Including Treatment And Baseline Covariates

Why Include Treatment and Baseline Covariates?

A shared-parameter factorization connects the outcome process and the response pattern through a latent effect. In an applied study, treatment and baseline characteristics can also be related to outcomes, response, and the distribution of that latent effect. Leaving them out of the factorization can conceal important assumptions: for example, a model may implicitly assert that response is unrelated to treatment after conditioning on the latent effect, even when treatment predicts response in the data.

The goal here is to place treatment and baseline covariates explicitly in the joint law. Doing so does not by itself establish a causal treatment effect, make missingness ignorable, or guarantee that the model is identifiable. It makes the conditional distributions and the assumptions they encode visible.

The first expression includes the distributions of baseline covariates and treatment assignment. The second conditions on their realized values and focuses on the latent effect, outcomes, and response pattern. The notation \(f\) can represent a probability mass function, a density, or an appropriate mixture of the two. For a longitudinal outcome, \(f(y\mid b,a,x)\) may itself be written as a sequence of conditional outcome distributions.

What Each Factor Says

The factor \(f(a\mid x)\) describes treatment assignment given baseline covariates. In a randomized study, it may be known from the randomization scheme; if assignment is independent of \(X\), it can be written \(f(a)\). In an observational study, assignment may depend on \(X\). When analysis conditions on treatment and covariates, this assignment factor is not part of the conditional law \(f(b,y,r\mid a,x)\), but it remains part of the full joint distribution.

The outcome factor \(f(y\mid b,a,x)\) allows outcomes to be associated with treatment and baseline characteristics, as well as with the latent effect. For example, a regression model may include treatment, age, disease severity at baseline, and interactions between treatment and baseline severity. The response-pattern factor \(f(r\mid y,b,a,x)\) allows response to depend on treatment, baseline characteristics, the latent effect, and outcomes. Conditioning on the complete \(Y\) in this selection-style factorization permits response probabilities to depend on outcomes that are missing; that possibility is one way the model can represent nonignorable missingness.

Finally, \(f(b\mid a,x)\) specifies the distribution of the latent effect across treatment and baseline-covariate groups. Its form must match the meaning assigned to \(B\). If \(B\) represents a pre-treatment subject characteristic, a model may posit \(f(b\mid a,x)=f(b\mid x)\), particularly under randomized assignment. If it represents a latent quantity that can vary with treatment, dependence on \(A\) may be appropriate. Neither choice follows automatically from the factorization.

The product above is a chain-rule decomposition for the ordering \(X,A,B,Y,R\). As written, it does not require a conditional-independence assumption beyond the stated conditioning sets: each factor is a conditional distribution in that ordering. Substantive restrictions arise when a factor is simplified—for example, replacing \(f(r\mid y,b,a,x)\) by \(f(r\mid y,b)\) asserts that response is conditionally independent of treatment and baseline covariates given \(Y\) and \(B\).

A Chain-Rule Justification

The factorization can be derived by repeatedly applying the definition of conditional probability. For variables with a joint density or mass function, start with the chosen ordering and write

Each equality follows from \(f(U,V\mid W)=f(U\mid W)f(V\mid U,W)\). Thus, adding \(A\) and \(X\) to the conditioning sets is not merely a notational adjustment: it specifies which covariates each conditional distribution may use. A more restrictive model can then be stated by removing a variable from a factor, but that removal is an assumption to defend, not a consequence of the chain rule.

Regression Models Show Where Covariates Enter

One convenient way to specify factors is with regression models. For a binary outcome, a possible conditional mean model is \(P(Y=1\mid A,X,B)=\operatorname{expit}(\beta_0+\beta_A A+\beta_X X+\beta_B B)\), where \(\operatorname{expit}(u)=1/(1+\exp(-u))\). A response model might use a separate linear predictor, \(P(R=1\mid Y,A,X,B)=\operatorname{expit}(\gamma_0+\gamma_A A+\gamma_X X+\gamma_B B+\gamma_Y Y)\). These are model choices, not requirements of the factorization. Continuous or categorical covariates can be represented with suitable terms, and longitudinal models can include visit-specific predictors and histories.

The response equation above conditions on \(Y\), including its value when \(R=0\). Consequently it specifies a complete-data response mechanism, not a regression that can be fitted by simply substituting an unobserved outcome. The factorization identifies where the variables belong conceptually; the treatment of unobserved values in a likelihood is a separate calculation.

Assumptions, Diagnostics, and Alternatives

Before interpreting fitted factors, check that their assumptions match the study and the role of each variable. The model presumes that \(A\) and \(X\) are defined and measured as intended, that the conditional distributions are well-defined in the relevant covariate groups, and that the chosen dependence on \(B\) adequately represents the shared latent structure. If treatment effects vary by baseline severity, a model containing only an additive treatment term may be too restrictive. If response depends on treatment after conditioning on \(Y,B,X\), a response factor that omits \(A\) will not represent that dependence.

Useful diagnostics include examining whether treatment and response are sparse within important covariate groups, checking that the model’s predicted outcome and response rates are compatible with observed summaries, and assessing whether the chosen functional forms and interactions are plausible. In longitudinal settings, assess predictions by visit and response history rather than relying only on a pooled response rate. A lack of overlap—for example, almost no treated participants at some values of \(X\)—can make comparisons or parameter estimates unstable. Diagnostics can reveal poor fit or weak information, but they cannot establish an untestable assumption about dependence on outcomes that are missing.

There are two important comparisons. First, the conditional shared-parameter expression models outcomes and response through factors linked by \(B\); omitting \(B\), or imposing conditional independence after conditioning on \(B\), changes the model’s assumptions. Second, a pattern-mixture factorization organizes the joint law by response pattern rather than by a response mechanism conditional on complete outcomes. These are alternative organizations of a joint distribution, but restrictions and parameterizations can make the resulting models substantively different. Neither becomes equivalent to the other merely by adding treatment and covariates to its notation.

Including treatment and baseline covariates also does not by itself solve confounding. In an observational study, a causal interpretation generally requires additional assumptions about treatment assignment and measured confounders. Likewise, specifying a response model that depends on unobserved outcomes does not make that dependence empirically verifiable from observed data alone. A careful analysis separates the factorization, the scientific assumptions, and the target interpretation.