Robust Generalized Linear Model
Also known as: robust GLM, GLM with robust estimation, robust quasi-likelihood model, M-estimator GLM
A Robust Generalized Linear Model fits the standard GLM family — linear, logistic, Poisson, and others — using M-type estimating equations that down-weight outlying or influential observations. The result is coefficient estimates and standard errors that remain stable even when a minority of data points deviate sharply from the assumed distribution.
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 a robust GLM when you have strong subject-matter reason to believe that a small fraction of observations may be outliers, contaminated measurements, or data-entry errors, and you cannot justify simply deleting them. It is especially valuable in biostatistics, epidemiology, and social sciences where extreme responses are common. It suits binary, count, and continuous outcomes where the GLM link structure is theoretically appropriate but outlier contamination is suspected. Do not use it as a substitute for proper data cleaning: if outliers reflect real phenomena they should be modelled explicitly, not down-weighted. Also avoid it when the contamination fraction is large (above ~20%) because the bounded-influence property degrades, or when the sample is very small and the tuning constant choice dominates inference.
Strengths & limitations
- Protects coefficient estimates and standard errors from the disproportionate influence of a small fraction of outliers or leverage points.
- Applies across the entire GLM family — Gaussian, binomial, Poisson, Gamma, and more — with no need to change the link structure.
- The sandwich covariance estimator remains consistent under mild model misspecification, giving reliable inference even with overdispersion.
- The tuning constant c provides a continuous spectrum from full robustness to classical MLE, allowing the analyst to calibrate the efficiency-robustness trade-off.
- Unlike simply excluding outliers, the method retains all observations and makes the down-weighting transparent and reproducible.
- Efficiency relative to classical GLM MLE can be notably lower when the assumed distribution actually holds and there are no outliers.
- The tuning constant c must be chosen, and results can be sensitive to this choice; there is no universally agreed default.
- The bounded-influence property breaks down when the contamination fraction exceeds roughly 10–20%; high breakdown-point alternatives (e.g., S-estimators for GLMs) are needed in more severe contamination scenarios.
- Implementation is less standardised than classical GLMs: package availability and defaults differ across software environments.
- Interpretation of down-weights requires careful communication; non-statistician audiences may misunderstand why some observations are treated differently.
Frequently asked
How does a robust GLM differ from using robust standard errors on a classical GLM?
Robust (sandwich) standard errors on a classical GLM correct the variance estimate but leave the coefficient estimates themselves unchanged and still influenced by outliers. A robust GLM changes the estimating equations so that both the coefficient estimates and their standard errors are protected from outlier influence.
How do I choose the tuning constant c?
A common default is c = 1.345 for Huber's ψ applied to Pearson residuals, giving approximately 95% asymptotic efficiency under a correct Gaussian model. In practice, fit the model at several values of c and examine whether the key coefficients are stable; instability signals influential observations worth investigating.
Can I use a robust GLM for binary outcomes?
Yes. Cantoni and Ronchetti (2001) developed the framework specifically for the full GLM family including the logit link with binary data. The Pearson residual and leverage weights are computed using the binomial variance function, and the ψ function bounds outlier influence on the logistic scale.
What software implements robust GLMs?
In R the robustbase package provides glmrob() covering Gaussian, binomial, Poisson, and Gamma families. The robust package and the rqlm package offer additional options. In Stata, gmm-based robust GLM estimation is possible but less turnkey. Python users can use statsmodels with robust covariance options, though full bounded-influence GLM is more limited.
When should I prefer a robust GLM over simply removing outliers?
When outliers cannot be verified as errors and may represent real but extreme units (e.g., super-spreaders in epidemiology), deletion introduces selection bias. A robust GLM retains all data and makes the reduced influence explicit, whereas ad-hoc deletion is difficult to justify and harder to reproduce.
Sources
- Heritier, S., Cantoni, E., Copt, S., & Victoria-Feser, M.-P. (2009). Robust Methods in Biostatistics. Wiley. ISBN: 978-0470027264
- Cantoni, E., & Ronchetti, E. (2001). Robust inference for generalized linear models. Journal of the American Statistical Association, 96(455), 1022–1030. DOI: 10.1198/016214501753209004 ↗
How to cite this page
ScholarGate. (2026, June 3). Robust Generalized Linear Model. ScholarGate. https://scholargate.app/en/statistics/robust-generalized-linear-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
- Robust Logistic RegressionStatistics↔ compare
- Robust Multiple linear regressionStatistics↔ compare
- Robust Poisson RegressionStatistics↔ compare
- Robust RegressionStatistics↔ compare