The causal inference literature is focused on model-building for large observational studies with many covariates. In contrast, confirmatory clinical trials are usually analyzed by fairly simple methods, specified in advance, with few covariates. I aim to give practical advice on what causal estimands can be estimated or approximated by simple, prespecified methods. I believe there are only three kinds:

1. The intent-to-treat effect using retrieved dropouts;

2. Composite outcomes incorporating dropout as an outcome in itself;

3. The effect in a subset defined by principal stratification, especially the total direct effect.