Working Papers
Unawareness, Impatience, and Mental Cost in Mortgage Refinancing Decisions, with Haizhen Lin, Ruli Xiao, and Jun Zhu
Abstract: This paper develops a dynamic discrete choice model to quantify the distinct roles of inattention, mental costs, and myopia in mortgage refinancing decisions, using the Home Affordable Refinance Program (HARP) as an empirical setting. We allow borrower awareness to evolve over time as an absorbing latent state, identified through variation in the lagged DMA-level refinancing rate as an attention shifter. Reduced-form evidence documents widespread failure to refinance despite substantial financial incentives and shows that local attention proxies significantly predict refinancing behavior. Structural estimates from the limited awareness model reveal moderate borrower myopia (monthly discount factor of 0.84) and a differential mental switching cost of approximately $3,950 for first-time homebuyers relative to repeat buyers. Counterfactual decompositions identify the closing fee and mental cost as the two dominant individual frictions (each approximately +4.5 pp), with attention contributing a smaller +0.5 pp at the estimated parameters. Jointly removing all three frictions raises the predicted refinancing rate by 11.7 pp, indicating substantial scope for cost-reduction and process-streamlining policies.
Organizational Complementarities in Diagnostic AI Adoption: Evidence from HeartFlow FFRct, with Avi Goldfarb
Abstract: Artificial intelligence has diffused into clinical medicine far more slowly than its technical promise would suggest. We study this puzzle through HeartFlow FFRct, a noninvasive diagnostic AI for coronary artery disease. Despite FDA clearance, Medicare reimbursement, clinical guideline endorsement, and no required capital investment, only 14 percent of capable U.S. hospitals had adopted HeartFlow eight years after approval. We argue that slow diffusion reflects not implementation costs but a mismatch between what the technology does and how hospitals are organized to capture value from it. HeartFlow improves the accuracy of a routing decision between intensive and non-intensive cardiac care, creating two organizational frictions: it reduces billable invasive procedures, giving fee-for-service physicians reason to resist it; and it shifts value from diagnostic testing toward downstream treatment, which only vertically integrated hospitals can retain. These frictions interact. Salaried compensation matters most when the hospital owns downstream cardiac surgery; cardiac surgery capability matters most when physicians lack incentives to preserve procedure volume. Using the 2023 AHA Annual Survey for 2,393 hospitals, we find that the interaction between salaried physicians and cardiac surgery capability raises adoption by roughly eight percentage points, more than half the mean adoption rate. The estimate is stable across specifications and absent for placebo technologies. These results suggest that for diagnostic AI in healthcare, the binding constraint on adoption may be organizational design rather than prediction quality.
