Dong Huang

This paper shows that delegating unpopular actions to an external agent can protect firms' customer relationships from the adverse consequences of such actions. This blame-shifting benefit of outsourcing is one that firms should consider when setting their organizational boundaries. I study an online consumer lender in China that randomly assigns delinquent borrowers to in-house or third-party debt collectors, who differ in collection harshness as measured by their propensity to call borrowers' social contacts. Harsher collection raises short-run repayment under both arrangements, but only in-house harshness deters borrowers—particularly those of higher credit quality—from returning to the lender after repayment. Call transcripts show that harsher collection comes with more threatening communication strategies in both arrangements, yet it provokes borrower anger only when the collector is in-house.

How good is AI at persuading humans to perform costly actions? We study calls made to get delinquent consumer borrowers to repay. Regression discontinuity and a randomized experiment reveal that AI is substantially less effective than human callers. Replacing AI with humans six days into delinquency closes much of the gap. But borrowers initially contacted by AI have repaid 1% less of the initial late payment one year later and are more likely to miss subsequent payments than borrowers who were always called by humans. AI's lesser ability to extract promises that feel binding may contribute to the performance gap.

Using transaction data from a large non-fungible token (NFT) trading platform, this paper examines how selection-neglect interacts with extrapolative beliefs, possibly accelerating the boom and delaying the crash in the recent NFT bubble. We document extrapolative beliefs about increasing prices and an upward sample selection bias in observed transaction prices. We show that agents prone to selection-neglect formed more optimistic beliefs and traded more aggressively during the boom. When market liquidity declined, selection bias intensified and delayed the crash onset. Finally, investors with sophisticated bidding behavior showed less selection-neglect, responded more strongly to declining market conditions, and performed better.

The most widely used U.S. house price indexes estimate a single price path for each metropolitan area and assume that the homes that sell are representative of the wider housing stock. We develop a Bayesian repeat-sales model that relaxes both assumptions. It corrects a repeat-sales index for selection on the return since purchase using a nonparametric selection function and it assigns properties to latent sub-markets, each with its own loading on a common factor, idiosyncratic drift, and selection function. Across the 20 Case-Shiller metros, uncorrected indexes overstate 1990–2020 appreciation in every metro, by 0.2 to 1.9 percentage points a year. For New Haven County, the model identifies three sub-markets that are geographically concentrated without using location data. Selection bias is concentrated in a small, high-volatility city segment. Membership in this segment changes at the city boundary in ways tract demographics do not explain.

NFTs provided an extraordinary real-time laboratory for bubble economics: returns were exceptionally right-skewed, illiquidity pervaded even the most active platforms, and a handful of trades drove aggregate performance. Investors extrapolating from realized returns without recognizing selection bias and survivorship faced a substantial risk of disappointment. As our data and simulations confirm, successful NFT investing during the bubble required an almost perfect confluence of timing, liquidity, and luck.