Research in Progress
-
Past Automation and Future AI: How Weak Links Tame the Growth Explosion (November 2026)
[ Slides ][ Online Appendix ]Abstract
How much past growth is due to automation, and what does this imply for the future of AI? We perform task-based growth accounting and show that capital task productivity has grown about 5 percentage points per year faster than labor task productivity in the U.S. economy. The main benefit of automation is using rapidly improving machines instead of slowly-improving humans on more tasks. With these historical measures, we calibrate a growth model in which the production of goods and ideas is endogenously automated. Growth accelerates and can explode. However, the acceleration is slowed by
weak links,
tasks still performed by labor. -
The Weak-Link Asymmetry: AI’s Peril and Promise (November 2026)
Abstract
In a
weak links
model of production, where tasks are complements, the gains from automation tend to arrive slowly because output is constrained by the least productive tasks. But the same complementarity makes attacks more harmful and easier to execute. For example, with an elasticity of substitution across tasks of 0.2, automating tasks worth 1 percent of GDP with a tenfold increase in productivity raises output by less than one percent, while cutting their productivity tenfold lowers output by 68 percent on impact. When a larger AI sector simultaneously makes AI models more capable and attacks more likely, the positive productivity spillover is negligible compared with the potential damage, so the market overinvests in AI. In a weak-link economy, the costs and benefits from AI can be extremely asymmetric. -
Risky Insurance: Life-Cycle Insurance Portfolio Choice with Incomplete Markets (September 2026)
Abstract
We study consumer demand for savings, life insurance, annuities, and long-term care insurance using new survey data and a structural life-cycle model. Individuals perceive substantial nonpayment risk across all three insurance products, and these beliefs predict ownership. Embedding the measured beliefs into an incomplete-markets model alongside additional real-world insurance features, we match empirical patterns of low participation. Relative to a no-insurance benchmark, access to existing imperfect insurance reduces median wealth by 16% and generates a modest 0.6% welfare gain. Eliminating nonpayment risk substantially increases ownership, reduces median wealth by 27%, and raises the welfare gain to 2.3%.
-
Risk Markups (November 2025)
[ Slides ]Abstract
We study optimal policy when markups reflect compensation for risk instead of market power. Although markups correctly capture the private cost of risk, they are socially inefficient. This calls for a subsidy, as in the market-power perspective. However, uninsurable risk also leads entrepreneurs to dynamically overaccumulate too large a share of wealth to self-insure. Therefore, an income effect makes relatively impoverished workers oversupply labor. In the long run this effect dominates and it is optimal to tax labor and reduce aggregate output, in sharp contrast to the common wisdom derived from the market-power perspective.
-
Identification of Marginal Treatment Effects using Subjective Expectations (September 2026)
Abstract
We develop a nonparametric method that identifies marginal treatment effects (MTEs) using survey data on subjective expectations instead of instruments. Subjective treatment probabilities rank people by their latent propensity to be treated if they forecast treatment the same way. Combined with expected treatment-contingent outcomes, the ranking identifies the full MTE function if those expectations are correct on average. Both results hold under common forms of reporting error. We apply the method to study childbirth and female labor supply in Denmark. Estimated MTEs are roughly constant across ranks of fertility beliefs, so there is little selection into childbirth on unobserved determinants of work after birth. The survey-based average treatment effect on the treated nearly equals the event-study estimate from administrative data at every horizon, evidence that the identifying assumptions hold. Furthermore, eliciting beliefs under a counterfactual policy measures its effect before implementation, accounting for how the policy changes selection into treatment.
-
Beliefs and Realities of Work and Childcare After Childbirth (July 2026)
Abstract
When women plan for life after childbirth, they form beliefs about work, childcare, and how their careers will unfold. These expectations shape key decisions but are formed under deep uncertainty. We use a 2019 state-contingent survey of 11,000 Danish women linked to administrative data to compare pre-birth beliefs to realized outcomes. Mothers accurately anticipate their eventual return to work but underestimate the duration of the career interruption. This miscalibration stems from two belief errors—about partner leave and own labor supply—which interact and persist even among second-time mothers, with implications for labor supply, planning, and policy design.
-
Due Diligence: Endogenous Information Acquisition and Offer Quality in Search and Matching (August 2026)
Abstract
In many search markets, offers must be accepted or rejected before their quality is known. We build a search and matching model in which suppliers privately choose offer quality and searchers choose costly due diligence before accepting. How carefully searchers screen determines the mix of high- and low-quality offers. This composition channel is absent when information and offer quality are exogenous. In the static model, improving low-quality offers makes accepting one less bad, so searchers screen less carefully and suppliers post more low-quality offers. The mix worsens enough that more searchers remain unmatched and their welfare falls. Raising the value of remaining unmatched actually reduces the number of unmatched searchers, because searchers screen more carefully and suppliers post more high-quality offers. In the dynamic economy the composition channel survives, and a classic market tightness force can reinforce or overturn it. A planner who treats information and offer quality as exogenous can subsidize entry where a tax is optimal.
-
The Value of a Job (Work in Progress)
Abstract
We estimate the present discounted value of earnings for a particular worker having a job at a particular firm, using a nonparametric statistical model that nests many rich structural models. We assume a stationary Markov structure conditional on a vector of idiosyncratic states, where states determine payoffs and transitions. We use rich employer-employee matched data from Denmark, cluster workers and firms into types, define a parsimonious set of state variables, and estimate type- and state-specific payoffs and transition probabilities directly from their empirical counterparts. We then compute values by iterating over a Bellman equation. We decompose the value into its components, including earnings on the current job, the probability of staying or moving to a new job of a particular type and the earnings growth associated with staying or moving, and transitions to nonemployment with associated payoffs. We also compare the distribution of job value changes upon job-to-job moves to the wage change distribution and show how patterns of worker and firm sorting differ when using values instead of wages.