Questions: Marginal Structural Models for Longitudinal Data

Question 1

In an HIV cohort study, CD4 count is measured at every visit. CD4 predicts both ART receipt (doctors prescribe when CD4 is low) and mortality (low CD4 signals disease progression). A researcher includes CD4 as a covariate in a standard Cox regression to estimate the causal effect of ART on survival. What is the fundamental problem with this approach?

• A. Including CD4 is fine as long as the model uses robust standard errors to handle correlation
• B. CD4 is affected by prior ART, so conditioning on it blocks part of the causal pathway — but omitting it leaves confounding; standard regression cannot resolve this dilemma
• C. CD4 is not a true confounder because it is an intermediate variable, so it should always be excluded
• D. The problem is measurement error in CD4, not the causal structure — better measurement would fix the bias

Answer: (B)

This is the defining challenge of time-varying confounding affected by prior exposure. CD4 is simultaneously a confounder (it predicts treatment and outcome) and a mediator (prior ART raises CD4). Conditioning on it blocks part of the causal effect of ART on mortality; omitting it leaves the confounding unaddressed. Standard regression, regardless of complexity or model choice, cannot handle this structure. The correct diagnosis is option B — the problem is causal, not statistical.

Question 2

In a marginal structural model, inverse probability weighting creates a 'pseudo-population.' What is the key property of this pseudo-population that enables causal inference?

• A. Every patient receives the same weight, eliminating all individual variation in the data
• B. All patients are assigned the most common treatment, making groups directly comparable
• C. Treatment assignment is no longer associated with measured confounders — it is as if treatment were randomly assigned
• D. Patients with extreme covariate values are excluded to reduce variance

Answer: (C)

The IPT weights rebalance the data so that, within the weighted pseudo-population, confounders no longer predict treatment. A person with low CD4 who received ART (the expected treatment) gets a low weight; a person with high CD4 who received ART (unexpected) gets a high weight. The result is a dataset where CD4 no longer distinguishes who got ART — mimicking the covariate balance you would expect from randomization. You can then fit a standard regression model in this weighted dataset and obtain a consistent causal estimate.

Question 3

A marginal structural model estimates the 'marginal' causal effect of a treatment, meaning the effect is averaged over the distribution of confounders in the target population rather than being conditional on specific covariate values.

• T. True
• F. False

Answer: True

The word 'marginal' in MSM refers specifically to this marginalization: the estimated effect is not 'the effect of treatment for patients with CD4 = 200' but the effect averaged across all patients in the study population with their observed covariate distributions. This is distinct from conditional models, which estimate effects holding covariates at specific values. The marginal treatment effect is analogous to what you would observe in a randomized trial — it is the population-average treatment effect.

Question 4

When time-varying confounders are present, a researcher who adds enough covariates and interaction terms to a standard regression model will eventually obtain an unbiased estimate of the causal effect of a time-varying treatment.

• T. True
• F. False

Answer: False

This is a common misconception. The problem is not model mis-specification in the usual sense — it is structural. When a confounder is affected by prior treatment, conditioning on it (regardless of how flexibly the model is specified) partially blocks the causal pathway you are trying to estimate. No amount of additional covariates or interaction terms fixes a structural problem in the causal graph. Marginal structural models with IPT weighting are specifically designed to handle this structure; standard regression models are not.

Question 5

Why does conditioning on a time-varying confounder that is affected by prior exposure create bias in a standard regression model, even though failing to condition on it also creates bias? How do marginal structural models escape this dilemma?

Model answer: If you condition on the confounder (e.g., CD4), you partially block the causal path from prior treatment to the outcome, inducing 'collider bias' by conditioning on a downstream effect of treatment. If you omit it, the confounder creates classical confounding bias. You are stuck either way. MSMs escape this by using IPT weighting to remove the association between confounders and treatment assignment before fitting any regression — the confounder is 'balanced away' in the pseudo-population rather than conditioned on in the model.

The key is that IPT weighting acts on the treatment mechanism rather than the outcome model. By reweighting observations to make treatment independent of confounders, MSMs achieve the effect of randomization without conditioning on the confounders in the regression. The confounder is no longer a predictor of treatment in the weighted data, so it is no longer a source of confounding — and because it was never included as a covariate in the outcome model, it cannot block the causal pathway either.