Experiment on Tax Compliance in Eswatini

2019-10-07

2019-12-16

For active taxpayers:
• Probability to declare more than previous year
• Amount of tax declared on top of what declared previous year
• Relative percentage increase in amount declared
• Probability to amend previous year’s declaration + amount + % share
• Probability to register additional tax types

For nil-filers:
• Probability to declare a positive amount and switch to active filing
• Amount of tax declared
• Probability to be below the threshold of taxable income and still pay zero taxes (tax avoidance)
• Probability to amend previous year’s declaration + amount + % share
• Probability to de-register from SRA

For non-filers:
• Probability to file a tax return
• Amount of tax declared
• Probability to be below the threshold of taxable income and still pay zero taxes (tax avoidance)
• Probability to file declarations for previous year(s) + amount
• Probability to de-register from SRA

• Declare : 0-1 dummy variable for filing a tax return
• Nilfile : 0-1 dummy variable for filing a nil tax return
• Log(Tax+1) : log transformation of tax declared – plus 1 to include zeros
• HIS(Tax) : inverse hyperbolic sine-transformation of tax declared
• Deregister : 0-1 dummy variable for deregistering from the authority
• Threshold : 0-1 dummy variable for filing non-nil and being below the threshold for which Tax==0
• Costs_Turnover : continuous 0-1 variable for the share of reported costs over turnover

Additional outcomes from administrative data (N=39,000):
• Spillovers on VAT declaration
• Increase in costs/deductions/allowances reported as opposed to more income reported (tax avoidance)
• Taxpayers’ reaction: contacting the SRA call center/taxpayer hot line
• Increased e-tax registration

Additional outcomes from survey data (N=1,000):
• Change in tax morale
• Change in perception of enforcement
• Change in tax knowledge (tax law, fines, penalties)
• Change in trust towards actors

• Etax_registration : 0-1 dummy variable for registering for e-tax (non-filers only)
• Declare : 0-1 dummy variable for filing a VAT return
• Nilfile : 0-1 dummy variable for filing a nil VAT return
• Log(Tax+1) : log transformation of VAT declared – plus 1 to include zeros
• HIS(Tax) : inverse hyperbolic sine-transformation of VAT declared

A randomized controlled trial (RCT) will be set up in order to gauge causal impacts. The inherent advantages of an RCT is to make sure that the experimental groups are draw at random, and are well balanced on observable characteristics, such that any difference arising between them can be attributed to the intervention.

Assignment to a given nudging group will be random, as well as the assignment to the survey sample.

The intervention consists in the reception of a “nudge” from the tax authority, in the form of a letter, enclosed in an SRA-labelled envelope, sent using registered mail. The registered mail, despite being more expensive, will make sure to identify whether the recipient actually collected the letter. The letter will be in the local post office for 30 days and, if not collected, returned back to SRA. The post office will report on the delivery status of each letter, indicating, if the taxpayers collected the letter, also the date of collection. On top of that, I will also know the take-up rate of the different nudges.

The nudges will communicate different messages, targeted to the different types of (non-)compliant taxpayers, framed according to the most up-to-date evidence from behavioralism. The actual number of different nudges tested per group depends on the size of that group, meaning that for the smaller group (nil-filers) I will be powered enough to test two alternative messages, plus control. The content of the nudges is summarised as follows:

I. Active taxpayers
a. Deterrence T1: information on what the Income Tax Law states on income undeclared
b. Deterrence T2: targeting a subgroup for which VAT data is available and discrepancy occurs (Table 1), which will receive a nudge equal to T1 + some additional text through which SRA informs that it is able to cross-checks the recipient’s returns and identify a discrepancy + a table summarizing the discrepancy amount
c. Fiscal exchange T3: it does not mention any deterrence appeal, but stresses the societal consequences of not paying one’s true tax liability, with a catchy messages, such as ‘we all suffer from noncompliance’ + inform how tax money were allocated in order finance a list public commodities, and how their taxes were essential in process of building new public services (infrastructures, roads, streetlight, etc.) + states a general invite to file (“Are you going to support the building of a better Eswatini?”)
d. Control C: no nudge

II. Nil-filing taxpayers
a. Deterrence T1: as above
b. Deregistration T2: provides a detailed procedure to deregister – a “manual” on how to do it and lists all documents required to finalize the deregistration (for example, the importance of deregistering from Government first) – a blank tax deregistration form as an attachment
c. Control C: no nudge

III. Non-filing taxpayers
a. Deterrence T1: as above
b. Fiscal exchange T2: as above, PIT only
c. Social norms and peer comparison T3: provides the share of filers in the administrative area of the non-filer, PIT only
d. Compliance costs/taxpayer assistance T4: shows a step-by-step procedure on how to file + attach a blank tax form + provides a link to file online
e. Deregistration T5: as above
f. Control C: no nudge

The assignment of the treatment status (see 3.3) will be done using STATA. The following guidelines will be followed:
• Randomization will take place separately for CIT and PIT payers and for each of the 3 taxpayer types
• A seed will be set in STATA so that the code is replicable
• Randomization will be done using the following strata:
o Region: 4 strata
o Old versus newly registered taxpayers (dummy): 2 strata
o Size (dummy): top 10% dummy based on turnover declared in the previous year if the taxpayer already declared, or on the estimated turnover reported at registration for newly registered taxpayers: 2 strata
o Individual taxpayer type (dummy, for PIT only): sole trader vs director of company
• Ideally, I plan to check for balance after randomization and re-run the codes multiple times until the highest balance is achieved

The survey sample in Table 2 will be derived using the same randomization protocol.

Individual taxpayers

No

0.

39,000 individual taxpayers for the nudge experiment, plus 1,000 individual taxpayers to be assigned to the survey sample.

Drawing from the study sample figures in 3.1.2, the sample will be randomly allocated in the following way (also see Appendix Table A2 for a better display of the treatment groups):

• Active:
o CIT: T1 1026 - T2 422 - T3 1027 - C of T1 and T3 1,041 + C of T2 477
o PIT: T1 2,552 - T3 2,553 - C 2,571 (T2 not available here)
• Nil-filers:
o CIT: T1 449 - T2 451 - C 454
o PIT: T1 716 - T2 715 - C 734
• Non-filers:
o CIT: T1 1,055 – T4 1,056 – T5 1,055 – C 1,065
o PIT: T1 1,382 – T2 1,382 – T3 1,377 – T4 1,377 – T5 1,370 – C 14,255

Given the sample composition described in 3.3, the following minimum detectable effects are derived, using power calculations with α=5% and power equal to 80%. CIT Active: 9.8% increase in log(tax+1) Nil-filers: 24% reduction in nil-filing Non-filers: 10% reduction in non-filing PIT Active: 1.3% increase in log(tax+1) Nil-filers: 19% reduction in nil-filing Non-filers: 4% reduction in non-filing Pooling CIT and PIT for nil-filers only: Reduction of 15% in nil-filing The study might be underpowered for CIT and PIT nil-filers when taken separately. I will show results for each subsample and for a pooled on (CIT+PIT). Below, I compare the expected MDE with the findings from other similar studies: Active: Business in Ethiopia, Shimeles et al (2017): 38% increase from deterrence T and 32% increase from persuasion T Business in Costa Rica, Brockmeyer et al (2018): doubling the log-tax declared from deterrence T Business in Rwanda, Mascagni et al (2017): 45% increase in amount of tax declared from reminder T Business in Bangladesh, Chetty et al (2014): 17% increase in 17% in VAT payments from peer group comparison T Non-filers: Business in Costa Rica, Brockmeyer et al (2018): doubling the filing probability from deterrence T For what concerns survey outcomes such as tax morale, perceived deterrence level and perceived peers’ evasion, I do not believe to be powered enough to capture small effects. Between rounds attrition and experiment attrition (taxpayer failing to collect the letter) may reduce power. At the same time, it is enough to say that survey outcomes will not be the main outcomes to be measured, but will be rather looked at in order to test some additional exploratory hypothesis (set D in sec. 4.2).