Other Events and Webinars

REECAP offers irregular workshops and webinars on thematic (such as the future of the CAP) and methodological issues (such as power analysis and open science).

Dates in 2025/2026

REECAP online methods workshop on "Covariate Selection in Randomized Controlled Trials"

April 29th, 2026, 10:30 - 12:00 CET

Presenter: Romain Espinosa, CNRS research fellow at CIRED

Romain will present his paper which develops a new framework for post-randomization model selection in randomized controlled trials. Covariate imbalance frequently arises in realized samples, creating uncertainty about whether to adjust for observed covariates when estimating treatment effects. Romain proposes a new decision rule that minimizes the Mean Squared Error (MSE) of the average treatment effect. The framework introduces a finite-sample statistical test comparing the full specification with any nested model and demonstrates that no finite-sample test can directly compare two nested models. To overcome this limitation, asymptotic MSE-based tests are derived (Wald, Likelihood Ratio, Lagrange Multiplier, and Wild Percentile Bootstrap), allowing comparison between any pair of models. Monte Carlo simulations show that all tests are valid, but the LR and LM tests exhibit higher power, and the Wald and bootstrap tests are overly conservative. Last, he addresses the practical implementation of these tests and the well-known issues that model selection poses for inference. Calibrating the selection rule on pilot data or registered priors resolves the known inferential distortions caused by data-dependent model selection.

Link to paper

Register to attend the webinar

REECAP methods workshop on "Machine Learning and Economic Experiments"

March 20th, 2026, 10:00 - 12:00 CET

Presenters: Alessandro Varacca and Hugo Storm

In this workshop, an introduction to and two applications of machine learning for economic experiments will be discussed followed by a question and answer session.

Hugo will provide an introduction to probabilistic machine learning and a workflow for experimental design, centred on the data-generating process that uses these methods. Key advantages of these approaches are that they allow for testing the entire experiment design and the empirical approach prior to data collection. Additionally, they offer the opportunity to jointly analyse separate parts of the experiment.

Alessandro will discuss how machine learning methods can be leveraged to estimate treatment effect heterogeneity in experimental studies. While traditional methods typically exploit rigid treatment-covariate interaction terms, this approach can yield unstable results and model misspecification, particularly when the number of interactions grows. Flexible functional forms estimated through machine learning can easily fix these issues and provide a coherent framework to both discover and estimate heterogeneous treatment effects.