Managing The Gig Economy Via Behavioral And Operational Lenses

Wichinpong Park Sinchaisri, University of Pennsylvania


This dissertation combines tools from operations management, econometrics, machine learning, and behavioral sciences to (i) study how on-demand workers learn and make decisions in complex environments, (ii) develop tools to help improve their decision-making, and (iii) inform the design of better policies to manage human-centered operations. Recent technologies create and accelerate new work arrangements that provide workers with flexibility in their schedule and choice of service. At the same time, the decisions a worker faces have become more complex. Platforms offer competing dynamic incentives, and the independent nature of gig work means that workers do not experience the benefits of learning from colleagues. The following three chapters investigate how behavioral operations management can be utilized to better manage the gig economy.

(i) Behavioral and Economic Drivers of Decisions.We empirically investigate how on-demand workers decide on when to work and \emph{for how long} depending on varying financial incentives and personal goals. Using the comprehensive data from a ride-hailing industry partner, we develop an econometric framework that addresses empirical challenges such as sample selection bias and endogeneity. Our results demonstrate that, while workers exhibit positive income elasticity as predicted by standard income theory, their decisions are significantly influenced by their cumulative earnings (more likely to stop working when reaching their income goal) and recent work duration (tend to stay working after long hours of work or exhibit "inertia"), more akin to the behavioral theory of labor supply. Inertia captures both the formation of work habits and the tendency to stay with the focal platform, suggesting that, amidst intensifying competition among platforms, platform loyalty could be induced through optimal incentive design. Thus, we propose a heuristic to optimize incentive allocation and demonstrate through counterfactual simulations the monetary and capacity benefits of accounting for our behavioral insights.

(ii) Dynamic Decisions and Multihoming Behavior. We leverage proprietary data from our ride-hailing industry partner and the publicly available trip record data to develop and estimate a structural behavioral model of gig workers' sequential dynamic decisions of when and where to work in the presence of alternative work opportunities. Our major contributions are in the modeling and estimation of dynamic decisions with temporal and spatial components and dynamic outside options, and the development of an efficient simulation-assisted machine learning-based estimation framework. Our results characterize gig workers' forward-looking behavior and heterogeneous cost of working. We find that workers are strategic in their choice of initial service location to ensure high utilization and are prone to multihoming behavior when facing longer idle times. Then, we study how the firm can influence multihoming behavior among workers. Our counterfactual analyses demonstrate the effectiveness of strategies commonly used in practice and offer insights that can help retain workers during high demand or nudge them to quit during low demand.

(iii) Improving Human Decision-Making with Machine Learning. We propose a novel machine-learning algorithm to automatically extract best practices from the trace data and infer simple tips that can help workers learn to make better decisions. We use an approach based on imitation learning and interpretable reinforcement learning and consider simple if-then-else rules that modify workers' strategy in a way that most improve their performance, capture useful insights that are challenging for workers to learn by themselves, and are simple enough for workers to understand. To validate our approach and test the performance of our algorithm, we design a virtual kitchen-management game and conduct large-scale pre-registered behavioral studies on Amazon Mechanical Turk. Our experiments show that rules inferred from our algorithm are effective and significantly outperform rules from other sources at improving performance and speeding up learning among workers. In particular, we help workers identify optimal early actions that help them improve in the long term and discover additional optimal strategies beyond what is stated by our algorithm.