Daniel Chen

Gig economy workers make strategic decisions about where and when to work. These decisions are central to gig economy operations and are important policy targets both to firms that operate ridehail and delivery platforms and to regulators that oversee labor markets. We collaborate with a driver analytics company to empirically measure two types of strategic behavior: multihoming, an online change between platforms, and repositioning, a physical change between locations. Using a comprehensive dataset that tracks worker activity across platforms, we estimate a structural model to analyze how workers optimize their earnings and respond to earnings-based incentives to switch platforms or locations. We show that workers are highly heterogeneous in their preferences and find multihoming especially costly, both in absolute terms and relative to the cost of repositioning. Through counterfactual simulations, we show that firms and regulators can substantially improve system efficiency by enabling workers to freely multihome: workers' hourly earnings increase by 2.0% and service levels increase by 53.1%. In contrast, the existing equilibrium is similar to a system without multihoming, in which hourly earnings increase by 1.3% but service capacity decreases by 4.1%. Additionally, we show that policies to limit traffic congestion by increasing travel costs should include incentives to ensure that workers remain able to efficiently reposition. An increase to repositioning costs by $1 per mile increases hourly earnings by 2.3% but substantially decreases service capacity by 29.6%.

Machine learning is a growing method for causal inference. In machine learning settings, prediction errors are a commonly overlooked problem that can bias results and lead to arbitrarily incorrect parameter estimates. We consider a two-stage model where (1) machine learning is used to predict variables of interest, and (2) these predictions are used in a regression model for causal inference. Even when the model specification is otherwise correct, traditional metrics such as p-values and first-stage model accuracy are not good signals of correct second-stage estimates when prediction error exists. We show that these problems are substantial and persist across simulations and an empirical dataset. We provide consistent corrections for the case where unbiased training data is available for the machine learning dataset.

Algorithms control pricing and match customers and workers in the gig economy. Despite their prevalence, algorithms face several critiques: they lack transparency, can be biased, and can be inefficient. We collaborate with a driver analytics company to empirically analyze these issues using a comprehensive dataset on completed trips in the Chicago ridehail market. We show that algorithmic pricing varies significantly after controlling for distance, duration, location, and time of day: customer fares have a residual standard deviation of 36% ($6.16 against an average of $17.12 per trip), while worker earnings vary 29% ($3.77 against an average of $12.86). We find that algorithms are not biased against low-income locations, as ridehail platforms offer lower prices and higher service levels relative to taxis. However, algorithms lose efficiency from two sources: inefficient routing from competition between platforms decreases worker utilization by 1%, and incentives for workers to decline trips further decrease utilization by 11%. We model workers’ strategic responses to variation in pricing and estimate counterfactuals on the effects of minimum wage and transparent pricing policies.

Online ratings are a prominent feature of e-commerce and social media platforms, but crowdsourced ratings are often unreliable signals of quality. We show that network information can significantly improve signal quality by mitigating difficulties from bots, strategic users, and heterogeneous preferences. First, using theoretical models, we illustrate several advantages of network-based ratings for converging to accurate ratings and preventing herding to undesirable outcomes. Second, using large empirical datasets from e-commerce and social media, we show that improvements from network ratings are large and practical to implement.