Sensitivity Analysis for Confounding
What you'll learn: How to evaluate whether an unmeasured confounder could be strong enough to invalidate your causal conclusions.
The Core Problem
You've controlled for every confounder you can measure, but there's always a nagging question: What if there's an unmeasured confounder we missed? Sensitivity analysis answers: "How strong would that hidden variable need to be to flip our conclusion from causal to non-causal (or vice versa)?"
Think of it like testing the fragility of a bridge: you're asking, "How much extra weight would it take to collapse?" If the answer is "just a feather," your bridge (conclusion) is fragile. If it would take "ten trucks," you can be more confident.
How It Works
Sensitivity analysis quantifies two things about a hypothetical unmeasured confounder:
- How strongly it would need to relate to the treatment (exposure)
- How strongly it would need to relate to the outcome
You then ask: "Is a confounder with both of those strengths plausible in my domain?"
For example, suppose you find that coffee increases productivity. Sensitivity analysis might reveal that an unmeasured confounder would need to be associated with coffee drinking with a correlation of 0.6 and with productivity with a correlation of 0.7 to overturn your result. You'd then ask: "Do I know of any realistic variable that strong that I didn't measure?" (Maybe sleep quality? Personality type?)
Why This Matters
You can't prove the absence of unmeasured confounders, but sensitivity analysis tells you whether reasonable ones would be deal-breakers. It makes your argument transparent: "My conclusion holds unless there's a confounder this strong."
Key Takeaway: Sensitivity analysis quantifies how robust your causal conclusion is to potential unmeasured confounding, helping you assess whether your findings are fragile or resilient.