Reducing Slack in Informal Transit: The Impact of Temporary Subsidies on Urban Mobility

Our intervention aims to increase service frequency on selected minibus routes by providing subsidies to drivers who agree to depart within a set number of minutes after the previous vehicle, even if not all seats have been filled. In the status quo, minibuses typically wait in the taxi park until all 14 seats are filled and only then depart, leading to passenger wait times that are highly counter-cyclical (much higher during off-peak periods). We expect that drivers who accept the payment will depart more frequently and more regularly, but with fewer passengers.

We will offer subsidies during a selected four-hour time window at each stage, Monday - Friday, for a period of 3 weeks. We will implement the intervention at multiple taxi parks around Kampala during three separate phases.

Drivers who accept the payment must depart within the specified time period (set for each route) after the previous vehicle. They must only begin loading passengers after the previous vehicle has departed, and they cannot turn any passengers away after they begin loading. Drivers must also avoid "holdbacks" (stopping for an extended period of time within a short distance of the taxi park to continue filling seats after accepting the subsidy and departing); however, picking up additional passengers waiting along the route is allowed (as is common practice). Drivers are instructed that "mystery passengers" will be randomly assigned to audit trips, and that non-compliant drivers will be excluded from eligibility for future payments.

We determined the payment amounts, target frequency, and subsidization time window using baseline data we collected on frequency, headway times, and passenger counts for each route. The subsidy amount is approximately 20,000 UGX per trip but will be slightly lower or higher, depending on the stage and time of day. The payment amount is calibrated based on the route and period of the day to compensate drivers for the revenue shortfalls we expect them to incur by departing before their vehicle is full. The target frequency is set for each route typically represents a 50% reduction in passenger wait times and 50% increase in trip frequency, on average, during the treatment time window. To select the four-hour time period of the day during which to subsidize more frequent departures, we chose time intervals where frequency was typically low and/or inconsistent.

2025-08-18

2025-12-19

The first outcome is trip frequency during the four-hour target window. The intervention aimed to increase this by subsidizing drivers to leave more often, and our first analysis will measure the effect on the number of departing trips per hour. This depends on take-up, as some drivers may refuse the incentive and wait until they fill up with passengers.
The second outcome is ridership or demand. We will measure whether and how much ridership of the route increases due to the intervention (more frequent departures).
The third outcome is queue length. If drivers leave more frequently due to the intervention, this would lead to a mechanic effect to reduce queue length. However, this may be partially compensated if other drivers join the queue in treatment routes.

We will analyze whether the intervention has lasting effects after the incentives end, i.e. the treatment routes continue to have frequent departures.

We will first conduct descriptive surveys with stage leaders to build a roster of stages, the routes at each stage, and fares for each route. Prior to the three-week intervention period, we will collect baseline data on arrivals, departures, and passenger counts for two weeks, covering all routes at each stage.

We will then use the baseline data to restrict the intervention sample to stages with route(s) where average headway times are between 30 - 120 minutes during any 4-hour time window during the day. Stages with multiple routes will be eligible for the sample if at least one route (or a merger between multiple routes, which is common at certain stages) meets these criteria. Stages will be excluded from the sample if observed average frequencies for any route do not meet these criteria and/or if stage leaders refuse to participate. Additional stages may be excluded if the quality of guide data is deemed poor (based on the degree of correspondence between guide and surveyor-collected snapshots, and whether this has improved during baseline after feedback is delivered to the guide), since it will not be possible to measure our outcomes accurately at these stages.

While treatment group assignments will be made at the stage level, payment amounts and eligibility will be route-specific. For stages with multiple routes, any departure along an eligible route, including departures merged with a non-treatment route, will be eligible for a payment; departures serving only ineligible routes will not be offered payment.

For each phase, we first select the sample of routes as described above, then define the route-specific intervention, and afterwards randomize stages into control and treatment, stratified as described below. This ensures we have counterfactual intervention information for control routes (e.g. time period when they would have been treated).

Not available

Eligible stages will be randomly assigned to treatment / control in-office by computer in R.

We will randomize the intervention at the stage level, with approximately 25 treatment and 25 control stages. (Treatment stages have at least one route where payments will be offered to drivers, whereas control stages have no routes eligible.)

We will stratify the sample by baseline frequency and route length (distance between origin and destination).

Yes

30 treatment and 30 control stages.

The number of trips that will ultimately be observed cannot be estimated prior to analyzing baseline data due to variance in both the number of departures per route and the number of routes per stage.

30 treatment and 30 control stages. Each stage has an average number of routes between 1 and 2.

We computed power for the first two primary outcomes. We assume 60 branches, our treatment pays for 50% reduction in headway and there is >=70% take-up of incentives. We are powered >99% powered to detect the frequency change (primary outcome 1). For ridership (primary outcome 2), assuming an elasticity of ridership with respect to average wait time of -0.2, power is 77%. For elasticity -0.3, power is >99%. Hence, under these assumption, we are powered to detect ridership elasticities a bit larger (more negative) than -0.2.