Marginal standardization and g-computation
What question does this help answer?
It converts a model coefficient into the quantity a clinician can act on: an absolute risk under each condition, and a risk difference or risk ratio between them. An odds ratio rarely answers "how many fewer events"; this does.
How it works in plain language
Fit the model as usual. Then take the whole dataset, set the exposure to "treated" for everyone, and average the predicted risks. Repeat with the exposure set to "untreated" for everyone. The two averages are standardized risks over the actual covariate distribution of the study population, and their difference is the marginal effect. Nothing about the fitting changed; the contrast is constructed afterwards.
What the standardization steps look like
Why this solves the non-collapsibility problem
An odds ratio is non-collapsible: the adjusted odds ratio differs from the unadjusted one even with no confounding, purely because of how odds behave when averaged. Risk differences and risk ratios do not have that property. Standardizing the model's predictions gives a marginal estimate on a collapsible scale, which is both more interpretable and more comparable across studies.
When not to use it
It does not rescue a model whose adjustment set is wrong. Every causal claim made from standardized risks rests on the same no-unmeasured-confounding assumption the coefficients rested on, stated no more loudly for being expressed in absolute terms.
Why Mallard may recommend it
Mallard may recommend marginal standardization when the estimand names a risk difference or risk ratio but the natural model is logistic, which is a common combination: the model is chosen for the outcome type and the estimand is chosen for the reader.
Key assumptions and design requirements
- The model is correctly specified enough for its predictions to be usable, including interactions that matter.
- The adjustment set is sufficient; standardization inherits every assumption the model made.
- Standard errors come from the bootstrap or the delta method, since the model's own do not carry through the averaging.
- The population being standardized to is stated: the whole sample, the treated, or an external population are different targets.
- Predictions are not extrapolated into covariate regions where nobody was actually observed under one of the conditions.
What to watch out for
The absolute risks it produces are only as good as the model and the adjustment set behind them, and presenting them as a causal contrast asserts that adjustment was sufficient. Standard errors need the bootstrap or the delta method, because the ordinary model standard errors do not carry through the averaging step.
Clinical example
Question: What is the absolute reduction in 30-day readmission from a discharge pharmacy review? Analysis: Fit an adjusted logistic model, predict each patient's risk with the review set on and then off, average each, and report the difference with bootstrap confidence intervals as the standardized risk difference.
- Implementation of G-computation on a simulated data set: demonstration of a causal inference technique (opens in a new tab) · Am J Epidemiol · doi 10.1093/aje/kwq472
- Principles of confounder selection (opens in a new tab) · Eur J Epidemiol · doi 10.1007/s10654-019-00494-6