G-Computation With Treatment-Confounder Feedback

Why Covariate Distributions Must Follow the Intervention

In a longitudinal study, a treatment can affect a later covariate, and that covariate can in turn affect both subsequent treatment and the outcome. This is treatment-confounder feedback. The later covariate is a confounder of the next treatment-outcome relationship, but it is also part of the pathway through which earlier treatment can affect the outcome.

For example, an initial treatment may change a patient's disease severity. Severity then influences whether the patient receives treatment at the next visit and also predicts the final outcome. To estimate the outcome under a specified treatment history, g-computation must account for the severity distribution that would arise under that history. It does not generally suffice to adjust for severity while keeping it distributed as observed in the study.

Use \(L_t\) for covariates measured just before treatment \(A_t\), with \(t=0,\ldots,K\), and \(Y\) for the outcome measured after \(A_K\). Write \(\bar L_t=(L_0,\ldots,L_t)\) and \(\bar A_t=(A_0,\ldots,A_t)\). A treatment regime is a fixed sequence \(\bar a=(a_0,\ldots,a_K)\); \(Y^{\bar a}\) denotes the outcome that would be observed if treatment were set to that sequence. Here the focus is on how the distribution of the covariates evolves under that intervention.

Identification and Propagation

The g-formula identifies a regime-specific mean under consistency, no interference between individuals, sequential exchangeability, and positivity. Informally, sequential exchangeability says that, conditional on the measured treatment and covariate history just before each treatment, treatment assignment is independent of the relevant future counterfactuals. Positivity requires that the treatment specified by the regime has positive probability at histories that can arise under the regime. These assumptions are stated for the intervention and target population of interest.

Let \(H_t=(\bar L_t,\bar A_{t-1})\) be the observed history before \(A_t\). For a fixed regime, the identifying formula is

The displayed sum is for discrete covariates; integration replaces summation for continuous covariates. Each transition distribution is evaluated with the treatment history set to the regime. In particular, the factor for \(L_{t+1}\) uses \(A_t=a_t\), not the treatment value that happened to be observed for a person in that history.

To see why the distributions must be propagated, define \(q_0(l_0)=p(l_0)\). For each \(t=0,\ldots,K-1\), define the distribution of the next covariate history by

For discrete covariates this is the joint probability of the history \(\bar l_{t+1}\); summing it over earlier covariate values gives the marginal distribution of \(L_{t+1}\). Under the intervention, consistency means that the covariate transition for people with history \(\bar l_t\) and treatment \(a_t\) follows the corresponding observed conditional distribution. Sequential exchangeability permits that observed conditional distribution to identify the transition among those who would have that history under the intervention. Positivity ensures the needed conditional distribution is represented in the observed data.

The recursion follows by the probability chain rule: the probability of \((\bar L_t=\bar l_t,L_{t+1}=l_{t+1})\) under the intervention equals the probability of \(\bar L_t=\bar l_t\) under that intervention times the conditional transition probability. Applying this identity at each time gives the product in the g-formula. Finally, conditional expectation gives

and consistency and sequential exchangeability identify the conditional outcome mean in this expression with \(E[Y\mid \bar L_K=\bar l_K,\bar A_K=\bar a]\). Substituting the propagated history distribution yields the stated formula. \(\square\)

Worked Calculations

Interpretation, Diagnostics, and Alternatives

In a parametric g-computation analysis, conditional models are used to estimate the covariate transitions and the outcome mean; their fitted distributions are then combined according to the g-formula. The central requirement for treatment-confounder feedback is structural: each later covariate distribution must be allowed to depend on the earlier treatment history that precedes it. A model or calculation that holds a post-treatment confounder to its observed pooled distribution generally targets no specified intervention.

Useful checks include comparing fitted and observed covariate distributions across treatment histories, checking that predicted probabilities are valid, and checking whether the intervention requires treatment-covariate histories with little or no data support. Compare results across plausible model specifications: large changes can indicate sensitivity to functional form or extrapolation. Good fit to observed outcomes alone cannot establish that the intervention-specific covariate transitions are correct.

A conventional outcome regression that adjusts for a time-varying confounder can block part of the effect of earlier treatment, while omitting it can leave confounding of later treatment. The g-formula addresses this by conditioning on covariates in the transition and outcome distributions, then averaging over the covariate histories induced by the intervention. Inverse-probability weighting takes a different route by reweighting observations to represent treatment histories; it does not perform this same explicit propagation of covariate distributions.

Neither approach removes the need for identification assumptions or adequate data support. Parametric g-computation can be biased if the transition or outcome models are misspecified, and positivity violations can make estimates depend heavily on extrapolation. The formula describes the target under its assumptions; it does not guarantee that the available data can estimate that target reliably.