Stage 3 – The models the claim is argued with

The statistical methods module fits every model that the exercise will later offer you, on the data of the case, and reads the outputs line by line, so that by the end each code on the specification menu stands for a choice you understand. It is the longest module on the site, and it repays a slow pass.

Work through	What you do	Timing
The statistical methods behind the course	Sections 1 to 5: why the predictor is measured at the country-year, a first linear model, cluster-robust standard errors, two-way fixed effects, and the difference between a within-country and a between-country comparison.	2 hours
The statistical methods behind the course	Sections 6 to 9: multilevel models and where the degrees of freedom come from, survey weights, the same model stated as a posterior, and a bounded outcome fitted with beta regression.	90 minutes
The statistical methods behind the course	Section 10: putting results from different estimators and different outcome scales on one axis, which is the conversion that the class chart uses.	30 minutes
Weights	The three decisions that go by the name ‘weight’: design weights inside a cell, precision weights across cells and population weights across countries, taken apart on the course panel, with C6HJR’s published coefficient refitted under every combination. Only one of the three changes the estimand.	45 minutes
Missing answers	What a respondent who mentions no frame at all does to a share built from mentions, the rules that the menu offers for them (score zero, drop, split, impute), and how multiple imputation, run with two different packages, compares on this case with the other three rules.	60 minutes
The items behind the scales	The instrument behind every outcome on the menu, taken apart: which items form which scale, how they hang together, whether the answers recover the published four dimensions and what the two composites gain, and what that means for the frame set and the scoring rule you are about to choose. Run the exploratory analysis yourself from the correlation matrix supplied with the module.	45 minutes
Binary outcomes	One item of the battery as the outcome: a logit and a probit side by side, odds ratios against average marginal effects, errors clustered by country-year and a random intercept, and the national unemployment rate on a single item with and without fixed effects. The page applies the generalised linear model at the level beneath every scale. If you built the respondent file, you can rerun every model on the page yourself.	60 minutes

By the end of the stage you can read the output of every model that the exercise will offer you, and you know how estimates from different estimators and different outcome scales are put on one axis.