Improving the Effectiveness of the IRS’ Automated Substitute for Return Program

The ASFR program is a non-filer/return delinquency compliance program that enforces filing compliance on taxpayers who owe significant amounts of income tax but have not filed their individual income tax return(s). Figure 1 of the attached IRB protocol presents the traditional ASFR timeline in black and the experimental components in red and green, described below. ASFR first attempts to get the taxpayers to file their tax returns by sending notices to delinquent taxpayers informing them that they are entering the ASFR program and asking them to file their own return by a specific date. The first notice is called the “30-day Letter.” If the taxpayer does not file their return after receiving the 30-day Letter, then ASFR sends a second notice, called the “90-day Letter.” If the taxpayer does not file their return within an IRS-designated time frame after receiving the 90-day letter, then ASFR prepares a “Substitute for Return” and issues Statutory Notices of Deficiency. In this instance ASFR determines and assesses taxpayers’ liability by computing tax, penalties, and interest based upon internal information available to the IRS.

We will select a random sample of taxpayers in the ASFR program to receive a repeat of the 30-day Letter (Control Group 1), receive one of the four new notices (treatment groups), or receive no additional letter (Control Group 2). If a taxpayer is randomly selected into a treatment group or Control Group 1 then they will receive the additional notice in between the 30- and 90-day Letters.

2021-06-01

2022-05-31

1. Tax return filing rate
2. Timeliness of filing
3. Amount of tax collected

The timeliness of filing is the number of days or weeks between treatment and tax return receipt by the IRS. The amount of tax collected may be measured in dollars or as a percentage of the assessed tax liability.

We will draw a random sample of approximately 48,000 taxpayers from the population of taxpayers who are in the ASFR program. In total, the IRS will distribute approximately 40,000 letters in the course of this study. The target sample size is 8,000 per letter group (the four treatment groups plus Control Group 1). In addition, the study comprises a control group of approximately 8,000 randomly selected taxpayers in the ASFR who do not receive an additional letter, but still receive the traditional 30- and 90-day Letters (Control Group 2). If a taxpayer is randomly selected into a treatment group or Control Group 1 then they will receive the additional notice in between the 30- and 90-day Letters.

We will then use prior years’ tax returns from taxpayers in our random sample to map them to their immediate family members to estimate spillover effects of the letters among family members. We can link our sample of taxpayers to their adult children (based on the taxpayer ID numbers of dependents listed in prior years), to their parents (based on the taxpayer ID numbers of taxpayers who claimed one of our sample taxpayers as a dependent on a previous tax return), and to their siblings (based on the taxpayer ID numbers of other dependents claimed by their parents on a previous return). We can then estimate spillover effects by comparing the rate of filing, the timeliness of filing, and the amount of tax collected among family members of taxpayers across the treatment groups relative to family members of taxpayers in each of the two control groups.

Randomization done in office by a computer at the IRS.

Individual

No

48,000

48,000

Approximately 48,000 taxpayers in the full sample. Target sample size is 8,000 per group, with 6 groups (the four treatment groups plus two control groups).

We consider a binary indicator of whether an individual filed their tax return as the outcome variable. We assume a baseline response rate of 0.08 to a single letter in the ASFR program, which is conservative given the total tax return filing rate within the ASFR program (which involves two letters). With a sample size of approximately 8,000 per group, we would be able to detect a treatment effect as small as 1.46 percentage points (18% of the baseline mean) or, equivalently, as small as 5.4% of a standard deviation of the outcome variable.