The Effects of Moral Licensing and Moral Cleansing in Contingent Valuation and Laboratory Experiments on the Demand to Reduce Externalities

Environ Resource EconDOI 10.1007/s10640-014-9872-y

The Effects of Moral Licensing and Moral Cleansing inContingent Valuation and Laboratory Experiments onthe Demand to Reduce Externalities

Benjamin Ho · John Taber · Gregory Poe · Antonio Bento

Accepted: 18 December 2014© Springer Science+Business Media Dordrecht 2015

Abstract Recent field experiments show that peer information can induce people to reducetheir production of negative externalities. Related work in psychology demonstrates thatinducing feelings of relative culpability in one domain can induce spillover pro-social behav-ior in another domain. We use a contingent valuation and parallel lab experiment to explorepatterns of cross-domain responses to norm-based interventions. Asymmetric responsesbetween those whose impacts are above or below the norm are found to be robust acrossdecision settings. Substantial heterogeneity in responses is observed across a number ofdimensions not explored in large field experiments, raising questions about the universalityof peer-information effects and the design of such programs.

Keywords Culpability · Moral licensing · Moral cleansing · Guilt · Peer information ·Green electricity

1 Introduction

Recent large-scale field experiments demonstrate that peer comparisons and social-normnudges are effective tools for inducing the conservation of privately purchased goods thatcollectively create negative public externalities. Randomized residential electricity experi-ments that have monitored energy use and informed households of their personal consumption

Electronic supplementary material The online version of this article (doi:10.1007/s10640-014-9872-y)contains supplementary material, which is available to authorized users.

B. Ho (B)Economics Department, Vassar College, 124 Raymond Ave, Poughkeepsie, NY 12604, USAe-mail: [email protected]; [email protected]

J. TaberFERC, Washington, DC, USA

G. Poe · A. BentoCornell University, Ithaca, NY, USA

123

http://dx.doi.org/10.1007/s10640-014-9872-y

B. Ho et al.

levels relative to a neighborhood norm provide evidence that energy consumers significantlyreduce their energy consumption relative to a control group that does not receive such com-parative information (Ayres et al. 2013; Allcott and Mullainathan 2010; Costa and Kahn2010, 2013; Allcott 2011). Such behavioral change-based interventions, as opposed to moretraditional price instruments, can indeed be powerful, especially amongst specific groupsof the population. Ferraro and Price (2013), for example, study the effects of providingnon-price interventions for household water use and find that, at least in the short run, thesocial-comparison effect is equivalent to that which would be expected if average prices wereto increase by 12–15 %; in a study of residential electricity consumption, Ayres et al. (2013)estimate that non-price, peer comparison intervention induce the equivalent consumptionresponse as a 17–29 % price increase.

While the average treatment effect has been shown to be significant, it is apparent that thereis variation in response patterns to norm-based interventions. Notably, in a localized studyof 290 households, Schultz et al. (2007) demonstrate that some households actually increasetheir energy consumption when they are informed that their baseline consumption is belowthe average of their peer group. In this same study, high-energy users significantly decreasedtheir electricity consumption levels relative to the baseline, as expected from the focus theoryof normative behavior (Cialdini et al. 1991). This asymmetry in treatment effects has beenreplicated, to an extent, in large scale field experiments with observations ranging from 75,000to 600,000 households. However, rather than observing a strong perverse boomerang effectwhere peer information increases consumption, there more commonly seems to be a zero,or muted negative, effect on consumption patterns of low-use households. Allcott (2011)estimates that social-norm treatment effects are not significantly different from zero for thelowest three deciles of baseline electricity users, but that there is a significant mean treatmenteffect in high-use households ranging from about −3.7 % for the 8th decile to over −7 % inthe 10th decile. Ayres et al. (2013) similarly find no significant treatment effect on two out oflowest three deciles of baseline electricity use (the second decile had a significant treatmenteffect of approximately +1 %), while consumption levels significantly decline by about −3 to−7 % for the top three baseline energy deciles. In a regression framework, Ferraro and Price(2013, p. 70) estimate that the “social norm effect for our high user group is approximately94.1 % greater (5.28 vs. 2.72 % relative reduction) than for our low user group—a differencethat is significant at the p < 0.005 level.” In all, while strong boomerang effects may notbe evident, there does appear to be an important asymmetry in responses to social-norminterventions between households with above and below norm consumption levels.

Moreover, although responsiveness to norm-based messages have been demonstrated in anumber of domains (e.g. Frey and Meier 2004; Cialdini et al. 2006; Salganik et al. 2006; Caiet al. 2009) recent research in the energy-social norms literature suggests that non-pecuniaryeffects may not be as universal as previously thought. Different socio-economic groups mayhave heterogeneous responsiveness to peer information. In interpreting these results, Costaand Kahn (2010) argue that:

behavioral economists have underestimated the role that ideological heterogeneityplays in determining the effectiveness of energy conservation “nudges”… we findthat liberals and environmentalists are more responsive to these nudges than the aver-age person. In contrast, for certain subsets of Republican Registered voters, we findthat the specific “treatment nudge” that we evaluate has the unintended consequenceof increasing electricity consumption. (p. 2)

In this paper we show that asymmetric and heterogeneous responsiveness of a spillover effectfrom peer information is manifested in both contingent valuation and laboratory economics

123

The Effects of Moral Licensing and Moral Cleansing

experiments. Along the lines of Bateman et al. (2008), who demonstrated parallelism betweencontingent valuation responses, and “inconsistencies…found in everyday decisions involvingreal commitments” (p. 125), we argue that evidence of convergent behaviors across methodslends validity to each. Further, the survey application allows us to explore whether hetero-geneity in response patterns occurs in demographic and other respondent-specific dimensionsnot able to be explored in large-scale field tests. The laboratory experiment permits exoge-nous control of the individual’s impact, avoiding possible endogeneity effects that may arisein field and contingent valuation studies.

The contingent valuation study calculates the household carbon footprint of a nationallyrepresentative sample of consumers by asking questions about their energy-related consump-tion habits. A carbon footprint is defined to be the number of tons of carbon dioxide emissionsan individual is personally responsible for based upon his or her energy consumption deci-sions in a given year. We then induce feelings of relative culpability in the treatment group byproviding them with information about how their household’s carbon footprint compares toothers in the study and then elicit their demand for green electricity. In an effort to parallel thefield contingent valuation study, the laboratory experiment has student subjects purchase “pri-vate commodities” (analogous to electricity) that generate a negative externality (analogousto pollution) for a group in which they are a member. A treatment group is given informationabout the private, pollution-generating choices of others and the subjects are subsequentlygiven an opportunity to contribute to a fund that would reduce the negative harm createdby the externality. In the taxonomy of Harrison and List (2004) we present results from aframed field experiment coupled with a conventional laboratory experiment. In our use ofvalidating information treatments in online samples our approach is similar to recent workby Kuziemko et al. (2013) who take a similar approach in the domain of income inequality,though unlike their results, we find that the heterogeneity of the subject pool matters a lot interms of treatment effect sizes.

Beyond demonstrating convergent validity between field experiments, economic labo-ratory exercises, and contingent valuation responses and identifying further dimension ofresponse heterogeneity to social-norm nudges, our research contributes to the broader lit-erature on norm-based conservation incentives. First, in contrast to energy and water con-servation in which the psychological cues and economics savings are mutually reinforcing,our contingent valuation study of quantity demanded for “green electricity” and laboratoryexperiment study of contributions to a public good involve tradeoffs between private costsand societal or group gains. As such, our work extends the work of Shang and Croson (2009)and Chen et al. (2009) who show that some individuals are willing to bear additional mon-etary burdens in response to information about social norms. Second, much of the previousresearch on norm-based messaging has been confined to providing information about peerconsumption in the domain of the desired conservation activity. For example, studies that seekto encourage towel re-use in hotels, provide information about towel re-use habits of others(Goldstein et al. 2008). At the same time some limited research suggests that social-norminformation in one domain of decision-making affects decisions in other domains (Mazar andZhong 2010; Keizer et al. 2008). These studies have considered moral licensing—learningyou are more moral in one domain makes you less moral in another—and moral cleansing—learning you are less moral in one domain makes you more moral in another. Our researchspeaks to both and finds an asymmetric response. This asymmetry could produce a “moralrebound” effect—where acting pro-socially in one domain increases anti-social behaviorin another domain—that limits the effectiveness of social-norm based policy interventions.Therefore, understanding such response patterns could significantly improve the design ofinterventions and explain the limited effectiveness of past trials. More mundanely, our design

123

B. Ho et al.

speaks directly to the effect of carbon footprint calculators on the demand for carbon offsetsand green electricity.

Our main findings are that information about the behaviors of other people in one domainaffects public provision behavior in contingent valuation and lab experiments in a differentdomain. In effect inducing cold prickles encourages the seeking of warm glows. This effectof social information is asymmetrical—the moral licensing effect for individuals better thanthe norm is larger than the moral cleansing effect for those whose consumption and neg-ative externality effects are worse than the perceived norm. Finally, we demonstrate thatsystematic heterogeneity in responses to social norm nudges extends substantially beyondthe political/environmental dimensions explored in Costa and Kahn’s field experiment. As weargue in the concluding section, these findings, in conjunction with emerging field research,raise questions about the universal efficacy of nudges vis-à-vis pricing incentives.

The remainder of this paper is organized as follows. In the following section we reviewprevious economic and psychological conceptualizations of the notion of culpability or guiltin choice and valuation and how these concepts have been tested in laboratory and contin-gent valuation exercises. We then provide details on our experimental design and data. Inthe fourth section we provide empirical analyses of our experimental results with respectto asymmetry in response patterns above versus below norm respondents. The fifth sectionlends supporting evidence to the Costa and Kahn results, and expands the analysis of hetero-geneity to demographic and respondent-specific characteristics available from survey data.Conclusions and discussion are provided in the final section.

2 Background and Experimental Design

2.1 Background on Culpability

In this research we explore how the desire to prevent a public bad is affected by an individ-ual’s relative culpability, which we define to be the amount of social damage resulting froman individual’s actions relative to damages caused by others.1 Whereas the mechanisms thatmight induce conformity to a perceived social norm have been extensively studied in eco-nomics (see for example Bikhchandani et al. 1992; Ellison and Fudenberg 1993; Bernherim1994; Akerlof and Kranton 2000; Glaeser and Scheinkman 2002), the mechanism of culpa-bility has received less attention. Guilt has been explored in the psychology literature (seeBaumeister et al. 1994 for a review). Perhaps most famously, Carlsmith and Gross (1969)induced guilt in subjects by having them administer electric shocks to another person, a con-federate. Later, when subjects believe they have completed the experiment, they are asked todonate blood. Subjects who actually administered the shock are much more likely to agreeto donate, relative to subjects who merely observed the shocking.

Building from psychological foundations and psychological game theory (see Geanako-plos et al. 1989), Charness, Dufwenberg and co-authors construct a general theory of guiltaversion in which decision-makers experience guilt if they believe they let others down (e.g.Dufwenberg and Lundholm 2001; Charness and Dufwenberg 2006, 2007; Battigalli andDufwenberg 2007). With supportive results from “Trust Game” experiments, they proposethat this general theoretical framework can be extended to specific instances, such as publicgoods games and social norms, where it seems plausible that decision-makers are affected

1 Our focus is on relative culpability because pilot experiments found that information about one’s absolutelevel of social damage without comparison to one’s peers had no effect on behavior.

123


by guilt. In doing so these authors take care to distinguish the role of guilt aversion fromconformity: “A norm is a social moral expectation, a definition of which acts people in soci-ety will judge as right or wrong...Too many authors use “norm” just to mean “conformity inbehavior”. (Dufwenberg and Lundholm 2001, p. 511).

Andreoni’s (1995) prior research on public goods suggests that such motivations maydepend on whether the provision of the public good is framed positively or negatively. InAndreoni (1995), two groups of subjects participated in strategically identical public goodsprovision games, but with two separate framings. In one, the experiment was framed asproviding a public good so that subjects would be motivated by warm glow altruism; inthe other, the experiment was framed as avoiding a public bad, so that subjects would bemotivated by a desire to avoid a “cold prickle” of guilt. Sonnemans et al. (1998) conduct alike set of experiments in a threshold provision setting, alternatively framing the experimentsas provision to provide a public good and prevention of a public bad. In both the Andreoniand Sonnemanns et al. studies, the tendency to free ride was more prevalent in the negativeframing. Similarly, Solnick and Hemenway (2005) present informal survey evidence wherepositional concerns matter more for public goods rather than for public bads.

A related asymmetry in pro-social behavior in experiments comes from how initial prop-erty rights are interpreted based on the framing of the question. Grossman and Eckel (2012)and List (2007) argue that the same strategic choice yield differing amounts of pro-socialbehavior when the action is framed as giving versus taking. Korenok et al. (2013) and Bard-sley (2008) parameterize this interpretation and estimate models of social preferences givendifferent environments.

In the specific area of environmental norms, Bamberg and Moser (2006) conduct a meta-analysis of the literature on psychological mechanisms that promote pro-environmentalbehavior, finding that both social norms and guilt are important correlates to pro-environmental attitudes and behavior. Clark et al. (2003) find that participation in a greenelectricity program is correlated with self-reported altruism and pro-environmental attitudesas measured by the New Environmental Paradigm (NEP). Brouwer et al. (2008) test the “pas-senger pays principle” to find that air travelers’ perceived responsibility for climate change,awareness of the environmental impact of flying, and the frequency of flying were all pos-itively correlated with WTP for a per-flight carbon offset program. This notion of personalresponsibility in creating public harm is an extension of what Kahneman et al. (1993) refersto as an “outrage effect”, in which people are willing to pay more to avoid an environmentalproblem if they think it is human-caused than if they think that it is an outcome of nature(Bulte et al. 2005). Kahneman et al. (1993) and Brown et al. (2002), amongst others havedemonstrated this “outrage effect” on contingent valuation responses.

Our experiments complement the aforementioned literature by honing in on the indi-vidual culpability in contingent valuation and public goods experimental settings. We usepeer information to manipulate the norm in a sequential setting most similar to the framingexperiments of Andreoni (1995) and Sonnemans et al. (1998). Rather than split “Provisionof Public Good” and “Prevention of Public Bad” samples as done in these studies, how-ever, we employ a sequential framework: in the first stage of the experiment, we observeprivate decisions in a negative externality setting; the second stage involves a public goodscontributions game in which contributions mitigate the negative effects of decisions in thefirst stage. We expect two main outcomes. For those who learn they contribute more to thenegative externality than the perceived norm, i.e. have positive relative culpability, we expectthey will be more altruistic in the second. For those who experience negative culpability, bylearning they contribute less to the negative externality than the perceived norm in the firststage, we expect they will be less altruistic in the second. We find support for both of these

123

B. Ho et al.

effects, but that the former dominates. This “moral licensing” effect has been explored byMazar and Zhong (2010) who find that those who are given the opportunity to purchase greengoods are more likely to cheat on a subsequent exam. Similarly, in one field experimentaltest of the “broken windows” effect Keizer et al. (2008) show that observing others violateone social norm makes subjects more likely to violate other social norms. Our results furtherdemonstrate that the effect predominates in those pre-disposed to provide more public goodsin the second domain—for example Democrats, replicating in a lab and contingent valuationcontext the findings of Costa and Kahn (2010) who observed that the effect is limited toDemocrats2 in a field experiment on electricity conservation. We extend their work to showthat the heterogeneous effect likely exists along other dimensions as well.

2.2 Contingent Valuation Experiment

The broad objective of the contingent valuation survey was to gather information from par-ticipants that allowed us to calculate the household carbon footprint for each respondent andthen elicit their quantity demanded for green electricity program given information abouttheir own carbon footprint and, in some treatments, their carbon footprint relative to thoseof another survey participant. Participants for the online survey were recruited through TheStudyResponse Project, a nationwide panel of 95,574 people. The diversity of the sample,as seen in the summary statistics in Table 1 will be important for our analysis. Participantswere chosen at random and emailed the URL for the survey. For completing the survey,participants received $5. Invitations to participate were sent to 520 panelists (stratified to benationally representative by age and race), and we received 297 completed surveys3 for an81 % response rate.

There were four steps in the survey: (I) Eliciting demographic questions to calculate thesubject’s household carbon footprint; (II) Providing information about International Panel onClimate Change (IPCC) predictions on the impacts of climate change; (III) Showing subjectstheir estimated annual carbon footprint based on the input they provided; and (IV) Elicitingindividual demand for green electricity. For the control treatment, subjects were not providedany information about the carbon footprint of others. All other subjects received informationabout the carbon footprint of “Others like you who took this survey” (see Fig. 1. Subjectscompleted each question in order and were not allowed to go back.

Part I of the survey consisted of several web pages eliciting information about energy use,including housing characteristics (type, age, size of residence, and location), home energyuse (monthly electric and gas bill expenditures, type of fuel used to heat house, whetherthe household generates or purchases electricity); automobiles (number, models, use of eachvehicle) and transportation choices (use of public transportation, frequency of short and longdomestic flights, frequency of international flights). Subjects were also asked about whetherthey purchased carbon offsets and if so, how many had they purchased. Only 31 subjectsreported having purchased carbon offsets.

Subsequent to providing the above information, subjects were provided with three IPCCclimate policy scenarios and their anticipated consequences as presented below in Table 2.The purpose of this screen was twofold. First, we wanted to make respondents aware ofcurrent climate projections and relative policy options ranging from “Business as Usual”

2 The Democratic party in the Unites States is socially liberal or “left leaning”, while the Republican party issocially conservative or “right leaning”.3 An additional 105 surveys were collected concurrently for an alternative treatment that is not reported here.

123


Table 1 Summary statistics for contingent valuation experiment

By treatment group Treated: by culpability

Control Saw 11 tons Saw 26 tons High culpability Low culpability

Green elec. 143.33 143.40 107.68 152.26 96.40

Demand (kWh) (15.41) (12.30) (12.98) (12.87) (11.46)

CO2 Total 23.30 20.84 25.91 32.01 11.08

(2.35) (1.85) (2.64) (2.34) (0.67)

Relative culpability 9.84 −0.09 16.96 −9.38

(1.85) (2.64) (2.18) (0.72)

NEP 34.01 35.25 34.65 34.37 35.82

(0.81) (0.67) (0.82) (0.71) (0.75)

Liberal 0.75 0.75 0.66 0.73 0.69

(0.05) (0.04) (0.05) (0.04) (0.05)

Children 0.58 0.50 0.53 0.63 0.36

(0.06) (0.05) (0.06) (0.05) (0.05)

Female 0.57 0.49 0.49 0.47 0.52

(0.06) (0.05) (0.06) (0.05) (0.05)

Age 37.61 37.50 40.39 36.86 41.20

(1.17) (1.19) (1.40) (1.10) (1.50)

Income 81.6 62.4 66.7 77.7 47.0

(4.03) (5.75) (5.65) (4.83) (3.08)

College 0.53 0.53 0.46 0.55 0.43

(0.06) (0.05) (0.06) (0.05) (0.05)

Democrat 0.41 0.46 0.34 0.41 0.40

(0.06) (0.05) (0.05) (0.05) (0.05)

N = 79 112 83 111 84

Standard errors in parenthesesCO2 total: total CO2 footprintCulpability: total CO2 footprint minus 11 or 26 tons, depending on treatmentNEP: aggregate NEP valueLiberal: binary for liberal/conservative (1 if liberal)Children: binary for children in householdFemale: binary for gender (1 if female)Age: age of respondentIncome: annual Household income in thousands USDCollege: binary for education (1 if at least college education)Democrat: binary for party affiliation (1 if democrat)

to “Aggressive Emissions Reductions.” To a certain extent, this information also served toinduce an element of moral outrage for those concerned about climate change.

In Part III, respondents were provided with an estimate of the carbon generated fromtheir use of utilities and transportation and, after accounting for offset purchases, their esti-mated carbon footprint (“the total amount of climate changing greenhouse gas emissionscaused directly and indirectly by your household”) in tons of carbon per year. Carbon foot-prints were calculated using two algorithms. If participants knew their electricity and heatingexpenditures, information about average electricity and fuel prices in each state were used to

123

B. Ho et al.

Fig. 1 Information about carbon footprint presented in the survey

Table 2 Information about climate change presented in on-line survey

Climate options the IPCC has presented several options for reducing climate change, each with differentfinal levels of carbon and impacts on the global climate:

Business as usual Small emissions reductions Aggressive emissionsreductions

Mean percent changein carbon emissionsfrom 2000 to 2050

115 % increase 55 % increase 70 % decrease

Global averagetemperaturesincreases

8.8–11◦ (4.9–6.1 ◦C) 7.2–8.8 ◦F (4–4.9 ◦C) 3.6–4.3 ◦F (2–2.4 ◦C)

Sea level increases 12–24 inch(0.3–0.6 m) Millionsat risk of coastalflooding

10–24 inch(0.26–0.6 m)Millions at risk ofcoastal flooding

<17 inch (0.45 m)

Extinction risk More than 40 % ofspecies face somerisk

More than 40 % ofspecies face somerisk

30 % of species facesome risk

Crops and famine Crop productivity isexpected todecrease. Globalfood production isexpected todecrease, causing anincreased risk offamine

Crop productivity isexpected todecrease. Globalfood production isexpected todecrease, causing anincreased risk offamine

Crop productivitymay increase insome regions anddecrease in others.Increased risk offamine in someareas

Other effects Increase in intensityand frequency ofheat waves.Increased range fortropical diseases.Together, these willcause death andsickness, placing asubstantial burdenon health services

Increase in intensityand frequency ofheat waves.Increased range fortropical diseases.Together, these willcause death andsickness, placing asubstantial burdenon health services

Increase in intensityand frequency ofheat waves

123


determine annual consumption of electricity and fuel (If participants knew their fuel expen-ditures but not their fuel source for heating, a weighted average of all fuel sources for thestate was used.). Annual consumption of electricity was then converted into CO2 emissionsusing the average CO2 intensity for each state. Fuel consumption was converted into CO2

emissions using information about CO2 intensity for each fuel type. If participants did notknow their electricity and heating expenditures, we gathered information about their housingstructure and compared it to information about average energy consumption for houses ofsimilar age, type and size in their state, which was then used to calculate CO2 emissionsas above. Information about fuel prices, generation mix and average household energy con-sumption was obtained from the Energy Information Administration of the Department ofEnergy.

Information about participants’ cars and miles driven was directly computed based oncombined city/highway fuel economy information from the EPA for every make, model andyear of car from 1983 to 2009. For air travel, short flights were assumed to be 100 miles eachway, long flights 750 miles, and international flights 4,250 miles. Carbon offsets reducedthe carbon footprint by 168 pounds for every dollar spent, equivalent to prevailing rates atpopular commercial carbon offset retailers.

Median estimated carbon emissions for the sample were 17.9 tons per household peryear. For subjects in the control group, no other information was provided.4 Individ-uals in the treatment groups were informed that “Others like you who took this sur-vey in the past had a carbon footprint of xx tons per year” and whether their contri-bution was MORE or LESS than this value. The “xx” value was randomly assignedto be high (26 tons) or low (11 tons). For example, as depicted in Fig. 1, a sub-ject with an estimated carbon footprint of 18 tons and was assigned to the “See Low”group would be told that “Others like you who took this survey in the past had a car-bon footprint of 11 tons per year” and that “Your contribution to global warming isMORE than this average.” Similarly, a like individual who was assigned to the “SeeHigh” treatment was “Others like you who took this survey in the past had a car-bon footprint of 26 tons per year” and that “Your contribution to global warming isLESS than this average”. 26 tons and 11 tons were selected because they were thefootprint from actual sub-samples collected during pilot experiments that happened tobe near the 25th and 75th percentile of the total sample. This ensures that on aver-age about half of all of those treated were informed that they were relatively more cul-pable than others, while half received information that they were relatively less culpa-ble. As will be discussed below, the difference between the subject’s carbon footprintand the value associated with the reference individual provided a measure of relativeculpability.

Given this information the demand for green electricity was elicited using a modification ofa green electricity payment card used in Champ and Bishop (2001, 2006) in which individualswere given opportunities to buy blocks of energy measured in kilowatt hours. As shown inFig. 2 each block had a corresponding monthly and annual cost and estimated annual tonsof CO2 averted based on information available from the Energy Information Agency of theDepartment of Energy.

In Part IV, debriefing and demographic questions were asked, along with ten questionsdesigned to measure environmental concern drawn from the New Environmental Paradigm(NEP) scale (Dunlap and Van Liere 1978; Dunlap et al. 2000.) This scale is widely used in the

4 In pilot experiments, we also compared the results of a control group where no information about carbonfootprint was given to the current control group where the carbon footprint was given without peer comparisonand found no significant difference in behavior.

123

B. Ho et al.

Fig. 2 Elicitation question for contingent valuation in on-line survey

psychology and sociology literature to characterize an individual’s environmental concernbased on the extent to which they agree or disagree with various statements of environmentalconcern:

limits to growth, anthropocentrism, the fragility of the balance of nature, rejection ofthe idea that humans are exempt from the constraints of nature, and the possibility of aneco-crisis or ecological catastrophe. The modified NEP-scale is commonly used in thepsychology literature and aims at capturing the following five facets of environmentalconcern: The response categories range between 1 and 5 so that high scores correspondto a stronger pro-environmental attitude than low scores (with the ordering reversedfor the statements that reject the NEP-paradigm) (Ek and Soderholm 2008, p. 175)

Past studies of demand for green electricity have found the aggregated values across a seriesof NEP questions to be a significant, exogenous explanatory variable (Kotchen and Moore2007; Ek and Soderholm 2008). We also asked subjects their political party identifica-tion, and political orientation on a Likert scale that ranged from “Very Liberal” to “VeryConservative”.

Twelve observations in our data set were identified as outliers and excluded from analysis:ten of these observations were excluded because at least one component of their carbonfootprint was much greater than the rest of the sample, often an order of magnitude more.These observations were unrealistically high values, appearing to be incorrectly enteredresponses as to miles driven, airline flights, carbon offsets purchased, or housing information.The other two observations are repeated surveys. Removing these twelve observations halvesthe mean of the reported carbon footprint and reduces the standard deviation by an order ofmagnitude. Regressions with the outliers included returned the same qualitative results butwere largely insignificant.

2.3 Lab Experiment

We conducted a parallel experimental economics laboratory in which subjects purchase“private commodities” (analogous to electricity) that generate a negative public externality(analogous to pollution) for a group in which they are a member. The subjects are subsequentlygiven an opportunity to contribute to a fund that would reduce the negative harm created bythe externality, akin, we believe to the opportunity to purchase green electricity.

123


Subjects (n = 240) were recruited from a variety of undergraduate business and eco-nomics courses at Cornell University. Pen and paper experimental sessions were conductedin the Laboratory for Experimental Economics and Decision Research in cohorts rangingin size from 10 to 20. A session lasted approximately 45 min and average earnings were$14.41.

Subjects were randomly assigned into groups of five anonymous participants includingthemselves. Adapting elements of Plott’s (1983) seminal externality experiments,5 each indi-vidual was given a balance of $9 at the beginning of each of five rounds and a per-unit value(demand) function for a commodity that could be purchased at a cost of $1 (experimentaldollars were converted to real dollars at a rate of $15 experimental = $1 real.). Subjects ineach group were randomly assigned into high, low and medium demands (2 high, 1 medium,2 small) and the choices offered to individuals were presented (see Appendix for full exper-imental instructions). Subjects were asked to read all of the instructions before beginning,but received no information about the choices of others except for the information from theexperimental treatment.

In addition to private return for each commodity unit purchased, subjects were informedthat each unit purchased would impose a negative externality on the entire group,

Your group also shares a GROUP FUND. This group fund began with 300 experimentaldollars, and at the end of the experiment, any dollars in this group fund will be dividedequally between all members of the group. Your actions and the actions of other peoplein your group in Round 1 may have reduced the total amount of dollars remaining inthe group fund.

In Round [1–5], every unit of the commodity that you purchase decreases the numberof experimental dollars in the group fund by 1.25. (Because there are five people in yourgroup, every unit of the commodity that you purchase reduces the amount in the groupfund by 0.25 dollars per person. Likewise, every unit of the commodity purchased byeveryone else in the group reduces the amount in the group fund by 1.25 dollars andtherefore costs everyone else 0.25 dollars.)

Hence, the optimal private decision would be to purchase only those commodities witha value of $1.25 or higher. Examples were worked through with the entire session on awhiteboard at the front of the lab, and after each decision, subjects were asked to cal-culate and report their own private returns and the impacts of their private decisions onother members of the group. Subjects were asked to sum their commodity purchases overthe first five rounds and write this number down on a “passing sheet” which was sub-mitted to the experimental moderator. The experimental moderator passed these sheetsback to other subjects, who were then asked to record their own total purchases and theamount of total purchases that they saw on the sheet that was passed to them. Those inthe high culpability treatment received the sheet of someone else with low demand, thosein the medium culpability treatment received their own sheet, those in the low culpabil-ity treatment received the sheet of someone with high demand. Finally, subjects were eachgiven the opportunity to play a public goods with the same group, where they could pay$1 to increase the group fund by $1.25. Amounts were chosen so that students couldexactly offset their negative externality, although the words offset and externality were neverused.

5 For the sake of simplicity, we did not include the double auction, and focused only on the demand side ofthe experiment.

123

B. Ho et al.

3 Analysis and Results

3.1 Contingent Valuation Experiment

Our analyses of the contingent valuation and laboratory experiments break the sample intotreatment and control groups. In the contingent valuation “Treatment” group, subjects wereinformed about the carbon footprints of “Others like [them] who took this survey in thepast”, with others like them corresponding to the “See Low” (n = 111) and “See High”(n = 84) information described previously. Similarly, the “Treatment” group in the LabExperiment is organized by whether subjects were passed information from a subject with a“High” (n = 63), “Medium” (n = 29) or “Low” (n = 62) induced demand. No such relativeinformation was provided to the “Control” groups in the contingent valuation (n = 79) andlab (n = 64) experiments.

Averages for the control and treatment groups are provided in Tables 1 for the contingentvaluation experiments. In the contingent valuation experiment, the dependent values reportedare annual quantity demanded for green electricity. As these data are not conditioned on otherpossible covariates, some caution should be taken in interpreting the treatment effects. How-ever, it is particularly notable that in both cases, providing information appears to eithernot affect average contributions or has a negative effect relative to the control group. Thehigh culpability (11 ton) inducement yielded the same purchases of green electricity (214.0kWh/year) as the control (213.8 kWh/year). The low culpability inducement led people topurchase less green electricity (160.7 kWh/year). This would suggest that providing socialinformation tends to decrease purchases of the public good. The average level of green elec-tricity purchased of the full treatment group was (191.3 kWh/year). If these results generalize,then contingent valuation studies that fail to provide information about peers would on aver-age provide higher values than studies that provide such information, regardless of whetherthe individual is higher or lower than the norm. Such a result corresponds to the “brokenwindows” effect that observing others violate one social norm makes subjects more likely toviolate other social norms (Keizer et al. 2008).

Columns (4) and (5) show the summary statistics divided by those who saw peer informa-tion lower (“high culpability”) or higher (“low culpability”) than themselves. While the greenelectricity demanded in these columns cannot be cleanly interpreted because membershipin the low or high culpability treatment is endogenous and depends on the individual’s owncarbon footprint, dividing the dataset in this way will be useful when we turn to regressionanalysis to understand the asymmetry in behavior. However, we address the endogeneitydirectly in the lab experiment.

Econometric modeling reveals more about the structure of how subjects responded tothe peer information. In modeling the responses to the contingent valuation experiment,the dependent variable we use is “kWh per year of green electricity.” Given the discrete,ordered nature of the payment card response options, we adapt Cameron’s expenditure dif-ference model (1988) for the interval modeling format developed in Cameron and Huppert(1989), wherein circling a particular threshold value provides the midpoint of an intervalbounded from above by the midpoint between the selected value and the value above,and bounded below by the midpoint between the selected value and the value below.Assuming a logistically distributed contributions function, and letting E(contributions) =γ Z and var(contributions) = σ2 yields the following log likelihood function:

Ln (L) =∑n

i=1ln

[F ((γ Zi − tiU )/θ) − F ((γ Zi − ti L )/θ)

], (1)

123


where F(·) indicates the logistic distribution, Z is a vector of covariates, tiU is the upper boundof the interval selected, tiL is the lower bound, and the scale parameter θ = σ

√3/π . For the

contingent valuation data used here, the bounds are the midpoint of the ranges below andabove the selected value.

For the treatment group, we constructed a relative culpability variable measuring thedifference between the subject’s carbon and the “other” carbon footprint he/she was shown.

Culpability = Own carbon footprint – Observed carbon footprint of others (2)

In specifications where we include the control group which had no information about theirpeers, we set culpability to zero on the assumption that people assume their footprint is aboutthe same as others. However, subjects in the control condition may have felt culpability justby being asked about their own demographic information. For this reason, the demographiccontrols may pick up some of the culpability effect for those in the Control condition. Hencewe would expect including the controls in the regression to attenuate our estimates. Thus wefocus on the regression specifications that drop subjects in the control condition. Columns (1)through (5) in Table 3 report regressions on Control and Treatment groups separately. In theseregressions we also included controls for the subject’s own carbon footprint (CO2 Total), theNEP scale response summed over the 10 Likert scale NEP questions (NEP),6 and a self-reported political scale (Political Scale) variable extending from 0 (very liberal) to 6 (veryconservative), which has been recoded into a binary variable for liberal political leaning at themedian of the sample. These latter two variables comport with the environmental and politicalorientation variables in the Costa and Kahn study (2010). In addition, standard demographicand socio-economic variables of the type typically included in contingent valuation research(age, gender, children in household, income and education) are added as covariates.

Table 3 reports estimation results for Full Models with all the aforementioned covariatesand Short Models with only a subset of the variables. The vector of covariates was organizedinto three sub-vectors: (1) Estimation Variables (Constant, Theta); (2) Culpability Measures(Relative Culpability > (<) 0; Relative Culpability, CO2 Total); and (3) Demographic Vari-ables (NEP, Politics, Children, Age, Income, Education). For both the latter two groups theestimation strategy followed the pretest estimation procedure presented in Goldberger (1991)wherein Likelihood Ratio Tests were used to test the zero-null-vector hypothesis for the entiregroup (which was rejected in all cases). This was followed by a stepwise procedure in whichthe most insignificant coefficients were sequentially dropped. Coefficients were retained inthe short model if their corresponding p values were less than the cutoff value of 0.15. Further,CO2 Total was kept as a control variable in all estimations.

Looking at the relative culpability variable in Table 3 reveals that though on average,those who received peer information were willing to contribute less than those who did not,people are indeed positively and significantly influenced by relative culpability—those whowere induced to feel relatively more culpable were willing to pay more than those who wereinduced to feel relatively less culpable. Specifically, for each ton of CO2 a person is led tobelieve that she polluted more than others, her purchase of green power increases by around4.01 kWh (the coefficient on Relative culpability in Column 2 of Table 3). For context, theaverage culpability as calculated by Eq. (2) for someone who saw a lower footprint was 16.96tons, and the mean contribution for the control group was 213 kWh/year. In addition to theestimation variables, culpability, and CO2 Total, only the NEP covariate was retained in theShort Model.

6 The Cronbach alpha value for the subjects for the NEP questions was 0.7785, generally consistent with theliterature, and indicating that the NEP is a coherent metric.

123

B. Ho et al.

Table 3 MLE results for contingent valuation experiment

(1) (2) (3) (4) (5)Control Treated

Continuous culpability Conditional culpability

Full model Short model Full model Short model

Relative culpability>0 4.768*** 4.470**

(1.785) (1.797)

Relative culpability<0 3.277* 2.956*

(1.769) (1.773)

Relative culpability 4.012*** 3.697***

(1.378) (1.382)

CO2 total 0.935 −1.215 −0.868 −1.779 −1.414

(0.994) (1.433) (1.453) (1.664) (1.662)

NEP 8.693*** 5.481*** 6.113*** 5.379*** 5.989***

(2.775) (1.485) (1.400) (1.485) (1.407)

Liberal 3.383 10.09 9.878

(43.24) (23.61) (23.57)

Children 13.80 30.04 32.50

(40.04) (21.98) (22.28)

Gender −36.72 −30.33 −29.53

(40.77) (20.90) (20.91)

Age −0.159 1.502* 1.490*

(1.948) (0.851) (0.848)

Income −7.433 −8.766 −8.412

(9.666) (6.373) (6.387)

Education 54.82 21.62 21.06

(39.79) (21.58) (21.58)

Constant −110.7 −56.60 −47.99 −52.95 −41.83

(125.2) (66.67) (57.16) (66.78) (57.81)

Theta 87.89*** 78.40*** 80.41*** 78.25*** 80.25***

(8.963) (4.993) (5.099) (4.988) (5.094)

Observations 79 195 195 195 195

Log likelihood −160 −398.1 −402.1 −397.9 −428.27

Standard errors in parentheses *** p < 0.01; ** p < 0.05; * p < 0.1Regressions in this table all include controls for politics, children, gender, age, income and education

To better reconcile the regression results with the aggregate effects, we interact binaryvariables for those with positive culpability scores (those who are induced to feel more cul-pable than others) and those with negative culpability scores (those who are induced to feelless culpable than others) with the relative culpability measure. This is referred to as “Con-ditional Culpability” in Table 3. Columns (4) and (5) present the results and find suggestiveevidence that the impact of peer information is asymmetric. Those who are more culpablethan those they observed significantly increase their contributions by 4.470–4.768 kWh foreach ton of additional culpability. Looking at the relatively culpability for those who are less

123


Table 4 Additional specifications for MLE results for contingent valuation experiment

(6) (7) (8) (9) (10)Control and treatedcombined

25th to 75th percentilehouseholds

Main specw/outliers

Treated −34.88* −33.77 −50.12 −72.18*

(20.75) (23.60) (31.05) (39.87)

Relative 2.271** 5.871

Culpability>0 (0.979) (4.968)

Relative 2.475 −0.831

Culpability<0 (1.742) (3.649)

Relative 2.324*** 1.878 3.118**

Culpability (0.817) (1.971) (1.447)

CO2 total 0.583 0.603 −7.766** −8.591** −3.084**

(0.741) (0.768) (3.365) (3.463) (1.445)

NEP 6.559*** 6.567*** 8.833*** 8.968*** 4.950***

(1.315) (1.318) (1.899) (1.905) (1.577)

Constant −74.99 −75.51 −0.0804 4.276 7.849

(59.69) (59.93) (98.86) (99.02) (69.98)

Theta 81.76*** 81.77*** 75.03*** 74.86*** 84.73***

81.76*** (4.413) (6.210) (6.186) (5.237)

Observations 274 274 117 117 204

Log likelihood −560.9 −560.9 −236 −235.6 −429.3

Standard errors in parentheses *** p < 0.01; ** p < 0.05; * p < 0.1Regressions in this table all include controls for politics, children, gender, age, income and education

culpable (Relative culpability < 0), the effect is smaller and of lower insignifice, althoughthe difference between the coefficients for positive culpability and negative culpability areinsignificant (p value of 0.59 in an F-test). The lab experiment provides stronger evidenceon this asymmetry.

Table 4 contains alternative specifications for the data. Column (10) presents the mainspecification with all outliers reintroduced, and consistent with the effects of measurementerror, we get similar estimates but with magnitudes reduced. Columns (6) and (7) repeatour baseline specification but we combine the treatment and control. Combining the twoadds the implicit assumption that those in the control condition feel some constant amount ofculpability. If this assumption is violated, for example merely being asked to think about yourcarbon footprint induces feelings of culpability as well, then we would expect our coefficienton culpability to be attenuated as part of the effect of culpability would be picked up byfootprint variable.7 In fact this is what we see.

The culpability experiment introduces a further source of endogeneity in that a person’sculpability is also correlated with their household CO2 usage. Note that since we control foreach individual’s CO2 total, the coefficient on culpability is identified off the exogenously

7 For example, consider a simple OLS regression where βc is the coefficient on culpability, and β f is thecoefficient on footprint, and we estimate Y = βC T +β f f +ε, where T is an indicator for the treatment groupand f is the carbon footprint. If we assume those in the control group experience no culpability, we are thenestimating in the model Y = β f f + ε for those in the control so that the β f variable will pick up part of theculpability effect.

123

B. Ho et al.

assigned treatment group. There remains the concern that in this asymmetry, we are merelycapturing the difference between those with high footprint and low footprint in a way thatis not controlled for by the inclusion of the footprint variable (perhaps due to a non-linearrelationship). This endogeneity concern can be addressed by using the subject’s assignment toeither a high or low culpability treatment, as an instrumental variable, which is correlated withthe perceived culpability of the subject, but uncorrelated with any subject characteristics.8

Another way to partially address this endogenity is to restrict our sample to only subjectsbetween the 25th and 75th percentile of consumption.9 This way, for those who receivethe treatment, whether they experience positive or negative culpability is entirely randomlydetermined by the treatment. Columns (8) and (9) present these results, but due to the lossof statistical power, coefficients on culpability are similar but insignificant. To fully addressendogeneity concerns, we rely on the results from the lab experiment where footprint isexogenously assigned.

3.2 Lab Experiment

In order to better isolate the effect of culpability we rely on the results of a context-freelab experiment in which an individual’s culpability in producing a public bad is due to anexogenously induced demand for the private good. Since culpability depends only on ownconsumption levels and the observed consumption levels of others, the lab experiment allowsa degree of exogenous control over both components.

Table 5 presents the summary statistics for the lab experiment. Note once again, that eventhough positive culpability was induced for two of the three treatment conditions, as before,all conditions yielded less (or at most equal) altruistic behavior than the control (3.36 tokens),although only one was statistically significantly. On average, it appears that information onculpability leads to less altruistic behavior in both CV and experimental laboratory settings.

Since each unit of a subject’s consumption choice generates negative externalities onothers in the experimental session, we use their consumption choice as the analogue for“carbon footprint.” Also, in order to ensure the exogeneity of the culpability variable, we usethe expected target footprint he would have been induced to select if he were a completelyself-interested rationally maximized individual given the treatment condition he was in (highdemand, medium demand, low demand) instead of using the subject’s actual own “footprint”minus footprint of others.

Induced Culpability = Induced target footprint − Observed footprint of others

This measure of induced culpability is highly correlated (ρ = 0.7799) with actual culpa-bility which was defined as actual footprint minus observed footprint of others, but ensuresthat the culpability score is exogenous and not correlated with subject characteristics likealtruism, as is possibly the case in the CV experiment.

Table 6 presents the maximum likelihood estimates using a similar econometric modeland estimation strategy as the one used for the CV experiment:one minor diffence is thathere the quantity interval is bounded by the quantity selected and the next possible value.Similar asymmetric patterns emerge. In the continuous culpability model the coefficient onrelative culpability is not significant in the full sample, nor are any other covariates Whenthe estimation separates those who were either above or below the norms shown, those withrelatively high induced relative culpability provide significantly more to the public good in

8 These results are available from the lead author.9 This approach was suggested by an anonymous external reviewer.

123


Table 5 Summary statistics for laboratory experiment

Control Entire sample Treated

By induced demand By culpability

Small Medium Large High culpability Low culpability

Round6 3.36 2.54 2.67 3.38 2.99 2.45

Purchases (3.57) (3.14) (3.12) (3.42) (3.24) (2.93)

Relative – (4.84) 4.46 10.76 (4.08) 15.05

Culpability – (8.37) (7.97) (10.36) (12.21) (8.45)

Total purchases 20.20 14.23 21.15 25.08 13.73 25.99

(9.92) (4.68) (6.01) (10.96) (5.56) (7.72)

NEP 24.18 22.76 25.53 24.30 22.83 24.60

(5.49) (5.08) (6.02) (5.74) (4.98) (6.25)

Liberal 0.59 0.69 0.54 0.53 0.61 0.58

(0.49) (0.46) (0.50) (0.50) (0.49) (0.50)

Democrat 0.50 0.54 0.46 0.46 0.51 0.47

(0.50) (0.50) (0.50) (0.50) (0.50) (0.50)

Obs 86 95 48 97 75 73

Standard errors in parentheses

the short, but not the full model. There is an insignificant effect for those with less relativeculpability in both the full and short models for the treated sample. Note that we used amaximum likelihood model here to be consistent with the CV specification, but OLS andIV regressions using experimental assignment to instrument for culpability yielded similarresults. OLS and IV specifications were clustered by experimental group.

Here again we see the asymmetric effect when one sees higher others compared to seeinglower others in the short model (F-test of the difference in coefficients, p = 0.03). Thenegative, coefficients on culpability when culpability was less than zero is a bit puzzling.These effects however are largely insignificant, and appear to be driven by a few outliers whobehaved in ways hard to reconcile with most typical theories (e.g. consuming at a point thatwas welfare destroying for both themselves and their group as a whole).

Therefore, in our final two columns of Table 6 we report our preferred specification, wherewe focus on the treated to account for the attenuation effect, and drop 14 of 240 observationscorresponding to subjects that chose to consume more than what was even privately optimalin the first part of the experiment (i.e. they consumed at levels where the private cost exceededthe private benefit). As demonstrated, these results were qualitatively the same as those withoutliers included.

4 Heterogeneity in Responsiveness to Norms

Costa and Kahn (2010) noted the heterogeneous effect of the peer information experimenton Democrats versus Republicans. We confirm their findings in Table 7 by dividing the datainto self-identified “Democrats” (a relatively liberal party in the United States) and all others(Non-DEM). We extend their work by also considering heterogeneity in other dimensions,including number of children, gender, age, income, education, and NEP score, available

123

B. Ho et al.

Tabl

e6

ML

Ere

sults

for

labo

rato

ryex

peri

men

t

Con

trol

Tre

ated

Full

sam

ple

Tre

ated

(out

liers

excl

uded

)

Con

tinuo

uscu

lpab

ility

Con

ditio

nal

culp

abili

tyFu

llm

odel

Shor

tmod

elFu

llm

odel

Shor

tmod

elC

ontin

uous

culp

abili

tyC

ondi

tiona

lcu

lpab

ility

Con

tinuo

uscu

lpab

ility

Con

ditio

nal

culp

abili

ty

Tre

ated

−0.5

98−1

.440

**(0

.458

)(0

.597

)R

elat

ive

0.05

05*

0.05

020.

0577

*0.

06*

Cul

pabi

lity>

0(0

.030

1)(0

.030

6)(0

.032

0)(0

.04)

Rel

ativ

e−0

.049

6−0

.047

8−0

.088

0*−0

.03

Cul

pabi

lity<

0(0

.044

1)(0

.044

1)(0

.047

0)(0

.05)

Rel

ativ

e0.

0134

0.01

360.

0034

20.

02C

ulpa

bilit

y(0

.020

2)(0

.020

3)(0

.020

3)(0

.02)

Tota

l−0

.015

7−0

.024

3−0

.029

0−0

.025

90.

392

−0.0

152

−0.0

169

−0.0

1−0

.01

Purc

hase

s(0

.044

2)(0

.033

0)(0

.032

8)(0

.032

7)(0

.485

)(0

.027

3)(0

.026

9)(0

.04)

(0.0

4)N

EP

0.07

06−0

.018

4−0

.010

70.

0141

0.01

94−0

.05

−0.0

4(0

.079

7)(0

.041

0)(0

.040

5)(0

.038

2)(0

.037

4)(0

.05)

(0.0

5)D

emoc

rat

0.80

10.

482

0.47

50.

427

0.51

20.

472

0.87

0.74

(0.8

52)

(0.4

91)

(0.4

84)

(0.4

90)

(0.4

44)

(0.4

40)

(0.5

6)(0

.56)

Sess

ion

dum

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Con

stan

t2.

152

2.82

2*2.

452*

*2.

190

2.04

4*2.

832*

2.88

*2.

42(2

.741

)(1

.534

)(1

.124

)(1

.555

)(1

.135

)(1

.456

)(1

.61)

(1.6

3)T

heta

1.86

6***

1.55

4***

1.58

9***

1.54

3***

1.57

8***

1.77

3***

1.75

5***

1.55

***

1.54

***

(0.1

73)

(0.1

08)

(0.1

09)

(0.1

07)

(0.1

08)

(0.0

976)

(0.0

965)

(0.1

2)(0

.12)

Obs

.82

150

154

150

154

232

232

119

119

Log

likel

ihoo

d−2

14.1

−366

.7−3

79.4

−365

.4−3

78.2

−594

.6−5

92.3

−290

.11

−289

.27

Stan

dard

erro

rsin

pare

nthe

ses

***

p<

0.01

;**

p<

0.05

;*p

<0.

1

123


Table 7 MLE results for democrat/non-democrat split: contingent valuation experiment

Democrat Not democrat

Full model Short model Full model Short model

Relative 7.025** 5.714** 2.615* 2.263

Culpability (2.745) (2.430) (1.576) (1.655)

CO2 total −4.110 −2.944 0.265 0.627

(2.600) (2.482) (1.678) (1.759)

NEP 3.937 4.313 5.852*** 6.650***

(2.834) (2.762) (1.649) (1.609)

Liberal −75.72 16.15

(65.43) (25.65)

Children 37.82 29.40

(36.33) (27.22)

Gender 26.93 −52.25**

(37.27) (25.96)

Age 1.767 1.214

(1.670) (0.954)

Income 6.996 −16.94**

(12.28) (6.790)

College 11.43 32.50

(36.69) (26.52)

Constant 17.13 60.90 −56.27 −99.61

(137.0) (111.7) (75.35) (64.42)

Theta 82.49*** 86.83*** 70.31*** 74.68***

(8.173) (8.492) (5.884) (6.214)

Obs 80 80 115 115

Log likelihood −162.2 −165.2 −228.6 −235

Standard errors in parentheses*** p < 0.01; ** p < 0.05; * p < 0.1

for the relatively diverse contingent valuation study. Tables 8 and 9 decompose the effectof culpability by each of these demographic variables. Summary statistics and correlationtables are found in the appendix—note that although these demographic characteristics arecorrelated, the correlations are quite low, ranging from −0.11 (for attributes CO2 Total andGender) to 0.19 (for attributes Age and NEP)

We first note that our results are consistent with Costa and Kahn (2010). As shown inthe first two columns of Table 7, the coefficient on relative culpability for Democrats waspositive and significant, indicating that such individuals are responsive to social norm nudges.Indeed, in the regression this parameter dominates in the sense that the coefficients for theother explanatory variables are not significant. As shown in the last two columns, however,neither the coefficient for Culpability nor for the CO2 Total is significant: non-democrats arenot affected by our culpability inducement. Yet, coefficients for NEP and Political Scale aresignificant and consistent with expectations in the Non-Democratic regressions.

It is evident that this heterogeneity in response patterns extends to other dimensions.Table 8 parallels the approach used in Table 7. That is, we divide the sample into subsamples

123

B. Ho et al.

Table 8 MLE results for demographic subgroups for contingent valuation experiment (full regression)

Subgroup Culpability coefficient

Liberal 4.045**

Not liberal 3.767*

Children 5.546***

No children 2.625

Male 4.131**

Female 3.545*

Age>36.5 6.005***

Age<36.5 1.548

Income>4.7 7.220***

Income<4.7 0.669

At least college 5.332***

Less than college 2.436

NEP>34.5 5.34**

NEP<34.5 2.377

Democrat 7.025**

Not democrat 2.615*

Each row reports the coefficient on culpability for the full regression run on the sub-group of the populationspecified*** p < 0.01; ** p < 0.05; *p < 0.1

above and below the median, testing the null hypothesis that the coefficient on relativeculpability is equal to zero. Such a test is of particular interest to those who are providing thenudge. Our results show that induced culpability is effective for liberals but not non-liberals;for those with children but not for those without children; for men but not for women; forthose older than 36.5 but not those younger; for those above approximately $50,000 forincome but not for those below; for those with a college degree but not for those without;for those who are more environmentally conscious (NEP score > 34.5). For brevity we onlyreport the culpability coefficients for each estimation.

Table 8 decomposes the effect of culpability across different groups by considering aregression where culpability is interacted with all of our demographic variables, in effecttesting the null hypothesis that the subsamples described above respond the same to nudges.This is done by creating constant and slope shift parameters for each demographic charac-teristic (the short model was created by eliminating all interaction terms that had a p valueless than 0.15 in a Likelihood-Ratio Test). We find that when accounting for the entirety ofmeasured demographic variables that people that are liberal, with children, have a relativelyhigh income and high NEP scores behave significantly different than their counterparts.

A possible practical explanation for the patterns in Tables 7, 8 and 9 is that peer informationnudges work on those already inclined to give, such as democrats, liberals and those with highNEP scores, but do not work and may even backfire when preaching to those less inclined (seeMeier 2007a, b for a brief summary of related work on the importance of heterogeneity). It isalso possible that in the specific context of climate change, those who question the premiseof whether climate change is happening may be unresponsive.

We should be careful to note that this heterogeneity analysis should be seen as exploratoryand mostly provided to be suggestive for future work. However, the fact that such heterogene-ity exists appears quite robust. Awareness of this heterogeneity is important for increasing the

123


Table 9 MLE results for demographic subgroups for contingent valuation

Long model Short model

Relative culpability 4.564** 5.176***(2.185) (1.863)

Liberal (binary) 33.46 34.84(23.96) (24.22)

Culpability× liberal 2.180 2.242(1.448) (1.366)

Children (binary) 28.18 19.79(23.84) (23.31)

Culpability×children (23.84) (23.31)(1.327) (1.195)

Female (binary) −17.61(21.58)

Culpability× female −0.204(1.290)

Age (binary) 1.873** 1.867**(0.885) (0.881)

Culpability×age 1.050 0.521(1.515) (1.367)

Income (binary) −8.674 −8.179(6.892) (6.698)

Culpability× income −0.656* −0.532*(0.379) (0.310)

College (binary) 16.54(22.52)

Culpability×college 0.689(1.483)

NEP (binary) 50.79** 47.57**(22.85) (22.68)

Culpability×NEP −1.530 −1.654(1.187) (1.027)

CO2 (binary) −19.07(32.15)

Culpability×CO2 1.223(2.340)

Constant 82.19 77.15(58.16) (55.93)

Theta 78.84*** 79.85***(4.993) (5.037)

Observations 196 196Log likelihood −401.8 −403.5

All variables are binary, continuous variables (like age and income) were made binary by splitting at themedian. Standard errors in parentheses*** p < 0.01; ** p < 0.05; * p < 0.1

precision of estimates of the effect of peer information interventions, as well as for increasingthe cost effectiveness of future norm based interventions.

5 Conclusions

Using a contingent valuation framed field experiment coupled with a conventional lab exper-iment, we examine how peer information that induces culpability differs from the peer infor-mation interventions based on conformity that have been the traditional focus of the “nudge”

123

B. Ho et al.

literature. We demonstrate that there is important heterogeneity in how quantities purchasedrespond to culpability based peer information, and find similar patterns of heterogeneity forboth the online contingent valuation experiment and the context free lab experiment using aconvenience sample. We find that the culpability effect is larger when the information makessubjects feel good about themselves relative to when the information makes them feel guilty.We also find evidence that the effect of culpability comes mostly from those who may bemore inclined to act more pro-socially.

These results have potentially important implications for public policy. It provides evi-dence on the usefulness of culpability as a separate channel than the norm and conformitybased behavioral nudges commonly put forward. It also provides guidance to policy makerson how best to target such interventions on the sub-groups most likely to respond. How-ever, our results also suggest caution on over reliance on informational nudges. Strategiesthat induce culpability primarily affect individuals who were already more inclined to reduceenergy consumption in the first place. As a consequence, behavioral policies are less effectiveon the worst offenders and therefore can complement but cannot replace traditional policieslike taxes or regulations that affect the entire population.

Acknowledgments The author would like to thank William Schulze, as well as seminar participants at Cor-nell University, Peking University, Sydney University, Vassar College and the Economic Science Associationfor helpful comments.

References

Akerlof GA, Kranton RE (2000) Economics and identity. Q J Econ 115:715–753Allcott H (2011) Social norms and energy conservation. J Public Econ 95:1082–1095Allcott H, Mullainathan S (2010) Behavior and energy policy. Science 327:1204–1205Andreoni J (1995) Cooperation in public-goods experiments: kindness or confusion? Am Econ Rev 85:891–

904Ayres I, Raseman S, Shih A (2013) Evidence from two large field experiments that peer comparison feedback

can reduce residential energy usage. J Law Econ Organ 29(5):992–1022Bamberg S, Moser G (2006) Twenty years after Hines, Hungerford, and Tomera: a new meta-analysis of

psycho-social determinants of pro-environmental behavior. J Environ Psychol 27:14–25Bardsley N (2008) Dictator game giving: altruism or artefact? Exp Econ 11:122–133Bateman IJ, Munro A, Poe GL (2008) Decoy effects in choice experiments and contingent valuation: asym-

metric dominance. Land Econ 84:115–127Battigalli P, Dufwenberg M (2007) Guilt in games. Am Econ Rev 97:170–176Baumeister R, Stillwell A, Heatherton TF (1994) Guilt: an interpersonal approach. Psychol Bull 115:243–267Bernherim B (1994) A theory of conformity. J Polit Econ 102:841–877Bikhchandani S, Hirshleifer D, Welch I (1992) A theory of fads, fashion, custom, and cultural change as

informational cascades. J Polit Econ 100:992–1026Brouwer R, Brander L, Van Beukering P (2008) “A convenient truth”: air travel passengers’ willingness to

pay to offset their CO2 emissions. Clim Change 90:299–313Brown TC, Nannini D, Gorter RB, Bell PA, Peterson GL (2002) Judged seriousness of environmental losses:

reliability and cause of loss. Ecol Econ 42:479–491Bulte E, Gerking S, List JA, de Zeeuw A (2005) The effect of varying the causes of environmental problems

on stated WTP values: evidence from a field study. J Environ Econ Manag 49:330–342Cai H, Chen Y, Fang H (2009) Observational learning: evidence from a randomized natural field experiment.

Am Econ Rev 99:864–882Cameron TA (1988) A new paradigm for valuing non-market goods using referendum data: maximum likeli-

hood estimation by censored logistic regression. J Environ Econ 15:355–379Cameron TA, Huppert D (1989) OLS versus ML estimation of non-market resource values with payment card

interval data. J Environ Econ Manag 17:230–246Carlsmith J, Gross A (1969) Some effects of guilt on compliance. J Personal Soc Psychol 11:232–239Champ P, Bishop R (2001) Donation payment mechanisms and contingent valuation: an empirical study of

hypothetical bias. Environ Resour Econ 19:383–402

123


Champ P, Bishop R (2006) Is willingness to pay for a public good sensitive to the elicitation format? LandEcon 82:162–173

Charness G, Dufwenberg M (2006) Promises and partnership. Econometrica 74:1579–1601Charness G, Dufwenberg M (2007) Broken promises: an experiment. SSRN: http://ssrn.com/abstract=1114404Chen Y, Harper M, Konstan J, Li SX (2009) Social comparisons and contributions to online communities: a

field experiment on movielens. Am Econ Rev 100:1358–1398Cialdini RB, Kallgren CA, Reno RR (1991) A focus theory of normative conduct. Adv Exp Soc Psychol

24:201–234Cialdini RB, Demaine LJ, Sagarin BJ, Barrett DW, Rhoads K, Winter PL (2006) Managing social norms for

persuasive impact. Soc Influ 1:3–15Clark C, Kotchen M, Moore M (2003) Internal and external influences on pro-environmental behavior: par-

ticipation in a green electricity program. J Environ Psychol 23:237–246Costa D, Kahn M (2010) Energy conservation ‘Nudges’ and environmentalist ideology: evidence from a

randomized residential electricity field experiment. NBER Working Paper No. w15939Costa D, Kahn M (2013) Energy conservation ‘Nudges’ and environmentalist ideology: evidence from a

randomized residential electricity field experiment. J Eur Econ Assoc 11:680–702Dufwenberg M, Lundholm M (2001) Social norms and moral hazard. Econ J 111:506–525Dunlap R, Van Liere K (1978) The new environmental paradigm: a proposed instrument and preliminary

results. J Environ Educ 9:10–19Dunlap R, Van Liere K, Mertig A, Jones R (2000) New trends in measuring environmental attitudes: measuring

endorsement of the new ecological paradigm: a revised NEP scale. J Soc Issues 56:425–442Ek K, Soderholm P (2008) Norms and economic motivation in the Swedish green electricity market. Ecol

Econ 68:169–182Ellison G, Fudenberg D (1993) Rules of thumb for social learning. J Polit Econ 101:612–643Ferraro PJ, Price MK (2013) Using non-pecuniary strategies to influence behavior: evidence from a large-scale

field experiment. Rev Econ Stat 95:64–73Frey B, Meier S (2004) Social comparisons and pro-social behavior: testing ‘Conditional Cooperation’ in a

field experiment. Am Econ Rev 94:1717–1722Geanakoplos J, Pearce D, Stacchetti E (1989) Psychological games and sequential rationality. Games Econ

Behav 1:60–79Glaeser E, Scheinkman J (2002) Non-market interactions. In: Dewatripont M, Hansen LP, Turnovsky S (eds)

Advances in economics and econometrics: theory and applications, eighth world congress. CambridgeUniversity Press, Cambridge

Goldberger AS (1991) A course in econometrics. Harvard University Press, LondonGoldstein N, Cialdini R, Griskevicius V (2008) A room with a viewpoint: using social norms to motivate

environmental conservation in hotels. J Consum Res Inc 35:472–482Grossman P, Eckel C (2012) Giving versus taking: a ‘Real Donation’ comparison of warm glow and cold

prickle in a context-rich environment. Monash Economics Working Papers 20-12, Monash University,Department of Economics

Harrison G, List J (2004) Field experiments. J Econ Lit 42:1009–1055Kahneman D, Ritov I, Jacowitz KE, Grant P (1993) Stated willingness to pay for public goods: a psychological

perspective. Psychol Sci 4:310–315Keizer K, Lindenberg S, Steg L (2008) The Spreading of disorder. Science 322:1681–1685Korenok O, Millner EL, Razzolini L (2013) Impure altruism in dictators’ giving. J Public Econ 97:1–8Kotchen M, Moore M (2007) Private provision of environmental public goods: household participant in green-

electricity programs. J Environ Econ Manag 53:1–16Kuziemko I, Norton MI, Saez E, Stantcheva S (2013) How elastic are preferences for redistribution? Evidence

from randomized survey experiments. Natl Bur Econ Res (w18865)List J (2007) On the interpretation of giving in dictator games. J Polit Econ 115:482–494Mazar N, Zhong CB (2010) Do green products make us better people? Psychol Sci 21:494–498Meier S (2007a) A survey of economic theories and field evidence on pro-social behavior. In: Frey BS, Stutzer

A (eds) Economics and psychology: a promising new field. MIT Press, Cambridge, pp 51–88Meier S (2007b) Do women behave less/more pro-socially than men. Public Financ Rev 35(2):215–232Plott CR (1983) Externalities and corrective policies in experimental markets. Econ J 93:106–127Salganik MJ, Dodds PS, Watts DJ (2006) Experimental study of inequality and unpredictability in an artificial

cultural market. Science 311:854Schultz PW, Nolan JM, Cialdini RB, Goldstein NJ (2007) The constructive, destructive, and reconstructive

power of social norms. Psychol Sci 18:429–434Shang J, Croson R (2009) A field experiment in charitable contribution: the impact of social information on

the voluntary provision of public goods. Econ J 119:1422–1439

123

http://ssrn.com/abstract=1114404

B. Ho et al.

Solnick S, Hemenway D (2005) Are positional concerns stronger in some domains than in others? Am EconRev Pap Proc 95(2):147–151

Sonnemans J, Schram A, Offerman T (1998) Public good provision and public bad prevention: the effect offraming. J Econ Behav Organ 34:143–161

123

The Effects of Moral Licensing and Moral Cleansing in Contingent Valuation and Laboratory Experiments on the Demand to Reduce Externalities

Documents