Interactive research companion Chile · 2005–2024

Private School Competition
and Student Achievement in Chile

The evidence changes with the grade. Explore the estimates, the choices behind them, and the limits of what they tell us.

Explore the evidence

One paper. Different views of the evidence.

Reported manuscript results

The main explanatory variable is the private share of schooling: the fraction of students enrolled in private schools. The estimates show how average test scores across public and private schools combined vary with that share, after accounting for each model’s controls. The outcome is the average for the schooling market, not the gap between private- and public-school scores.

Why cream skimming cannot by itself explain the results: Moving higher-scoring students from public to private schools changes which sector they attend but does not change their combined average score. Pure sorting between sectors therefore cannot mechanically produce the gains shown here. (This argument assumes the within commune student population is relatively constant).

The paper uses several models to examine this relationship. Click the tabs below to compare panel estimates, results by grade, cohort value added, and cumulative exposure, or explore different definitions of the schooling market. Within each tab, choose a specification or click a chart row to see its estimate and explanation.

Panel estimates

Private-share coefficient · SIMCE points

Estimate ± 1 standard error
10.52Main panel estimate · standard error 2.9

Lines show one standard error, not a 95% confidence interval. Select a row to inspect its result.

Panel: commune fixed effects
10.52coefficient p < 0.01
Standard error
2.9
Observations
19,221
Clusters
336

Main panel specification, with commune and grade × subject × year fixed effects. Weighted by tested students; errors clustered by commune.

View source table ↗
What to take from this

The main pooled estimate is 10.52. With commune × grade fixed effects it falls to 5.42. The grade-specific estimates do not depend on that fixed-effect choice.

Run the evidence

Ask the paper. Inspect the calculation.

Paper2Agent

Ask a question that goes beyond the tables. The agent can change the cohort sample, compare communes with different initial private enrollment shares, check which communes influence the estimate, and examine alternative baseline-score adjustments. New calculations are exploratory and come with uncertainty, sample counts, and a downloadable record of the choices made.

A conversation with the evidence

Checking availability…

Use a model that supports tool calling. Browse OpenRouter models ↗

Use your own OpenRouter key or OpenAI key. Usage is billed to your provider account; no author access code is needed. Your key is sent over HTTPS through this site’s server to your selected provider for each request. This app does not save it in browser storage or a database. Clearing the key or leaving this page removes it from the form. Your question, recent conversation, manuscript excerpts and calculation results are sent to the provider; its retention policies apply.

Built with Paper2Agent’s conversion and verification workflow. Supported Python calculations were checked against independent Stata runs. This sandbox uses the paper’s matched commune-cohort data; panel and exposure results in the explorer remain reported estimates. Exploring specifications does not by itself establish causality. About Paper2Agent ↗

Read the evidence closely

Questions worth asking

Answers drawn from the manuscript.
This first edition uses curated explanations.

What does “private share” measure?

In the baseline, private share refers to subsidized-private enrollment in the public-plus-subsidized-private school market. Fully private schools and communes dominated by fully private schooling are excluded. Outcomes describe students attending schools in the commune, rather than necessarily living there. The analysis uses reported market definitions and tested-student weights.

Does this overturn Hsieh and Urquiola’s finding?

It refines the comparison by grade. The grade-4 long-difference estimate is 4.16 (SE 4.46), small and imprecise. Grade-10 long differences and cohort value added are stronger. Hsieh and Urquiola’s test-score evidence concerned grade 4.

Compare the grades ↑
Are the same students followed for six years?

No individual student panel is used. The design matches a commune’s grade-4 mean in year t to its grade-10 mean in year t + 6, separately by subject. Migration, cross-commune attendance, grade progression, and changing test participation can alter the observed cohort. The paper examines cohort fidelity and alternative market definitions.

Does the paper establish a causal effect?

The designs establish conditional relationships under different assumptions. The cohort analysis controls for prior achievement, but changes after grade 4 that affect both private enrollment and achievement can still confound the estimate. Sorting within a fixed tested population cannot change its mean; changes in who belongs to that population can. The robustness results help assess these explanations without eliminating all of them.

Do longer exposure windows identify when the gains occur?

Exposure estimates are consistent with cumulative schooling effects, but the measures correlate above 0.94. The comparisons do not sharply identify the timing of gains. Enrollment is observed during the testing gap, so exposure can be constructed even though the missing test waves cannot be restored.

Compare exposure definitions ↑
Sources & reproducibility

Every number has a source.

The explorer displays reported estimates from the manuscript. The Ask & run section computes cohort replications and exploratory comparisons on demand. Each new calculation includes its settings, sample counts, uncertainty, and source fingerprints.