Attacking Tax Evasion in Production Networks: Theory and Evidence from Paraguay

The intervention in the first experiment consists of sending notices to taxpayers suspected of underreporting
their taxable sales, based on the work of Carrillo, Pomeranz, and Singhal (2017). For
each taxpayer-month we calculated the sum of sales made by that taxpayer as reported by their clients in third-party information reports (“libros de compras”) annexed to the clients’ VAT returns. If this amount was larger than the total sales reported by the taxpayer by at least
10,000,000 Guaranies (around USD 7,000), that taxpayer-month was eligible to receive a notice.

2022-08-15

2024-01-31

Do firms that receive the intervention declare more sales? Less inputs? Pay more taxes?
Is the discrepancy between the seller's VAT reports and the sum of the buyers' purchases report reduced by the intervention? Is the discrepancy between the seller's bilateral sales reports and the buyer's bilateral purchases report reduced by the intervention?
Do the clients of firms that received the intervention change their behaviour?
Do the suppliers of firms that received the intervention change their behaviour?

Do firms update their tax declaration? When do they make these amendments? Do firms pay more taxes on the declarations for which there were discrepancies?

We randomly assign taxpayers to receive notices. The randomization method is optimized to allow for the detection of spillovers.

The experiment is carried out in 4 waves.

In waves 1 and 2, We pursue a saturation design, common in experiments seeking to measure spillovers (Baird,
Bohren, McIntosh, & ¨ Ozler, 2018). To do this, we first create a network of indirect linkages
formed by pairs of clients sharing a common supplier. The idea is that treating that common
supplier may have indirect effects on any of their clients, both those mentioned in the notification
as well as those not mentioned.
1. Create clusters of suppliers based on who their clients are. To do this, we first create a
network of clients where links are formed between pairs of clients who share a supplier. We
then partition this network of 185,119 clients into 6,486 clusters using the Leiden algorithm
(Traag, Dooren, & Nesterov, 2011). We then assign each of the suppliers to one cluster
based on which cluster of clients they supply to the most.
2. We randomly assign clusters to 0%, 50% or 100% saturation rates of eligible taxpayers.
3. We choose our sample of 500 taxpayers for wave 1 following the clusters’ assigned saturation
rates.
4. We chose our sample of 1,000 taxpayers for wave 2 from the remaining eligible taxpayers
following the clusters’ assigned saturation rates.

In waves 3 & 4 the treatment assignment protocol was refined to maximize statistical power. Our choice of experimental design and parameters has been guided by the work of a team of econometricians. They compared the statistical power of several different possible experimental designs given the structure of our network, to test for direct and spillover effects. They also searched for the optimal parameters under these candidate designs.

The resulting optimal design is a split-graph design. In that design, firms are treated at random, stratifying the randomization by type $x$ (determined by their degree and by their connections to other eligible firms, as discussed in subsection 3.2. of our pre-analysis plan (published here) and by anticlique (as defined in appendix A of the pre-analysis plan). The likelihood of being treated differs for each each stratum.

computer

taxpayer

No

5500 taxpayers

5,500 treated taxpayers, circa 100,000 untreated taxpayers. All observed in the administrative data from January 2021-present.

5,500 treated, remaining 29,000 eligible firms as controls