THE CODING OF OUTCOMES IN TAXPAYERS REPORTING DECISIONS. A. Schepanski The University of Iowa

Transcription

1 THE CODING OF OUTCOMES IN TAXPAYERS REPORTING DECISIONS A. Schepanski The University of Iowa May 2001 The author thanks Teri Shearer and the participants of The University of Iowa Judgment and Decision-Making Workshop for helpful comments. Correspondence should be addressed to: A. Schepanski, Henry B. Tippie College of Business, The University of Iowa, Iowa City, Iowa

2 1 THE CODING OF OUTCOMES IN TAXPAYERS REPORTING DECISIONS ABSTRACT: Prospect theory posits a coding operation by which decision makers internally represent outcomes in terms of changes from a neutral reference point. Kahneman and Tversky have shown that, by manipulating the way in which outcomes are described to decision makers, the implied reference point can also be manipulated, resulting in reversals of preference that can violate normative principles of decision making. In order to describe and predict behavior in naturally-occurring settings, however, a better understanding is needed of how reference points naturally arise. This study investigates the coding process that tends naturally to be used to frame the settlement outcome in a taxpayer s year-end filing decision. The study tests two competing explanations of the psychological processes underlying the coding process. The first of these, the current asset position hypothesis, has received empirical support in prior research. The competing hypothesis, the hedonic coding hypothesis, posits that the reference point adopted is based on principles of utility maximization. The results, which favor the hedonic coding hypothesis, extend our theoretical understanding of how reference points are determined in taxpayers compliance decisions, and provide relevant implications for taxing authorities, tax advisors, and others concerned with compliance in tax reporting. Keywords: reference point; hedonic coding; taxpayer compliance; prospect theory.

3 2 INTRODUCTION Kahneman and Tversky (1979) introduced prospect theory, a descriptive model of choice under uncertainty, in an attempt to explain certain discrepancies between observed behavior and expected utility theory. Prospect theory distinguishes between two phases in the choice process--an early phase of framing, or editing, and a subsequent phase of evaluation. The theory describes several operations that might occur in the framing phase, of which one, coding, is the subject of the present research. The coding operation refers to the process by which individuals internally represent outcomes in terms of changes from a neutral reference point. The nature of the neutral reference point is not precisely defined in prospect theory. According to Kahneman and Tversky (1979, 274), The reference point usually corresponds to the current asset position, in which case gains and losses coincide with the actual amounts that are received or paid. However, the location of the reference point, and the consequent coding of outcomes as gains or losses, can be affected by the formulation of the offered prospects, and by the expectations of the decision maker. Thus the same outcome could be coded in more than one way depending upon the reference point that is used by the decision maker in evaluating her choice. However, the reference points that seem to have been given the most attention in the literature are current asset position and expected asset position, where the latter includes the individual s current asset position plus the anticipated outcomes of certain future events. Kahneman and Tversky have shown that, by suitably altering the manner in which outcomes of prospects are labeled to the decision maker, reversals of preference can occur that violate normative principles of decision making. In their well-known Asian disease problem (Tversky and Kahneman 1981), for example, the reversal of preference (from predominantly risk averse in the condition where outcomes were labeled as lives saved to predominantly risk seeking in the condition where outcomes were labeled as lives lost ) was attributed by Kahneman and Tversky to the manipulation of the neutral reference point (from expected asset position in the lives saved condition to current asset position in the lives lost

4 3 condition). However, Kahneman and Tversky cannot predict which of these reference points would tend naturally to be adopted, because the very description of the alternative conditions entails labeling the outcomes as either gains (lives saved) or as losses (lives lost). Relatively little is known about how these reference points naturally arise (Yates and Stone 1992). While Tversky and Kahneman s Asian disease problem shows that individuals can employ both current asset position and expected asset position as reference points, prospect theory provides no guidance regarding the circumstances in which one or the other of these would tend naturally to be employed by individuals in decision making in the absence of explicit gain and loss labels. Yet an awareness of the processes governing the location of these reference points is crucial to our understanding of decision making behavior in naturallyoccurring settings. Kahneman and Tversky (1979) emphasize this point as follows: the location of the reference point, and the manner in which choice problems are coded and edited emerge as critical factors in the analysis of decisions (288). And as Fischhoff (1983, 104) notes, in order to predict behavior in less controlled situations, one must be able to anticipate how problems will be represented and what frames people will use to interpret them (see also, Payne, Laughhunn and Crum 1980). The taxpayer s reporting decision is one context in which the reference point might be expected to arise naturally. In this context, there are no explicit gain or loss labels. Hence, the processes that are used to code the settlement outcome will be those which tend naturally to occur in this decision context. Accordingly, this study builds on prior experimental research by testing between two competing explanations of the psychological processes that tend naturally to be employed by taxpayers in the coding phase of the compliance decision. The results extend our theoretical understanding of how reference points are determined in taxpayers compliance decision making, as well as providing relevant implications for taxing authorities, advisors, and others concerned with compliance in tax reporting. 1

5 4 The balance of the paper is organized as follows. The first section describes prior experimental research on the adoption of a neutral reference point by individual taxpayers making tax-reporting decisions. The conclusion that emerged from this research was that current asset position best described the reference point adopted by taxpayers when making compliance decisions. In the second section, the hedonic coding hypothesis, an extension of Thaler s (1985) principle of hedonic efficiency, is advanced as a competing explanation of the processes leading to the adoption of a neutral reference point. The third section reports upon an experiment designed to test between current asset position and the hedonic coding hypothesis in a tax compliance decision context. 2 The results, reported in the fourth section, are consistent with the hedonic coding hypothesis. The final section of the paper discusses the implications of these results. PRIOR EXPERIMENTAL RESEARCH If current asset position is the reference point taxpayers naturally employ in compliance decision making, then an implication of prospect theory is that overwithholding would lead to risk aversion and compliant behavior, whereas underwithholding would lead to risk seeking attitudes and noncompliant behavior. 3 Much of the early experimental research in taxpayer compliance, e.g., Jackson and Spicer (1986), Schadewald (1989), Schepanski and Kelsey (1990), and White et al. (1993), sought to test this implication of prospect theory. The general conclusion that emerged from these research studies is that overwithholding tends to foster greater compliance and underwithholding tends to foster greater noncompliance. However, these studies assumed that taxpayers routinely employ current asset position as the neutral reference point. Accordingly, the validity of their theoretical conclusion regarding the relationship between withholding position and compliance rests, in turn, upon the validity of the assumed reference point. If expected asset position better represents the reference point taxpayers tend naturally to employ, then prospect theory would predict that underwithholding, for

6 5 example, could have either a compliance decreasing effect (if the taxpayer s underwithheld position exceeded her expected underwithholding), or it could have a compliance increasing effect (if the taxpayer s underwithheld position was less than her expected underwithholding). Schepanski and Shearer (1995) is the only prior study to date to explicitly test between current asset position and expected asset position as alternative neutral reference points that might be employed to code tax-related outcomes. Preference behavior observed in that study was interpreted as favoring current asset position as better representing the reference point that taxpayers tend naturally to employ in tax-reporting decisions. Their results, therefore, support the conclusions from prior research. However, there are grounds for treating the conclusions of prior research with some skepticism. This is because an implication of prior research is that the use of current asset position is hard wired. That is, current asset position is always the reference point used, independent of surrounding circumstances. But this implication is contradicted in prior research by Tversky and Kahneman (1981). Their Asian disease study found that, depending upon how outcomes are labeled, individuals use either expected asset position or current asset position as reference points. What is lacking in the current literature is a theory that would identify the circumstances under which decision makers would select either current asset position or expected asset position as the reference outcome, in the absence of explicit gain and loss labels. The objective of the present study is to propose such a theory and to test it in controlled experimental conditions. This theory, referred to as the hedonic coding hypothesis, is described in the following section of the paper. This hypothesis posits that when individuals have flexibility regarding which reference point to adopt for decision making purposes, they will incorporate expectations or not incorporate expectations according to utility maximization principles. In contrast to the routine use of current asset position supported in prior research, the hedonic coding hypothesis attributes to the decision maker a more active approach to the decision framing process. As

7 6 described below, the hedonic coding hypothesis provides an alternative explanation for the results reported in Schepanski & Shearer (1995). HYPOTHESES Hedonic Coding. The reference points employed by taxpayers may be guided by a principle of hedonic efficiency. Thaler (1985) suggested this principle as a way of explaining individuals preferences for compound events. In the present research, the principle of hedonic efficiency is extended to the determination of reference points in taxpayers reporting decisions. Accordingly, when taxpayers are faced with a situation that is ambiguous with respect to the reference point to be adopted, they may code outcomes in a way that yields the highest utility. Because of the nonlinearities and asymmetries associated with the prospect theory value function, the principle of hedonic efficiency suggests, as discussed below, that in some circumstances a higher level of satisfaction (utility) can be achieved by including expectations in the reference point (i.e., by adopting expected asset position as the reference point), whereas in other circumstances satisfaction can be greater by excluding expectations from the reference point (i.e., by adopting current asset position as the reference point). To extend the principle of hedonic efficiency to taxpayers compliance decisions, it is assumed that the outcome coding process is conducted in the following two-step sequence. In the first step, the taxpayer selects a reference point (current asset position or expected asset position) to code the (riskless) outcome associated with the compliant alternative. In the second step, the reference point selected for the compliant alternative is assumed to be employed by the taxpayer to code the outcomes for the noncompliant alternative as well. This assumption is consistent with prospect theory s premise that a common reference point is employed to code outcomes associated with all available choice alternatives. Accordingly, the following discussion regarding the selection of reference points is couched in terms of the outcome associated with the compliant alternative.

8 7 Let x + x represent what the taxpayer views as the legally correct settlement amount at filing, where x is the expected settlement amount and x is the unexpected settlement amount. When x + x is positive, the taxpayer is said to be overwithheld, whereas if it is negative the taxpayer is referred to as being in an underwithheld position. The issue at hand is how the taxpayer would code the outcome x + x. One possibility is that the expected component is segregated from the unexpected component of the outcome and the two components are separately evaluated by the prospect theory value function. That is, v(x, x) = v(x) + v( x). The expected component, upon being evaluated (relative to current asset position), is then immediately incorporated into the neutral reference point, which, according to the theory, is assigned a value of zero. This converts the reference point from current asset position to expected asset position. Therefore, only the unexpected component, x, of the outcome from the compliant alternative is evaluated against expected asset position. Thus segregation implies a reference point based on expected asset position and the coding of the outcome x + x as v( x). Another possibility is that the expected component, x, is integrated with the unexpected component, x, and the combination evaluated by the prospect theory value function. This implies a reference point based on current asset position, and coding of the outcome for the compliant alternative, x + x, as v(x + x). Whether taxpayers segregate or integrate the expected and unexpected components is assumed to be guided by the principle of utility maximization.. The principle of hedonic efficiency, therefore, suggests the following hypothesis regarding the reference point taxpayers employ in making compliance decisions: When coding what is perceived as the legally correct settlement amount, the taxpayer will seek to maximize utility by incorporating expectations, or not incorporating expectations, into the reference point depending upon which of these possibilities yields the higher level of satisfaction. The reference point that

9 8 accomplishes this is then used to code the outcomes associated with the noncompliant alternative. These assumptions, in combination with those of prospect theory, 4 will be referred to in this manuscript as the hedonic coding hypothesis. This hypothesis can be used to generate predictions regarding the set of circumstances that would lead individuals to employ either expected asset position or current asset position as the reference point. There are six situations, or cases, to be analyzed when applying the hedonic coding hypothesis to taxpayers choice of reference point. The theoretical analysis for these six cases is presented in the following paragraphs. Case 1: Increase in Expected Refund (x >0; x>0; x + x>0). The hedonic coding hypothesis predicts that a higher level of satisfaction would be achieved by segregating x and x rather than integrating them. That is, v(x) + v( x) > v(x + x) because the concavity of the value function for gains implies that x (or x) would be more highly valued when it is evaluated near the neutral reference point than when it is evaluated further away from it that is, after receiving x (or x). Thus, the hedonic coding hypothesis predicts that expected asset position, rather than current asset position, would be the reference outcome employed. Case 2: Increase in Expected Payment (x < 0; x < 0; x + x < 0). For this case, a higher level of utility could be achieved by integrating x and x rather than segregating them. Since the value function in the domain of losses is convex, v(x + x) > v(x) + v( x) because x ( or x) is evaluated on a flatter portion of the value function and thus is less aversive when it is integrated with x (or x) than it would be if it were segregated from x. Accordingly, current asset position is the predicted reference outcome. The remaining four cases all involve different signs for the expected and unexpected components, one positive and the other negative. The reference point that will emerge as utility maximizing is uncertain for these cases. As discussed below, it will depend upon the

10 9 steepness 5 of the value function for losses relative to that for gains, and the absolute magnitudes of x relative to x. For the analysis of these cases, let us distinguish two possibilities with respect to the curvature of the value function one where the curvature for gains exceeds that for losses, denoted L<G, and another where the curvature for losses exceeds that for gains, denoted L>G. For L<G functions, the value function for losses is everywhere steeper than for equivalent gains, but that is not the case for L>G functions. For the latter, therefore, let us refer to the absolute magnitude of x or x as small if it is in the vicinity of the reference point, and large if it falls outside of this range. The terms small and large are used here to parallel prior research which supports prospect theory s assumption of a steeper value function for losses than for equivalent gains for outcomes in the vicinity of the neutral reference point (see footnote 4), but for outcomes outside of this range, the evidence is less clear. For L>G functions, therefore, the point at which small becomes large is the point where the value function for gains becomes steeper than for equivalent losses. Cases 3 and 4. The actual outcome in these cases is a refund (i.e., the taxpayer is overwithheld). Case 3 is a decrease in an expected refund (x > 0; x <0; x + x >0), and Case 4 is a decrease in an expected payment (x <0; x >0; x + x >0). In both cases the negative component is the smaller of the two in absolute magnitude. The question in both cases is whether evaluating the smaller negative component as a reduced gain (integration) is less aversive than evaluating it as a loss (segregation). L<G Functions. Since the value function for losses is everywhere steeper than for equivalent for gains, evaluating the smaller negative component as a reduced gain (integration) will always be less aversive than evaluating it as a loss (segregation). To see this, recall that if x and x are equal in absolute magnitude, integration (v(x + x) = v(0) = 0) is preferred to segregation (v(x) + v( x) < 0)--see footnote 5). From this position of equality of x and x,

11 10 assume, for expositional convenience, that the absolute magnitude of the negative component remains unchanged, hence its aversiveness when evaluated as a loss (segregation) remains unchanged, while the positive component is increased. Now the evaluation of the smaller negative component as a reduced gain (integration) occurs further from the neutral reference point than before the change, where the value function for gains is flatter, and thus less aversive than before the change. Hence integration everywhere dominates segregation and current asset position would be the utility maximizing reference point. L>G Functions. If the absolute magnitude of the negative component is small, then evaluating it as a reduced gain will be less aversive than evaluating it as a loss. This follows from the fact that the value function for losses is assumed to be steeper than for equivalent gains in this region of the value function, and so the analysis is the same as for L<G functions described above, regardless of the size of the positive component. Hence integration would be preferred to segregation and current asset position would be the preferred reference point. On the other hand, if both the positive and negative components are in the large region, where the value function for gains is steeper than for equivalent losses, the utility maximizing reference point is uncertain, and will depend upon the relative absolute magnitude of x and x. If x and x are equal to one another in absolute magnitude, segregation (v(x) + v( x) > 0) would be preferred to integration (v(x + x) = v(0) = 0--see footnote 5). From this position of equality, assume that the negative component is held constant while the positive component is increased. If the negative component is evaluated as a reduced gain (integration), it will occur further from the neutral reference point than before the change, where the value function for gains is flatter, and thus be evaluated as less aversive than before the change. With further increases in the positive component, eventually integration would be preferred to segregation. Accordingly, the larger the positive component relative to the absolute amount of the negative one, the more likely integration would be preferred to segregation and current asset position emerge as the utility maximizing reference point. Conversely, the more nearly equal the two components are in

12 11 absolute magnitude the more likely segregation would be preferred to integration and expected asset position emerge as the utility maximizing reference point. Thus, for example, if a taxpayer was expecting a large refund, but was entitled to only a relatively small one, expected asset position could emerge as the utility maximizing reference point, leading to risk seeking attitudes despite the taxpayer being overwithheld. Cases 5 and 6. The actual outcome in these cases is a payment (i.e., the taxpayer is overwithheld). Case 5 is a decrease in an expected refund (x>0; x<0; x+ x<0), and Case 6 is a decrease in an expected payment (x<0; x>0; x+ x<0). In both cases, the positive component is the smaller of the two and the issue is whether evaluating it as a reduced loss (integration) enhances utility more than evaluating it separately as a gain (segregation). L<G Functions. The utility maximizing reference point is uncertain and will depend upon the relative absolute magnitude of x and x. Let s begin the analysis at values of x and x where the two are equal to one another in absolute magnitude. At these values, integration is preferred to segregation because the value function is assumed to be everywhere steeper for losses than for equivalent gains. From any pair of (x, x) values equal to one another in absolute magnitude, assume that the absolute magnitude of the negative component is increased while the positive component remains unchanged. If the positive component is evaluated as a reduced loss (integration) it will occur at a point on the value function for losses further from the neutral reference point than before the change, where the function is flatter. Hence evaluating the positive component as a reduced loss will enhance utility less after the change than before the change. Accordingly, the larger the absolute amount of the negative component relative to the positive one, the more likely segregation would be preferred to integration and expected asset position emerge as the utility maximizing reference point. Conversely, the more nearly equal x

13 12 and x are to each other in absolute magnitude, the more likely integration would be preferred to segregation and current asset position emerge as the utility maximizing reference point. L>G Functions. Again, a distinction is made between outcomes in the vicinity of the neutral reference point ( small ) and those outside of this region ( large ). If the smaller positive component (x or x) lies in the small region the analysis is identical to that described above for L<G value functions. Hence the more nearly equal x and x are in absolute amount, the more likely current asset position is to emerge as the utility maximizing reference point. On the other hand, the larger the absolute size of the negative component relative to the positive one, the more likely expected asset position is to emerge as the utility maximizing reference point. For values of x and x that are in the large region, segregation is always preferred to integration and expected asset position would emerge as the reference point that maximizes utility. We again begin the analysis at any two values of x and x in this region equal to one another other in absolute magnitude. At these values of x and x segregation is preferred to integration because the value function is assumed to be steeper for gains than for equivalent losses. If, from any one of these (x, x) pairs, the negative component is increased while the positive one remains the same, the evaluation of the positive component as a reduced loss (integration) will occur further from the neutral reference point thus enhancing utility less than before the increase. Thus since segregation was preferred to integration before the increase, it will be even more preferred after the increase. Thus segregation dominates integration for values of x and x in this region, and expected asset position will emerge as the utility maximizing reference point. Thus current asset position would be expected to emerge as the utility maximizing reference point when the positive component is in the vicinity of the neutral reference point and differs from the absolute magnitude of the negative component by a small margin. Otherwise, expected asset position would emerge as the utility maximizing reference point. Thus, for

14 13 example, a taxpayer expecting to make a large payment but is required to make only a relatively smaller one is likely to exhibit risk averse attitudes, despite being underwithheld. These predictions of the hedonic coding hypothesis are summarized in Table 1 below Insert Table 1 about here Current Asset Position. This hypothesis assumes that the reference outcome employed by individuals will regularly and automatically include their current total net worth position, but it will not incorporate expectations regardless of whether or not a higher level of utility could be achieved by doing so. Thus, the adoption of current asset position is not contingent upon utilitymaximizing considerations, unlike the hedonic coding hypothesis. This noncontingent hypothesis will simply be referred to in this manuscript by the term current asset position. The context of the discussion will make it clear whether this term is referring to the current asset position hypothesis or to the reference outcome adopted under the hedonic coding hypothesis. Support for the current asset position hypothesis is given by Schepanski and Shearer s (1995) study, the only prior experimental research study cited to explicitly test between current asset position and expected asset position. Their results were interpreted as supportive of current asset position as best representing the reference outcome employed by taxpayers in taxreporting decision making. However, as discussed in the following paragraphs, the Schepanski and Shearer (1995) results are also consistent with the hedonic coding hypothesis. Subjects in the Schepanski and Shearer (1995) study were randomly assigned to one of four between-subject conditions, two control conditions and two treatment conditions. Subjects in all conditions were presented with a hypothetical tax-reporting situation that required them to choose between one of two tax-reporting alternatives. The compliant alternative involved a fixed dollar outcome (refund or payment due) whose probability was assumed to be 1.0. The noncompliant alternative involved two possible dollar outcomes depending upon whether the return was audited (in which case the dollar outcome was below that in the compliant alternative)

15 14 or not audited (in which case the dollar outcome exceeded that in the compliant alternative). Subjects in all conditions were advised of the likelihoods of each of the two dollar outcomes for the noncompliant alternative, and the expected values of the two choice alternatives were equal to one another. Dollar outcomes in each experimental condition were linear transformations of one another. Identical nonzero dollar outcomes involving additional payments were presented to subjects in two of the experimental conditions (one control condition and one treatment condition). For example, the outcome for the compliant alternative in these conditions was a payment of $250. Similarly, identical nonzero dollar outcomes involving refunds were presented to subjects in the remaining two experimental conditions. For the compliant alternative, for example, this outcome consisted of a refund of $1,000. Table 2 presents the manner in which the outcome associated with the compliant alternative was predicted to be coded in Schepanski and Shearer (1995) under each of the two hypothesized reference outcomes. The predicted coding for the noncompliant alternative is excluded for convenience Insert Table 2 about here Note that under the column headed Current Asset Position, the treatment condition involving an additional payment is obtained from the corresponding control condition by subtracting $1,250 from x (the expected portion of the outcome) and adding $1,250 to x (the unexpected portion of the outcome). Similarly, the treatment condition involving a refund is obtained from the corresponding control condition by adding $1,250 to x and subtracting this amount from x. This leaves unchanged the amount of the additional payment ( $250) and the amount of the refund ($1,000), but it changes the predicted outcome coding under the column Expected Asset Position. Accordingly, if current asset position better represents the reference point taxpayers tend naturally to employ in decision making, then preference behavior in the two experimental conditions involving refunds should not differ significantly from each other, and both

16 15 should exhibit significantly greater risk aversion than in the two experimental conditions involving payments. The payment conditions, in turn, should not differ significantly from each other. Alternatively, if expected asset position better represents the reference point taxpayers tend naturally to employ in decision making, preference behavior in the treatment condition involving an additional payment (refund) should not differ significantly from the control condition involving a refund (additional payment), but both experimental conditions coded as gains should exhibit significantly greater risk aversion than should the two experimental conditions coded as losses. Observed preference behavior in the treatment conditions was interpreted as favoring current asset position over expected asset position as better representing the reference outcome that taxpayers tend naturally to employ in tax-reporting decisions. As shown in Table 2, Schepanski and Shearer s (1995) treatment condition involving an additional $250 payment was made up of an expected component consisting of a $1,250 payment (x = $1,250) and an unexpected component which consisted of a reduction in the expected payment of $1,000 ( x = +$1,000). This condition corresponds to case 6 of the hedonic coding hypothesis. Since the values of x and x are nearly equal in absolute magnitude, the hedonic coding hypothesis favors current asset position as the reference point yielding greater utility for L<G value functions. This would be the case for L>G value functions as well if x is in the vicinity of the neutral reference point, which might well be the case here. Hence the design of this treatment condition can reasonably be regarded as favoring current asset position as the utility maximizing reference point. The treatment condition involving a $1,000 refund, as shown in Table 2, consisted of an expected component amounting to a $1,250 refund (x = +$1,250) and an unexpected component which reduced the expected refund by $250 ( x = $250). This experimental

17 16 condition corresponds to Case 3 of the hedonic coding hypothesis which, for L<G value functions, favors current asset position as the utility maximizing reference point. Even for L>G value functions, the hedonic coding hypothesis favors current asset position because this value of x is sufficiently small, both absolutely and relative to x, that evaluating it as a reduced gain (integration) is likely to be less aversive than evaluating it as a loss (segregation). Thus the hedonic coding hypothesis can account for the results in Schepanski and Shearer (1995). Accordingly, the present research, described below, seeks to extend prior research by testing between the current asset position hypothesis and the hedonic coding hypothesis as alternative explanations of the coding process employed by taxpayers when making tax-reporting decisions. Prospect theory is a maintained hypothesis in this study. If the experimental results suggest that decision makers segregate the expected and unexpected components of the outcome, this evidence, in conjunction with that from Schepanski and Shearer (1995), would be consistent with the theory that individuals can and do indeed employ either current asset position or expected asset position as reference outcomes in accord with utility-maximizing considerations as posited by the hedonic coding hypothesis. On the other hand, if the experimental results suggest that decision makers integrate the expected and unexpected components of the outcome, despite the measures taken (described below) to enhance the utility associated with segregating the components of the outcome, there would seem to be little logical basis for retaining the hedonic coding hypothesis as a serious theoretical possibility. METHOD Overview. Two considerations are central to the design of this study. First, the experiment is structured to allow the reference point that subjects tend naturally to employ to be inferred from their observed risk posture. Second, the experiment is designed to allow the following two uses of current asset position to be distinguished from one another: the routine employment of current asset position as the neutral reference point; and the contingent use of current asset position as the neutral reference point when a higher level of utility is associated with it (rather

18 17 than expected asset position), as posited by the hedonic coding hypothesis. This consideration, in turn, requires that the design of the experiment seek to maximize, to the extent possible, the utility associated with expected asset position, as discussed below. Table 3 summarizes the predicted reference point and risk posture for each of the six cases described above, under each of the two competing hypotheses. As this table Insert Table 3 about here illustrates, only in Cases 3 and 6 is it possible to distinguish between the two hypotheses, and only then when segregation yields more utility than integration. Accordingly, Cases 3 and 6 comprise the treatment conditions, and the design of the present research seeks to maximize, to the extent possible, the utility associated with segregating x and x. Specific design features included a personal bankruptcy feature, whose purpose was to induce insensitivity in the value function for losses, and the selection of values for x and x that would favor the use of expected asset position as the reference point if the hedonic coding hypothesis better represents the process of reference point selection. The experimental design, described below, is thus structured to give the hedonic coding hypothesis as favorable as possible an opportunity to be evidenced by the experimental data, without biasing the test against current asset position. These design features are discussed in more detail later in the manuscript. Experimental design and stimuli. Subjects were randomly assigned to one of four betweensubjects conditions: two control conditions (Unexpected Refund and Unexpected Payment) and two treatment conditions (Expected Refund based on Case 3; and Expected Payment based on Case 6). Subjects in all experimental conditions were presented with a hypothetical situation in which they are preparing their annual tax return at a point in time in the future. In each experimental condition, subjects were presented with dollar outcomes and related probabilities of these dollar outcomes for each of two tax-reporting alternatives a compliant choice alternative

19 18 involving a single dollar outcome whose probability was assumed to be certain; and a noncompliant choice alternative involving two possible dollar outcomes depending upon whether or not the deduction of uncertain legality claimed by the subject was denied by the IRS. These dollar outcomes and related probabilities in each of the four experimental conditions are presented in Table 4 below. The probabilities of each dollar outcome Insert Table 4 about here are given in parentheses where, for the noncompliant alternative, the probability of the lower dollar outcome was presented as Thus for the Unexpected Refund condition, for example, the compliant choice would entail claiming a refund of $100 to be received with probability 1.0. The noncompliant choice alternative would involve a 65% probability of receiving the claimed refund of $5,000 filed with the return. The noncompliant choice alternative would also involve a 35% probability of the claimed deduction being denied, and no refund being received from, nor additional payment made to, the IRS. Instructions. Subjects were instructed to assume that at the time they were preparing their annual tax return, they possessed a total net worth of $0, $5,000, $0, and $5,000 in the Unexpected Refund, Unexpected Payment, Expected Refund and Expected Payment conditions, respectively. This was done to ensure that final wealth states would be the same in all experimental conditions prior to subjects making their choices. Total net worth was described to subjects as the cash value of everything you own minus money you owe. Subjects in the Expected Refund and Expected Payment conditions were also advised that their net worth of $0 and $5,000, respectively, did not include an expected refund of $5,000, and an expected payment of $5,000, in these respective treatment conditions.

20 19 Subjects in all experimental conditions were asked to assume that, in the event they chose an alternative which gave rise to an outcome involving a payment which exceeded their total net worth, they would file for personal bankruptcy, thereby raising their total net worth to $0. This outcome could only occur in the event the noncompliant alternative was chosen and the claimed deduction was denied. In such an event, the IRS would levy the subject with an additional payment consisting of the additional tax, interest, and penalties amounting to $9,000 in the two Refund conditions and $14,000 in the two Payment conditions. Since losses were limited to the subjects total net worth, these foregoing amounts would be reduced by $9,000 in each of the four experimental conditions to produce the amounts shown in Table 4 ($0 in the Unexpected Refund and Expected Refund conditions, and $5,000 in the Unexpected Payment and Expected Payment conditions). This personal bankruptcy feature was included in the design of the experiment to induce insensitivity in the value function for losses in the neighborhood of subjects net worth position in each experimental condition. 7 Unexpected Refund Condition. Subjects were led to expect that their total tax liability would be about equal to the amount of taxes withheld, so that no refund or additional payment was expected (i.e., x = $0). Subjects then learned that a preliminary version of their return indicated that they would receive an unexpected refund of $100 (i.e., x = +$100). Subjects at this point learned that the rough draft of their return failed to include a subtraction item consisting of incurred business expenses, the deductibility of which is legally questionable. If this deduction were claimed, subjects were told that there was a 65% probability that their refund would amount to $5,000. They were also told that if the deduction were claimed, there was a 35% chance that the deduction would be disallowed, and they would be levied by the IRS for additional tax, interest, and penalties amounting to $9,000. Subjects were then advised that in this event, filing for personal bankruptcy would absolve debts sufficient to bring their total net worth back to $0. Subjects were then asked to indicate whether or not they would claim the business expense deduction on the final copy of their tax return.

21 20 Unexpected Payment Condition. Subjects were also led to expect that their total tax liability would be about equal to the amount of taxes withheld such that no refund or additional payment could be expected (i.e., x = $0). They then learned that a preliminary version of their return indicated an outcome of $0 no refund or additional tax due. At this point, subjects learned that their preliminary return included a business expense deduction of questionable legality. If this deduction were excluded from the return, the result would be an additional payment of $4,900 (i.e., x = $4,900). On the other hand, if this deduction were claimed, subjects were told that the return could be filed with no additional payment to be made, an outcome with a 65% probability. Subjects were also advised that there was a 35% probability that their claimed deduction would be denied, and they would receive a levy for additional tax, interest, and penalties amounting to $14,000. As in the Unexpected Refund condition, subjects then learned that filing for personal bankruptcy would raise their total net worth to $0 (i.e., an increase of $9,000). As shown in Table 4, outcomes for each choice alternative differed from those in the Unexpected Refund condition by an additive constant of $5,000, while the probabilities of each outcome remained the same. Subjects were instructed to indicate whether or not they would claim the deduction on the final version of the return to be submitted to the IRS. Expected Refund Condition. Subjects were led to expect their total tax liability to be $5,000 less than the total tax they had withheld during the year, so that a $5,000 refund was expected (i.e., x = +$5,000). After preparing a rough draft of their return which confirmed their expectation that they would be entitled to a refund of $5,000, subjects then learned that the return contained a legally-questionable business expense deduction which, if not claimed on the return, would reduce the refund to $100 (i.e., x = $4,900). Instructions for the noncompliant alternative paralleled those in the Unexpected Refund condition. They were then asked to indicate whether or not they would claim the legally-questionable deduction on the final draft of their tax return. As shown in Table 4, all outcomes and related probabilities were the same as those in the Unexpected Refund condition.

22 21 Expected Payment Condition. Subjects were led to expect that their total tax liability would be $5,000 less than the total tax they had withheld during the year, so that a $5,000 payment was expected (i.e., x = $5,000). Subjects then learned that a preliminary version of their return indicated that a payment of $4,900 would be required to file a compliant return instead of $5,000 (i.e., x = +$100). At this point, subjects learned that their return, like that of subjects in the Unexpected Refund condition, did not include a deduction for incurred business expenses whose deductibility is legally questionable. Instructions for the noncompliant choice alternative paralleled those in the Unexpected Payment condition. Subjects were then asked to indicate whether or not they would claim the questionable business expense deduction on their returns. Outcomes and probabilities for each choice alternative were the same as those in the Unexpected Payment condition (see Table 4). As noted earlier, two features were included in the design of the experiment to enhance the utility associated with segregating x and x such that, if hedonic coding better represents the outcome coding process employed by taxpayers, then expected asset position is more likely to be used as the neutral reference point. These features are summarized in the following paragraphs. The first was the personal bankruptcy feature discussed above, which was common to all experimental conditions. Since individuals losses were limited to their total net worth, it seems reasonable to hypothesize that the individual s value function for losses would be relatively flat, or insensitive, to outcome changes in this region. Accordingly, if subjects value functions exhibit greater steepness for gains than for equivalent losses (i.e., L>G functions), they are more likely to be observed for outcomes in this region. The second feature was the selection of values of x and x intended to complement the first feature in enhancing the likelihood of segregation. Accordingly, in the Expected Payment condition (based on Case 6), x was set at a large amount ( $5,000, equal to net worth) relative

23 22 to x (+$100) thus enhancing the likelihood that the smaller positive component would be evaluated as a gain rather than as a reduced loss, especially if the value function for gains is steeper than for equivalent losses at these values. 8 In the Expected Refund condition (based on Case 3), both x and x were set at large values, +$5,000 and $4,900 respectively, are likely to fall in the large region of L>G functions. As noted earlier, since the absolute difference between them is small, these values maximally favor segregation and the emergence of expected asset position as the utility maximizing reference point. 9 Response Scale. Subjects indicated their preferences on a 9-point rating scale where the numbers 6 through 9 represented preference for the noncompliant choice alternative, i.e., claiming the legally uncertain deduction on the return filed with the IRS. Each number represented a different verbal category as follows: 6 = slightly prefer, 7 = prefer, 8 = much prefer, and 9 = very much prefer the noncompliant alternative. The numbers 1 through 4 represented preference for the complaint choice alternative (i.e., not claiming the legally uncertain deduction on the filed return), where 1 = very much prefer, 2 = much prefer, 3 = prefer, and 4 = slightly prefer the compliant alternative. The midpoint of the scale (5) represented indifference and was labeled neutral. Efforts were made to personalize the hypothetical scenario in which subjects were placed and to make the expected and realized tax positions salient to subjects in a manner similar to that described in Schepanski and Shearer (1995, 178). Theoretical predictions. As shown in Table 4, subjects in the Unexpected Refund condition were presented with a pair of positive prospects, whereas those in the Unexpected Payment condition were presented with a pair of negative prospects. According to prospect theory, therefore, all nonzero outcomes should be coded as gains and as losses, respectively, in these control conditions. The first two rows of Table 5 present the predicted outcome coding in these control conditions for each hypothesized reference outcome, where and denote the predicted

24 23 preference or indifference. These evaluations of the prospects in the control conditions are independent of whether the reference point is better represented by current asset position or by expected asset position. This follows from the fact that the expected settlement amount at filing was $0 in both of the control conditions, so that a common reference point is implied by both current asset position and by expected asset position Insert Table 5 about here Note that no prediction is made regarding subjects preference behavior (risk averse, risk neutral, or risk seeking) in either of the control conditions or in either of the treatment conditions. The expected value of the noncompliant choice alternative exceeds that of the compliant choice alternative in all four experimental conditions, which would be expected to bias the responses of subjects in favor of the noncompliant alternative. In addition, the personal bankruptcy feature included in the experimental design, which was common to all experimental conditions, would be expected to bias the responses of subjects in favor of the compliant alternative. However, a prediction of preference behavior in each experimental condition is unnecessary to test whether current asset position or expected asset position better represents the reference point that individuals employed in this study. If subjects preference behavior in one control condition is observed to differ significantly from that in the other control condition, it is possible to test whether the reference point adopted by subjects is better represented by current asset position or by expected asset position by comparing subjects preferences in each of the treatment conditions with those observed in the control conditions as described below. If the reference point employed by subjects is better represented by current asset position, then the pair of prospects in the Expected Refund and Expected Payment conditions are predicted to be evaluated in the same manner as shown in Table 5 for the Unexpected Refund and Unexpected Payment conditions, respectively. This follows from the fact that the treatment conditions do not involve a manipulation of observed outcomes, only of expected