0 cumulative citations
View corpus contextTargeted physician placements increase primary‑care supply and advanced preventive use and lower emergency admissions and mortality, especially in disadvantaged and rural areas; mid‑century mortgage‑market integration raised homeownership, population and construction in capital‑scarce cities; and in bail settings most judges worsen outcomes when they override algorithmic recommendations, though a small share improve both accuracy and fairness.
Essays in Labor and Public Economics
openalex
quasi_experimental
medium evidence
7/10 relevance
Full text
usable extracted full text
Source
PDF
Using quasi-experimental variation across three essays, the dissertation shows that NHSC physician arrivals raise local primary care supply and advanced preventive utilization and reduce emergency hospitalizations and mortality; mortgage-market integration in the 1930s spurred homeownership, population and construction in capital-scarce cities and influenced fertility; and in bail decisions most judges underperform an algorithm when overriding it while a minority improve both accuracy and fairness.
Citation observations
Cumulative provider counts captured on specific dates; providers are never combined.
In this dissertation, I use tools from labor and public economics to study the effectiveness of policies aimed at improving access to public goods and services and how the benefits of these policies are distributed across key decision-makers and recipients. In the first essay, joint with Shreya Tandon, we estimate the causal effects of new primary care physician arrivals to underserved areas on net physician supply, healthcare utilization, and health outcomes of local residents using quasi-random variation in the timing of physician placements through the National Health Service Corps (NHSC) program. Our primary data include the roster of NHSC physician recipients and a sample of Medicare claims from 1999--2019. We find that physician arrivals increase overall primary care supply. Half of the non-inpatient healthcare provided by new entrants would have been satisfied by the incumbent primary care physicians, while the other half constitutes an increase in overall utilization. The additional utilization is concentrated in more advanced preventive care services, such as advanced testing and imaging. We also find an overall increase in elective inpatient procedures, particularly related to cardiovascular diagnoses. The greater healthcare utilization is associated with significant reductions in emergency hospitalizations and mortality. Health improvements persist for at least four years following physician arrival and are particularly pronounced for beneficiaries with chronic conditions and those living in more disadvantaged and rural locations. In the second essay, joint with Leonardo D'Amico, we examine how a persistent change in mortgage rates affects city growth, home prices and construction, and household formation. We study the revolutions in mortgage financing that took place in the U.S. between 1933 and 1940, which created a national mortgage market and allowed mortgage capital to move from the financial centers to the rest of the country. By digitizing city-level census data and a new sample of loan-level data, we show that differences in mortgage rates across cities went from nearly 300 basis points to just over 100 in only six years. This national mortgage market allowed initially capital-scarce places to grow more than initially capital-abundant ones. In the decades following the housing policies, cities where mortgage rates declined more as a result of the integration of mortgage markets saw higher growth in rates of homeownership, population, and housing construction. House prices moved only modestly, suggesting that, in 1940--50, housing supply responded robustly to higher demand for homes. These macro-level policies also influenced intra-household decision-making: we find that women who experienced lower mortgage interest rates during their childbearing years had more children. In the third essay, joint with Will Dobbie and Crystal Yang, we study a common feature of policies that deploy predictive algorithms in high-stakes settings: despite the availability of algorithmic recommendations, human decision-makers are typically kept in the loop and can exercise discretion to override them. We ask whether these discretionary overrides add valuable private information or merely reintroduce human biases and mistakes, in the context of bail decisions. We develop new quasi-experimental tools to measure the impact of human discretion over an algorithm on the accuracy of decisions, even when the outcome of interest is only selectively observed. We find that 90 percent of the judges in our setting underperform the algorithm when they make a discretionary override, with most making override decisions that are no better than random. Yet the remaining 10 percent of judges outperform the algorithm in terms of both accuracy and fairness when they make a discretionary override. We provide suggestive evidence on the behavior underlying these differences in judge performance, showing that the high-performing judges are more likely to use relevant private information and are less likely to overreact to highly salient events compared to the low-performing judges.
Summary
Main Finding
Victoria Angelova’s dissertation (Harvard PhD, 2026) contains three empirically rich essays showing (1) targeted physician placements increase primary care supply and generate persistent health improvements in underserved areas, (2) mid-20th century creation of a national mortgage market lowered mortgage-rate dispersion and spurred local housing, population and fertility responses in capital-scarce cities, and (3) in high-stakes, human-in-the-loop algorithmic settings (pretrial bail), discretionary human overrides usually worsen predictive performance but a minority of decision‑makers improve both accuracy and fairness. The work combines quasi-experimental variation, new digitized historical microdata, and new identification tools for selectively observed outcomes.
Key Points
- Essay 1 (Physician supply & health)
- Uses quasi-random timing of National Health Service Corps (NHSC) physician placements to identify causal effects.
- New physician arrivals increase local primary care supply; roughly half of the new entrants’ non-inpatient visits substitute for incumbents, half are net new utilization.
- Additional utilization concentrated in advanced preventive services (testing, imaging) and associated elective inpatient procedures (notably cardiovascular).
- Net effects: reductions in emergency hospitalizations and mortality that persist ≥4 years, with stronger benefits for beneficiaries with chronic conditions and for disadvantaged/rural areas.
- Essay 2 (National mortgage market)
- Digitized city-level census series and loan‑level data show mortgage-rate dispersion across cities fell sharply between 1933–1940 (from ~300 bps to ~100 bps), driven by federal housing policies and market integration.
- Cities that experienced larger mortgage-rate declines later exhibited higher growth in homeownership, population, and housing construction; house prices rose only modestly—consistent with supply response.
- Exposure to lower mortgage rates during women’s childbearing years is associated with higher fertility.
- Essay 3 (Algorithms and human discretion)
- Studies bail decisions with an implemented risk-assessment algorithm and judge discretion to override.
- Novel quasi-experimental tools developed to evaluate the marginal accuracy/fairness effect of human overrides when outcomes are selectively observed.
- Findings: ~90% of judges’ overrides reduce accuracy (most overrides are no better than random); ~10% of judges outperform the algorithm in both accuracy and fairness when overriding.
- High-performing judges appear to use relevant private information and are less reactive to highly salient events; heterogeneity in objective functions and private information matter.
Data & Methods
- Essay 1
- Data: NHSC roster of physician recipients; 1999–2019 sample of Medicare claims; provider/practice location data, local area characteristics.
- Identification: quasi-random timing of NHSC placement starts across zip codes — event-study / difference-in-differences frameworks, matched control zip codes; robustness checks (affiliation types, rural vs urban, alternative control groups).
- Outcomes: local PCP counts, outpatient/carrier claims by service types, elective inpatient procedures, emergency hospitalizations, mortality; subgroup analyses for chronic conditions.
- Essay 2
- Data: newly digitized city-level census series; NBER mortgage loan cards and a new sample of loan‑level data; building permits, house prices, homeownership, population by city (1920s–1950s).
- Identification: exploiting cross-city variation in initial mortgage rates and exogenous shocks to mortgage market integration following 1933–40 federal policies; event studies, semi-elasticity estimates, cohort exposure designs for fertility.
- Tests: decomposition by lender types (B&Ls, commercial banks), placebo/robustness checks, analysis of supply vs price responses.
- Essay 3
- Data: administrative pretrial case data (risk-assessment reports shown to judges), judge assignment quasi-randomness, misconduct outcomes (new criminal activity, failure-to-appear), judge survey.
- Methods: new quasi-experimental estimators to extrapolate counterfactual misconduct under algorithmic-only decisions when outcomes are only observed for released defendants; judge-level performance decomposition; instrumental/assignment-based designs to handle selective observation; robustness to alternative algorithmic counterfactuals.
- Mechanism probes: analysis of private information usage, judge surveys on preferences, response to salient adverse events.
Implications for AI Economics
- Human-in-the-loop governance
- The third essay provides direct, causal evidence on the value of human discretion in algorithmic decision-making: most discretionary overrides worsen predictive performance, but a small subset of decision‑makers add value. That implies policy/design choices about retaining human discretion should be targeted—not blanket retention or removal.
- Practical implication: deploy hybrid systems that (a) allow overrides only for identified high-performing decision-makers, (b) require justification/audit trails for overrides, or (c) route borderline cases to human experts whose performance is validated.
- Measuring algorithm vs human value when outcomes are selectively observed
- The dissertation introduces identification tools for evaluating algorithmic counterfactuals under selective observation (common in criminal justice, medical triage, credit denials). These tools are directly useful for AI economists evaluating impacts of predictive systems where outcomes are only observed conditional on decisions.
- Heterogeneity matters
- Heterogeneous judge performance shows that aggregate measures of algorithmic benefit can mask important distributional and implementation heterogeneity. AI-economics evaluations should estimate heterogeneity in human–algorithm interactions (by agent, case-type, context) and not rely solely on average treatment effects.
- Auditing and incentive design
- Evidence that high-performing judges use relevant private information suggests training, structured decision support, or incentives could improve human overrides. AI-economic policy design should consider screening, training, and incentive schemes to align discretion with welfare-improving private information use.
- Deployment in public services and welfare consequences
- The first essay’s findings on substitution vs new utilization and persistent health benefits illustrate the broader point that algorithmic triage or allocation of public goods (e.g., automated referrals, predictive outreach for preventive care) must account for: (a) local supply responses, (b) substitution externalities on incumbents, and (c) long-run welfare gains that may be concentrated among disadvantaged populations.
- Cost-benefit analyses for algorithmic interventions should incorporate downstream use increases (some net new demand improves welfare) and distributional effects.
- Financial-market and macro insights relevant to modern digital credit
- The mortgage essay implies that lowering frictions to capital allocation (e.g., via AI-driven underwriting, platform-based mortgage distribution) can have heterogeneous local growth and demographic effects. AI-enabled financial integration may spur housing supply and household formation in capital-scarce places, but impacts depend on supply elasticity and local institutions.
- Methodological contributions for AI economists
- The combination of event studies, matched control designs, digitization of archival microdata, and estimators for selective-outcome settings offers a toolkit directly applicable to the evaluation of AI systems in public policy contexts (healthcare triage, credit, pretrial risk tools).
- Policy recommendations for algorithmic systems
- Use validated algorithms where they demonstrably improve accuracy/fairness.
- Restrict or audit overrides; identify and empower high-performing decision-makers; provide structured channels for use of verifiable private information.
- Design evaluation pipelines that account for selective observation and heterogeneous agent performance.
- When algorithms affect public goods, model both substitution and induced utilization and measure long-run welfare and distributional impacts.
Limitations and future directions (brief) - External validity: the bail results are jurisdiction- and institution-specific; judge heterogeneity may differ elsewhere. - Measuring private information: more work needed to systematically identify what private signals humans use and whether they can be codified into algorithms. - Dynamic/adaptive responses: agents may change behavior in response to algorithm deployment (gaming, effort changes) — further study needed. - Integration with cost analyses: explicit welfare and budgetary trade-offs of algorithm deployment vs human staffing/training could be quantified.
Overall, Angelova’s dissertation contributes empirical evidence and methodological tools that are highly relevant to AI economics—especially for evaluating hybrid human–algorithm systems in public-sector settings, understanding heterogeneous effects of lowered frictions in markets, and designing governance for algorithmic discretion.
Assessment
Paper Typequasi_experimental
Evidence Strengthmedium — All three essays use rich administrative and historical data and exploit plausibly exogenous policy or placement variation rather than simple cross-section comparisons, producing credible causal claims; however none are randomized experiments, and each design faces potential threats (time-varying confounders or migration in the NHSC results, other concurrent policies or general equilibrium effects in the historical mortgage episode, and residual selection/measurement issues in the bail-setting analysis) that limit causal certainty and external validity.
Methods Rigorhigh — The author links large administrative datasets (NHSC rosters, Medicare claims, digitized census and loan-level data) and develops new quasi-experimental estimators to handle selective observation; the empirical approach is careful and multi-pronged, using long panels, subgroup analyses, and robustness checks, indicating strong econometric and data-construction rigor.
SampleMultiple datasets: (1) NHSC physician recipient roster linked to a sample of Medicare claims covering 1999–2019 and local area indicators; (2) digitized city-level decennial census data and newly digitized loan-level mortgage records covering the 1933–1940 mortgage reforms and subsequent decades for city-level outcomes (homeownership, population, construction, prices); (3) administrative court data on bail decisions where algorithmic recommendations and judge overrides are observed, with outcomes used to evaluate accuracy and fairness (selectively observed).
Themeshuman_ai_collab governance inequality
IdentificationAcross three essays the dissertation uses quasi-experimental variation: (1) plausibly exogenous timing of National Health Service Corps (NHSC) physician placements to identify local effects on primary care supply, utilization, and health outcomes using NHSC rosters linked to Medicare claims (1999–2019); (2) historical, policy-driven variation in city-level mortgage rates produced by 1933–1940 mortgage-market integration to identify long-run effects on city growth, housing, and household formation using digitized census and loan-level records; and (3) novel quasi-experimental tools that exploit variation in algorithmic recommendations and judge behavior to estimate the causal effect of judicial overrides on decision accuracy and fairness even under selective outcome observation.
GeneralizabilityHealth results derived from Medicare claims primarily reflect older beneficiaries and U.S. contexts and may not generalize to younger populations or non-U.S. health systems., NHSC placement mechanism and underserved-area context are specific to a particular U.S. federal program and may not generalize to other physician recruitment policies., The mortgage-market reform analysis concerns mid-20th century U.S. institutions; the magnitude and channels of effects may differ under modern mortgage markets and regulatory regimes., Bail-algorithm findings depend on the particular algorithm, legal context, and judge population studied and may not map directly to other high-stakes algorithmic settings or jurisdictions.
Claims (14)
| Claim | Direction | Outcome | Confidence & Evidence | Details |
|---|---|---|---|---|
| Physician arrivals increase overall primary care supply. Consumer Welfare | positive | primary care supply |
Reading fidelity
high
Study strength
medium
|
not reported
|
| Half of the non-inpatient healthcare provided by new entrants would have been satisfied by the incumbent primary care physicians, while the other half constitutes an increase in overall utilization. Consumer Welfare | mixed | non-inpatient healthcare utilization (substitution vs additional visits) |
Reading fidelity
high
Study strength
medium
|
Half
|
| The additional utilization is concentrated in more advanced preventive care services, such as advanced testing and imaging. Consumer Welfare | positive | utilization of advanced preventive care services (advanced testing and imaging) |
Reading fidelity
high
Study strength
medium
|
not reported
|
| There is an overall increase in elective inpatient procedures, particularly related to cardiovascular diagnoses, following new primary care physician arrivals. Consumer Welfare | positive | elective inpatient procedures (especially cardiovascular-related) |
Reading fidelity
high
Study strength
medium
|
not reported
|
| The greater healthcare utilization is associated with significant reductions in emergency hospitalizations and mortality. Consumer Welfare | positive | emergency hospitalizations and mortality |
Reading fidelity
high
Study strength
medium
|
not reported
|
| Health improvements persist for at least four years following physician arrival. Consumer Welfare | positive | persistence of health improvements (e.g., reduced mortality/hospitalizations) over time |
Reading fidelity
high
Study strength
medium
|
at least four years
|
| Health improvements are particularly pronounced for beneficiaries with chronic conditions and those living in more disadvantaged and rural locations. Consumer Welfare | positive | heterogeneous health improvements by chronic condition status and area disadvantage/rurality |
Reading fidelity
high
Study strength
medium
|
not reported
|
| Differences in mortgage rates across cities went from nearly 300 basis points to just over 100 in only six years (1933–1940). Fiscal And Macroeconomic | positive | cross-city dispersion in mortgage rates (spread) |
Reading fidelity
high
Study strength
medium
|
from nearly 300 basis points to just over 100
|
| The national mortgage market allowed initially capital-scarce places to grow more than initially capital-abundant ones; cities where mortgage rates declined more as a result of mortgage market integration saw higher growth in homeownership, population, and housing construction in subsequent decades. Fiscal And Macroeconomic | positive | growth in homeownership rates, population, and housing construction |
Reading fidelity
high
Study strength
medium
|
not reported
|
| House prices moved only modestly in 1940--50, suggesting that housing supply responded robustly to higher demand for homes after mortgage market integration. Fiscal And Macroeconomic | null_result | house prices (modest change) |
Reading fidelity
high
Study strength
medium
|
moved only modestly
|
| Women who experienced lower mortgage interest rates during their childbearing years had more children. Social Protection | positive | fertility (number of children / household formation) |
Reading fidelity
high
Study strength
medium
|
not reported
|
| Ninety percent of the judges in our setting underperform the algorithm when they make a discretionary override, with most making override decisions that are no better than random. Decision Quality | negative | judge decision accuracy when overriding algorithm (and randomness of overrides) |
Reading fidelity
high
Study strength
medium
|
90 percent of the judges
|
| The remaining 10 percent of judges outperform the algorithm in terms of both accuracy and fairness when they make a discretionary override. Decision Quality | positive | accuracy and fairness of judge overrides |
Reading fidelity
high
Study strength
medium
|
10 percent of the judges
|
| Suggestive evidence indicates high-performing judges are more likely to use relevant private information and are less likely to overreact to highly salient events compared to low-performing judges. Decision Quality | positive | use of relevant private information and sensitivity to salient events (behavioral correlates of override performance) |
Reading fidelity
medium
Study strength
speculative
|
not reported
|