Economics,
3(3):244–67.

Jackson, C. K. and Schneider, H. S. (2015). Checklists and worker behavior: A field experiment.

American Economic Journal: Applied Economics, 7(4):136–68.

Jaffe, A. B. and Stavins, R. N. (1994). The energy paradox and the diffusion of conservation

technology. Resource and Energy Economics, 16(2):91–122.

Jessoe, K., Papineau, M., and Rapson, D. (2018). Utilities included: Split incentives in commercial

electricity contracts. Technical Report 029, E2e.

Jessoe, K. and Rapson, D. (2014). Knowledge is (less) power: Experimental evidence from

residential energy use. American Economic Review, 104(4):1417–38.

Karjalainen, S. (2007). Gender differences in thermal comfort and use of thermostats in everyday

thermal environments. Building and Environment, 42(4):1594 – 1603.

Karjalainen, S. (2012). Thermal comfort and gender: A literature review. Indoor Air, 22(2):96–109.

Kema (2011). PJM empirical analysis of demand response baseline methods. Technical report,

Kema, Inc.

Kingma, B. and van Marken Lichtenbelt, W. (2015). Energy consumption in buildings and female

thermal demand. Nature Climate Change, 5:1054–1056.

Kox, M., van Eijk, L. T., Zwaag, J., van den Wildenberg, J., Sweep, F. C. G. J., van der Hoeven,
J. G., and Pickkers, P. (2014). Voluntary activation of the sympathetic nervous system and
attenuation of the innate immune response in humans. Proceedings of the National Academy of
Sciences, 111(20):7379–7384.

Kuminoff, N. V., Smith, V. K., and Timmins, C. (2013). The new economics of equilibrium sorting
and policy evaluation using housing markets. Journal of Economic Literature, 51(4):1007–1062.

96

Leffler, K. B. and Rucker, R. R. (1991). Transactions costs and the efficient organization of
production: A study of timber-harvesting contracts. Journal of Political Economy, 99(5):1060–
1087.

Levinson, A. and Niemann, S. (2004). Energy use by apartment tenants when landlords pay for

utilities. Resource and Energy Economics, 26(1).

McAnany, J. (2018). 2017 Demand response operations markets activity report: April 2018.

Technical report, PJM Demand Side Response Operations.

Monitoring Analytics (2018). State of the market report for pjm: January through september.

Technical report, Monitoring Analytics, LLC.

Muzi, G., Abbritti, G., Accattoli, M. P., and dell’Omo, M. (1998). Prevalence of irritative symptoms
in a nonproblem air-conditioned office building. International Archives of Occupational and
Environmental Health, 71(6):372–378.

Myers, E. (2017). Are home buyers myopic? Evidence from capitalization of energy costs.

Technical report, E2e.

Myers, E. (2018). Asymmetric information in residential rental markets: Implications for the

energy efficiency gap. Technical report, E2e.

Nagashima, K., Yoda, T., Yagishita, T., Taniguchi, A., Hosono, T., and Kanosue, K. (2002). Thermal
regulation and comfort during a mild-cold exposure in young Japanese women complaining of
unusual coldness. Journal of Applied Physiology, 92(3):1029–1035.

Nakano, J., ichi Tanabe, S., and ichi Kimura, K. (2002). Differences in perception of indoor
environment between Japanese and non-Japanese workers. Energy and Buildings, 34(6):615 –
621. Special Issue on Thermal Comfort Standards.

O’Connel, N., Pinson, P., Madsen, H., and O’Malley, M. (2014). Benefits and challenges of
electrical demand response: A critical review. Renewable and Sustainable Energy Reviews,
39:686 – 699.

Papineau, M. (2017). Energy efficiency premiums in unlabeled office buildings. The Energy

Journal, Volume 38(Number 4).

Parsons, K. C. (2002). The effects of gender, acclimation state, the opportunity to adjust clothing
and physical disability on requirements for thermal comfort. Energy and Buildings, 34(6):593 –
599. Special Issue on Thermal Comfort Standards.

Pellerin, N. and Candas, V. (2003). Combined effects of temperature and noise on human discomfort.

Physiology and Behavior, 78(1):99 – 106.

Pigou, A. C. (1920). The Economics of Welfare. Macmillan, London.
PJM (2013). 2012 Economic demand response performance report: Analysis of economic DR par-
ticipation in the PJM wholesale energy market after the implementation of order 745. Technical
report, PJM Interconnection.

97

PJM (2014a). Demand response and why it’s important. Technical report, PJM Interconnection.

PJM (2014b). PJM economic demand resource in energy market. Technical report, PJM State and

Member Training Department.

PJM (2017). Demand response. Technical report, PJM Interconnection.
Reiss, P. C. and White, M. W. (2005). Household electricity demand, revisited. The Review of

Economic Studies, 72(3):853–883.

Reynolds, S. J., Black, D. W., Borin, S. S., Breuer, G., Burmeister, L. F., Fuortes, L. J., Smith,
T. F., Stein, M. A., Subramanian, P., Thorne, P. S., and Whitten, P. (2001). Indoor environmental
quality in six commercial office buildings in the midwest United States. Applied Occupational
and Environmental Hygiene, 16(11):1065–1077. PMID: 11757903.

Rich, E. (2017). FERC order 825 - economic demand response. Technical report, PJM Market

Settlements Subcommittee.

Ringsdorrf Jr., W. and Cheraskin, E. (1982). Vitamin C and tolerance of heat and cold: Human

evidence. Orthomolecular Psychiatry, 11:128–131.

Schellen, L., Lichtenbelt, W. D. V. M., Loomans, M. G. L. C., Toftum, J., and Wit, M. H. D. (2010).
Differences between young adults and elderly in thermal comfort, productivity, and thermal
physiology in response to a moderate temperature drift and a steady-state condition. Indoor Air,
20(4):273–283.

Taylor, N. A., Allsopp, N. K., and Parkes, D. G. (1995). Preferred room temperature of young vs
aged males: The influence of thermal sensation, thermal comfort, and affect. The Journals of
Gerontology: Series A, 50A(4):M216–M221.

US Census (2013). American housing survey 1997-2013.
Veiga, A. and Weyl, E. G. (2016). Product design in selection markets. The Quarterly Journal of

Economics, 131(2):1007–1056.

Walawalkar, R., Blumsack, S., Apt, J., and Fernands, S. (2008). An economic welfare analysis of

demand response in the PJM electricity market. Energy Policy, 36:3692 – 3702.

Weisburd, S., Bird, D., and Ben-Porath, R. (2018). Adverse selection and moral hazard in the

leasing market: Are buybacks the solution? Technical report, CEPR.

Wooldridge, J. (2010). Econometric analysis of cross section and panel data. MIT Press, Cam-

bridge, Mass, 2 edition.

Young, A. J. (2010). Homeostatic responses to prolonged cold exposure: Human cold acclima-
tization, pages 419–438. John Wiley and Sons, Inc., supplement 14: handbook of physiology
edition.

98