Day 3 – The methods in depth
Day 3 slows down everything that day 2 had to take at speed. The sessions follow the order in which the ideas build on one another. The causal graph settles what a model should contain, so it comes first and the models follow. The multiverse is what results when the choices behind those models are enumerated instead of picked, and simulation is how you find out what a design recovers when the answer is known in advance. The day closes on the two skills that the workshop names on a slide but never gives you time to practise: where research materials are archived, and how work is versioned.
| Work through | What you do | Timing |
|---|---|---|
| Causal graphs | Confounders, mediators and colliders, each one fitted as well as described; graphs built from R formulas with dagify(); the EU-frames graph, whose adjustment set turns out to be what two-way fixed effects already absorb; what fixed effects cannot fix; then a graph for a question of your own. |
Session 1, 90 minutes |
| Slides, day 3: the methodological tutorial, then The statistical methods behind the course | Nine of the fourteen axes of the menu, a slide or two each, with what changing each one does to an estimate. Then every model that day 2 fitted, output by output: a first linear model, pooled OLS with cluster-robust errors, two-way fixed effects, random effects and the difference between within and between, and where the degrees of freedom come from. | Session 2, 90 minutes |
| Slides, day 3: the multiverse tutorial and inference, then The multiverse and its critics and Reading a specification curve | Where multiverse thinking came from, the purposes of a multiverse, and the decomposition of the variance by fork. Why hundreds of estimates on one dataset are not hundreds of tests and what a curve-level test would have to do. Then the module and the reading: the three purposes that a multiverse can serve, and the 2025–26 critical debate read against the course grid. | Session 3, 90 minutes |
| Simulation as a method, then slides, day 3: archiving and registering, with Reproducibility repositories and Git and GitHub | Data with a recipe you wrote yourself: plant a parameter and find out whether the model recovers it, build a multiverse over data with no effect in it at all. Then where the materials behind the case came from and what their licences allow to be shared, where a finished reanalysis is deposited now that the Open Science Framework is changing, what registration and forking each mean, and version control from first principles. The data sources module is the reading for the first of those. | Session 4, 90 minutes |
By the end of the day you have a causal graph drawn for a question of your own, every model that day 2 fitted read output by output, a recovery check on a parameter whose true value you set yourself, and an answer to the question of what may be committed to a repository and what may not.