Treatment-Confounder Feedback Across Two Visits

What Makes Confounding “Feedback”?

With repeated treatment, one covariate can occupy two roles at different points in time. It can be changed by an earlier treatment and then help determine both a later treatment decision and the outcome. This pattern is called treatment-confounder feedback. The word “feedback” describes the ordering of causal relationships; it does not mean that the outcome itself feeds back into treatment.

Consider two treatment decisions. Let \(A_1\) be treatment at the first visit, \(L_1\) a covariate measured after \(A_1\) but before the next decision, \(A_2\) treatment at the second visit, and \(Y\) the later outcome. Let \(L_0\) denote covariates measured before \(A_1\). The essential structure is \(A_1 \to L_1\), \(L_1 \to A_2\), and \(L_1 \to Y\). The last two arrows make \(L_1\) a confounder for the later treatment-outcome relationship in the causal model; the first makes it treatment-affected.

The timing is crucial: \(L_1\) must be measured after \(A_1\) and before \(A_2\). A covariate measured only after \(A_2\) cannot, in this ordering, determine the second treatment decision. Baseline covariates \(L_0\) may affect all later variables, but they are not themselves treatment-affected by \(A_1\).

A Probability Model That Exhibits Feedback

The following model makes the three required causal links explicit. It is a deliberately simple binary example, not an assumption that clinical data generally follow these particular probabilities. Let \(U_L,U_A,U_Y\) be mutually independent Uniform\((0,1)\) variables, also independent of \(A_1\), and let \(A_1\) be Bernoulli with probability \(0.5\). Define:

Here \(I\{\cdot\}\) equals one when its condition holds and zero otherwise. Each threshold is a probability between zero and one for all binary inputs. The independent exogenous variables make the conditional probability calculations below valid; merely saying that equation errors have mean zero would not, by itself, justify them.

Sequential Structure and Its Implications

A longitudinal data-generating process can be represented by its conditional distributions in temporal order. For the variables in this example, the chain rule for probabilities gives the following factorization (with \(L_0\) included when observed):

This is first a probability identity, not proof of a causal model. To derive it, repeatedly apply \(p(x,z)=p(x)p(z\mid x)\): order the variables as \(L_0,A_1,L_1,A_2,Y\), and condition each new variable on those that precede it. A causal interpretation additionally requires correct temporal ordering and assumptions about the mechanisms that generate the variables.

In the probability model above, the factorization also lets us derive how the first treatment is associated with the distribution of the second treatment through \(L_1\). By the law of total probability,

When \(A_1=0\), this is \(0.1(0.8)+0.6(0.2)=0.08+0.12=0.20\). When \(A_1=1\), it is \(0.3(0.2)+0.8(0.8)=0.06+0.64=0.70\). These totals combine the treatment-specific \(L_1\) distribution with the later-treatment mechanism. They are not a comparison that isolates a causal effect of \(A_1\) on \(A_2\) through \(L_1\), since the model also includes a direct \(A_1\) term in the \(A_2\) mechanism.

This calculation illustrates why the full history matters for describing the treatment process: the distribution of \(L_1\) changes across first-treatment groups, and \(L_1\) also predicts later treatment. The calculation does not, on its own, establish that the model is suitable for an observational study.

Distinguishing Feedback from Similar Patterns

A time-varying covariate is not automatically a treatment-confounder feedback variable. To see the distinctions, compare the following two fully specified mechanisms. As in the previous example, independent Uniform\((0,1)\) variables implement each Bernoulli probability, so the stated conditional probabilities also have the corresponding causal interpretation within each model.

These comparisons show why the definition has two parts. “Affected by past treatment” is not enough, and “predicts treatment and outcome” is not enough. The causal timing and both roles must be present together.

Assumptions, Diagnostics, and Limits

In an observational study, the causal diagram is a substantive hypothesis, not something verified just by finding associations in a data set. Interpreting \(L_1\) as a cause of \(A_2\) and \(Y\) requires that the visit schedule and measurement times are understood, that the covariate precedes the second treatment, and that relevant common causes have been considered. For causal comparisons, investigators also need assumptions about how treatment is assigned given measured history; measured \(L_1\) alone does not guarantee that all confounding has been controlled.

Useful checks begin with the design and data collection. Confirm that \(L_1\) was measured after the first decision and before the second, and that its definition is comparable across treatment groups. Examine whether \(L_1\) varies with \(A_1\), and whether it predicts \(A_2\) and the outcome within relevant histories. These are diagnostic checks for compatibility with the proposed structure, not proofs of causal arrows: unmeasured causes, selection, measurement error, or model misspecification can produce or conceal the same associations.

A further warning sign is inadequate overlap. If some combinations of \(A_1\) and \(L_1\) almost never occur, the data may provide little information about later treatment comparisons within those histories. Even a correctly described feedback structure does not ensure that all causal effects can be estimated from the available data.

The main alternatives are distinct causal structures, not competing names for the same pattern. A baseline confounder precedes both treatment visits and is not changed by \(A_1\). A time-varying prognostic covariate may be changed by \(A_1\) but fail to influence \(A_2\). A later-treatment confounder may influence \(A_2\) and \(Y\) but be unaffected by \(A_1\). Only the pattern combining all three links is treatment-confounder feedback. Identifying that structure does not itself select an estimator or establish that a particular statistical model is valid.

A common mistake is to call any covariate that predicts later treatment and outcome a feedback confounder. That reasoning omits whether the covariate was caused by earlier treatment. The correction is to check the temporal order and each causal link separately: \(A_1\) to \(L_1\), \(L_1\) to \(A_2\), and \(L_1\) to \(Y\). Another mistake is to infer a causal change from a conditional regression coefficient without assumptions about the assignment mechanisms. As the first example makes clear, its probability claims were justified by a fully specified model with independent exogenous variables, not by the form of an equation alone.