Loss Aversion, Transaction Costs, or Audit Trigger? Learning about Corporate Tax Compliance from a Policy Experiment with Withholding Regime

We analyze firm's tax choices facing a withholding and enforcement regime with a focus on three salient mechanisms of bunching: (i) transaction costs, (ii) withholding threshold as a reference point for taxpayer that creates a kink due to loss aversion, and (iii) withholding threshold as a reference point for audit (audit trigger model). The transaction costs model predicts that none of the firms that bunch at the withholding threshold would declare higher taxes when withholding rate is increased, as was the case in Ecuador in 2007. Evidence from a triple-difference research design shows the opposite. A prospect theoretic model with the power value function of Kahneman and Tversky (1979) does not generate bunching at the withholding threshold. While linear prospect theory (LPT) can generate bunching under certain conditions, it also yields testable predictions that are not consistent with the behavior of a significant proportion of firms. Under the LPT, given an enforcement and withholding regime, if a firm bunches in one year it should also bunch in all the following years, or if it unbunches in a following year, it should declare taxes less than the withheld amount. The evidence from panel data on the universe of all corporations in Ecuador shows very low persistence in bunching: conditional on bunching at least once, only 3-4 percent firms bunch every year before changes in the withholding rate, and among the firms that unbunch 35-40 percent declare taxes more than the withheld amount, thus contradicting the LPT model for a substantial proportion of the firms. Using the Sasabuchi t test as developed by Lind and Mehlum (2010), we find that the relation between probability of bunching and assets of a firm is inverted-U which is consistent with the audit trigger model. The evidence suggests that the behavior of the firms cannot be captured by a single model. The strength of enforcement is important in determining bunching in an LPT model which suggests cross-country differences in the role played by loss ...