Social Preferences and Tax Compliance

We designed and conducted an experiment to understand the role of social preferences and social norms in shaping tax compliance.

2019-04-15

2019-06-07

We plan to use the tax records to construct proxies for tax evasion.

We will not know which of the proxies will have meaningful variation until after we gain access to the administrative records. For example, one proxy for evasion that we are planning to use is based on employed individuals, whose wages are reported to the tax agency by the employee as well as by the employer. One proxy of tax evasion is then given by the gap between the amount reported by the employee and the amount reported by the employer. We are planning to construct other proxies too. For example, we can compare the number of dependents claimed by the subject in their tax forms to the actual number of dependents that they have as reported by third-parties. And, if we can find other sources of third-party reporting, we can construct other proxies in a similar spirit. If multiple proxies are viable, we will analyze them separately. Last, we will try to exploit the fact that self-employed individuals have more wiggle room than employed individuals when it comes to under-reporting. For example, if a particular piece of information increases the income reported by the self-employed but does not change the income reported by the salaried workers, that would suggest that the piece of information reduced tax evasion.

There are two types of analysis: non-experimental and experimental. They are both intended to shed light on similar questions, but using complementary approaches.

The non-experimental analysis consists of measuring the relationship between tax compliance and social preferences, as measured by online experiments with real stakes. Subjects have to play a series of 13 games. Most of the games are meant to measure social preferences (e.g., altruism, trust, lying aversion).

The goal of the analysis is to measure if these different outcomes are significantly correlated (and with the expected sign) to the actual tax evasion of the individuals. For example, are individuals who are more lying-averse in laboratory-type experiments more likely to evade taxes in the real world? Are individuals who are more altruistic in laboratory-type experiments less likely to evade taxes in the real world? Are individuals who say that cheating on taxes in unacceptable less likely to evade taxes in the real world?

We are planning to use standard methods for joint hypotheses testing. Additionally, we will also consider more modern machine learning methods (e.g., the PCA-LASSO approach).

As a benchmark for social preferences, we included other games that measure preference parameters that are unrelated to social preferences (e.g., risk-aversion, time discounting). As an additional benchmark, the survey also collects data on some beliefs and values that could be related to tax compliance, such as attitudes towards tax cheating and preferences for redistribution.

The experimental analysis consists of an information-provision experiment embedded in the survey and which was designed to assess the role of social norms. The survey elicits two beliefs: (i) The share of salaried employees that under-report their salary; (ii) The share of value-added taxes that goes under-reported by firms on average. After providing their prior beliefs, subjects are randomized into one of the following four groups:

(a) 25% of the sample receives no information.
(b) 25% of the sample receives information about the under-reporting by employees.
(c) 25% of the sample receives information about the under-reporting by firms.
(d) 25% of the sample receives the two pieces of information (i.e., about firms and employees).

After this stage, we re-elicit the same two beliefs. The information-provision experiment will generate exogenous variation in these posterior beliefs. For example, consider two individuals who believe that 40% of employees under-report their salaries. Due to the information experiment, one of them is provided a message that 15% of employees under-report their salaries, and the other subject is not given any information. As a result, at the end of the survey, the individual who was given the information will end up with lower beliefs about the share of employees under-reporting their salaries. The goal is to measure whether that shock to beliefs affects:

(1) Survey outcomes elicited at the end of the survey. Most importantly, we include a question related to attitudes about tax cheating. The hypothesis is that, due to social norms, a lower belief about the tax evasion of others will make it less acceptable to evade taxes.

(2) Actual tax evasion in the subsequent months, according to the administrative data. The hypothesis is that a lower belief about the tax evasion of others will make it harder for the individual to evade taxes in the future.

Note that our plan goes beyond measuring the average treatment effect of providing information. Instead, we want to measure the causal effect of beliefs on behavior. Note also that, if we are able to collect enough survey responses, we may have enough power to study the effects of the two beliefs (about individuals and about firms) separately. However, we may need to combine these two beliefs into a single belief.

Note also that we will have the administrative data for the year before our intervention as well as the year after our intervention. As a result, we can use pre-treatment outcomes for falsification tests (i.e., did the information affect behavior before the information was provided?). We'll be able to use the pre-treatment outcomes as control variables, which will help us to increase power.

The pre-treatment outcomes will also be important for heterogeneity analysis. This is important because it is likely that information about tax evasion will have a different effect on individuals who already evaded vs. individuals who never evaded.

Real-time randomization programmed into the Qualtrics platform.

Individual.

No

The final number of respondents will depend on the response rate to the survey. We are sending survey invitations to about 80,000 individuals. Based on a small pilot, our best guess is that we will collect around 5,000 responses.

We expect a minimum of 2,000 individuals

The four groups in the information-provision experiment have equal likelihood of being chosen, resulting in four treatment groups with equal number of observations.