As regular blog readers will know, I’m a big fan of randomisation. In the context of tax audits, this is particularly useful. Though politically controversial, random audit experiments like the US TCMP have taught us a lot about who underreports tax. And now a new two-stage experiment in Demark is revealing other lessons. Perhaps Australia – which has never had a random audit of personal income taxpayers – could follow suit.
Unwilling or Unable to Cheat? Evidence from a Randomized Tax Audit Experiment in Denmark (gated stable link, ungated unstable link)
Henrik Kleven, Martin Knudsen, Claus Kreiner, Soren Pedersen and Emmanuel Saez
This paper analyzes a randomized tax enforcement experiment in Denmark. In the base year, a stratified and representative sample of over 40,000 individual income tax filers was selected for the experiment. Half of the tax filers were randomly selected to be thoroughly audited, while the rest were deliberately not audited. The following year, “threat-of-audit” letters were randomly assigned and sent to tax filers in both groups. Using comprehensive administrative tax data, we present four main findings. First, we find that the tax evasion rate is very small (0.3%) for income subject to third-party reporting, but substantial (37%) for self-reported income. Since 95% of all income is third-party reported, the overall evasion rate is very modest. Second, using bunching evidence around large and salient kink points of the nonlinear income tax schedule, we find that marginal tax rates have a positive impact on tax evasion, but that this effect is small in comparison to avoidance responses. Third, we find that prior audits substantially increase self-reported income, implying that individuals update their beliefs about detection probability based on experiencing an audit. Fourth, threat-of-audit letters also have a significant effect on self-reported income, and the size of this effect depends positively on the audit probability expressed in the letter. All these empirical results can be explained by extending the standard model of (rational) tax evasion to allow for the key distinction between self-reported and third-party reported incomes.
Update: On a similar theme, here’s Tim Harford on why policymakers should do more randomised trials.
Andrew Leigh's Archived Blog 2004-2010 
Interesting that it was carried as part of the implementation of policy rather than as an experiment. Perhaps you can use that idea as a way to fund randomisation trials for experimental purposes.
I am confused. Whilst it is lovely of our Danish compatriots to have produced this public good for us, am I alone in understanding that the result affirms the ATO’s approach?
To wit, the ATO targets audit activity in three primary ways:
market-sector review: The ATO periodically chooses a market group with particular characteristics for general review, and subjects to audit those whose review indicates that they have greater risk or that there are obvious understatements of tax.
risk-based review: The ATO periodically processes returns and other data in order to come up with ‘high-risk taxpayers’ (ie excessive gearing, certain disclosures don’t match industry patterns, high book-to-tax volatility/differences, high related-party transactions, etc). In addition, in this category, the ATO sometimes determines generic risk factors (ie ‘Losses more than 3x taxable income) and subjects all taxpayers of that class to a review.
specific-issue audit: The ATO also sometimes becomes aware of a specific issue, either from a review of another taxpayer, from AUSTRAC or from a foreign tax authority, and subjects this to a specific issue audit focused on that issue.
How would you recommend using randomisation, greater than is already the case, to supplement this?
Also, the whole point of both the tax industry, the ATO and the business community is to ‘de-randomise’ tax audits given the considerable cost these impose. A certain level of avoidance is cheaper than a certain level of audit activity, so there is a need to calibrate the level of audit at a deterrent-but-not-punitive level.
Tim Harford’s piece is very curious – he advocates for experiments, but only of a particular kind (Randomised Controlled Trials), using a non-experimental example to demonstrate good practice.
It was not through RCTs that we learned that changing advice about infant sleeping position is effective in reducing the incidence of Sudden Infant Death Syndrome. It was a systematic process of bringing together evidence from many studies, including retrospective and prospective epidemiological studies, pathological studies and case studies, identifying a number of possible contributing factors, and ruling out other possible causes (such as vaccinations). The SIDS example shows both the value of drawing on a diverse set of evidence and how it is possible to develop effective policy even when the evidence is not definitive.
Andrew, are you suggesting we need a large RCT of infant sleeping positions to test the effectiveness of the current policy? Or in this case would other evidence be sufficient? And if that is the case, what’s the logical link between the SIDS example and the conclusion that we need more randomised trials (rather than more evidence using the designs that are methodologically and practically appropriate)?