MerQuR+Statistics Glossary

Statistics Glossary → Instrumental variable and endogeneity

Instrumental variable and endogeneityconcept

An approach for consistently estimating a causal effect when a predictor is correlated with the error term (endogenous), using a variable that affects that predictor but influences the outcome only through it.

Endogeneity arises from omitted variables, reverse causation (the outcome also affects the predictor) or measurement error in the predictor, and it makes the OLS coefficient biased and inconsistent. A valid instrument must meet two conditions: relevance (it must be strongly related to the endogenous predictor) and the exclusion restriction (it must affect the outcome through no path other than the predictor and be uncorrelated with the error). In two-stage least squares (2SLS), the endogenous predictor is first predicted from the instruments, and the outcome is then regressed on that prediction.

Relevance can be checked with the first-stage regression; a first-stage F statistic below about 10 signals weak instruments, which lead to serious bias and misleading confidence intervals. The exclusion restriction cannot be tested directly with the data and has to be defended on theoretical grounds. When there are more instruments than endogenous variables, overidentification tests (Sargan, Hansen J) can be applied. With heterogeneous effects, the IV estimate reflects not the average effect for the whole population but that for the subgroup moved by the instrument (the LATE). The term 'endogenous variable' in structural equation models simply means a variable explained within the model and is a different sense.

An everyday example

The effect of education on earnings is studied, but ability raises both years of schooling and earnings, so education is endogenous. The researcher uses the distance from the person's childhood home to the nearest university as an instrument: being close makes it easier to stay in education, but is not expected to affect earnings directly. If this assumption holds, the effect of education on earnings can be estimated from the differences in schooling created by distance.

In MerQur

Instrumental Variables Regression (2SLS, LIML; weak-instrument, Wu-Hausman and Sargan/Hansen J diagnostics); the GMM/S2SLS estimation of the spatially lagged dependent variable in the SAR model.

Analyses where this concept appears

Instrumental Variables Regression (2SLS)SAR (Spatial Autoregressive Lag)

Related concepts

Two-stage least squares (2SLS) and LIML