Course contentsShow
← Data Science
Lesson 1465 of 2,14536. Causal Inference BasicsFree lesson

What Makes a Valid Instrument

Learning the relevance and exclusion restriction criteria that instruments must satisfy for causal identification.

What Makes a Valid Instrument

You'll learn the two critical conditions an instrumental variable must satisfy to help uncover causal effects when experiments aren't possible.

The Core Problem Recap

In the previous lesson, you learned that endogeneity (when your treatment variable is correlated with the error term) breaks standard regression for causal inference. Instrumental variables offer a clever workaround—but only if they meet strict criteria.

The Two Sacred Requirements

A valid instrument must satisfy two non-negotiable conditions:

1. Relevance (First Stage)

The instrument must be strongly correlated with the treatment variable. Think of it like a "nudge"—the instrument must actually push people toward or away from the treatment. If your instrument barely affects whether someone gets treated, it's too weak to be useful.

Analogy: Imagine trying to predict coffee consumption using "distance to nearest coffee shop." If distance barely affects whether people drink coffee, it's not relevant enough.

2. Exclusion Restriction (Exogeneity)

The instrument must affect the outcome only through its effect on the treatment, and it cannot be correlated with other factors (the error term) that influence the outcome. This is the crucial, untestable assumption.

Analogy: Distance to a coffee shop should only affect productivity through coffee consumption—not because people near coffee shops also live in trendy neighborhoods with better job opportunities.

Why This Matters

If either condition fails, your causal estimates will be biased. Relevance can be tested statistically. Exclusion restriction requires strong theoretical justification and domain knowledge—you can't prove it with data alone, only argue it convincingly.

Key Takeaway: A valid instrument must be (1) strongly relevant to treatment and (2) affect the outcome exclusively through treatment, with no backdoor paths to the outcome.