Replacing What Could Be Repaired: A Structural Analysis of Two-Stage Diagnostic Decisions in Managing Shared-Bike Returns
Hailong Cui, Jingxuan Geng, Guangwen Kong(Alphabetical Order)
Revising for submission to Management Science
Bike-sharing platforms face significant challenges from high maintenance costs, driven by heavy usage and inefficiencies in diagnostic decision-making. Using task-level data from a leading bike-sharing platform, we develop a structural estimation model to analyze two-stage diagnostic decisions made by inspectors (stage 1) and workers (stage 2). These decisions are modeled as a strategic interaction governed by a Bayesian Nash Equilibrium (BNE). To address the computational complexity of Maximum Likelihood Estimation with BNE constraints, we employ machine learning to approximate BNE. We identify systematic overtreatment tendencies among inspectors and workers, resulting in a higher false positive rate than that under the firm’s optimal decisions and thus inflating maintenance costs. Our counterfactual analyses show that higher part costs, reducing workers’ piece-rate wages, adopting structured job matching, and prioritizing worker training can substantially reduce costs. Transitioning from a two-stage to a one-stage process lowers diagnostic accuracy and increases costs, although optimizing wages narrows this gap. This framework provides actionable insights for mitigating inefficiencies in multi-agent diagnostic decision systems and is generalizable to other credence goods industries, such as heavy equipment maintenance and healthcare, where diagnostic errors have significant financial, operational, and health implications.