Separating Exposure Effects From Time-Varying Confounding

Why a Longitudinal Exposure Association Can Mislead

In an observational treatment course, dose is often changed in response to a patient's evolving condition. A patient whose disease worsens may receive a higher dose; that worsening may also predict a worse outcome. Consequently, patients with higher observed exposure can have poorer outcomes even when the drug is beneficial. The reverse distortion is also possible: patients who improve may remain on treatment longer, making continued exposure appear especially effective.

The difficulty is not merely that a prognostic covariate was omitted. Disease status at an intermediate visit may be affected by earlier treatment and then influence both later treatment and outcome. Adjusting for it in a conventional regression can remove part of the earlier treatment effect, while failing to adjust for it can confound the later treatment effect. The analysis must represent the treatment and disease history in the correct time order.

Here, \(A_0\) denotes treatment or exposure at the first decision, \(L_0\) measured baseline disease status, \(L_1\) disease status after the first decision, \(A_1\) treatment at the second decision, and \(Y\) a later continuous efficacy outcome, with larger values representing greater benefit. This tutorial uses two decision times to make the reasoning explicit; the same logic extends to longer histories.

Define the Causal Question and Its Assumptions

For example, one may ask for the mean outcome if every eligible patient followed the regimen \(A_0=1,A_1=1\), compared with the mean if every patient followed \(A_0=0,A_1=0\). These are hypothetical interventions on treatment choices, not claims that patients in the observed data actually followed those regimens. An exposure-based intervention must also be scientifically meaningful: a concentration target may not be well-defined if different dosing strategies produce different effects beyond that concentration.

Identification from observational data requires assumptions. Consistency means that a patient's observed outcome equals the potential outcome under the treatment history the patient actually received, and that treatment versions relevant to the question are sufficiently well specified. Sequential exchangeability means that, conditional on measured history, treatment assignment at each decision is independent of the relevant future potential outcomes. In this setting, baseline history \(L_0\) must suffice to control confounding of \(A_0\), while history including \(A_0,L_0,L_1\) must suffice to control confounding of \(A_1\). Positivity means that each treatment choice in the regimes being compared has positive probability for every history represented in the target population. The analysis also assumes no interference between patients and adequate measurement and modeling of the relevant variables.

These assumptions cannot be established merely by fitting a model. In particular, sequential exchangeability is not guaranteed by adjustment for the variables available in a data set: important unmeasured disease severity, adherence, or clinician judgment can leave residual confounding. The causal interpretation is conditional on the stated assumptions.

Deriving the Longitudinal g-Formula

Let \(Y^{a_0,a_1}\) denote the potential outcome under the fixed regimen \(A_0=a_0,A_1=a_1\), and let \(L_1^{a_0}\) denote the intermediate disease state that would occur under first treatment \(a_0\). For discrete \(L_0\) and \(L_1\), the mean outcome under the regimen is identified by:

The baseline distribution \(P(L_0=l_0)\) is essential: it describes the target population before treatment, rather than the subset of patients who happened to receive \(a_0\). Omitting \(L_0\) from the conditional models or averaging over the treated group's baseline distribution can give the wrong regime mean.

The formula follows by averaging the outcome under the regimen over the disease state that would arise under its first treatment, and then over baseline status. Specifically, iterated expectation gives \(E[Y^{a_0,a_1}]=\sum_{l_0,l_1}E[Y^{a_0,a_1}\mid L_0=l_0,L_1^{a_0}=l_1]P(L_1^{a_0}=l_1\mid L_0=l_0)P(L_0=l_0)\). Consistency and sequential exchangeability identify the first conditional mean with the observed outcome mean among patients with the specified treatment and history, and identify the intermediate-state distribution with the observed conditional distribution under \(A_0=a_0\). This yields the displayed formula. For continuous covariates, the corresponding sums are integrals.

In practice, g-computation estimates the conditional distribution of \(L_1\) given \(A_0,L_0\), and the conditional mean of \(Y\) given \(A_0,L_0,L_1,A_1\). It then predicts outcomes under each regime for the baseline population, averages those predictions, and contrasts the regime means. This approach retains the pathway through which early treatment changes later disease status; it does not simply hold \(L_1\) fixed when comparing early treatment.

Weighting as an Alternative Adjustment

Inverse-probability weighting offers a different way to address measured time-varying confounding. Each observed treatment history is weighted by the inverse of its probability conditional on the measured history. Histories that were unlikely under the observed treatment process receive more weight, creating a weighted population in which measured history is less associated with treatment. For two decision times, an unstabilized treatment weight for an individual's observed history is:

Under the assumptions already stated and correctly estimated treatment probabilities, a weighted comparison can estimate regimen effects. Unlike ordinary regression that conditions on \(L_1\) as a fixed covariate, the weighting approach can account for its role as a confounder of \(A_1\) while avoiding direct adjustment that blocks the pathway from \(A_0\) through \(L_1\). G-computation and weighting rely on different models—the former for outcome and intermediate-state distributions, the latter for treatment probabilities—and can be used as complementary analyses. Neither method repairs unmeasured confounding or poor data.

Diagnostics, Limitations, and Common Pitfalls

Before interpreting an adjusted estimate, inspect whether the data support the comparison. Check treatment probabilities across measured histories, the number of patients following each relevant history, and whether model predictions for \(L_1\) and \(Y\) are plausible. For weighting analyses, inspect the distribution of weights, extreme values, and the effective sample size; a small effective sample size indicates that a few observations dominate. Weight truncation may reduce variance but changes the analysis and can introduce bias, so report it as a sensitivity analysis rather than a cure for positivity violations.

For g-computation, assess model fit and sensitivity to reasonable alternative specifications, including nonlinear relationships and interactions when clinically justified. Compare results from plausible outcome and intermediate-state models. A model can fit the observed data reasonably yet extrapolate poorly to a treatment history scarcely represented in a subgroup. If estimates change substantially under modest modeling choices, the data may not support a stable causal conclusion.

A common mistake is to add \(L_1\) to a single outcome regression and interpret the coefficient of \(A_0\) as the total effect of early treatment. This conditions on disease status that may itself have been changed by \(A_0\), thereby blocking part of the effect pathway; it may also create additional bias under some causal structures. The correction is to define the regimen contrast first, then use a longitudinal method such as g-computation or inverse-probability weighting under explicit assumptions. Another mistake is to treat a measured exposure concentration as if it were randomized: disease progression, adherence, clearance, and dose changes can all shape observed exposure. Adjustment can address only relevant confounders that were measured adequately and modeled appropriately.

The methods also have limitations. Treatment probabilities and outcome models are estimated from finite data; uncertainty should reflect that estimation, for example through appropriate resampling. Exposure may be measured with error, and dose or concentration may not uniquely define a feasible intervention. Informative dropout requires additional modeling or weighting assumptions. These analyses estimate effects for a specified target population and treatment contrast, not a universal drug effect.