Fabián Duarte

Universal coverage systems remove monetary barriers to care, but they leave non-monetary barriers in place. This paper examines how much of the burden of such a barrier falls on those who live farthest from care, in a system where monetary barriers have been substantially removed by institutional design. I exploit staggered lockdowns in Chile, which raised the cost of reaching a healthcare facility without closing it, and find that distance, not price, determines who bears the loss. Copayment levels, which the system varies from zero to 20% across income tiers, show no detectable effect within any distance category, while the loss of guaranteed care deepens monotonically with distance to hospital, from 5.7% in the nearest quartile of municipalities to 13.3% in the farthest. The disruption concentrates at the screening stage, where asymptomatic patients must initiate contact, while patients already in treatment show substantially smaller declines. Screening-dependent cancer mortality rises 8.8% at an eighteen-month lag in non-metropolitan regions, while mortality from non-screening cancers shows no effect at any lag, consistent with the causal chain from missed screening to excess deaths. These findings suggest that equalizing financial access, without addressing geographic access, leaves the binding constraint on equitable care unresolved.

This paper studies what happens to a private health insurance market when a regulatory crisis undermines institutional credibility. Using administrative data on 62 million individual-quarter observations from Chile's dual public-private system (2016 to 2025), I study how a 2022 Supreme Court ruling triggered market contraction through both exit and entry margins. Two sequential shocks, COVID-19 affecting access and the ruling affecting price, separately identify two channels of adverse selection. Sicker enrollees are less price-sensitive, and working-age sick enrollees value private coverage more when public-system access deteriorates. New entry into the private system collapsed by 75% peak-to-trough and adds roughly 30% to the net enrollment response to premium increases, a margin that exit-only models miss. The crisis created adverse selection where mild favorable selection had prevailed, while switching costs, a welfare-reducing friction in stable markets, acted as a stabilizing force by retaining sick enrollees.

Are the uninsured treated at more inopportune times than the insured? Clinically appropriate care can follow one of two pathways, a planned daytime course of treatment, or an unplanned off-hours course driven by acute worsening of a health condition. Health insurance facilitates entry into the first pathway by clearing financial and administrative barriers to scheduled appointments. We test this hypothesis using natality data on all US births from 2011 to 2024, roughly 48 million births. In unadjusted analyses, uninsured mothers are nearly 20% more likely to deliver during early-morning hours, between midnight and 6:59 AM, than insured mothers. This difference persists after adjustment for mothers' and fathers' demographics, but is almost fully attenuated after adjustment for procedure and setting variables, consistent with a scheduling-access interpretation. These findings suggest that health insurance coverage shapes not only access to health care, but also the timing of that care.

We estimate the price elasticity of demand for private health insurance in Chile. Private health insurance plans in Chile adjust the base premium according to a policyholder's age and sex, in turn generating price discontinuities at certain ages. Using a multigroup multiperiod difference in difference method, we estimate the effect of these price increases on plan exit. We find an overall price elasticity of approximately −0.078. In addition, we find that high-income individuals are more price sensitive than low-income individuals and men less price sensitive than women.

In most contributory social insurance systems, protection against illness is attached to the formal employment relationship, so a worker who loses the job also loses the insurance. This paper asks when that transfer of risk happens. Using linked administrative records from Chile's public health insurer, we follow a matched sample of 216,187 individuals at monthly frequency (2016 to 2019) and estimate the effects of hospitalization on healthcare use, sickness absence, formal employment, and taxable income. Treatment cohorts are defined by the calendar month of admission in a staggered difference-in-differences design with cohort-specific matching, allowing for the health deterioration that precedes admission. Because sickness benefit payments are recorded in taxable income, the data show the insurance operating, and income does not fall during sick leave. The losses appear once it stops. Averaged over the three post-admission years, formal employment is 5.8 percentage points lower with no return to baseline, and income is 8.6% of the baseline mean lower per month, accumulating to 3.4 months of earnings by the end of the window. Earnings conditional on employment show no decline, so the loss runs entirely through exit from formal work. The binding gap in employment-linked social insurance is not income replacement during illness but employment continuity after it.

We examine whether transitions into and out of overweight or obesity are associated with subsequent academic outcomes and whether these associations differ between childhood and adolescence. Using linked administrative data, we follow Chilean students measured in first grade in 2014, fifth grade in 2018, and ninth grade in 2022. For each interval, we compare students who changed weight status with peers who shared the same status at the beginning of the interval, controlling for prior academic outcomes. Entering overweight or obesity is associated with lower attendance for both sexes during both intervals and with a lower within-school percentile rank among girls, 1.08 percentile points between first and fifth grade and 1.24 points between fifth and ninth grade. Severity appears more relevant than timing, since transitions into obesity are associated with larger declines than transitions into overweight. Associations are consistently smaller for externally scored standardized tests than for teacher-assigned grades, a pattern consistent with, but not sufficient to establish, an evaluative channel. Because the adolescent interval includes the COVID-19 school closures, developmental stage and pandemic exposure cannot be separately identified, so the findings should be interpreted as associations rather than causal effects.

Private health insurers in Chile sold 4,374 individual plans in 2016 across six firms. This paper asks what determines which plan a household buys when the observable terms of the contract, the premium and the out-of-pocket cost of care, are known for every plan. We combine the universe of plans, their coverage rules, and the risk-rated premium each household would face with the distribution of new enrollees across insurers and household types, and compare three demand models that see the same menu, an expected-utility model, a nested logit with insurer-specific intercepts, and a large language model instructed to choose as the household would. Contract terms alone do not explain choice. On the full menu, only the model with six insurer intercepts reproduces enrollment shares, and its coefficient on coverage is zero once the intercepts are in. The intercepts, however, are not preferences. Restricting each household to a consideration set of five plans, drawn from those sold in its region with probability that falls with the premium relative to income and rises with the size of the insurer's catalogue, allows a logit with no insurer intercepts to match shares within 3.3 percentage points. Given the same five plans, the language model, with no fitted parameters, comes within 6.3 points and reproduces the age composition of each insurer's enrollees within 1.8 years. Households appear to choose well among what they see, and what they see is governed by how many plans each firm puts in the market and by what they can afford.

We use large language models to classify occupational task exposure to generative AI for 95,434 employed individuals in the Chilean national household survey (CASEN), producing 477,170 task-person classifications. Each task is scored on three dimensions: exposure, displacement of baseline human work, and augmentation. Because the classification is applied to a nationally representative survey rather than to occupational codes alone, exposure can be linked to individual earnings, informality, job quality and poverty status, allowing the distributional incidence of AI exposure to be characterized across the labor force rather than across occupations.