Tax compliance under tax regime changes

The difference between predicted and observed compliance does not only occur in studies based on survey or field data but also for experimental studies where “in most cases the level of tax compliance was higher than predicted” (Torgler 2002, p. 677).

This is an important feature when introducing a tax regime change because it allows us to directly compare the two regimes. We hold the income potential constant over time and avoid, for instance, learning dynamics in answering the trivia questions that could have arisen from a new income determination in every period. The random component in income determination gives us controlled variability over income. Other design choices would have been possible, especially a design without a random component and a quiz every round. Practical reasons (the shorter time requirement) and the stationarity of the income variable were important arguments in favor of our choice.

For reasons of parsimony and analytical clarity we chose very straightforward tax regimes. Furthermore, they make it much easier for subjects to understand the task.

Our experimental program induced the expected income E[Y] of the population to be slightly skewed to the right and, hence, the expected revenues of the two tax regimes were not completely identical in the experiment. All our conclusions are unaffected by this feature.

In the instructions we only confronted participants with tables including the tax amounts not making average and marginal tax rates explicit. Thus we avoided any possible confusion from tax rate terminology (“marginal tax rates” versus “average tax rates”), which is not always familiar to non-economists. Of course, we include individuals’ marginal tax rates in our econometric analysis of compliance below.

Like in many other experiments, we choose an auditing probability that is considerably higher than the one in the real world. This is to account for the fact that several real-world leveraging effects of auditing such as potential social disapproval after being caught cheating or increased auditing scrutiny after once being caught cheating are not separately modeled in our experiment.

In addition to the hypotheses developed below we have considered a possible impact of social preferences and fairness perceptions. Theoretically, social preferences such as inequity-aversion might be relevant. For example, above-average income earners could increase compliance under progressive taxation in order to reduce inequality. Furthermore, the perceived relative fairness of tax regimes could play a role. Spicer and Becker (1980) have provided evidence that people who believe that the tax system treats them unfairly relative to others tend to engage in more tax evasion. However, none of these aspects showed any significance in our experiment: There is neither a kink in compliance rate when income exceeds the average income. Nor were we able to establish an empirical link between the perceived fairness of a tax regime (covered by our post-experiment survey questions) and compliance behavior following a tax reform.

Simulations are available on request.

A motive of “loss repair” could also do the trick. Andreoni et al. (1998) explain the unexpectedly negative effect of audits on compliance with the intention to get back some of the money foregone after a fine. In analogy, a tax reform confronting the individual tax payer with losses should lead to more evasion motivated by a compensation strategy.

Providing subjects with instructions for the first two parts right away helps to make the claim that later endowments will depend on the performance in the quiz more credible.

Note that this is a conservative test on the level of matching group averages.

The task consists of ten decisions between Option X and Option Y, where both options include a lottery. Option X is the relatively safer option because both possible lottery outcomes are between the outcomes of option Y. Throughout the decisions, the payoffs are fixed, but the probability of receiving the higher payoff increases by 10 percentage points from 10 % in decision 1 to 100 % in decision 10 in both options. Depending on the subject’s risk attitude, the subject should, moving down the decisions, switch at some point from Option X to Option Y (or in the unlikely case of being extremely risk-loving always choose Option Y). From the switching point, risk attitudes can be calculated. The exact numbers and how the task was incentivized can be found in Appendix B. The method has been used several times before in the experimental literature on tax evasion (see, for instance, Cummings et al. 2009).

Summary statistics for all regression variables are presented in Appendix A.

For our participants, the switching point for the risk measure is, on average, 6.7 for men and 7.1 for women. Notice that in other regression models we also used another variable for risk attitudes that reflects the potentially non-linear relationship between risk attitudes and compliance. When we use a dummy variable for extreme risk aversion (for having a switching point greater than 8), the risk coefficient in the regression becomes larger in magnitude. This is exactly what one would expect.

Note, however, that by construction the second regime dummy is highly correlated (correlation coefficient +0.87) with the period indicator. This means that we cannot disentangle whether the decreasing tax honesty over the course of the experiment is simply a consequence of time and experimental experience, or whether it is also affected by the regime change.

In our setup, subjects know their earnings potentials when they choose between the two systems. A design that lets them decide behind the veil of ignorance would also be interesting, but less natural in the context of our entire experimental setup.

Regression results are available on request.