Advertisement

Environmental and Resource Economics

, Volume 73, Issue 3, pp 923–955 | Cite as

Does Absolution Promote Sin? A Conservationist’s Dilemma

  • Matthew Harding
  • David RapsonEmail author
Article

Abstract

This paper shows that households signing up for a green program exhibit an intriguing behavioral rebound effect: a promise to fully offset customers’ carbon emissions resulting from electricity usage increases their energy use post-adoption by 1–3%. The response is robust across empirical specifications, and is consistent with an economic model of rational energy consumption. Our results provide a cautionary tale for designing green product strategies in which the adoption of a product may lead to unexpected consequences.

Keywords

Carbon offsets Behavioral rebound Green marketing Energy consumption 

1 Introduction

To our knowledge, this is the first study to systematically analyze the impact of the introduction of a voluntary carbon offset program on household consumption behavior. We document an intriguing behavioral response to an offset program by evaluating the implementation of a large scale, utility-run carbon offset program in the field. The premise behind the program is that for a relatively small charge a household can entirely offset its carbon footprint from energy consumption by purchasing an “offset”. The amount collected by the utility is invested in third party environmental projects such as planting trees or building methane capture plants. The combined effect on enrollees is to decouple their energy use from its carbon emissions, while slightly increasing their energy price.

“Green” products, such as the present offset program, have the potential to provide substantial social benefits by deploying scarce abatement resource towards optimal use. This study focuses on the demand for carbon offsets and its impact on the related demand for electricity. The supply side of the carbon offset market has additional implications on other important industries, including forestry and agricultural. González-Ramírez et al. (2012) review in some detail the implications of carbon offsets on agricultural policy and highlight the design features of these programs from an agricultural perspective. According to some estimates, offset markets increase farm income by 5% (Baker et al. 2010). Quantifying the impact on agricultural markets depends however on a number of factor such as the price of the offset and the productivity of the carbon sequestration technology when applied in agriculture (Graff-Zivin and Lipper 2008). Moreover, farmers shifting supply towards offsets could also have a price impact on US commodities (Brown et al. 2010).

Why consumers exhibit a willingness to pay for carbon offsets is not completely understood, and some drivers of adoption may lead to unforeseen consequences. In such cases, the social benefit may be eroded.1 We will show that adoption in the program we study is associated with an increase in household energy use post-adoption. If one interprets this effect as causal, which these authors consider to be the simplest explanation, it has implications for the effectiveness of voluntary carbon offset programs. We present a simple economic model which rationalizes this causal interpretation. Conservation is shown to be a substitute to carbon offsets, and the measured effects on energy usage are economically and statistically significant at 1–3% of monthly usage during months after enrollment. The behavioral rebound that we observe is robust to a wide array of alternate empirical specifications.

The design of the offset program ensures that the behavioral rebound will not lead to increased GHG emissions as long as the carbon offset market functions properly. However, there are reasons to believe that carbon offset markets are exposed to a high degree of information asymmetry and monitoring costs, and thus are likely to perform imperfectly. For example, it is nearly impossible to verify whether abatement investments that underpin offsets are “additional” even if abatement itself can be verified. This is due to the near impossibility of knowing the correct counterfactual. There is an observational equivalence between abatement activity that is additional and activity that is not, and only the agent implementing the investment knows (or may know) the circumstances under which the investment would or would not have occurred. Furthermore, the carbon offset market is exposed to the potential for fraud. There are numerous news stories about fraud in the carbon offset market, especially when the projects are planned in developing countries and not well supervised.2

Under a poorly-performing carbon market, a behavioral rebound along the lines of what we document has the potential to undermine the primary objective of the offset program. Furthermore, the relevance of this behavioral effect is likely to be even more general. Consumers of other products that are, or are perceived as being, environmentally-friendly (e.g. rooftop solar, green electricity, energy efficiency, electric vehicles, etc) may exhibit similar behavior, as has been documented by Kotchen and Moore (2007) and Herberich et al. (2011). As such, this paper contributes a new data point to the ongoing discussion of how second-best environmental policies create incentives that undermine environmental objectives.

The interplay between environmental programs, economic incentives and behavioral factors has been explored and documented across disciplines. While economists have focused mostly on the evaluation of real world examples such as utility programs, the broader psychology and sociology literature focused mostly on smaller experiments (Clayton et al. 2015). Ebeling and Lotz (2015) show that in a field experiment and in an Amazon Turk experiment that households defaulted into a green power program had a ten times higher adoption rate. Pichert and Katsikopoulos (2008) also find in a laboratory experiment that most consumers would not switch when defaulted into a green power program. Similarly, in an experimental intervention, Litvine and Wüstenhagen (2011) show that nudges can overcome consumer reticence to invest in green power programs due to uncertainty about the final impact of their purchase.3

A common theme in the psychological literature is that monetary incentives may be less effective than psychological nudges when consumers are asked to make environmental choices, particularly when the price of environmental amenities is low. Bolderdijk et al. (2013) find that in some cases consumers prefer to be seen as “green” rather than “greedy” in a series of experiments. Behavioral factors have also been shown to be important when consumers choose “green power”. The importance of behavioral considerations has now been documented both in experiments (laboratory and field experiments) and in observational studies. When consumers choose to invest in green power programs they ensure that a fraction of their electricity is provided from renewable sources such as solar power. This is a scenario similar to that of carbons offsets since purchasing green power is simply a mechanism to subsidize investment in alternative energy resources. The consumer trusts the utility to use the funds collected today through the green power premium to build new sustainable power plants which will provide clean energy at some point in the future. Kotchen and Moore (2007) provide evidence that environmental concern and altruistic attitudes are correlated with green program enrollment. Jacobsen et al. (2012) show that households enrolling in a utility-sponsored green electricity program exhibit a measurable behavioral change from purchasing just a small amount of green electricity (but not from purchasing more), which provides evidence that non-traditional factors may be at work. Consumers also choose goods with environmental benefits such as energy efficient durables. These purchase decisions have also been shown to be influenced by behavioral factors. Herberich et al. (2011) find that in the purchase of compact fluorescent lightbulbs (CFLs), social pressure appears to operate on the extensive margin (whether or not to purchase a CFL) but not on the intensive margin (how many CFLs to buy).

The design features of the present program both create the possibility of a particularly strong behavioral effect while at the same time ensuring that the environmental outcome is in no danger of being reversed by a behavioral rebound as long as the offset market is functioning. The positive energy rebound may be induced in this setting due to the very low price associated with enrolling in the offset program. This need not generally be the case. There surely exists an offset price that is high enough to induce a financial response that outweighs any behavioral incentives. In the present context, the price of offsets was set equal to the marginal cost of procuring those offsets. On one hand, the price may well be near what can be expected from a profit-neutral voluntary offset program. On the other hand, in practice we may see carbon offsets being part of a portfolio of green power programs which also include investments in renewable energy. In that case the bundle may be offered to consumers at higher prices, and the financial incentive to reduce energy use could more than off-set any behavioral rebound effects.

Either way, while there is almost certainly an improvement in environmental outcomes in our setting, one can envision circumstances under which poorly functioning markets would cause such a rebound to result in inefficiencies. We readily admit that our setting does not allow us to fully eliminate a reverse-causation interpretation of the results (that people enroll in the carbon offset program due to an anticipated near-term increase in their electricity usage). Nonetheless, the possibility that a behavioral rebound is causing the increase in energy demand should be of concern to policymakers and researchers alike, if only to consider measures that will help to achieve the desired outcomes. Thus we view the simple goal of documenting this behavioral feature of the setting as a contribution to the literature and an area that would benefit from further study.

The paper proceeds as follows. Section 2 presents a simple, stylized model which captures the basic economic intuitions. Section 3 describes the program before discussing the data and our sampling framework. Section 4 estimates the short- and long-run impacts of the program, including a dynamic event study of energy use before and after adoption. We also provide a series of econometric robustness checks. Section 5 concludes. Supplementary background material on the voluntary carbon offset market is provided in the Appendix.

2 Conceptual Framework

2.1 An Economic Model of Energy Consumption

A very simple economic model of energy consumption allows us to capture the main stylized predictions about individual behavior when the choice of a carbon offset is offered. Suppose a household i consumes electricity \(x_i\) and other goods \(y_i\). Take \(y_i\) to be the numeraire, and assume for simplicity that utility is quasi-linear. Further assume that aggregate consumption of electricity is associated with a utility penalty. Concretely, let utility be given by
$$\begin{aligned} u(x_i) + y_i - \delta _i c\left( x_i,\sum _{-i} x_{-i}\right) \end{aligned}$$
(1)
where \(u' > 0\), \(u''<0\), \(\frac{\partial c}{\partial x_i}>0\), \(\frac{\partial ^2 c}{\partial x_i^2}>0\), and \(\delta _i > 0\) is a parameter determining the perceived externality from the aggregate consumption of electricity. We define \({\bar{X}}=\sum _{-i} x_{-i}\) in subsequent notation, which reflects the lack of strategic interaction between actors.4

In our model, consumption of electricity imposes a social cost in the form of GHG emissions via c(.). While pollution and global warming are global problems, the extent to which an individual is aware of them or incorporates them into her utility function can vary from individual to individual. This is captured by the parameter \(\delta _i\) which is heterogeneous across the population.5 The presence of \({\bar{X}}\) in c() allows a flexible interpretation of i’s perception of cost that includes (but differentiates between) her own contribution to social cost and the externalities associated with the behavior of others. Either way, we assume that individuals are heterogeneous with respect to the \(\delta \) parameter and that the extent to which they internalize the social cost is relatively fluid. \(\delta \) can also be influenced by advertising and informational campaigns.

For simplicity also assume that \(u'(0) = \infty \) so the household will always optimally set \(x_i > 0\). The household’s problem is to maximize utility subject to the budget constraint \(px_i+y_i\le m_i\), where p denotes the price of electricity.6 Let the value function be denoted by \(V_1(p,m,\delta _i)\).

Now consider the introduction of an optional carbon-offset program that eliminates the negative effect of \(x_i\), at an incremental per-unit price of \(\pi \). If the household decides to adopt the program, the modified problem will be to maximize
$$\begin{aligned} u(x_i) + y_i - \delta _i c(0,{\bar{X}}) \end{aligned}$$
(2)
subject to \((p+\pi )x_i + y_i \le m_i\). Conditional on adoption, a consumer contributes \(\pi x_i\) dollars to the carbon offset program. Assuming the program is implemented correctly, the utility invests this contribution in environmental programs designed to capture an equal amount of emissions leaving the household emissions neutral (at least as far as the footprint resulting from residential electricity consumption is concerned). We assume that \(\pi \) is chosen so as to exactly balance a household’s cost of emissions with the costs of offsetting the same emissions. Thus, the carbon offset program reduces aggregate emissions by the amount previously contributed by the household, \(x_i\). Note also that the extent of the reduction in this utility penalty resulting from enrollment in the offset program depends on the gradient of \(c(x_{i},{\bar{X}})\) with respect to \(x_{i}\), which we denote \(c_{1}\). Let the value function to this problem be denoted by \(V_2(p,\pi ,m,\delta _i)\).7 We now present a series of simple predictions resulting from this stylized model.

Proposition 1

  1. 1.

    As \(\delta _i\) increases, the household is more likely to adopt.

     
  2. 2.

    As \(\pi \) increases, the household is less likely to adopt.

     

The first result follows from the Envelope Theorem, \(\partial V_1 / \partial \delta _i = -c(x_i,{\bar{X}})\) and \(\partial V_2 / \partial \delta _i = -c(0,{\bar{X}})\). Furthermore since \(x_i > 0\), \(c_{1}>0\) and \(c_{2}>0\), it follows that \(\partial (V_1-V2) / \partial \delta _i< 0\). Hence adoption becomes more attractive in \(\delta _i\). The second result is immediate because \(V_1\) is unaffected by \(\pi \), while the Envelope Theorem implies that \(\partial V_2 / \partial \pi = -x_i < 0\).

This proposition implies that we would expect households who incur a higher disutility from the social cost of aggregate emissions to be more likely to sign up for the program. In the empirical sections we will attempt to quantify this by various indicators of the extent to which individuals are identified as “environmentalists” or engage in activities such as camping and other outdoor recreational activities. Empirically identifying drivers of an individual’s perception of the social cost of emissions has important implications for the design of marketing strategies for green goods (Jacobsen 2011). This proposition also shows that as long as \(\delta _i>0\) even individuals with very low disutility from the social cost of carbon emissions may join the program if the price is sufficiently low. In principle, we would expect a broad cross-section of the population to join a carbon offsetting program as long as some awareness of the social cost is present. In reality however, it may be hard for any program to generate sufficient interest even when potential adopters are aware of the social cost, given the countless other demands on individual time and attention whereby programs such as these may have very low salience.

Proposition 2

  1. 1.

    For sufficiently small \(\pi \), if a household decides to adopt, consumption of electricity increases post-adoption.

     
  2. 2.

    If two households decide to adopt, the household with large \(\delta \) will experience a larger increase in post-adoption consumption.

     

The first part of this proposition follows since the first-order condition to the household’s problem without adoption is \(u'(x_i) - \delta _i c_{1} = p\) whereas the first-order condition under adoption is \(u'(x_i) = p+\pi \). The solutions coincide if \(\pi = \delta _i c_{1}\). Any smaller \(\pi \) will lead to an increase in consumption after adoption as a consequence of \(u''<0\).

For the second part of this proposition, note that post-adoption consumption is independent of \(\delta \) so it is sufficient to show that pre-adoption consumption is smaller for larger \(\delta \). This follows immediately from Topkis’ Theorem applied to the pre-adoption objective function, observing that \(\partial ^2 U/\partial x_i \partial \delta _i = -c_{1} < 0\).

This proposition implies that, conditional on adoption, a household signing up for a carbon offset program may, in fact, increase energy use. While this may seem counter-intuitive, it follows from the fact that adoption depends on awareness of the social cost of carbon emissions. Any program claiming to mitigate these emissions will encourage households to consume more. The social cost of emissions acts as a dampening mechanism on individual consumption. When this mechanism is removed we would expect consumption to increase to the level that would have been chosen if the disutility from the social cost of emissions were set to zero. Naturally, we also have to consider the small impact of the price increase resulting from the introduction of the offsets on overall demand.

The net benefits of a perfectly functioning carbon offsetting program will still be positive. Because offset contributions are proportional to use, the net environmental cost of a household’s energy will be zero regardless of the level of post-adoption consumption. However, if we believe the growing body of anecdotal evidence which suggests that the offset market is at times dysfunctional, the net effect can in fact be negative. Proposition 2 also states that when comparing post adoption changes in usage for households with different utility costs of carbon emissions, households with higher pre-adoption costs should increase consumption more post-adoption than households which place less value on the social cost of emissions.

2.2 Behavioral and Psychological Explanations

This very simple economic model captures the economic intuition behind the main result of this paper: households may actually increase their energy consumption (at least temporarily) when adopting a carbon offset program. There are several different models from behavioral science and psychology which are consistent with the empirical evidence provided in this paper.

Altruism and Warm Glow Preferences At first glance, the provision of carbon offsets shares many similarities with charitable giving (DellaVigna et al. 2012), which could be employed as a theoretical framework for offsets. The one important difference to charitable giving, however, is the fact that consumption and offsets are linked and may potentially be subject to feedback effects from one to the other (see Kotchen 2009). Given the existence of a trade-off between the consumption of the private good and the public good, the individual may wish to offset her negative impact on the public good by contributing directly to the public good. Kotchen (2009) shows that this model admits a Nash equilibrium where the provision of offsets is positive above a certain wealth threshold. It does however have the property that if the private good is “more green” then the overall environmental impact can be negative.

Guilt The availability of offsets may allow individuals to alleviate their guilt while continuing or even increasing their consumption of the socially undesirable good. In a different context Gneezy and Rustichini (2000) find that when asked to pay a fine for delivering their children late for day care, parents actually arrived even later. The opportunity to buy an indulgence at a relatively affordable price may actually decrease the level of individual virtue. Similar results have been documented in responses to enrollment in green electricity programs (e.g. Kotchen 2006; Kotchen and Moore 2007; Jacobsen et al. 2012).

Moral Licensing Consumers often appear to justify actions which may not conform to their self-image (Ayal and Gino 2011). Moral licensing–engaging in self-licensing based on past ethical actions–can be thought of as a mechanism to solve the ethical dissonance experienced by the consumer. According to one definition, moral licensing “occurs when past moral behavior makes people more likely to do potentially immoral things without worrying about feeling or appearing immoral” (Monin and Miller 2001). Recently the psychological literature has devoted a substantial amount of effort to document the presence of moral licensing through a variety of laboratory and small scale field experiments (Effron and Monin 2010; Kouchaki 2011; Merritt et al. 2010). Since consumer choices, especially in the environmental arena, reflect social and moral values, this appears to be a promising mechanism that could explain the psychology behind the adoption of carbon offsets and subsequent change in energy consumption (e.g. Jacobsen 2010).

3 Offset Program Description and Data

3.1 PG&E’s ClimateSmart Program

In June 2007, Pacific Gas and Electric (PG&E), a large California-based utility, launched the ClimateSmart program (hereafter CS program) that offered its customers a means to offset the greenhouse gas (GHG) emissions associated with their usage of electricity. Customers choosing to opt into this program pay an extra $0.00254 per kilowatt-hour (kWh).8 This price was set by the California Public Utilities Commission (CPUC) and the per kWh charges correspond to an implied price of $9.71 per short ton of \(\text {CO}_{2}\)-equivalent emissions.9 The price remained constant throughout the program implementation. The program was initially launched as a demonstration program which expired on December 31, 2009. It was dormant during 2010 and was formally closed at the end of 2011.

By the end 2009, PG&E customers had contributed approximately $4.9 million through CS program charges, $2.2 million of which were charged to residential customers which are the focus of this analysis. The remaining contributions come from commercial, agricultural or government entities as well as direct contributions made by PG&E shareholders. These funds were used to invest in GHG emissions abatement projects, offsetting an amount of GHG equal to those associated with the energy use generating the CS program revenue. Marketing activities were subsidized via a small surcharge to PG&E’s broad customer base, which is substantially larger than the number of CS program adopters.

As the CS program progressed, PG&E solicited offers for emissions reduction projects. Each of these was subsequently verified by a reputable third-party certification organization called the Climate Action Reserve. As a result of its certification efforts, PG&E offsets associated with the ClimateSmart program are generally viewed as being very high quality.10 Examples of projects include the reduction of tree harvesting in several Northern California forests and methane capture from a major diary farm and from several landfill projects. The quantity is equivalent to saving 137 million gallons of gasoline, or taking around 225 thousand cars off the road in California for a year.11

From the customer acquisition perspective the CS program has been a qualified success. Through 2010, the program had enrolled approximately 30,000 commercial and residential customers. However, when compared to the 6 million potential customers in the PG&E service territory, the adoption rate was low. Geographically, the top five cities for CS program sign-ups were San Francisco (11.0%), San Jose (6.3%), Oakland (5.7%), Berkeley (2.9%) and Sacramento (2.2%). Retention in the program was strong with less than 0.2% active de-enrollments per month.12

Throughout the program duration, PG&E customers were able to enroll through the company’s website. The precise description of the program was easily available to the customers online. Figure 1 shows an example of online marketing which highlights the nature of the carbon offsetting program. The marketing information also highlights the fact that erollees should expect a 100% offset as a result of joining the program.
Fig. 1

Example of marketing material explaining the nature of the carbon-offset program

The timing of adoption of the CS product varied in a significant and predictable fashion as function of a variety marketing activities. For the purpose of our analysis, we focus on households enrolling after August 2008, after which point marketing activities were very limited, while the overall awareness of the existence of the program was high. During this period there was no directed outreach from PG&E to individual households, but households were exposed to a television and on-line media campaign centered on the message that a typical California home emits the same amount of GHG over the course of a year as an SUV. It is important to note that the marketing did not target a specific geography during the late period of the program (Fig. 2).

We focus exclusively on this period since it provides a setting virtually free from the confounding effects of marketing framing. Enrollment required households to invest effort to identify the program and then enroll in it. The impetus for such effort is idiosyncratic, so enrollment times during this period were fairly random. In the event study analysis which follows, the distribution of adoption timing allows us to flexibly control for the presence of time-varying factors that may otherwise generate confounding concerns. The necessity to seek out the program also selects on households who are, as revealed by their behavior, actively engaged on the margins of environmental concern and energy use.

3.2 Data

The sample used in this paper consists of residential CS program adopters purchasing electricity from PG&E. We restrict attention to customers on the E1 electric rate, which is the most common residential rate.13 Our sample consists of 748 customers who enrolled in the CS program between August 2008 and November 2009 (the last month for which new enrollments were processed under the original program) and 13,449 control households, which are also on the E1 rate schedule but never enrolled in the CS program. The sample of control households was drawn at random from the full PG&E database and stratified to reflect the different areas in which PG&E operates and thereby accurately reflect the distribution of customers. This sample was obtained after removing various outliers as explained below.
Fig. 2

Histogram of adoption timing

For each CS program customer we know the date of enrollment. Unfortunately, we do not know if or when a customer de-enrolled because it does not affect their presence in the billing dataset unless they also disconnect their electricity service. The total de-enrollment rate averaged 1–2% per month, and happened automatically as a customer moved. To avoid bias from unobserved de-enrollments, we limit our sample to those customers for which we have near-complete billing data between June 2006 and December 2010. We thus eliminate recent movers, who could arguably have different consumption patterns especially in the months before and after a move. All households in our sample have at least 50 (out of 55 possible) months of billing data.14

While average household electricity usage in PG&E territory ranges from 600 to 700 kWh per month, the distribution of electricity consumption in PG&E territory is extremely skewed to the right, with the top 1% of users consuming in excess of 4781 kWh per month. We truncate the sample to remove the top 1% of users. Furthermore, it is common to see negative billing amounts due to various adjustments. Since it is difficult to interpret these observations, and as they may potentially influence our results, we have also eliminated households with any monthly usage of less than 1 kWh or a billed amount of less than $5. The results that we report subsequently are robust to re-inclusion of these households.

One important limitation of our electricity usage data is that we don’t observe the actual billing cycle for each household. In practice billing periods do not overlap perfectly with calendar months. Thus, a usage amount for say November 2007 may include up to two weeks of December 2007. The billing cycle differs from household to household and induces measurement error in the dependent variable at the monthly level.

For each household in our sample we observe its ZIP code of residence (772 unique zip codes) as well as a number of demographic and lifestyle characteristics obtained from a third-party data provider, the marketing services firm, Acxiom.15 In Table 1 we present the summary statistics for our sample. For each variable we report the mean and standard error for both the group of customers who sign up for the CS program and the control group of non-adopting customers. We also report the p-value of the t-test for the difference in means between the two groups for each variable. It is immediately obvious that the two groups of customers are different from a statistical point of view, though perhaps by not as much as we would have guessed initially.
Table 1

Summary statistics for ClimateSmart adopters and non-adopters

 

E1 rate customers

 

Adopters (\(\text {N}=748\))

 

Non-adopters (\(\text {N}=13,449\))

 

Mean

SD

p-value

 

Non-adopters

Adopters

Non-adopters

Adopters

 

(Average) kWh*

635.507

598.053

360.500

340.831

0.01

(Average) Bill*

109.731

96.712

97.345

81.307

0.00

Age

56.545

51.313

14.306

12.770

0.00

College

0.294

0.295

0.456

0.457

0.93

HHIncome $80k\(\,+\)

0.480

0.539

0.500

0.499

0.00

Children

0.471

0.455

0.499

0.498

0.48

Working woman

0.467

0.453

0.499

0.498

0.48

HH size

2.877

2.513

1.471

1.398

0.00

Home owner

0.966

0.957

0.180

0.204

0.18

Environmental

0.097

0.184

0.296

0.388

0.00

Green living

0.615

0.575

0.487

0.495

0.03

Charity

0.370

0.400

0.483

0.490

0.10

Charitable

0.547

0.487

0.498

0.500

0.00

Outdoors

0.551

0.623

0.497

0.485

0.00

Wildlife

0.073

0.110

0.261

0.313

0.00

Camping

0.266

0.352

0.442

0.478

0.00

Home age

38.547

42.937

23.403

27.393

0.00

Heating

0.273

0.298

0.446

0.458

0.13

Cooling

0.111

0.095

0.314

0.293

0.18

Sqft \(2500\,+\)

0.161

0.146

0.367

0.354

0.35

Home value $500k\(\,+\)

0.161

0.206

0.368

0.405

0.00

Pool

0.139

0.111

0.346

0.314

0.03

ClimateZone

6.670

5.864

4.459

4.199

0.00

*The average kWh and bill amounts for adopters are computed for the pre-adoption periods

The mean electricity usage for adopters is smaller than that for the non-adopters (598 kWh vs. 635 kWh). Similarly adopters paid an average monthly bill of $96 while non-adopters paid average monthly bills of $110. The lower than average energy consumption for adopters explains the overall very low additional cost of signing up to the CS program for the adopting households. In our sample the mean monthly contribution to the CS program was $1.35 (with a standard deviation of $0.79). The maximum observed monthly contribution in our sample was only $8.94.

Since we have eliminated movers from our sample and since younger people are more likely to be mobile, the average household head age in our sample is 51 years for adopters and 56 years for non-adopters. At the same time adopters tend to be wealthier and less likely to have children. In terms of housing characteristics, adopters tend to live in smaller but more expensive homes. They are also less likely to have a swimming pool. California is divided into climate zones which reflect the specific climate in the area. Cooler, coastal climate zones are labeled with lower numbers while hotter more arid zones are labeled with higher numbers.16 Table 1 shows that adopters are more likely to live in coastal areas where they are substantially less likely to require air conditioning. This also corresponds with the location of the bigger cities. Overall the demographic profile of adopters suggests that adopters are younger, wealthier households who live in more urban, coastal areas (e.g. the San Francisco Bay Area).

Our data also include information on lifestyle choices made by households based on marketing characteristics determined by Acxiom. We report the propensity of households to have expressed interest in environmental or wildlife issues (“Environmental”), those making environmentally healthy product purchase decisions or donating funds to environmental causes (“Green Living”), as well as two variables relating to charity. “Charity” captures whether a household is interested in their community and/or involved in local charitable organizations or activities, and “Charitable” is a model based variable designed to predict households with a high propensity to donate to charitable causes.

These variables were acquired at the start of the program and are not affected by the later participation of a household in the CS program. Furthermore, all demographic and lifestyle choice variables are proprietary to the third party provider, Acxiom, and the precise details of the data construction and predictive models for some of these variables have not been made available to us. These variables are, however, widely used in the private sector and have become a highly valued asset in the marketing activities of firms. Thus, while questions regarding their validity and extent of measurement error remain, we lean towards trusting these variables.17 Overall, we find that adopters strongly identify with environmental principles as revealed by their interests from subscriptions to their enjoyment of camping and the outdoors. But at the same time they appear to be less green in their actions.

4 Impact of Adoption

In this section we examine how energy use changes in response to adoption of the carbon offset program. We begin by estimating a difference-in-differences model, which reveals the nature of the long-run effect. We then examine short-run effects using first-differences model. Finally, coefficients from an event study provide a flexible representation of the dynamic effects, both before and after adoption. In all specifications, we use the aforementioned pseudo-balanced panel of customer electricity usage data from June 2006 to December 2010. Results of these analyses are unchanged if we use the larger, unbalanced panel.

4.1 Difference-in-Differences

The difference-in-differences approach compares electricity use across households and over time relative to when the CS program was adopted. Specifically, we estimate the following model:
$$\begin{aligned} ln(k)_{it} = \beta I_{it}^{cs} + \alpha _{t} + \gamma _i+ \epsilon _{it} \end{aligned}$$
(3)
where \(ln(k)_{it}\) is the log of monthly electricity consumption (kWh) for household i in month t, \(I_{it}^{cs}\) is an indicator variable equal to one if household i is enrolled in the CS program in period t, and zero otherwise, \(\alpha _{t}\) and \(\gamma _{i}\) are month-by-year and household fixed effects, respectively.18 The error term, \(\epsilon _{it}\), accounts for factors that idiosyncratically perturb electricity usage at the monthly level, and may or may not be observable to the household. We also estimate Eq. 3 for heterogeneous treatment effects, interacting \(I_{it}^{cs}\) with a variety of demographic variables.

It is worth pausing to discuss the interpretation of the results that estimation of this equation will yield, particularly in light of the fact that the CS program is voluntary, and that households must select into it. One possible goal would be to estimate the causal effect of the CS program on electricity usage in the context of random assignment. That is, were we to randomly assign some households to “treatment” (CS) and others to “control”, we could estimate the population average treatment effect (PATE) on consumption. The identifying assumption in this case would be a standard orthogonality condition: conditional on time-invariant household characteristics and aggregate period-level effects, \(E[I_{it}^{cs} \epsilon _{it}]=0\). While it may well be the case that this condition holds in our setting (i.e. if adoption of the CS program is orthogonal to electricity use), it’s easy to imagine situations under which the condition is not satisfied. For example, a household may feel guilty about an anticipated increase in consumption, and enroll in the offset program as a result.

However, the effect that we seek to estimate is the presence of a behavioral rebound effect on enrollees—the population average treatment effect on treated households (PATT). The voluntary nature of the program is appropriate for the desired hypothesis testing, since the self-selecting nature of the act of enrolling is the very treatment that one might expect to induce the behavioral rebound.19 Any change in consumption after adoption (or, for that matter, before), may under very reasonable conditions be related to the enrollment decision. Thus, what we are interested in estimating is the effect of selection into the program on usage in surrounding periods (i.e. the PATT).

The assumptions under which the Eq. 3 will retrieve the PATT when estimated using our data are plausible. If households are exposed to a stochastic shock that compels them to enroll in the offset program, and if that shock does not directly alter their electricity use, then \({\hat{\beta }}\) will be a consistent estimate of the PATT. Many such unobserved shocks are possible—for example, the local newspaper prints a story about the causes and effects of global climate change, compelling them to investigate ways to reduce their carbon footprint.

Estimates of \(\beta \) are displayed in Table 2. Each column provides estimates from a separate regression which explores the degree to which the effect varies with the demographic variables. We find that low-income adopters exhibit an increase in electricity use of 11.6% in the long run. No other demographic interaction term produces a change that allows us to reject zero long-run effect with 95% confidence.
Table 2

Difference-in-differences and first difference estimates

Column

Difference-in-differences

First difference

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

(10)

Base case

0.006

\(-\,0.001\)

   

0.018*

0.018*

   

(0.008)

(0.007)

   

(0.009)

(0.010)

   

Age (young)

  

0.024

    

0.060**

  
  

(0.016)

    

(0.024)

  

Age (MidL)

  

\(-\,0.009\)

    

0.016

  
  

(0.013)

    

(0.017)

  

Age (MidH)

  

\(-\,0.001\)

    

\(-\,0.008\)

  
  

(0.015)

    

(0.015)

  

Age (old)

  

\(-\,0.02\)

    

\(-\,0.001\)

  
  

(0.021)

    

(0.027)

  

Sqft (H)

   

\(-\,0.002\)

    

0.039

 
   

(0.017)

    

(0.031)

 

Sqft (M)

   

0

    

0.031*

 
   

(0.012)

    

(0.017)

 

Sqft (L)

   

0.008

    

0.013

 
   

(0.015)

    

(0.015)

 

Income (H)

    

0.017

    

\(-\,0.009\)

    

(0.023)

    

(0.029)

Income (M)

    

0.002

    

0.017

    

(0.026)

    

(0.034)

Income (L)

    

0.116**

    

− 0.04

    

(0.052)

    

(0.064)

Household FEs

Y

Y

Y

Y

Y

Y

Y

Y

Y

Y

Month-by-Year FEs

Y

 

Y

Y

Y

Y

 

Y

Y

Y

MoY-by-County FEs

 

Y

    

Y

   

R-squared

0.093

0.234

0.099

0.093

0.093

0.128

0.203

0.137

0.128

0.128

Observations

414,253

407,862

369,706

414,253

414,253

395,456

389,363

352,972

395,456

395,456

Number of IDs

14,442

14,219

12,887

14,442

14,442

14,442

14,219

12,887

14,442

14,442

Each entry is a point estimate (with standard error in parentheses) from a regression of ln(kWh) on an adoption indicator interacted with the demographic categorical indicator noted in the row header. * Significant at the 0.10 level, **Significant at the 0.05 level, ***Significant at the 0.01 level. Standard errors clustered at the HH level

4.2 First Differences

If persistence of any change in electricity usage due to the CS program adoption is low, then results in Table 2 will understate the near-term response. The treatment effect estimated from a model of first differences will be identified off of changes in behavior during the treatment assignment period, and thus provides an estimate of the effect in the short run (one month). Let \(d ln(k_{it}) = ln(k_{i,t})-ln(k_{i,t-1})\). We estimate the following first differences specification:
$$\begin{aligned} d ln(k_{it}) = \beta CS_{it} + \alpha _{t} + \epsilon _{it}, \end{aligned}$$
(4)
where \(CS_{it}\) is an indicator variable equal to 1 in the month when household i adopts the CS program and 0 otherwise. Again, this is estimated on the pseudo-balanced panel of customer electricity usage data from August 2008 to December 2010. Time-invariant household characteristics are eliminated by differencing, although we estimate treatment effects by demographic category. We also include aggregate time controls at the month-by-year level (not differenced).

Estimates of Eq. 4 are shown in right half of Table 2. For adopters, carbon offsets appear to be a substitute for conservation. The overall effect in first differences for these households is a 1.8% increase, statistically significant at the 95% level.20 This effect is once again driven by young households, who exhibit a 6.0% increase in electricity use in the post-adoption month.

4.3 Dynamic Effects: Event Study

Results in Sects. 4.1 and 4.2 show an initial increase in consumption of 0–5% (depending on the demographic group) and an attenuation of the effect in the long run as an implication of the difference-in-differences results. While it does not appear that the CS program had an economically significant impact in the long run, it may still have substantial transitory effects. In order to exploit the heterogeneous timing of adoption, we conduct an “event study” analysis which allows us to analyze the dynamic effects of program adoption (see21). The use of the event study methodology requires us to separately identify the effect of program adoption over time from the seasonality in consumption due to weather patterns.

A strength of the event study approach is its ability to control directly and indirectly for unobserved time-varying aggregate effects. Direct controls come in the form of calendar month dummy variables, but the innovation of the approach is to re-align the time series into proximity to the event of interest (in our case, enrollment in the ClimateSmart Program). Enrollment that occurs in different months for different households reduces the avenues of contamination that are possible from time-varying confounders. The approach is most reliable when adoption occurs at random intervals throughout the time series. Indeed, over the period from August 2008 to the end of 2009, households adopt at very different points in their annual seasonal consumption pattern. We confirm that the actual adoption over time is as close as possible to a random uniform pattern by constructing a time-to-adoption variable, \(\tau _i\), as the number of months between July 2008 and the month of adoption. We estimate a duration model for \(\tau _i\) conditional on the observable demographics, choose a simple Weibull specification and regress \(ln(\tau _i)\) on the list of demographics, lifestyle variables and home characteristics discussed above.22

Table 3 reports the estimated coefficients from several models which differ in their included set of covariates. While a few covariates appear to be statistically significant, they are typically not significant across different specifications. Moreover the extent to which they explain the timing of adoption is doubtful since the model fit is extremely poor, with an adjusted \(R^2\) of between 0 and 0.02 and thus even the statistically significant variables are economically not significant. This provides encouraging support to our assertion that adoption occurred more or less at random and timing of adoption cannot be predicted from the observables.
Table 3

Weibull model for the time to adoption during the period July 2008 to November 2009

Column

(1)

(2)

(3)

(4)

(5)

(6)

(7)

Age

\(-\,0.00120\)*

 

\(-\,0.00128\)*

\(-\,0.00148\)**

\(-\,0.00178\)**

\(-\,0.00138\)*

\(-\,0.00170\)**

 

(0.001)

 

(0.001)

(0.001)

(0.001)

(0.001)

(0.001)

College

0.00296

 

0.00139

0.00101

\(-\,0.00251\)

\(-\,0.00013\)

\(-\,0.00545\)

 

(0.018)

 

(0.018)

(0.018)

(0.019)

(0.018)

(0.019)

HHIncome $80k\(\,+\)

0.01199

 

0.01192

0.01304

0.00448

0.01418

0.00651

 

(0.018)

 

(0.018)

(0.018)

(0.020)

(0.018)

(0.019)

Children

\(-\,0.03054\)

 

\(-\,0.03247\)

\(-\,0.02837\)

\(-\,0.03617\)

\(-\,0.02981\)

\(-\,0.04006\)*

 

(0.020)

 

(0.021)

(0.021)

(0.022)

(0.022)

(0.023)

Working woman

0.01238

 

0.00676

0.00570

0.00075

0.00651

0.00191

 

(0.017)

 

(0.018)

(0.018)

(0.019)

(0.018)

(0.020)

HH size

0.00247

 

0.00297

0.00156

0.00388

0.00163

0.00365

 

(0.007)

 

(0.008)

(0.008)

(0.008)

(0.008)

(0.008)

Home owner

\(-\,0.00918\)

 

\(-\,0.01555\)

\(-\,0.01701\)

\(-\,0.19112\)***

\(-\,0.01929\)

\(-\,0.18918\)***

 

(0.049)

 

(0.050)

(0.052)

(0.055)

(0.052)

(0.055)

Environmental

  

\(-\,0.02187\)

\(-\,0.03080\)

\(-\,0.02953\)

\(-\,0.02442\)

\(-\,0.03084\)

   

(0.021)

(0.022)

(0.023)

(0.026)

(0.027)

Green living

  

0.02725

0.01756

0.02232

0.01801

0.02467

   

(0.021)

(0.027)

(0.028)

(0.027)

(0.029)

Charity

   

0.03123

0.04128*

0.03440*

0.04473*

    

(0.020)

(0.023)

(0.021)

(0.023)

Charitable

   

0.00129

\(-\,0.00645\)

0.00343

\(-\,0.00935\)

    

(0.025)

(0.026)

(0.026)

(0.028)

Outdoors

     

\(-\,0.04651\)*

\(-\,0.06227\)**

      

(0.024)

(0.025)

Wildlife

     

\(-\,0.02156\)

\(-\,0.00612\)

      

(0.029)

(0.033)

Camping

     

0.04533**

0.06281***

      

(0.020)

(0.022)

Home age

 

\(-\,0.00033\)

  

\(-\,0.00009\)

 

\(-\,0.00016\)

  

(0.000)

  

(0.000)

 

(0.000)

Heating

 

0.00952

  

\(-\,0.00050\)

 

0.00179

  

(0.017)

  

(0.019)

 

(0.019)

Cooling

 

0.00302

  

0.00589

 

0.00630

  

(0.026)

  

(0.030)

 

(0.029)

Sqft 2500\(\,+\)

 

\(-\,0.00660\)

  

\(-\,0.02373\)

 

\(-\,0.01991\)

  

(0.024)

  

(0.027)

 

(0.028)

Home value $500k\(\,+\)

 

0.03602*

  

0.05643**

 

0.05747***

  

(0.019)

  

(0.022)

 

(0.022)

Pool

 

0.00228

  

\(-\,0.00012\)

 

\(-\,0.00875\)

  

(0.023)

  

(0.025)

 

(0.026)

log(kWh)

      

0.00471

       

(0.020)

Constant

3.07218***

3.00769***

3.07198***

3.07887***

3.26251***

3.08860***

3.24835***

 

(0.061)

(0.018)

(0.063)

(0.064)

(0.070)

(0.065)

(0.151)

Observations

516

569

516

516

447

516

447

Adjusted R-squared

0.00

0.00

0.00

0.00

0.01

0.01

0.02

Robust standard errors in parentheses. *Significant at the 0.10 level, **Significant at the 0.05 level, ***Significant at the 0.01 level

We can therefore proceed to estimate the following equation:
$$\begin{aligned} ln(k)_{it} = \displaystyle \sum _{j={\underline{m}}}^{{\overline{m}}}\xi _{j}D^j_{it} + \alpha _{t} + \gamma _i+ \epsilon _{it} \end{aligned}$$
(5)
where \(D^j_{it}\) are a set of indicator variables set equal to one if, in calendar month t, household i is j months after its CS program adoption month. The event window is defined as \(j \in [{\underline{m}},{\overline{m}}]\), and we normalize the coefficient of event month prior to adoption to zero. Additional indicators corresponding to “outside the event window” allow us to fully capture the dynamic effects of treatment. The underlying assumption here is that, conditional on time-invariant household characteristics and aggregate month-specific shocks, all households that are j months away from enrolling in the offset program are identical (in expectation). We take a flexible approach to determining the event window, allowing \({\overline{m}}=-{\underline{m}}=24\) and therefore defining the largest possible event window as comprising 2 years on either side of the adoption event. The specification thus includes an indicator variable for \(+\,25\) which captures all periods more than 24 months from the adoption of the CS program. Similarly we include an indicator variable for \(-\,25\), which captures all periods more than 24 months before adoption. Estimating the event study using a large window allows us to “zoom” in or out as desired to most clearly present the graphical results. The above specification for the estimation of the response pattern relative to the time of adoption implicitly models the response as a piecewise linear function of relative time to adoption, with no restrictions on the variation or pattern of the response over time.
We estimate Eq. 5 and plot the resulting estimates in event time (where 0 denotes the month prior to adoption) in Fig. 3, using an 18-month event window. The continuous line denotes the estimates for the event period dummies while the dashed line corresponds to the confidence bounds for the estimates on the indicator variables. The estimates exhibit three features that provide interesting context to the previous results. In the months leading up to enrollment, households are engaged in increasing conservation efforts; consumption decreases by 4% over that 6 month period. At the time of adoption, conservation ends and households increase consumption by approximately 3%. This “rebound” is persistent, as can be seen by the relatively flat consumption profile in the six months after enrollment. The higher usage of adopters over the control group then levels out and becomes increasingly imprecise as distance from the event grows.
Fig. 3

Estimation of the dynamic effect of adoption in event time

The true short-term effect of enrollment on usage—the difference in usage under enrollment (observed) versus that under non-enrollment (unobserved)—depends on what one believes to be the appropriate counterfactual. One’s interpretation may thus be influenced by whether the measured trend before adoption is statistically significant. If the downward trend of enrollee usage continues in the absence of enrollment, this would lead the “true” effect to be larger than the point estimates of the event time dummy coefficients (which are estimated relative to zero). An F test comparing the restricted model where the slope before adoption is constant to the unrestricted model rejects the restricted model with a p-value of 0.15. While not overwhelmingly significant, it does lend support to the interpretation that households are engaged in conservation behavior before adoption, after which time they exhibit a rebound effect. It is unlikely the case that the adoption of the carbon offset leads to a temporary reduction in consumption during the adoption month and that households returns to their pre-adoption level afterwards.

Our difference-in-differences and first-difference specifications suggest that responses may be heterogeneous across demographics. We investigate this further by interacting the event-time indicators with observable demographics. Results from these specifications are presented in Table 4. It is clear from the increased statistical significance that the nonparametric controls for aggregate calendar-month variation that are inherent in the event study approach are effectively soaking up residual variation that was present in the earlier specifications. It is to be expected that the coefficients on event month 1 do not precisely match the first-difference coefficients, since the specifications differ on at least two important dimensions. The event study controls differently for calendar time effects, and normalization of the event time zero coefficient potentially influences each of the event-period coefficient estimates
Table 4

Event study: Heterogeneous dynamics

Event study heterogeneous dynamic program effects (demographics)

 

Age (Young)

Age (MidL)

Age (MidH)

Age (Old)

Sqft (L)

Sqft (M)

Sqft (H)

Income (L)

Income (M)

Income (H)

Months from event

\(-\,6\)

0.079***

0.028

0.038

0.101**

0.045

0.051*

\(-\,0.02\)

0.039

0.033

0.077***

(0.030)

(0.030)

(0.029)

(0.040)

(0.027)

(0.027)

(0.042)

(0.045)

(0.028)

(0.022)

\(-\,5\)

0.097***

0.021

0.02

0.105**

0.04

0.056**

0.002

0.012

0.042

0.074***

(0.033)

(0.030)

(0.028)

(0.043)

(0.028)

(0.027)

(0.039)

(0.046)

(0.028)

(0.022)

\(-\,4\)

0.047

0.023

\(-\,0.004\)

0.101**

0.013

0.033

\(-\,0.009\)

\(-\,0.038\)

0.044

0.041**

(0.029)

(0.028)

(0.027)

(0.046)

(0.027)

(0.025)

(0.038)

(0.048)

(0.028)

(0.020)

\(-\,3\)

0.045

0.050**

\(-\,0.003\)

0.083*

0.022

0.029

0.025

0.036

0.037

0.043**

(0.028)

(0.024)

(0.025)

(0.043)

(0.024)

(0.023)

(0.035)

(0.041)

(0.026)

(0.018)

\(-\,2\)

0.036

0.009

\(-\,0.011\)

0.04

0.008

0.023

0.021

0.015

0.026

0.007

(0.022)

(0.021)

(0.020)

(0.035)

(0.019)

(0.020)

(0.032)

(0.035)

(0.021)

(0.014)

\(-\,1\)

\(-\,0.002\)

0.029*

0.018

0.042

0.009

0.019

0.057*

0.031

0.023

0.016

(0.016)

(0.016)

(0.018)

(0.032)

(0.016)

(0.014)

(0.032)

(0.027)

(0.016)

(0.013)

0

0

0

0

0

0

0

0

0

0

0

1

0.051**

0.018

0.017

0.041

0.032**

0.033**

0.041

0.016

0.033**

0.030**

(0.022)

(0.015)

(0.016)

(0.031)

(0.015)

(0.016)

(0.029)

(0.024)

(0.017)

(0.013)

2

0.055***

0

0.043**

0.057*

0.021

0.031*

0.028

0.013

0.03

0.039**

(0.021)

(0.021)

(0.018)

(0.034)

(0.019)

(0.018)

(0.031)

(0.028)

(0.020)

(0.015)

3

0.081***

0.014

\(-\,0.001\)

0.065*

0.022

0.028

0.036

\(-\,0.02\)

0.031

0.050***

(0.028)

(0.025)

(0.024)

(0.039)

(0.024)

(0.022)

(0.037)

(0.035)

(0.024)

(0.019)

4

0.059*

\(-\,0.001\)

\(-\,0.02\)

0.051

\(-\,0.005\)

0.011

\(-\,0.007\)

\(-\,0.026\)

0.004

0.037*

(0.032)

(0.028)

(0.025)

(0.037)

(0.026)

(0.024)

(0.038)

(0.041)

(0.026)

(0.022)

5

0.072**

\(-\,0.002\)

0.004

0.095**

0.023

0.024

\(-\,0.011\)

0.017

0

0.060***

(0.031)

(0.027)

(0.025)

(0.041)

(0.025)

(0.024)

(0.039)

(0.039)

(0.026)

(0.021)

6

0.076**

\(-\,0.002\)

0.008

0.098**

0.037

0.03

\(-\,0.032\)

0.02

\(-\,0.008\)

0.070***

(0.031)

(0.029)

(0.027)

(0.039)

(0.027)

(0.024)

(0.040)

(0.043)

(0.027)

(0.021)

*Significant at the 0.10 level, **Significant at the 0.05 level, ***Significant at the 0.01 level. Standard errors clustered at the HH level. All specifications include household and month-by-year controls

For the most part the results conform to the pattern observed in the base case, but with varying degrees of pre-adoption conservation and event-month rebound. In the six months preceding adoption, households of various demographic groups conserve electricity by 2–10%. Much of the amount conserved is reversed in the months following adoption, as enrollees increase electricity usage relative to non-adopters.

These results once again demonstrate that the strongest increase in consumption at event time is exhibited by young households, who change consumption by 5% in that period. This effect persists for several months. Heterogeneity along demographic lines also reveals a treatment-period increase for wealthy households and those in small or medium dwellings. These effects are approximately 3% in magnitude, and likely reflect a strong correlation between the demographic variables (age, income and home size) and geographic location (the San Francisco Bay Area, where adoption rates were highest). In some cases the effect dissipates after quite quickly (e.g. in the mid-to-high age and middle-income categories), consistent with the difference-in-differences and first differences specifications. In others it appears to linger, and even grow over time (e.g. older and higher-income households).

4.3.1 Robustness Checks

In order to bolster the claim that our results are not being driven by some unobserved factors, we also perform several placebo tests. Any difference-in-differences type strategy requires that there be no systematic variation in the outcome variable that is spuriously correlated with treatment. In the context of an event study, it is difficult to imagine how such fluctuations would survive the time controls and conversion to event time. Nonetheless, we implement an array of placebo tests on adopting households to reinforce this point. We estimate placebo treatment effects using samples partitioned into the pre-adoption and post-adoptions months. We implement these placebo tests by drawing from 50 simulated samples with randomly assigned placebo treatment status and placebo adoption date. Specifically, for each simulation we randomly assign 10% of adopting households into placebo treatment status, and separately randomly assign to them a placebo adoption month. We repeat this separately for months before and after the actual adoption decision by the households in the sample. The difference-in-difference and first-differences coefficient estimates and standard errors are averaged and presented in Table 5. Each of the coefficient estimates is economically and statistically indistinguishable from zero. In Fig. 4 we present estimates of the event study coefficients that arise from Placebo regressions using this methodology. It is clear that there are no measurable effects of the placebo on monthly electricity consumption. These results provide additional evidence that our main treatment effect estimates are not artifacts of spurious correlations in the data.
Table 5

Placebo tests on adopting households

Sample specification

Adopter households,

Adopter households,

 

pre-adoption periods

post-adoption periods

 

DinD

FD

DinD

FD

Placebo treatment

0.000

0.001

0.003

0.006

(0.007)

(0.008)

(0.023)

(0.025)

HH FEs

Y

Y

Y

Y

Month-by-year FEs

Y

Y

Y

Y

R-squared

0.135

0.126

0.106

0.114

Observations

242,473

231,006

16,493

15,623

Placebo effect estimates calculated from the mean of \(\text {NS}=50\) simulation draws. Means of the standard errors (clustered at the household level) across simulation draws are reported in parentheses. *Significant at the 0.10 level, **Significant at the 0.05 level, ***Significant at the 0.01 level

Fig. 4

Placebo event study (100 Placebo simulation draws)

In addition, we performed an extensive array of robustness checks. These are too numerous to present in the manuscript, but the alternate specifications produced results that are qualitatively and quantitatively consistent with our main results. These included additional model specifications involving consumption levels as opposed to log consumption, alternative normalizations of individual consumption observations, and several different models involving inverse probability weights.23 We retrieve similar results when estimating our baseline specifications on alternate samples, such as the re-inclusion of electricity usage outliers and the use of an unbalanced panel that allows for the inclusion of households that may have recently moved. In the “Appendix”, we also show that our results are robust to other approaches to seasonality corrections, such as multiple time-varying interactive effects and H-P filtering.

5 Discussion and Conclusions

In this paper we document a behavioral effect with potential consequences for the deployment of environmental programs: selecting into a good behavior may lead to increased energy consumption. This area has recently attracted attention in a number of fields under the broad headline of “incongruous actions”. These actions may be particularly unfortunate for the valuation of environmental projects, where it is often known as a “rebound effect”. While the cost of carbon emissions is in principle offset by the consumers through their participation in the program, there remains uncertainty as to its eventual success due to imperfect abatement and the ongoing doubts over fraud and market imperfections.

Our particular focus in this paper is voluntary carbon offsets. These are provided to individuals if they contribute a surcharge on their energy use in order to mitigate the environmental impact of energy production through investment in carbon abatement projects. We document two main empirical findings. First, many individuals who voluntarily sign up for carbon offsets actually increase their consumption following adoption. Since offsets increase the marginal cost of electricity, the driving force behind this substitutability must be non-monetary. The results are consistent with an intuitive behavioral rebound effect which can be explained by a number of cognitive mechanisms documented by behavioral scientists in related areas. Second, households selecting into the offset program have meaningfully different characteristics than those that do not. The uniquely detailed data available to us at household level allows us to investigate individual heterogeneity and the extent to which it matters for the adoption of environmentally beneficial programs and their subsequent usage.

To our knowledge, this study provides the first evidence of the behavioral rebound effect from a large-scale field deployment of carbon offsets. Thus it confirms the relevance of a growing body of laboratory and field evidence on the unforeseen effects of green programs, which also has implications for program design. To the extent that markets are subject to inefficiency, these sorts of unexpected behavioral responses may decrease the value of the abatement investments.

Footnotes

  1. 1.

    An extensive literature exists on the so-called “rebound effect”, which broadly addresses the potential general equilibrium effects of energy efficiency and other conservation policies (Borenstein 2015; Gillingham et al. 2016). Additionally, recent psychological literature has devoted a substantial effort to document the presence of moral licensing through a variety of laboratory and small scale field experiments (Effron and Monin (2010); Kouchaki (2011); Merritt et al. (2010)). Since consumer choices, especially in the environmental arena reflect social and moral values, this appears to be one promising mechanism that could explain the psychology behind the adoption of carbon offsets and subsequent change in energy consumption (Jacobsen (2010); Kotchen (2009)). Green markets also affect social welfare, underlying the importance of understanding consumer choice in this setting (Kotchen 2006).

  2. 2.

    A particularly poignant story of fraud in the market for “environmental indulgences” was reported by the Christian Science Monitor in April 2010, which revealed how the Vatican was convinced to purchase carbon offsets that would have lead to the Vatican becoming the first carbon free state, but which were never implemented. The purchased offsets were meant to be used for the planting of millions of trees in Hungary. As it turned out the trees were never planted and the Hungarian company abruptly closed down at the end of 2007. See http://www.csmonitor.com/Environment/2010/0420/Carbon-offsets-How-a-Vatican-forest-failed-to-reduce-global-warming.

  3. 3.

    While the experimental literature seems to find that behavioral factors substantially increase the adoption of green power programs, we should caution that the effect may not be universally effective. In an unpublished experiment, one of the authors worked with a major utility and sent 50,000 letters encouraging utility customers to sign up for a green power program using several well-documented behavioral nudges such as social pressure. The experiment did not generate a single adoption but did lead to three complaint letters being sent to the utility and the experiment was not published. This might indicate that publication bias also plays a role in the claimed successes of nudges.

  4. 4.

    The stock of greenhouse gases is large enough that any individual’s contribution is infinitesimal. We nonetheless include it in order to keep the model applicable to a broad class of closely-related environmental considerations (e.g. effects on local criteria pollutants).

  5. 5.

    We interpret the parameter \(\delta \) as a behavioral parameter but remain agnostic about the precise behavioral/psychological mechanism influencing it. It is possible that \(\delta \) reflects the degree of understanding or awareness of the social cost of pollution which is a function of education, social, religious and political beliefs. Similarly, since a large share of the cost is likely to be incurred in the future, variation in the \(\delta \) parameter may reflect individual inter-temporal discount rates and concern for future generations.

  6. 6.

    For the purpose of our stylized model we ignore non-linear electricity pricing, which is common in the residential electricity market, but does not affect the results of interest.

  7. 7.

    Note that as a consequence of our stylized model we are ignoring income effects. Basic economic intuitions tell us that, all else equal, higher income households are more likely to sign-up for the program. As we shall see later in the paper this is indeed the case. It is easy to derive the income effect in the current framework under a suitable reformulation of the utility function. This however does not change the other implications of the model and we abstract from the income effect in favor of notational simplicity.

  8. 8.

    The baseline price of electricity is determined by usage with five tiers ranging between $0.12233 and $0.34180 per kWh.

  9. 9.

    The price is well-within the range of per ton prices usually encountered for carbon offsets in the offset markets, but substantially lower than common estimates of the social cost of carbon which typically is measured in the $30–50 per ton range, depending on the chosen discount factor.

  10. 10.

    Climate Action Reserve describes their protocol in detail in a program manual that is publicly available: http://www.climateactionreserve.org/how/program/program-manual/.

  11. 11.

    Source: International Carbon Bank and Exchange, California Energy Commission.

  12. 12.

    Further details on the CS program are available through the detailed annual reports issued by PG&E. The figures quoted are from the report for 2010 which is publicly available at: http://www.pge.com/myhome/environment/whatyoucando/climatesmart/programdetails/.

  13. 13.

    This rate is not time-of-use dependent. Throughout this analysis we do not consider low-income households. These households are on special variants of this electric rate, and PG&E took extra steps to ensure that they were aware of being in the program. The program was available to all customers on an opt-in basis but low-income customers on special rates were subsequently de-enrolled by the utility.

  14. 14.

    It is quite common for anomalies to appear in any electricity billing dataset. For example, billing errors may occur that create unrealistic patterns (say, zero dollars in June but twice the seasonally-adjusted expectation for July). We remove these months from the data, but see no justification for dropping the entire household time-series due to such brief idiosyncratic billing events. As such, we allow for a small number of absent months from our pseudo-balanced panel. In any case, results are robust to changes in these minor sampling assumptions.

  15. 15.

    Data on individual households was purchased by the utility as part of the regular business processes. Given the availability of complete address information in their administrative database we do not think selection to be a major factor.

  16. 16.

    For example, San Francisco is in zone 3 while zone 12 corresponds to the Northern Central Valley which experiences substantially hotter summers than zone 3.

  17. 17.

    One of the authors of this paper purchased data on himself from the same provider and the purchased information found the information to be accurate, although it slightly underestimated his environmentalist credentials.

  18. 18.

    We have also used month and year dummies to control for seasonality, which impose a lower computational burden as that specification requires fewer parameters. The results were almost identical. Later we shall discuss more robustness checks to account for seasonality in our specifications.

  19. 19.

    Further, from a program evaluation perspective, our setting is also appropriate. Under no realistic implementation of this sort of program would the allocation of offsets be random. The same is not true for allocation of permits under a cap-and-trade regime, a very different setting than what we study here.

  20. 20.

    Note that effects in this direction would imply an upward sloping demand curve if price were the driving factor.

  21. 21.

    Jacobson et al. (1992).

  22. 22.

    The results are not affected by the choice of the Weibull model. More complicated duration models do not change the results.

  23. 23.

    A specification that included household by month-of-year controls produced results that were essentially noise. However, such an over-specified model does not leave adequate residual variation to credibly identify an effect.

  24. 24.

Notes

Acknowledgements

Funding was provided by Stanford Precourt Institute for Energy (US).

References

  1. Ayal S, Gino F (2011) Honest rationales for dishonest behavior. Exploring the causes of good and evil, chap. Honest rationales for dishonest behavior. APAGoogle Scholar
  2. Baker JS, McCarl BA, Murray BC, Rose SK, Alig RJ, Adams D, Latta G,Beach R, Daigneault A (2010) Net farm income and land use under a USgreenhouse gas cap and trade. In: Agricultural and Applied EconomicsAssociation. Policy Issues 7: April 2010. p 5Google Scholar
  3. Bloomberg (2012) State of the voluntary carbon market 2012. New energy finance, ecosystem marketplaceGoogle Scholar
  4. Bolderdijk JW, Steg L, Geller ES, Lehman P, Postmes T (2013) Comparing the effectiveness of monetary versus moral motives in environmental campaigning. Nat Clim Chang 3(4):413CrossRefGoogle Scholar
  5. Borenstein S (2015) A microeconomic framework for evaluating energy efficiency rebound and some implications. Energy J 36:1–21Google Scholar
  6. Brown TR, Elobeid AE, Dumortier JR, Hayes DJ (2010) Market impact of domestic offset programsGoogle Scholar
  7. Clayton S, Devine-Wright P, Stern PC, Whitmarsh L, Carrico A, Steg L, Swim J, Bonnes M (2015) Psychological research and global climate change. Nat Clim Chang 5(7):640CrossRefGoogle Scholar
  8. Costa D, Kahn M (2013) Energy conservation “nudges” and environmentalist ideology: evidence from a randomized residential electricity field experiment. J Eur Econ Assoc 11(3):680–702CrossRefGoogle Scholar
  9. DellaVigna S, List J, Malmendier U (2012) Testing for Altruism and social pressure in charitable giving. Q J Econ 127(1):1–56CrossRefGoogle Scholar
  10. Ebeling F, Lotz S (2015) Domestic uptake of green energy promoted by opt-out tariffs. Nat Clim Chang 5(9):868–871CrossRefGoogle Scholar
  11. Effron D, Monin B (2010) Letting people off the hook: When do good deeds excuse transgressions? Pers Soc Psychol Bull 36:1618–1634CrossRefGoogle Scholar
  12. Gillingham K, Rapson D, Wagner G (2016) The rebound effect and energy efficiency policy. Rev Environ Econ Policy 10(1):68–88CrossRefGoogle Scholar
  13. Gneezy U, Rustichini A (2000) A fine is a price. J Legal Stud 29:1–18CrossRefGoogle Scholar
  14. González-Ramírez J, Kling CL, Valcu A (2012) An overview of carbon offsets from agriculture. Annu Rev Resour Econ 4(1):145–160CrossRefGoogle Scholar
  15. Graff-Zivin J, Lipper L (2008) Poverty, risk, and the supply of soil carbon sequestration. Environ Dev Econ 13(3):353–373CrossRefGoogle Scholar
  16. Harding M, Lamarche C (2011) Least squares estimation of a panel data model with multifactor error structure and endogenous covariates. Econ Lett 111(3):192–199CrossRefGoogle Scholar
  17. Herberich D, List J, Price M (2011) How many economists does it take to change a light bulb? A natural field experiment on technology adoption, Working PaperGoogle Scholar
  18. Hodrick R, Prescott E (1997) Post war business cycles: an empirical investigation. J Money Credit Bank 29:1–16CrossRefGoogle Scholar
  19. Hsieh D, Manski C, McFadden D (1985) Estimation of response probabilities from augmented retrospective observations. J Am Stat Assoc 80:651–662CrossRefGoogle Scholar
  20. Jacobsen G (2010) Do environmental offsets increase demand for dirty goods? Evidence from residential electricity demandGoogle Scholar
  21. Jacobsen G (2011) The Al Gore effect: an inconvenient: truth and voluntary carbon offsets. J Environ Econ Manage 61:67–78CrossRefGoogle Scholar
  22. Jacobsen GD, Kotchen MJ, Vandenbergh MP (2012) The behavioral response to voluntary provision of an environmental public good: evidence from residential electricity demand. Eur Econ Rev 56(5):946–960CrossRefGoogle Scholar
  23. Jacobson L, Lalonde R, Sullivan D (1992) Earnings losses of displaced workers. Am Econ Rev 83:685–709Google Scholar
  24. Kotchen M (2006) Green markets and private provision of public goods. J Polit Econ 114:816–834CrossRefGoogle Scholar
  25. Kotchen M (2009) Voluntary provision of public goods for bads: a theory of environmental offsets. Econ J 119:883–899CrossRefGoogle Scholar
  26. Kotchen M, Moore M (2007) Private provision of environmental public goods: household participation in green-electricity programs. J Environ Econ Manage 53:1–16CrossRefGoogle Scholar
  27. Kouchaki M (2011) Vicarious moral licensing: the influence of others’ past moral actions of moral behavior. J Person Soc Pychol 101:702–715CrossRefGoogle Scholar
  28. Litvine D, Wüstenhagen R (2011) Helping “ light green” consumers walk the talk: results of a behavioural intervention survey in the Swiss electricity market. Ecol Econ 70(3):462–474CrossRefGoogle Scholar
  29. Manski C, Lerman S (1977) The estimation of choice probabilities from choice based samples. Econometrica 45:1977–1988CrossRefGoogle Scholar
  30. Merritt A, Effron D, Monin B (2010) Moral self-licensing: when being good frees us to be bad. Soc Personal Psychol Compass 4(5):344–357CrossRefGoogle Scholar
  31. Monin B, Miller D (2001) Moral credentials and the expression of prejudice. J Personal Soc Psychol 81:33–43CrossRefGoogle Scholar
  32. Pesaran H (2006) Estimation and inference in large heterogeneous panels with a multifactor error structure. Econometrica 74(4):967–1012CrossRefGoogle Scholar
  33. Pichert D, Katsikopoulos KV (2008) Green defaults: information presentation and pro-environmental behaviour. J Environ Psychol 28(1):63–73CrossRefGoogle Scholar
  34. Ravn M, Uhlig H (2002) On adjusting the Hodrick–Prescott filter for the frequency of observations. Rev Econ Stat 84(2):371–376CrossRefGoogle Scholar

Copyright information

© Springer Nature B.V. 2019

Authors and Affiliations

  1. 1.Department of Economics and Department of StatisticsUniversity of California - IrvineIrvineUSA
  2. 2.Department of EconomicsUniversity of California - DavisDavisUSA

Personalised recommendations