Randomization Method
To randomly allocate SMEs and self employed taxpayers into the different treatments, the tax authorities provided us with data from 172.172 SMEs and self employed for the year 2016 (i.e., pre-experiment data). Specifically, we received annonymized information about taxpayers' monthly VAT payments, monthly social security payments for workers and firm size. Thus, while the tax authority interacts with the taxpayers, i.e., sends out mailing, performs audits and collects the data, we, the researchers, performed the randomization process taking into account the capabilities and interests of the tax authorities, i.e., number of audits that can be performed and number of mailings that can be sent out.
Precisely, the tax authorities allow us to send out 68.200 mailings and are willing to perform 2210 audits. Importantly, these audits will be performed anyway and therefore do not increase nor decrease the overall likelihood of a firm in the sample to be audited in Bulgaria compared to 2016.
Moreover, to cover the range of audit probabilities the tax authorities are interested in, we are able to assign 5000 SMEs and self employed to the ambiguity treatment, 5200 firms to the 1% audit probability treatment, 5000 firms to the 10% audit probability treatment, 2000 firms to the 40% audit probability treatment, 1180 firms to the 60% audit probability treatment and 10.000 firms to the control group, and 10.000 firms to each of the moral appeals treatments and 10.000 firms to the survey group. We are limited by strategical considerations and manpower restrictions of the tax authorities to only include 100 taxpayers in the 0% audit treatment and 100 firms in the 100% audit treatment (treatment 4 and 9).
Given these considerations and restrictions and in light of our goal to have comparable firms in each treatment group, we randomized as follows. First, we ranked taxpayers according to mean VAT payments in the year 2016 in ascending order. Second, we divided the taxpayers in deciles; that is, we created 10 groups where the first group consisted of the 10% taxpayers with the largest VAT payments and the 10th group consisted of the 10% taxpayers with the lowest VAT payments. Third, within each decile we randomly assigned a number to each firm and ranked firms within each decile by this random number. Fourth, based on this random ranking in each decile, we assigned the taxpayers to treatment groups by assigning n/10 firms to treatment group X, where X is one of the treatment groups and $n$ is the total number of firms we intended to assign to group X. Since we do this within each decile, we in total assigned 10 * n/10 = n firms to group X.
This procedure ensures that an equal number of taxpayers from each decile is assigned to the respective treatments or control group. Since we want the survey to be comparable to the taxpayers in the treatments and representative for the population of SMEs in Bulgaria, we also included the taxpayers who will be invited to fill in the survey into the randomization process.