Explain the concept of endogeneity and how it can affect the results of causal inference studies.

Official answer

In a regression model, endogeneity arises when an explanatory variable is correlated with the error term. Omitted common causes, some forms of measurement error and simultaneous determination of variables can violate the conditions needed for a causal ordinary least-squares interpretation.

For example, a support team might contact the customers already most likely to cancel. A regression comparing contacted and uncontacted customers can mix the contact effect with the selection process.

Choose an identification strategy that addresses the actual source:

• Random assignment can remove baseline confounding of assignment, subject to the experiment's design and follow-up assumptions.
• An instrumental variable needs relevance and a defensible exclusion/exogeneity argument; correlation with treatment alone is insufficient. Weak or invalid instruments can mislead, and the identified effect may concern a specific population.
• Difference-in-differences needs a credible counterfactual parallel-trends assumption, plus appropriate treatment timing and no-anticipation conditions. Passing a pre-trend test does not prove that assumption.
• Unit fixed effects remove time-invariant additive unit factors, not all time-varying confounding or reverse causality.

State the target effect, causal assumptions and sensitivity checks. Adding controls or choosing a more flexible prediction model does not automatically eliminate endogeneity.

Sources

Cunningham: instrumental-variable assumptions; difference-in-differences identification.