Heterogeneous Treatment Effect Marginal Structural Model (HTE-MSM)
Heterogeneous Treatment Effect Marginal Structural Model · Also known as: HTE-MSM, heterogeneous MSM, subgroup MSM, effect-modified marginal structural model
The Heterogeneous Treatment Effect Marginal Structural Model extends the classic MSM framework of Robins, Hernan, and Brumback to estimate how treatment effects vary across subgroups or individual-level moderators. By weighting observations with inverse probability of treatment weights (IPTW) and interacting the treatment with effect modifiers in the weighted outcome model, the approach produces subgroup-specific or continuous causal effect estimates from observational data.
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 HTE-MSM when you have observational data, want to estimate causal treatment effects, and have substantive reason to believe the effect differs across measured subgroups or moderators. It is especially appropriate with time-varying treatments or confounders, or when structural equation approaches are needed. The method requires a rich set of measured confounders (no unmeasured confounding assumption), adequate overlap in propensity scores across moderator levels, and a sufficiently large sample to estimate both the propensity model and the interaction-enriched outcome model with precision. Do not use it when key confounders are unmeasured, when propensity score overlap fails for certain subgroups, or when sample sizes are too small to support reliable estimation of interaction terms.
Strengths & limitations
- Handles time-varying treatments and time-varying confounders that are affected by prior treatment — a setting where conventional regression adjustment fails.
- Recovers subgroup-specific causal effects from observational data by combining IPTW re-weighting with interaction modelling.
- Parameters have explicit causal (counterfactual) interpretations under the identifying assumptions, unlike naively stratified regressions.
- Stabilised weights reduce variance relative to standard IPTW, making the method more efficient than unweighted subgroup analyses.
- Compatible with any parametric or semi-parametric outcome model, including logistic, linear, or survival models.
- Relies on the no-unmeasured-confounding (exchangeability) assumption, which cannot be tested from the data and may fail in practice.
- Positivity assumption requires that every unit has a non-zero probability of receiving each treatment level within every subgroup; violations cause extreme weights and instability.
- Interaction terms for heterogeneous effects require substantially larger samples than main-effect-only MSMs; power drops quickly when moderators are continuous or numerous.
- Propensity model misspecification biases both the main causal effect and the heterogeneity estimates simultaneously.
- The number of separately identifiable subgroup effects is limited by sample size and overlap; many interaction terms risk overfitting.
Frequently asked
How does HTE-MSM differ from a standard MSM?
A standard MSM estimates the population-average causal effect of treatment. HTE-MSM adds interaction terms between treatment and pre-specified effect modifiers to the weighted outcome model, allowing the causal effect to vary across subgroups or along a continuous moderator. The IPTW weighting step is identical; the difference is in the outcome model specification.
Why not just stratify by subgroup and run a separate MSM in each stratum?
Stratification can work when subgroups are large enough, but it loses statistical efficiency, makes cross-stratum comparisons informal, and may fail the positivity assumption within small strata. The interaction-based HTE-MSM borrows strength across the full sample while still estimating subgroup-specific effects through the interaction coefficient.
What sample size is needed?
Much larger than for a main-effect MSM, because estimating interaction terms is inherently power-hungry. A rough heuristic from clinical trials literature is at least 10–15 events or outcome observations per interaction term, but simulation studies suggest even more may be needed in weighted observational settings. Formal power analysis or simulation before data collection is advisable.
Can this be combined with doubly robust methods?
Yes. Augmented IPTW (AIPW) or targeted maximum likelihood estimation (TMLE) can be used in place of pure IPTW to obtain doubly robust heterogeneous treatment effect estimates that remain consistent if either the propensity or the outcome model — but not necessarily both — is correctly specified.
Does HTE-MSM handle multiple time points?
Yes; in longitudinal settings the product-of-ratios IPTW formula extends naturally across time periods, and the interaction with the moderator is included in the weighted pooled-over-time outcome model. This is the original motivation for the MSM framework — handling time-varying confounding that standard methods cannot address.
Sources
- Robins, J. M., Hernan, M. A., & Brumback, B. (2000). Marginal structural models and causal inference in epidemiology. Epidemiology, 11(5), 550-560. DOI: 10.1097/00001648-200009000-00011 ↗
- Hernan, M. A., & Robins, J. M. (2020). Causal Inference: What If. Chapman & Hall/CRC. link ↗
How to cite this page
ScholarGate. (2026, June 3). Heterogeneous Treatment Effect Marginal Structural Model. ScholarGate. https://scholargate.app/en/causal-inference/heterogeneous-treatment-effect-marginal-structural-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.
- Doubly Robust EstimationCausal inference↔ compare
- Heterogeneous Treatment Effect Difference-in-DifferencesCausal inference↔ compare
- Inverse Probability WeightingCausal inference↔ compare
- Marginal Structural ModelCausal inference↔ compare
- Propensity Score WeightingCausal inference↔ compare