My research is in applied microeconomics, with projects spanning
marriage and the family, education, and economic history. Each project below leads
with the question it asks — the published papers add the main figure and finding.
Published
If you marry your cousin, do your children pay for it their whole lives — not just in infancy?
Figure 2. Left: share of offspring surviving beyond each age, when parents are first cousins vs. not. Right: the estimated survival gap, with 95% confidence band.
Yes. Tracing couples married over a century ago through U.S. genealogical records, the children of first cousins live about 2.2 fewer years (measured from age 5) than the children of those same parents’ siblings — and once under-reported infant deaths are counted, the loss is at least 3.2 years. The figure makes the pattern visible: the left panel plots the share still alive at each age, and the first-cousin curve stays below the comparison curve the whole way; the right panel plots the estimated survival gap, which keeps widening with age (to roughly 4–5 percentage points) rather than closing after childhood.
Can banning cousin marriage loosen the grip of family ties and push people into the modern economy?
Figure 4. Effect of cousin-marriage bans by birth decade relative to the ban, comparing people from high- vs. low-cousin-marriage surnames — urbanization, farm residence, urban residence, and interstate migration (95% confidence bands).
Using 19th-century U.S. state bans and a genetics-based measure of cousin marriage built from millions of marriage records, the bans moved people from high-cousin-marriage families off farms (−8 pp), into cities (+3 pp), and into higher-paying, more varied occupations (~+4% income). Each panel of the figure is an event study — one outcome plotted by birth decade relative to the ban (0 marks the ban year). The estimates sit flat and near zero before the ban and only move after it, exactly the signature you’d expect if the ban itself caused the shift: farm residence turns down while urbanization, urban living, and interstate migration turn up. The pattern looks cultural — the effect of loosened kinship ties — rather than genetic.
Does a census-taker’s handwriting quietly decide which people history lets us trace?
Figure 3. Both the linkage rate (left) and the share of links later validated (right) rise with handwriting legibility — consistently across all eight linking algorithms.
It does. Measuring handwriting legibility across the 1940 U.S. Census and using enumeration-district boundaries for identification, moving from illegible to perfectly legible records would raise the rate at which people can be linked across censuses by 5–10 percentage points. In the figure, each line is one of eight different linking algorithms, and legibility runs along the x-axis: both panels slope steadily upward, so as handwriting gets clearer, both the linkage rate (left) and the share of links later validated (right) improve — for every algorithm. Legibility, not just the past itself, shapes the historical record we can recover.
How much measurement error is buried in historical census data — and is it random?
Figure 5. Correcting name and age errors with genealogical profiles raises the linkage rate (left) and lowers the share of invalid links (right), across census years — before vs. after correction.
Comparing census records against genealogical profiles, the paper finds that about one in two records has an error in a name or age — and, crucially, the errors aren’t random noise: they’re systematically more common for people with less human capital. The figure shows what correcting them buys, decade by decade: in the left panel the “after correction” line sits above “before,” so more people get linked, and in the right panel it sits below, so fewer of those links are false — a net 20–36% more valid links. (Reassuringly, the errors have only a negligible effect on estimates of intergenerational mobility.)
Why does the very same apartment sell for less just because it’s on the top floor?
Figure 1. Resale price by floor level (relative to the building), for 5-, 10-, 15-, 20-, and 25-story buildings. Prices rise with height, then drop sharply at the top floor.
Across 1.9 million Korean apartment resales, top-floor units sell at a 1.3–8.3% discount. Each panel of the figure is a different building height, and the shape repeats in all of them: the price premium climbs steadily floor by floor and then drops sharply at the very top floor — that last point, sitting well below the trend, is the discount. The reason isn’t the view or the climb; it’s energy: top floors swing through more extreme indoor temperatures and cost more to heat and cool (their residents even own more air conditioners).
Can paying couples to marry across caste lines reduce caste divisions — and who pays the price?
Illustration of the reported result — original figure to be added.
A ₹10,000 incentive raises marriages between Scheduled-Caste men and non-Scheduled-Caste women by about 4% in rural India — the upward bar in the figure. But the chart’s second, downward bar captures the catch: the same policy lowers the welfare of rural women. So integration is bought partly at their expense — a trade-off the paper models and then measures.
If you can make a living without your family’s land or connections, does it change who you marry?
For most of American history, making a living ran through the family — land, credit, insurance, and work were organized around kin, and many couples married within their extended family networks. Using family trees that link millions of Americans between 1750 and 1950, this project measures how far outside their own kin network each couple found their spouse, and follows counties as expanding transportation networks brought manufacturing jobs — work requiring nothing but one’s own labor — within reach.
Can we teach a machine to correct the errors in century-old census transcriptions?
Much of what we know about long-run social mobility and migration comes from linking the same person across successive U.S. censuses, yet those links hinge on names hand-copied from century-old manuscripts — names that different genealogy firms often transcribe inconsistently. This project trains a machine-learning model to re-read the original handwritten census images from 1930 and 1940, learning from the cases where two firms already agree and re-transcribing the cases where they disagree. The aim is to see whether cleaner, machine-corrected transcriptions let researchers follow more people reliably across their lives — recovering more individuals across census rounds, with fewer false matches.
Are some countries better natural laboratories for testing whether policies work than others?
Whether a policy’s effect can be detected at all depends on the setting: how many regions a country has, how steady their underlying trends are, and how much data exists. Drawing on population censuses from countries across the world, this project studies hypothetical policies aimed at women’s participation in the labour force, comparing regions before and after over successive census rounds. It treats each country-and-period as a potential natural laboratory for policy evaluation and asks what makes some far better than others.
