Robust Probit Model
Robust Probit Regression Model · Also known as: probit with robust standard errors, sandwich-SE probit, heteroscedasticity-robust probit, M-estimation probit
The Robust Probit Model estimates the probability of a binary outcome using the probit link function while protecting inference from misspecification of the error distribution or heteroscedasticity. Coefficients are obtained via maximum likelihood; standard errors are then replaced by the sandwich (Huber-White) estimator, which remains consistent even when the assumed error variance is incorrect.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use robust probit when the outcome is binary and you have concerns about heteroscedasticity or mild non-normality in the error structure — for example, when residual plots or a Breusch-Pagan analog test flags non-constant variance. It is especially valuable in cross-sectional social science, health, and economic datasets where textbook assumptions rarely hold perfectly. It is preferable to standard probit whenever sample size is sufficient for sandwich asymptotics (roughly n > 50). Do not use it as a substitute for correctly modelling clustered data — if observations share a group, use cluster-robust standard errors or a multilevel probit instead. It is also not appropriate when the latent-variable assumption itself is severely violated; consider semiparametric binary models in that case.
Strengths & limitations
- Coefficient estimates remain consistent under mild misspecification of the error distribution.
- Sandwich standard errors are valid under heteroscedasticity without requiring re-estimation of the model.
- Widely supported in statistical software (Stata, R glm with vcovHC, Python statsmodels) with minimal extra effort.
- Average marginal effects computed from robust probit are comparable to those from logit, facilitating cross-model checks.
- Asymptotically valid inference without assuming a fully correctly specified parametric error distribution.
- In small samples (n < 100) sandwich standard errors can be downward biased; small-sample corrections (HC3) are advisable.
- Robust standard errors protect inference but do not correct coefficient bias if the conditional mean itself is misspecified.
- Unlike fully Bayesian or M-estimation approaches, the sandwich correction does not change predictions or model fit statistics.
- Interpretation still requires marginal effects rather than raw coefficients, adding a post-estimation step.
Frequently asked
How does robust probit differ from standard probit?
The point estimates (coefficients) are identical — both use maximum likelihood. The difference lies in the standard errors: robust probit uses the sandwich (Huber-White) estimator, which remains valid under heteroscedasticity or mild distributional misspecification, whereas standard probit relies on the expected information matrix that assumes a correctly specified model.
Should I use robust probit or robust logit?
For most applications the choice between probit and logit is minor — both give similar marginal effects and sandwich corrections apply equally. Probit is preferred when you have a theoretical reason to assume normally distributed latent errors; logit has a closed-form odds-ratio interpretation and is slightly faster to estimate. Robust versions of both protect inference similarly.
How do I interpret the coefficients of a robust probit?
Probit coefficients measure the change in the probit index (Xβ) per unit increase in a predictor. They are not directly interpretable as probability changes. Compute average marginal effects — the average across the sample of φ(Xβ̂)·β̂_j — to get the estimated change in the probability of y=1 per unit change in xj.
When should I use cluster-robust rather than heteroscedasticity-robust standard errors?
Use cluster-robust standard errors when observations are grouped (e.g., students within schools, patients within hospitals) and errors within a cluster may be correlated. Heteroscedasticity-robust (HC) standard errors only correct for non-constant variance across individual observations but assume independence. Ignoring clustering when it exists substantially understates standard errors.
What sample size is needed for sandwich standard errors to be reliable?
Sandwich estimators are asymptotic; in practice n ≥ 100 is generally recommended for HC1/HC2 corrections, and n ≥ 50 with HC3 (which applies a leverage-based correction). Below n = 50 consider bootstrap confidence intervals as an alternative.
Sources
- Wooldridge, J. M. (2010). Econometric Analysis of Cross Section and Panel Data (2nd ed.). MIT Press. ISBN: 978-0262232586
- White, H. (1982). Maximum Likelihood Estimation of Misspecified Models. Econometrica, 50(1), 1–25. DOI: 10.2307/1912526 ↗
How to cite this page
ScholarGate. (2026, June 3). Robust Probit Regression Model. ScholarGate. https://scholargate.app/en/statistics/robust-probit-model
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Generalized Linear ModelStatistics↔ compare
- Logistic RegressionResearch Statistics↔ compare
- Robust Logistic RegressionStatistics↔ compare
- Robust RegressionStatistics↔ compare