MerQuR+Statistics Glossary

Statistics Glossary → Robust (sandwich) standard errors (HC0-HC3)

Robust (sandwich) standard errors (HC0-HC3)concept

Coefficient standard errors that remain consistent even when the error variance is not constant or the model's variance structure is misspecified.

Classical standard errors assume the model's variance assumption is correct (e.g. in regression, equal error variance for all observations). The sandwich estimator takes its name from the shape of its covariance formula: (X'X)⁻¹ X'ΩX (X'X)⁻¹, where the 'meat' in the middle is estimated from the observations' own squared residuals instead of an assumed constant variance. White's original form is HC0; HC1 adds a degrees-of-freedom correction (n/(n - p)); HC2 and HC3 inflate the residuals according to their leverage (e²/(1 - h) and e²/(1 - h)²). HC3 is recommended in small samples.

Robust standard errors do not change the coefficient estimates, only their uncertainty; if the model is wrong about the mean structure (e.g. an omitted variable), they do not fix that. When observations are dependent in clusters (pupils in the same school, repeated measurements of the same person), heteroscedasticity-robust HC types are not enough and cluster-robust standard errors are needed. GEE and the Cox model use the sandwich estimator with the same logic against within-cluster dependence. With very small samples or few clusters, sandwich standard errors can come out too small.

An everyday example

A researcher studies the effect of household income on holiday spending. Spending by low-income households is similar, while that of high-income households is very spread out. The classical standard error for the income coefficient is 0.010, while the HC3 standard error is 0.016; the confidence interval widens, but the coefficient itself (0.045) does not change.

In MerQur

Available as 'Standard error type' (Classic, HC0, HC1, HC2, HC3) in multiple linear regression, the same option (Classic, HC0, HC1) and 'Cluster variable' in logistic regression, 'Robust SE (HC)' in two-way ANOVA, Robust (sandwich) under 'Covariance / scale' in Poisson / negative binomial regression, 'Standard error' (Robust, Model-based, Mancl-DeRouen) in GEE, and the clustered options in panel data regression. In confirmatory factor analysis and structural equation modeling MLR gives Huber-White sandwich SEs; SAR offers the GMM-S2SLS White robust SE, the Cox, time-dependent Cox, Survey-PHREG and cluster-robust Cox models use a sandwich variance, GARCH gives robust standard errors, and instrumental variables, quantile and Tobit regression, item response theory and conditional logit also offer sandwich or robust standard errors.

Analyses where this concept appears

Two-Way ANOVAMultiple Linear RegressionLogistic RegressionPoisson / Negative Binomial Regression (Count Regression)Generalized Estimating Equations (GEE)Panel Data RegressionInstrumental Variables Regression (2SLS)Propensity Score AnalysisDifference-in-Differences (DiD)Confirmatory Factor Analysis (CFA)Structural Equation Modeling (SEM)Item Response Theory (IRT / Rasch)Quantile RegressionTobit Regression (Censored Regression)Conditional Logit (Choice Model)SAR (Spatial Autoregressive Lag)Cox Regression (Proportional Hazards Model)Time-dependent Cox RegressionSurvey-PHREG (Weighted Cox)Cluster-Robust Cox (Marginal and Frailty Models)GARCH (Volatility Model)

Related concepts

Cluster-robust (clustered) standard errorsHeteroscedasticityStandard error