Juan C. Yamin

How much should a small ramp-up change the experiment that follows?

When and How to Pilot — solo-authored · arXiv:2607.16982

Digital experiments often run in two waves: a small ramp-up precedes the full experiment, and the ramp-up data can inform the main-wave traffic split. Splitting evenly ignores what the ramp-up learned about which arm is noisier; plugging noisy variance estimates into the Neyman formula can overreact badly — with a small ramp-up it can assign a live arm almost no traffic. My rule is a closed-form traffic split that moves from an even split toward the optimal allocation only as far as the ramp-up's evidence has earned, and reports a finite-sample certificate bounding what the move can cost.

• Covers A/B/n tests with shared control, stratified splits across user segments or markets, and ramp-ups that observe a fast proxy while the primary metric arrives with delay.
• Answers the sizing question too: with an experimentation budget of B observations, the pilot should be roughly B^2/3; for a 100,000-observation budget, that is about 2,000 pilot observations.
• In calibrations to Upworthy headline tests and Criteo's 13.9-million-user ad-experiment archive, the rule avoids feasible Neyman's severe small-ramp-up losses — and in the Criteo calibration captures about 90% of the attainable precision gain by 4,000-user ramp-ups.

Targeting scarce resources when information is imperfect

Poverty Targeting with Imperfect Information — solo-authored · arXiv:2506.18188

The paper studies a broader prioritization problem: how to allocate a fixed budget when decisions must be based on estimates rather than direct observations. In the antipoverty-targeting application, I show that the standard plug-in rule, which treats estimates as truth, is systematically improvable: another feasible rule can do at least as well in every case and better in some cases. I then build a nonparametric empirical Bayes rule that first learns how much to trust each estimate — using the same data, the same prediction model, and the same budget.

• In simulations on household surveys from nine African countries, the rule reaches nearly 80% more of the target population than plug-in ranking (45.6 vs. 25.5 per 1,000) and outperforms plug-in OLS and machine-learning benchmarks.

Applied causal research from question to data to interpretation

Political economy, development, administrative data

Alongside my methods work, I have carried traditional applied-economics projects from institutional question to data construction, identification, robustness, and policy interpretation. In coauthored research conditionally accepted at the American Political Science Review, I study whether partisan alignment across levels of government facilitates corruption in Colombia: the project combines legal and institutional analysis, administrative education records, close-election regression discontinuity, validation and robustness checks, and evidence on ghost enrollment, discretionary hiring, outsourcing, and electoral fraud risk.

Earlier at Universidad de los Andes, I coordinated multi-year data collection, digitization, replication, and empirical analysis across development and political-economy projects. That work included RA contributions to Acemoglu, Fergusson, and Johnson's Review of Economic Studies paper on population growth and conflict, as well as research assistance on Acemoglu and Robinson's Non-Modernization: Power-Culture Trajectories and the Dynamics of Political Institutions (Annual Review of Political Science) and Moscona, Nunn, and Robinson's Segmentary Lineage Organization and Conflict in Sub-Saharan Africa (Econometrica).