Audit Selection Under Weak Fiscal Capacity
Last registered on July 23, 2018


Trial Information
General Information
Audit Selection Under Weak Fiscal Capacity
Initial registration date
July 11, 2018
Last updated
July 23, 2018 1:20 AM EDT

This section is unavailable to the public. Use the button below to request access to this information.

Request Information
Primary Investigator
World Bank
Other Primary Investigator(s)
PI Affiliation
Paris School of Economics
PI Affiliation
World Bank
Additional Trial Information
In development
Start date
End date
Secondary IDs
Should low-income countries leave discretion to tax inspectors to select firms for audit or should selection be determined by risk-scoring algorithms? Does the optimal amount of discretion depend on the quantity of third-party information available and its ease of access for tax inspectors? In a context with weak fiscal capacity, leveraging inspectors’ private information could be valuable but opens the door to discrimination. This project tests which selection method (discretionary inspector selection or algorithm selection based on risk scores) is most effective in detecting firm non-compliance and increasing audit yield, in partnership with the tax administration of Senegal.
External Link(s)
Registration Citation
Bachas, Pierre, Anne Brockmeyer and Bassirou Sarr. 2018. "Audit Selection Under Weak Fiscal Capacity." AEA RCT Registry. July 23.
Experimental Details
Our core treatment is the use of a data-driven algorithm to select firms for tax audits. The algorithm builds a risk score for each firm, drawing on data from corporate income tax, value-added tax and personal income tax declarations (employer withholding records), and external data from customs (imports/exports) and public procurement contracts, for the period 2012-2016. The risk score relies on two types of risk indicators: discrepancies and anomalies. Discrepancy indicators flag taxpayers whose self-reported information according to their tax declarations deviates from third-party reported information. For instance, a discrepancy indicator is created for a firm which reports sales lower than third-party reported sales. Anomaly indicators flag firms with unusual behavior relative to their peers. An example would be a firm with a relatively low profit rate, which might be associated with evasion. Anomaly indicators are constructed within firm size and sector groups, so that taxpayers are compared to similar peers. For full audits, the algorithm disregards firms which were audited in the previous two years, as per the instructions from the tax authority.

We compare the performance of audits selected by the algorithm to that of audits proposed by tax inspectors and randomly selected audits. First, inspectors provide their proposals without knowledge of the algorithm. Second, we select an identical number of firms through the algorithm. We rank firms by their risk score and select the highest ranked firms for audit. The process is described in more detail further below.

Our main objective thus is to compare the return of audits selected through different methods.
Intervention Start Date
Intervention End Date
Primary Outcomes
Primary Outcomes (end points)
There are two sets of primary outcomes: the first one refers to outcomes of the audits, and the second set refers to the inspector practices.

Audit outcomes:
-financial yield of the audit (penalties, additional tax payment, inspector bonus)
- contested amounts and payment made
- time taken to confirm the evaded amount and penalty, time to payment after notification,
- type of infraction discovered.

Inspector practices:
-time to conclude audit,
-taxes and years investigated,
-self-reported difficulty of the audit
-self-reported challenges encountered during the audit
Primary Outcomes (explanation)
The monetary variables are measured in absolute values or as a share of the firm's turnover. For outcome Y, we will consider the following measures: Y, 1(Y>0), 1(Y>p50), 1(Y>p90), ln(Y), arcsinh(Y).
Secondary Outcomes
Secondary Outcomes (end points)
Secondary outcome are measured through survey data (and the administrative data used to measure primary outcomes), and include:
-firm satisfaction: taxpayer satisfaction with the audit process; audit costs to taxpayer; taxpayer perception of the tax authority; taxpayer reports on bribes, gifts, attempts of collusion.
-inspector satisfaction: tax inspector satisfaction with the selection process
-medium-term outcomes: firm behavior in future tax periods, medium-term cost of audits for firms
Secondary Outcomes (explanation)
Same as for primary outcomes.
Experimental Design
Experimental Design
There are two main groups of audits in the experiment: the "full audits" (team work) and the "short audits" (individual tasks). All firms subject to a full audit are excluded from the pool for selection of short audits.

Full audits:
The full audits were selected according to one of two methods: the data-driven algorithm or the tax authority. The tax authority selected a specified number of firms for audit, and we ran the algorithm to select the same number of firms within each tax office. Given capacity constraints, there are no randomly selected full audits. For each firm selected via the algorithm, we provided the respective inspector with information on the three main reasons why the firm was selected (the "flags"). This is similar to the "risk criteria" that the inspectors need to provide when nominating a firm for audit according to the discretionary procedure. The sequencing of full audits was determined by the tax authority based on their considerations of the complexity of the case and capacity constraints. However, we provided managers with a suggested sequencing, in which we randomly alternated inspector-selected and algorithm-selected audits. Moreover, we encouraged tax office managers to strive to alternate between inspector-selected and algorithm-selected audits in their annual planning.

Short audits:
Short audits are more numerous than full audits and thus allow for more experimental sub-treatments to test for how selection methods and information availability influence audit outcomes. For this type of audit, we have three kinds of audits: audits selected by tax inspectors, audits selected by the algorithm and also randomly-selected audits (which can all overlap with each other). The algorithm-selected and randomly selected audits were randomly assigned to individual tax inspectors within tax offices. Each tax inspector is thus provided with a list of firms to audit.

Cross-randomized with the selection method, we provided an information treatment for the audit cases. The three treatments are 1) no additional information on the case, 2) the three most important risk indicators detected by the algorithm, or 3) the three most important risk indicators and the taxpayer's data (tax declarations and third-party reports for the last four years) organized in an excel document. Note that we run the algorithm for all firms and thus obtain the risk score and risk indicators also for firms selected by the tax inspectors, so that even if the firm was not selected by the algorithm, there is a risk score assigned to it. Moreover, in the case of "short audits", the sequencing of the audits is randomized for each inspector, so that the inspector cannot choose which firm to audit first, among those that were assigned to him/her.

Experimental Design Details
Not available
Randomization Method
Randomization in an office by a computer.
Randomization Unit
Was the treatment clustered?
Experiment Characteristics
Sample size: planned number of clusters
Sample size: planned number of observations
There are two main groups of audits that are run in this experiment: the "full audits" and the "short audits". There are 317 "full audits" and 798 "short audits".
Sample size (or number of clusters) by treatment arms

There are 317 "full audits", and they were selected in two different ways:
-algorithm: 167 audits
-tax authority: 166 audits
-overlap of the two: 16 audits

There are 798 "short audits", and they were selected as follows:
-algorithm: 285 audits
-tax authority: 278 audits
-random: 202 audits
-overlap between tax authority and algorithm: 33 audits
Minimum detectable effect size for main outcomes (accounting for sample design and clustering)
IRB Name
Paris School of Economics
IRB Approval Date
IRB Approval Number