Experimental Design
Research questions:
a) How do environmental shocks impact productivity and the cost of effort when effort is not incentivised on the margin, i.e. under a fixed wage contract?
b) How do different incentive contracts impact resilience? Do these effects vary with the worker’s resilience to weather shocks under the fixed wage contract?
c) Are these impacts heterogeneous across places, i.e., higher baseline variance vs lower baseline variance, and across different environmental determinants such as air pollution and heatwaves. What is the optimal contract under different distributions of environmental shocks?
d) How does contract performance change over time, as a function of shock realizations? Is this due to entry/exit, learning or both?
e) How is the surplus from different contracts allocated between workers and firms?
Design:
We leverage an at-scale nationwide randomized experiment conducted in partnership with the largest mobile money service provider in Ghana (Annan and Raymond 2026). The implementing partner provides its services through retail agents, who are compensated using a linear contract that provides a bonus payment per mobile money transaction (status quo).
The existing RCT randomly assigns agents to one of five contracts (see Annan and Raymond for further detail):
Individual Performance-Based Schemes:
1. Simple Linear Contract (Status Quo): Workers receive bonuses for every unit of output.
2. Threshold Contract: Workers receive a flat payment upon reaching a predetermined threshold.
3. Pure Franchising Scheme: Akin to two-Part Tariff, workers receive a boost in the per-unit commission in exchange for an upfront fee.
Relative Comparison Schemes:
4. Tournament Scheme: Workers compete in tournaments where top performers in each locality receive an additional bonus.
Control Scheme:
5. Flat Bonus/week Contract: This contract is included (i) to check whether respondents pay attention and understand the options presented and (ii) to evaluate the implications of the provider MTN switching to non-performance-based compensation schemes.
Randomization is at the local market (community) level, meaning every agent (and manager assigned) within a community receives the same treatment [Footnote: Randomization is proportional to managerial rankings, such that assignment is conditionally random. We will account for this in the analysis].
Contracts were designed to be ex ante expenditure neutral from the perspective of the firm.
Specifications:
a) To answer the first research question, we will regress outcome(s) on weather and/or pollution, with (month) and agent fixed effects, restricted to the flat wage contract. This specification will build on prior work estimating the impact of weather shocks in this setting (Annan et al., 2024). Specifically, we will allow for non-linearities in the response, and will incorporate (distributed) lags, and higher order terms to test which specification fits best. We will also include forecasts and leads in some specifications.
In separate analyses, we will estimate individual-specific environmental sensitivity to characterize the heterogeneity in the response function, and its determinants (for instance, baseline profitability). Specifically, we will estimate the relationship between outcomes and environmental variables in the pre-period (under the linear contract), and again in the post-period for the set of agents assigned to the flat wage contract. We will use the relationship between the relationships in the estimate weather sensitivity among agents who show up in both analyses to extrapolate to the rest of the sample. This will provide a sample-wide estimate of the cost of environmental shocks on worker productivity and cost of effort.
b) To answer the second research question i.e. how contracts incentivizing effort impact resilience, we will use the previous specification, incorporating all treatments and time periods, and add treatment indicators, alone and in interaction with the environmental variables. We will estimate the resilience effect of each type of contract (linear, tournament, threshold, and franchising) relative to the fixed wage for different environmental shocks (e.g. heatwaves vs. air pollution) and test for equality across treatment effects.
c) We will test for heterogeneity across locations and over time. First, we will estimate whether certain types of contracts are more resilient in more volatile vs. less volatile areas., which address the third research question. Second, we will test whether resilience improves overall with time on a new contract, and whether improvements are a function of shocks, consistent with learning, which will answer our fourth research question.
d) Finally, we will decompose the surplus associated with different contracts between workers and the firm, using the estimated effort costs from step (1) combined with sales volumes and bonuses. To calculate the payoffs to the firm, we will estimate their net revenue as a function of treatment and environmental realizations. This will also allow us to estimate how different contracts perform as weather variability increases.
We will use lasso to include additional baseline controls in all specifications for improved power. In the future, we hope to characterize effects on customers and the role of demand-side mechanisms for affecting contract performance.
Data:
We will combine administrative and survey data from Annan and Raymond (2026) with environmental data on weather and pollution.
Weather data: Temperature and precipitation will come from Hersbach et al. (2020)’s ERA5 reanalysis product, which is available hourly at a 31.5 km^2 resolution. If finer resolution data become available, we will use that instead.
Pollution data: We will use daily surface-level PM2.5 as modeled by Westervelt et al. (2025) at a 1km^2 resolution. The current modeled output covers all of West Africa. We will revert to Ghana-specific data if/when it is made available.
References:
(1) Annan and Raymond (2026) [https://doi.org/10.1257/rct.13865-2.0].
(2) Westervelt, D. M., Amooli, J. A., & Anand, A. (2025). Twenty Years of High Spatiotemporal Resolution Estimates of Daily PM2. 5 in West Africa Using Satellite Data, Surface Monitors, and Machine Learning. ACS ES&T Air.
(3) Hersbach, H., Bell, B., Berrisford, P., et al. (2020). The ERA5 global reanalysis. The Quarterly Journal of the Royal Meteorological Society, 146(729), 1999-2049.