The Origins of Tax Compliance and State Capacity: Tax Intervention

We randomly assign the city’s neighborhoods to central tax collection, conducted by agents of the provincial tax ministry, or local tax collection, conducted by local city chiefs. We also implement two hybrid collection interventions and cross-randomized information treatments to elicit the mechanisms through which central and local tax collection shape citizen compliance.

T1. Central tax collection (C): Agents of the provincial tax ministry (DGRKOC) complete all steps of the property tax campaign (census and tax collection). Collectors work in teams of two and each team is assigned to two neighborhoods per month. Every month collectors are re-randomized in teams of two.

T2. Local tax collection (L): Local chiefs complete the steps of the campaign. These chiefs act as intermediaries between citizens and the government and can be thought of as the bottom link in the chain of the city-level government bureaucracy. They are typically in charge of: (1) organizing and enforcing weekly public good provision (Salongo), (2) communicating citizens’ grievances to government authorities, and (3) mediating in local disputes. This position is appointed for life to an individual who already lives in the neighborhood. Often, it is given to an individual who is well-known in the neighborhood. To make treatments comparable, each chief is asked to pick an assistant, so each neighborhood assigned to local taxation is visited by a team of two.

T3. Central X Local (CXL): Central and local collectors complete all steps of the campaign together. Central collectors are re-randomized to selected chiefs each month.

T4. Central Plus Local Information (CLI): This arm is identical to the central tax collection treatment, except that, after completing the census, the central collectors meet with the avenue chief in the neighborhood to transfer knowledge about the capacity and willingness to pay of all individuals in that neighborhood.

T5. Control: In a small number of neighborhoods, individuals are supposed to pay themselves at the tax ministry (the old system up to 2016). Two agents from the tax ministry visit each household in these neighborhoods, conducting a census that is identical to that administered in treatment neighborhoods except that individuals are informed they should pay at the tax ministry rather than pay collectors themselves.

In addition, we randomize the following information interventions at the individual level. We implement two deterrence messages, two fiscal exchange (public goods) messages, one trust message, and one control message:

I1. Central deterrence. The central version says that refusal to pay the property tax entails the possibility of audit and investigation by the provincial tax ministry.

I2. Local deterrence. The local version of the deterrence message says that refusal to pay the property tax entails the possibility of audit and investigation by the neighborhood chief (chef de quartier). Note that this is a higher-rank chief relative to those who are collecting taxes in Local neighborhoods.

I3. Central public goods. The central version of the flyer says that the provincial government will be able to improve infrastructure in the city of Kananga only if citizens pay the property tax.

I4. Local public goods. The local version of the flyer is exactly the same, except that it mentions each citizens’ locality instead of Kananga.

I5. Trust. The trust message reminds citizens that paying the property tax is a way of showing that they trust the state and its agents.

I6. Control. The final flyer is the control. It simply says: “It is important to pay the property tax.”

2018-06-15

2019-03-31

The primary outcome is tax compliance (i.e., whether citizens pay the property tax).

Secondary outcomes include 1) bribe payment to tax collectors, 2) targeting by collectors, 3) citizens' views of the local government, 4) citizens' views on chiefs, 4) civic engagement, and 7) formalization.

The unit of randomization is the neighborhood, or polygon, each of which was identified on a satellite map with boundaries like roads, ravines, or other natural features that would be easily identifiable from the ground. There are 364 neighborhoods in total in Kananga. We excluded the 8 neighborhoods that were part of the logistics pilot mentioned above. This leaves 356 neighborhoods for the full randomization.

We employ a block-randomized design, stratifying on three variables:

1. Geographic stratum: We group the city neighborhoods into twelve strata that take into account (a) geographic regions of the city and (b) whether neighborhoods are deemed the “city center” or “the periphery” by the tax ministry. Two of these strata correspond to downtown. This ensures balance on the extent to which a neighborhood is central/peripheral, which is a good proxy of the importance of the corresponding locality chief and the degree of enforcement of tax collection.

2. Treatment status in the 2016 tax campaign: This is a dummy capturing whether the neighborhood had been assigned to treatment in the 2016 tax campaign, studied in Weigel (2018a).

3. Past experience of local chiefs collecting taxes: We create temporary strata based on the two variables above and, for each of these, find the median proportion of chiefs (ranked 1-5 according to the chief selection procedure) per neighborhood who ever collected taxes and split each temporary stratum into two additional substrata around the median.

The information interventions are randomized at the individual level.

Randomization was done by a computer.

For the taxation interventions, the unit of randomization is the polygon (neighborhood). The information interventions are randomized at the individual level.

Yes

356 polygons (neighborhoods).

For our primary outcome tax payment we will use administrative data to evaluate the effect of the various treatment arms on compliance. For this analysis, we have the universe of compounds in Kananga, approximately 47,122 according to data from Weigel (2018). For other outcomes, we will rely on an endline survey, with a sample size at least as large as our baseline sample (N=4,343)

Central tax collection: 110 neighborhoods
Local tax collection: 110 neighborhoods
Central x Local collection: 51 neighborhoods
Central Plus Local Information: 80 neighborhoods
Control: 5 neighborhoods

Based on the results of the pilot, the power calculations assume a minimum detectable effect (MDE) of 0.1 percentage points for the Central, 0.12 pp for Local, 0.12 pp for CLI, and 0.14 pp for CXL. We also assume the assignment of neighborhoods to different treatment arms shown in Table 4 and an average of 200 households per neighborhood. Power is above 0.9 for most hypothesis tests. However, for some of the comparisons –in particular, for detecting the difference between (1) C vs. CLI, (2) L vs. CLI, (3) C vs. CXL, (4) L vs. CXL, and (5) CXL vs. CLI– power is lower than 0.8. Power is low for the comparison between the local treatment and central with information, but this comparison is not of theoretical interest in this study. A test of joint orthogonality of coefficients rejects the null with probability 0.97. See Pre-Analysis Plan for details.

The Origins of Tax Compliance and State Capacity: Tax Intervention