Special topics: Sensitivity and Bounds

When you just don’t know, sometimes you should offer all possible answers

In this session we consider how to to gauge the robustness of causal inferences to things we know we don’t know, but suspect might matter. This process is called sensitivity analysis and is not practiced as much as you might hope.

This is distinct from what are referred to as ‘robustness checks’ because rather than investigating alternative functional form of or additional measured variables, it asks: how big do the effects of all unmeasured confounders have to be before the effect I currently have disappears.

We will also look at tools to bound causal effects. Suprisingly often observed data will imply logical constraints on how big or small a causal effect could possibly be. Applying these can often be informative, even without a traditional causal identification.

Readings

Cinelli and Hazlett (2020) ‘Making sense of sensitivity: extending omitted variable bias’

Lecture

Link