Statistics Glossary → Two-stage least squares (2SLS) and LIML
Two-stage least squares (2SLS) and LIMLconcept
Two methods that estimate the effect of an explanatory variable consistently, with the help of instrumental variables, when that variable is correlated with the error term (endogenous).
Endogeneity arises from reverse causation, an omitted confounder or measurement error, and makes ordinary least squares biased. An instrument is a variable that is related to the endogenous variable (relevance) but affects the outcome only through that variable (exclusion). In the first stage of 2SLS the endogenous variable is regressed on the instruments and exogenous variables, and its fitted values are kept. In the second stage the outcome is modelled with these fitted values in place of the endogenous variable. For correct standard errors, the method's own formulas should be used rather than running the second stage by hand with OLS.
If the instruments are weak, explaining the endogenous variable poorly, the 2SLS estimate is biased toward OLS and its confidence intervals are unreliable. A common rule is a first-stage F statistic above 10. LIML (limited-information maximum likelihood) is a k-class estimator with κ ≥ 1 that is less biased than 2SLS when instruments are weak and numerous; in an exactly identified model (as many instruments as endogenous variables) the two give the same result. The exclusion assumption cannot be tested directly with the data; overidentification tests (Sargan, Hansen) provide only a partial check.
An everyday example
A researcher wants to measure the effect of years of education on earnings, but ability affects both, so a simple regression is biased. Whether a university was located near the person's childhood home is used as an instrument: proximity increases schooling but is assumed to have no direct effect on earnings. 2SLS first predicts education from proximity, then uses this prediction in the earnings model.
In MerQur
In instrumental variables regression, the 'Estimator' option chooses between 2SLS (two-stage least squares) and LIML (limited-information maximum likelihood). The SAR model can also be estimated with GMM/S2SLS (spatial two-stage least squares) besides maximum likelihood.
Analyses where this concept appears
Instrumental Variables Regression (2SLS)SAR (Spatial Autoregressive Lag)
Related concepts
Instrumental variable and endogeneitySeemingly unrelated regression (SUR) and 3SLS