The Social Fabric of Corruption: A Field Experiment in Congo's Public Transport

Last registered on January 10, 2025

Pre-Trial

Trial Information

General Information

Title
The Social Fabric of Corruption: A Field Experiment in Congo's Public Transport
RCT ID
AEARCTR-0015135
Initial registration date
January 08, 2025

Initial registration date is when the trial was registered.

It corresponds to when the registration was submitted to the Registry to be reviewed for publication.

First published
January 10, 2025, 1:48 PM EST

First published corresponds to when the trial was first made public on the Registry after being reviewed.

Locations

Primary Investigator

Affiliation
UNIVERSITY OF CHICAGO

Other Primary Investigator(s)

PI Affiliation
University of Antwerp

Additional Trial Information

Status
Completed
Start date
2015-08-10
End date
2015-09-10
Secondary IDs
Prior work
This trial does not extend or rely on any prior RCTs.
Abstract
This studyexplores the role of the social environment in the transaction costs of corrupt exchange. Through a field experiment in the Congo's capital, we randomly assign public transport drivers to reroute their operations on certain days but not in others, targeting routes where they lack \textit{relationships} with the officers overseeing the line. On these rerouted days, the drivers are significantly less likely to have pre-existing connections with the officers on the line; they also generate lower revenue net of operating costs (joint profit). By analyzing two cross-randomizations, we find that this reduction in joint profit stems from the absence of the drivers' \textit{relationships} with the officers, and that those provide drivers with the security to take actions that increase passenger demand and fares without risking extortion. These relationships are formed through large initial investments when rerouting is expected to persist for multiple days. Our findings highlight how social relationships lower the costs of sustaining corrupt arrangements and suggest that expectations of corruption profit opportunities shape social relationships.
External Link(s)

Registration Citation

Citation
sanchez de la sierra, Raul and Kristof Titeca. 2025. "The Social Fabric of Corruption: A Field Experiment in Congo's Public Transport." AEA RCT Registry. January 10. https://doi.org/10.1257/rct.15135-1.0
Experimental Details

Interventions

Intervention(s)
We provided a randomized encouragement in some days and not others for ten drivers to drive on lines where they did not have relationships with the officers on the street (henceforth, \textit{foreign lines}).


We also borrow from a practice commonly used by private drivers, in which high-ranking officers intervene in negotiations with officers on the street. Assistance from ACCO allowed us to leverage connections with high-ranking officers and extend their protection to drivers on randomly selected days. To separately identify not having a relationship vs. investing in a new one, we also cross-randomized whether the driver was rerouted for three consecutive days, or one.

Intervention (Hidden)
Intervention Start Date
2015-08-10
Intervention End Date
2015-09-10

Primary Outcomes

Primary Outcomes (end points)
Revenue on a trip (number of passengers, fee per passenger)
Operating costs (gas expenses, repairs)
Bribe payments made (amount and type of bribe)
Number of trips made
Primary Outcomes (explanation)

Secondary Outcomes

Secondary Outcomes (end points)
Connection between driver and each officer interacted with (acquaintance, friendship, enemy)
Secondary Outcomes (explanation)

Experimental Design

Experimental Design
The Experiment

In this section, we present our experiment design. We conducted the experiment in August and September 2015. Independently for each driver, we randomly assigned minibus-days for the driver to drive in foreign lines those days.

Rerouting to Foreign Lines

Each driver had a driver-specific set of lines they considered home lines. The rest of the lines were considered, for each driver, their corresponding foreign lines.

To determine which lines were home and foreign for each of the drivers, we organized meetings with the drivers prior to the experiment. The information was elicited from drivers' own reports, which was straightforward given the saliency of the home vs. foreign line distinction. On average, drivers had 1.3 lines that they considered home, and 8.1 foreign lines.

For each driver, the rerouting treatment consisted of an encouragement to drive through lines considered foreign during a block of contiguous days. The control consisted in letting the driver choose his own route in the set of his home lines.

Due to initial uncertainty over the extent of funding, the study was initially planned to last working 23 days, covering the period from August 10th 2015 to September 4th 2015 (henceforth, first wave). The randomization procedure in this first wave was applied for each separate driver as follows.

Step 1: Partitioning into blocks. We partitioned the first 23 days into nine blocks of contiguous days (henceforth, "driving blocks"). In that window, each driver had seven three-day driving blocks and two one-day driving blocks.

Step 2: Assignment to home vs. foreign. First, 3 three-day-long driving blocks were randomly assigned to home line. The remaining four were randomly assigned to rerouting in a foreign line. All the days within each three-day block had the same home vs. foreign assignment. Second, 1 one-day-long driving block was randomly assigned to home, and the other foreign.

Step 3: Selection of the home and foreign lines. In blocks assigned to home, the driver was asked to drive on their main home line. In each block assigned to foreign, we randomly selected one unique foreign line for all days of the block from the pool of foreign lines for that driver without replacement.

Step 4: Block sequence randomization. In blocks assigned to home, the driver was asked to drive on their main home line. In each block assigned to foreign, we randomly selected one unique foreign line for all days of the block from the pool of foreign lines for that driver without replacement.

Upon securing additional funding, a second wave of study was added, covering four additional days, September 7-10th 2015:

Step 5: Second randomization wave. The remaining four days were partitioned into 4 one-day driving blocks; then, two of these were randomly assigned to home driving, and the was randomly assigned to foreign. Among days rerouted to foreign, foreign lines were randomly selected as in the first wave, among the pool of the driver's foreign lines that had not been previously selected, in any wave, for the driver.


This randomization procedure was independently applied to ten drivers each with one main home line, thus totaling for 23+4=27 days for ten drivers, 270 driver-day observations. Since the foreign/home assignment clusters are randomly assigned across 13 driving blocks of contiguous days each for each driver, the clusters, in two waves, the experiment is a stratified clustered randomization of 130 clusters, stratified within 20 randomization strata (the driver-waves). The sample consist in 13,092 events of crossing an intersection produced by 1,935 trips (comparable to, for example, to 304 trips in Olken Barron).

Drivers were compensated for participating in the experiment, and compensation was orthogonal to treatment status. This compensation was decided ex-ante per weeks-long conversations with the drivers. The compensation, 500 USD, was calculated based on the expected losses from driving on foreign lines on some days and not on other days. Only after the end of the study, drivers withdrew the funds. The compensation was presented as the team's acknowledgment that the requirement of rerouting would otherwise have been financially costly on the drivers. If a driver failed to comply with the randomization schedule, they risked exclusion from the experiment. A supervisor verified implementation each day. Data collection, following the same strategy as in the pre-experimental data, is detailed in Appendices A1-A3. We recorded no violation of the assignment.

This design would call for reasonable spillover concerns. First, given the horizon of the experiment, it is unlikely that officers meeting new minibuses will crowd out their cooperation with drivers with whom they had a relationship; vice versa, it is unlikely that drivers with new officer ties would change their behavior with the officers in their home lines within the span of the experiment. We note that it is theoretically possible that drivers who create ties in a more profitable line may have a shorter horizon of interaction in their initial home line; reassuringly, this would tend to reduce cooperation in their home line, thus tend to weaken the estimated value of relationships. Second, informational spillovers across drivers and officers are unlikely given that relationships are bilateral, and the presence of protection for a driver or whether the driver is home or foreign driving does not carry information that is germane to how to interact with other drivers. Finally, our interventions are unlikely to have any effect on the equilibrium, given the experiment drivers' number pales in comparison to the hundreds operating daily.

Cross-Randomization: Third-Party Driver Protection

Independently for each driver, we provided, on randomly selected days, third-party protection from the portfolio of existing third-party protection. Drawing on pre-existing relationships we had established in a prior study, we identified a set of high-ranking police officers and of street-level officers who regularly offered their protection.
Involvement of the non-Congolese members would have triggered expectations of high payment. We also worked with ACCO to produce a sticker for display on the minibus' windshield that could be removed at the end of the day.

The randomization procedure to third-party driver protection was as follows. First, we assigned the police protection. For each driver: (a) among the three-day driving blocks, we randomly assigned some three-day blocks to benefit from third-party police protection in two days of the block (high intensity) or instead one day of the block (low intensity). (b) among the one-day driving blocks, we randomly assigned some one-day blocks to receive the police third-party protection, and the rest to remain without third-party police protection. This procedure ensured strict balance of home/foreign assignment for third-party police protection. This procedure was independently and identically repeated for each driver.

Second, we determined what exact source of police protection was given each time a driver had been assigned for third-party police protection as follows. For each driver, among the days assigned to third-party police protection, we randomly assigned the source of police protection across such days in two steps. In a first step, we randomly assigned the police protection day to be sourced from either high-ranking (colonels or majors) or street-level police protection (the police officers who sat with the driver in the minibus, henceforth "police escort"). In a second step, we randomly selected some days assigned to high-ranking protection for it to be provided by a colonel, and the rest by a major. This procedure was independently and identically repeated for each driver.

Third, we determined the allocation of ACCO sticker protection as follows. Stratifying by minibus-driving-block-protection status, each of two days in the three-day driving blocks that had the same police protection status has a 0.5 probability of being assigned an ACCO sticker. The remaining day, as well as the one-day driving block, has a 0.43 chance of getting a sticker.

Cross-Randomization: Time Horizon of Rerouting

As described in Section 2, for each driver, nine of thirteen driving blocks comprised three days and the remaining four comprised one day (step 1), the assignment to foreign vs. home was stratified by whether the driving block was three vs. one day (step 2), and the sequencing of the blocks across the study was randomized (step 4). Thus, the sequencing of shorter vs. longer driving blocks was random and strictly balanced by block duration.

Including this additional randomization strata as fixed effects is inconsequential for the estimation. Thus, the time-horizon of driving was cross-randomized over the assignment to foreign vs. home. While three-day blocks are arguably not a very long horizon, the duration was determined in qualitative preparative interviews, which revealed that horizons as short as three days could be sufficient to alter the drivers' inter-temporal incentives.

Experimental Design Details
The characteristics appear balanced across foreign vs. home.
By design, assignment to foreign line rerouting is balanced, while assignment to any type of third-party protection is balanced. For this reason, we include randomization strata (minibus-wave) fixed effects when analyzing rerouting.
Randomization Method
Randomization done in office by a computer
Randomization Unit
driver-block of days

Block of days can be 1-day long, or 3-day long.






Was the treatment clustered?
Yes

Experiment Characteristics

Sample size: planned number of clusters
There are 27 days, partitionned into 13 blocks (7 3-day blocks and 6 1-day blocks) over ten drivers, resulting in 270 driver-days and 130 clusters.
Sample size: planned number of observations
2160 driver trips 270 driver-days
Sample size (or number of clusters) by treatment arms
Foreign driving: 149 driver-days
Home driving: 120 driver-days

Police protection: 134
No police protection: 135

Sticker protection: 127 driver-days
No sticker protection: 143 driver-days

"Long-run" driving block: 210 driver-days
"Short-run" driving block: 60 driver-days

Minimum detectable effect size for main outcomes (accounting for sample design and clustering)
MDE with ICC=0 (intra correlation cluster within driver block): The MDE is approximately 0.15 times the standard deviation of the outcome variable. MDE with ICC=.1: The MDE is approximately 0.16 times the standard deviation of the outcome variable. MDE with ICC=.5: The MDE is approximately .23 times the standard deviation of the outcome variable. Calculations for ICC=.1 as example: Total number of observations (n): 269 driver-days. Number of clusters (k): 130. Average cluster size (m): n / k = 269 / 130 ≈ 2.07 driver-days per block. Intracluster correlation (ICC): 0.1. Step 1: Effective Sample Size (ESS) The design effect (D) adjusts for clustering: D = 1 + (average cluster size - 1) * ICC = 1 + (2.07 - 1) * 0.1 = 1.107. Effective sample size (ESS) = n / D = 269 / 1.107 ≈ 243. Step 2: Standardized Effect Size For a two-sided test with alpha = 0.05 and 80% power, the critical values are: z_alpha/2 = 1.96, z_beta = 0.84. Standardized minimum detectable effect (MDE) = sqrt(ESS) * (z_alpha/2 + z_beta). MDE = sqrt(243) * (1.96 + 0.84) = sqrt(243) * 2.8 ≈ 15.59. Result with ICC of 0.1: The MDE is approximately 0.16 times the standard deviation of the outcome variable.
IRB

Institutional Review Boards (IRBs)

IRB Name
University of California Berkeley
IRB Approval Date
2015-06-30
IRB Approval Number
CPHS # 2015 - 06 - 7686

Post-Trial

Post Trial Information

Study Withdrawal

There is information in this trial unavailable to the public. Use the button below to request access.

Request Information

Intervention

Is the intervention completed?
No
Data Collection Complete
Data Publication

Data Publication

Is public data available?
No

Program Files

Program Files
Reports, Papers & Other Materials

Relevant Paper(s)

Reports & Other Materials