What is Marginal Structural Model?
Marginal Structural Model is a model for the marginal distribution or mean of counterfactual outcomes under specified treatment interventions, commonly estimated after inverse-probability weighting removes measured treatment-confounder associations.
Quick Facts
| Specification | Official Specification |
|---|
How It Works
Specify the intervention and causal contrast first
An MSM can compare static regimes, dynamic rules, cumulative exposure, or treatment histories, but each coefficient must map to a declared counterfactual quantity. The analysis should state the target population, treatment versions, follow-up horizon, outcome scale, and whether the contrast is per-protocol or assignment-like. A compact regression formula is not a substitute for that protocol.
Weights construct a pseudo-population, not new evidence
For longitudinal treatment, a stabilized weight multiplies time-specific numerator probabilities divided by treatment probabilities conditional on measured history; censoring weights may be multiplied in as well. Under Consistency, sequential Exchangeability, Positivity, and correct nuisance models, the weighted data can identify marginal strategy effects. Robins and colleagues' tutorial explains MSM construction and interpretation.
Diagnose the nuisance models and the causal model separately
Report weight mean, tails, truncation, effective sample size, covariate balance at each decision, treatment-history support, and sensitivity to propensity and censoring models. Then assess whether the MSM's functional form represents the intended strategy contrast. Stable-looking weights cannot rescue a misspecified causal model, and a flexible MSM cannot repair unmeasured confounding.
Key Characteristics
- Models marginal counterfactual outcomes under declared interventions
- Separates the causal outcome model from its weighting estimator
- Can represent static, sustained, or dynamic treatment strategies
- Handles treatment-confounder feedback under sequential assumptions
- Often combines treatment and censoring weights
- Requires overlap, balance, effective-sample-size, and specification checks
Common Use Cases
- Comparing always-treated and never-treated longitudinal strategies
- Estimating cumulative treatment effects with time-varying confounders
- Analyzing per-protocol effects in a Target Trial Emulation
- Combining treatment and censoring adjustment in observational follow-up
- Reporting population-average effects rather than conditional coefficients
Example
Loading code...Frequently Asked Questions
Is a Marginal Structural Model the same as inverse-probability weighting?
No. The MSM specifies how counterfactual outcomes vary across interventions. Inverse-probability weighting is one way to estimate its parameters by balancing measured treatment histories. An analysis can use weights without a meaningful MSM, and MSM parameters can be estimated by other valid procedures.
Why is the model called marginal and structural?
Marginal means the modeled intervention outcome is averaged over a target population rather than conditioned on a fixed covariate profile. Structural means the parameters refer to counterfactual outcomes under interventions, provided the causal identification assumptions hold.
When is an MSM useful for Time-Varying Confounding?
It is useful when covariates affect later treatment and outcome while also being affected by prior treatment. Sequential weights can account for observed treatment-confounder feedback without directly conditioning the final outcome model on those post-treatment covariates.
What assumptions identify an MSM effect?
Typical requirements include well-defined treatment strategies, Consistency, no interference or an explicit interference model, sequential Exchangeability given measured history, Positivity for each eligible decision, correct treatment and censoring models, and a suitable MSM functional form.
Which MSM diagnostics should be reported?
Report treatment-history support, propensity distributions, stabilized-weight mean and tails, truncation rules, effective sample size, time-specific covariate balance, censoring behavior, alternative nuisance models, and sensitivity to unmeasured confounding and MSM specification.