A Field Experiment on Tax Avoidance

We study households who were about to have the opportunity to protest the tax assessments of their homes. We sent letters to 100,432 of those households to invite them to participate in a survey. Within the survey, we conduct an information-provision experiment.

We elicit beliefs on, and provide information about, three topics:
1) The share of households who filed a protest among the richest-1% of households.
2) The expected savings from filing a protest.
3) The share of households who filed a protest among households that are similar to the respondent's household (within $50K of the respondent's own home value).

2022-04-22

2022-05-16

The main outcome is a dummy variable indicating whether the household protested or not by the deadline (May 16th, 2022).

Due the nature of the setting and the timing of the protest process, we expect our information to affect primarily whether the households protested directly. For that reason, we will focus on direct protests as the primary outcome. We plan on using direct protests in previous years as falsification/placebo outcomes, in the spirit of event-study analysis.

Right after the information-provision, the survey includes a few questions that could be used as outcomes. A sample of the full survey instrument is attached to this application. The most important survey outcome is the intention to protest. This outcome can potentially pick up transitory effects of the information and also allows us to compare the stated intentions to protest in the survey versus the revealed actions in the administrative data. We have another way of measuring the intention to protest indirectly, through the usage of aid provided at the end of our survey (more details below).

There is another secondary outcome that is quite important. At the end of the survey, we use a willingness-to-pay (WTP) exercise to measure whether the household wants to help other households file a tax protest. We elicit this WTP twice (in random order): the WTP to help a household in the richest 1% of households, and the WTP to help a household in the poorest 1% of households. These two outcomes allows us to test whether the beliefs affect not only the household's intention to file a protest, but also the household's views as to weather rich and poor households should protest.

Our intervention provides all respondents with help for filing a protest, in the form of a website with step-by-step instructions. This data allows for a secondary outcome that indicates whether the household had the intention to protest: a dummy variable that takes the value 1 if the household accessed the website with the information on how to protest. We are able to collect this data because the access to the help website requires to introduce the validation code included in the letter. We are able to compare the protest probability for households who visit the webpage with instructions and those who do not.

We sent letters to a sample of 100,432 households from Dallas County (Texas) who were about to have the opportunity to protest the tax assessments of their homes. A sample of the letter is attached to this pre-registration. Our letters included an invitation to participate in an online survey. A sample of the survey instrument is attached to this pre-registration as well.

Since each subject must enter a survey code, we can link their survey responses back to administrative records (including, but not limited to, whether they filed a protest in 2022 or in any of the previous years).

We want to measure how the information provided inside the survey affected the subsequent decisions to file a protest. The main focus of our information-provision experiment is not to investigate if providing information (relative to not providing information) has an effect on the average probability of protesting. When provided with the information, some individuals may update their beliefs down and others may update the beliefs up, so those effects may cancel each other out. Instead, our main focus is to measure the causal effects on beliefs, by exploiting how individuals update relative to their prior beliefs. With that goal in mind, we will use econometric models and robustness checks used in some of our previous work such as:

Cullen, Z. and Perez-Truglia, R. (2018). How Much Does Your Boss Make? The Effects of Salary Comparisons. Journal of Political Economy, Vol. 130 (3), pp. 766-822
Nathan, B.; Perez-Truglia, R. and Zentner, A. (2020). My Taxes are Too Darn High: Why Do Households Protest their Taxes? NBER Working Paper No. 27816.
Giaccobasso, M.; Nathan, B.; Perez-Truglia, R. and Zentner, A. (2022). Where Do My Tax Dollars Go? Tax Morale Effects of Perceived Government Spending. NBER Working Paper No. 29789.
Bottan, N. and Perez-Truglia, R. (2020). Betting on the House: Subjective Expectations and Market Choices. NBER Working Paper No. 27412.

We elicit beliefs on, and provide information about, three topics:
1) The share of households who filed a protest among the richest-1% of households.
2) The expected savings from filing a protest.
3) The share of households who filed a protest among households that are similar to the respondent's household (within $50K of the respondent's own home value).

The main hypothesis is the "trickle-down tax avoidance": the higher the perception of the share of the richest-1% households who file tax appeals, the higher the likelihood that the respondent files a appeal. This is precisely why we measure belief 1) above.

The other two beliefs are intended to disentangle causal mechanisms.

One potential mechanism is that when finding out that the richest-1% of households are likely to protest, respondents infer from that information that they could be saving more money if they protested themselves. In other words, households may think that if the richest-1% protest, it must be because filing a protest is the financially smart thing to do. The belief in 2) above is intended to disentangle this mechanism.

Another potential mechanism is that households do not care about the share of households protesting among the richest-1% specifically, but more generally about the behavior of households that are similar to them. That is, households could be reacting to the information about the richest-1% simply because they are extrapolating from that information. The belief in 3) above is intended to disentangle this second mechanism.

We have a lot of additional survey data and administrative data for heterogeneity analysis that may help to disentangle causal mechanisms. For example, we can reproduce the analysis separately for Democrat vs. Republican households or by age and race.

Randomization done in office by a computer.

Household

No

100,432 households will be invited to the survey. These households are a subsample of the universe of all homeowners in Dallas County, Texas. We arrived to this subsample by applying a number of filters (e.g., focusing on single-family homes, excluding business properties). We originally selected 100,616, but 198 were dropped by the mailing company by using the NCOA (Change of Address) system. We would like to clarify that some additional households will need to be dropped from the subject pool due to reasons that we cannot observe ex-ante.

- Most importantly, some households may respond to the survey after they protested or after the window to submit a protest had been closed (in which cases, by construction, the information provided in the survey could not affect their choices).
- Some households may sign with a tax agent before they respond to our survey, in which case they will be excluded from the sample.
- Following prior information-provision studies, we'll exclude subjects who are outliers in terms of their prior misperceptions. If a respondent is extremely off in his or her prior belief, that is a sign that the household is either not paying attention or misunderstood the question.

Since we do not know the response rate in advance, the number of subjects participating in the survey can be anywhere from 0 to 100,432. Based on the response rate in a previous experiment (3.7%, from Nathan et al., 2020 and 3.6% from Giaccobasso et al., 2022), our best guess is that we'll have around 3,600 responses. References: Giaccobasso, M.; Nathan, B.; Perez-Truglia, R. and Zentner, A. (2022). Where Do My Tax Dollars Go? Tax Morale Effects of Perceived Government Spending. NBER Working Paper No. 29789. Nathan, B.; Perez-Truglia, R. and Zentner, A. (2020). My Taxes are Too Darn High: Why Do Households Protest their Taxes? NBER Working Paper No. 27816.

We cross-randomized subjects to receive three pieces of information, with 50% probability each. So 1/8 of the sample will receive all three pieces of information, 1/8 of the sample will receive the first piece of information only, 1/8 will receive the second piece of information only, ..., and 1/8 of the sample will receive no information.

Since we do not know the response rate yet, we cannot estimate the minimum detectable effect size in advance.