# «Forthcoming, American Economic Review Abstract We conduct an experiment assessing the extent to which people trade off the economic costs of ...»

Wooldridge 2006). From Equation (2), adding a stochastic error and rearranging terms, each participant i’s (i=1,…261) latent utility difference between truthtelling and lying at direct economic ECOSTj is given by Yij* C i bECOST j ij. (3) Under utility maximization, an observed realization of TRUTHFUL CHOICE, Tij, is related

**to Yij* by the following mechanism:**

economic costs of truthfulness. (In Section IV, we explore sources of variation in C i and we discuss that, besides the direct effect of reducing the attractiveness of truthfulness, ECOST may have an indirect effect through the total costs of lying.) The standard errors correct for possible serial correlation and heteroskedasticity by clustering at the individual level. (Recall that participants went through all five economic cost situations.) Column (1) of Table III shows the results of this analysis. Consistent with the fact that many individuals did, in fact, tell the truth, the constant term is positive. Importantly, ECOST is a highly significant determinant of the relative attractiveness of truthfulness and lying for an individual. Indeed, the implied marginal effect of ECOST is powerful: A 30cent increase in ECOST was associated with a 16.9% decrease in truthtelling.14 Together with the observations made in Table II, this finding supports Hypothesis HET.

** TABLE III ABOUT HERE**

Column 2 of Table III adds individual-level controls. The main result for ECOST remains unchanged. We observe some interesting additional findings. First, women appeared to be more likely to tell the truth, as did students in fields outside of psychology and economics.

Second, given the decision-theoretic nature of the experiment, altruistic and distributional concerns, as well as attempts to live up to others’ expectations so as to avoid guilt, should not have affected behavior in this experiment. Yet, we note that 35HURTS does enter significantly in this baseline regression. This suggests one of two possibilities. Either participants’ altruistic concerns drove behavior, or 35HURTS was correlated with some If is normally distributed, one obtains the probit model. As is typical in econometric applications, the two models yield virtually identical inferences.

In the real world, managers are indeed faced with substantial cross-sectional and time-series variation in the economic cost of truth-telling. Our results are consistent with findings by Bergstresser and Philippon (2006), who showed that the use of discretionary accruals to manipulate reported earnings was more pronounced at firms where CEO compensation depended more on the stock price.

general differences in preferences that were in turn correlated with intrinsic costs of lying and thus reflected variation in a variable omitted in Column 2. As Section IV shows, the latter explanation is corroborated by the data.

The conclusions we draw from these main results are simple but important. Hurkens and Kartik (2009) demonstrated that Gneezy’s (2005) data would be consistent with a population of pure opportunists, who always lie, and pure ethical types, who always tell the truth (as in Koford and Penno 1992). However, the present evidence of changeability in truthtelling behavior and of significant sensitivity to economic costs associated with truthfulness rejects Hypothesis TYP and is in direct contrast to the implications of a typebased model.

** IV. Sources of heterogeneity in total costs of lying**

We have established that the participants in our experiment showed more variation in their total costs of lying than if they had belonged to one of just two fixed extreme types. In this section, we expand on these findings by considering various potential sources of the variation among individuals in total costs of lying. Moreover, this extension allows us to provide further evidence refuting the type-based model; in particular, we document that there is also heterogeneity within individuals (across situations) in total costs of lying.

**A. Enhanced model**

We consider two constituent sources of heterogeneity in total costs of lying. First, these costs are driven by individuals’ Intrinsic Costs Of Lying, for which we use the term ICOLi.

Second, we also allow the situation, that is, the economic costs of truthfulness or the extrinsic incentives for lying, ECOSTj, to influence the total costs of lying. Adjusting notation, we posit that total costs of lying may vary both among and within individuals, and we now write C ij C ij ICOLi, ECOST j . Since all participants encountered the same ECOST situations, the evidence on heterogeneous total costs of lying provided in Section III necessarily implies that there is heterogeneity in ICOLi. Next, the intrinsic costs of lying and the economic incentives for lying (economic costs of truthfulness) may enter Cij separably or non-separably. Indeed, whether intrinsic preferences and extrinsic incentives interact in determining total preferences for a certain action has implications that extend beyond the scope of the current study; see, for example, Bowles and Polanía-Reyes (2011) for a discussion of how incentives to contribute to public goods may affect social preferences. In order to capture both possibilities, we consider, for parsimony, a simple parametric specification C ij ICOLi, ECOST j 0 1 ICOLi 2 ECOST j 3 ICOLi ECOST j, (6)

coefficient thus allows us to test whether a candidate measure of ICOL helps explain heterogeneity among individuals in total costs of lying. Moreover, specification (6) allows for two channels through which heterogeneity in total preferences for truthfulness within individuals enters. First, it seems reasonable to postulate that Cij is increasing in ECOST so

only tells us that Cij is not increasing in ECOST at a rate greater than marginal utility b.

Second, heterogeneity of preferences within individuals for truthfulness can unambiguously be detected by considering the interaction term IE. Under the discrete-choice model’s assumptions, a significant interaction term provides further evidence against the notion, posited by the type-based model, that there are two fixed types.15 Specifically, a positive interaction term arises if the intrinsic costs of lying are more important in determining total preferences for truthfulness when the stakes (ECOST) are higher. An equivalent interpretation (useful in settings where an agent can choose the size of the lie) is that individuals with stronger intrinsic costs of lying perceive “larger” lies, which yield larger economic benefits, as less attractive and will, thus, tell “smaller” lies. A negative interaction term instead arises if the source of the intrinsic costs of lying is relatively less influential at higher stakes. If the coefficient on the interaction is zero, the economic costs of truthtelling are perceived identically by all agents, regardless of the strength of their intrinsic costs of lying, ICOLi. In that case, all agents’ utilities would react identically to changes in the economic costs, even though heterogeneous ICOLi would imply that some would report the truth while others would lie at a given ECOST.16 C. Results We first consider three possible sources of intrinsic costs of lying, ICOLi, for which survey measures are available to us, and we then discuss other possible drivers of behavior. The Alternatively, if the true utility function has a separable form but the assumption of weak exogeneity of the error term ij does not hold, then the interaction term in the model may serve as an instrument to correct for ij. The interaction correlation between the explanatory variables (ICOLi and ECOSTj) and the error term term can serve as an instrument because it arises as one of the terms in the second-order Taylor-series expansion of the random utility function (with violated weak exogeneity). It captures jointly the effects of both explanatory variables. One anonymous Referee provided an example in which 90% of individuals make deterministic decisions based on an additively separable utility function, i.e., they tell the truth when Ci=ICOLi is greater than ECOST, and they lie when ICOLi is smaller than ECOST. 10% of individuals make decision errors; that is, they tell the truth although ECOST is greater than ICOLi, and they lie even when ICOLi is greater than ECOST. In simulated data, the Referee showed that a logit regression (which is a misspecified model under the assumptions made) may yield a significant interaction term on ECOST and intrinsic costs of lying, even in this setting. In the Referee’s example, the error term is not independent of the observable variables. Thus, the interaction term becomes significant in this setting because it is an instrumental variable for an omitted variable. We conclude that, even if the true utility function is separable, researchers may well want to use a non-separable reduced form because this specification is robust to the violation of the assumption of weak exogeneity of the error term, such as occurs in decision errors of the form proposed by the Referee.

When testing for non-separability (that is, for the significance of the interaction term), we consider coefficients, rather than marginal effects, from the logit regressions. Recall that marginal effects in a logit regression are given by β' X 1 β' X β, where is the logistic cumulative distribution function giving the initial probability of truthfulness. Those with high (low) ICOL have high (low) initial probabilities of truthfulness. Thus, the highest marginal effects of ECOST on behavior are likely to be found in the middle range, and smaller marginal effects are likely to be found among those with high intrinsic costs.

Analyzing coefficients instead allows us to consider the hypothetical case of participants who would display identical initial probabilities of reporting the truth.

descriptive statistics for EXTDECEIT, SELFDECEIT, and PV shown in Table IV indicate that there is wide variation in these three variables, suggesting that they could potentially explain the observed variation in truthtelling behavior. In Table III, we test whether this is the case. We allow each possible source to affect behavior both separably from economic costs and jointly by way of an interaction.

** TABLE IV ABOUT HERE**

First, participants may have developed an interest in impressing the experimenter by appearing honest and non-greedy (e.g. Fischbacher and Heusi 2008); this would act like a preference for truthfulness. Given the design of the experiment, in which we took great care to make the responses anonymous, this is very unlikely to have occurred. Indeed, EXTDECEIT is not significant in any of the regressions, and neither is the interaction term with ECOST.

Second, it is possible that participants deceived themselves by making the “right” choices. However, SELFDECEIT is also not significant in any of the regressions, and neither is the interaction term with ECOST.

Third, we consider the possibility that moral values were a source of the intrinsic costs of lying. While many moral concepts are potentially relevant, we focus on protected values (PV). The literature that has developed the theory of these values emphasizes that protected values are non-consequentialist and induce a resistance to engaging in actions that would violate moral values, reducing the attractiveness of any financial gains obtained through such actions.17 That is, the economic costs of truthfulness matter less to those who hold stronger protected values of truthfulness; those people are trade-off resistant. This idea naturally translates into a functional form for Ci that is non-separable into intrinsic (moral) costs of lying and economic costs of truthfulness.

Column (3) of Table III shows that PV of truthfulness was a highly significant predictor of behavior in the experiment. A one-point increase in PV was associated with a 17.4% See, for example, Baron and Spranca (1997); Tetlock, Kristel, Elson, Green, and Lerner (2000); and Tanner, Medin, and Iliev (2008). The source of protected values is modeled by Bénabou and Tirole (2011) as a need of agents to invest in their identity. For the strongest form of PV, “sacred” values and taboos, see in particular section V of their paper.

increase in the probability of truthtelling, holding the other variables at their means. In Column (4), we obtain a positive, significant coefficient on the interaction term between PV and ECOST. This is evidence that, conditional on the correctness of the discrete choice (logit) model’s specification, the data are consistent with non-separability of the economic incentives and this measure of intrinsic costs of lying. That is, the data confirm that there is heterogeneity within individuals’ total costs of lying, again inconsistent with the type-based model’s assumption.

Note that E IE PVi is negative even when evaluated at PV = 6. Thus, in the crossˆ ˆ section of participants, the presence of a strong protected value of truthfulness lessened, but did not eliminate, the relevance of the economic costs associated with the earnings management decision. With PV in the regression, the significance of the demographic controls vanishes. It is also noteworthy that, as soon as we include the interaction term with PV, 35HURTS is no longer significant.18 Finally, as shown in Column (5), we also find that our results continue to hold in the subsample without the free-truth situation.

We emphasize that, despite these findings, one cannot conclude that PV has a stronger claim to organizing the data than plausible alternatives. For example, participants may be driven by non-consequentialist preferences that attach expressive utility to low-stakes acts or decisions that substantiate or confirm personal identity. This expressive-preferences concept was developed in the political science literature to explain why citizens vote despite an apparent lack of economic incentive (Buchanan 1954; Tullock 1971).