Feedback Between Response and Time-Varying Covariates

Why Feedback Changes the Modeling Question

A time-varying covariate can be more than a predictor measured repeatedly. Disease severity, for example, may affect the next clinical response and influence the next treatment decision; treatment may then change subsequent severity. The response and covariate therefore evolve in a feedback system. A model that uses the recorded covariate history as if it were fixed can still describe response conditional on that history, but it does not, by itself, describe how the history arose or how it would change under a different treatment process.

The essential step is to define the order of events. At occasion \(k\), let \(C_k\) denote a time-varying covariate, \(Y_k\) the response, and \(A_k\) the administered dose or treatment decision. Let \(E_k\) denote exposure during the relevant interval, as determined by the dosing and PK history. A useful ordering is: observe \(C_k,Y_k\); choose \(A_k\); then observe \(C_{k+1},Y_{k+1}\). A model must follow the actual study timeline rather than impose this ordering when the measurements and decisions occurred differently.

This distinction is about the target of the model, not a rule that one approach is always superior. For forecasting a response conditional on a fully observed history, a conditional model may be appropriate. For predicting future covariates and responses together, or comparing treatment strategies that would change future histories, the covariate and treatment processes generally cannot be ignored.

Write the Time Order and Transition Model

Define the history available just before the decision at occasion \(k\) as \(H_k=(C_0,\ldots,C_k,Y_0,\ldots,Y_k,A_0,\ldots,A_{k-1})\), augmented as needed by baseline characteristics and prior exposure. The treatment mechanism describes how the decision is made from that history. The state-transition model describes how the next covariate and response arise after the current decision.

Here \(q\) is the conditional distribution of treatment given the available history. The transition distribution may be modeled jointly, allowing the next covariate and response innovations to be associated. Alternatively, its density can be factored by the chain rule, for example into a distribution for \(C_{k+1}\) and a distribution for \(Y_{k+1}\) conditional on \(C_{k+1}\). Such a factorization is a modeling choice about the conditional distributions; it is not an assertion that the transitions are independent.

For illustration, suppose higher values of \(C_k\) and \(Y_k\) both indicate greater disease burden. A simple pair of transitions is:

The term \(\beta_{CY}Y_k\) lets prior response predict later covariate values, and \(\beta_{YC}C_k\) lets the covariate predict later response. Treatment can affect covariate dynamics through \(\beta_{CA}\), and exposure can affect response through \(\beta_{YE}\). These equations are examples, not universal biological laws. The model must match the scientific scale, timing, and plausible mechanisms. If the innovations \(\epsilon_{C,k+1}\) and \(\epsilon_{Y,k+1}\) are assumed independent, that assumption needs justification; otherwise their joint distribution can allow dependence.

Likelihood: Sequential Decisions and State Transitions

For a joint model of states and treatment decisions, the sequential likelihood follows from the chain rule. Let \(S_k=(C_k,Y_k)\). Conditional on baseline information, a first-order history specification gives:

This expression assumes that the chosen history contains the information needed to specify each treatment decision and state transition. If it does not, the history must be expanded or the process modeled differently. The factor \(q(A_k\mid H_k)\) is included when the distribution of treatment is part of the joint model. It can be omitted from a likelihood for state transitions conditional on treatment only under an appropriate conditioning assumption.

The important qualification is that conditioning on an entire observed dose sequence is not automatically equivalent to conditioning on the current dose at each decision. If later doses depend on earlier covariate or response values, selecting individuals or trajectories with a particular full sequence can change the distribution of earlier states. Independent transition errors do not remove this selection. A conditional product of state-transition densities is justified only if the dose history is exogenous or noninformative for the transition innovations and conditioning on the full dose sequence leaves the transition distributions unchanged. Otherwise, the joint treatment mechanism must be represented, or the conditional state distribution must account for the selection induced by the observed sequence.

Worked Examples

When Is a Fixed Covariate History Appropriate?

A response model that conditions on the recorded covariate and exposure histories can be useful for describing or predicting response for those histories. This is often called a conditional analysis: the covariate path is treated as supplied information rather than modeled as an outcome. It can be adequate when that is the stated target, the response model is correctly specified conditional on the path, and the relevant sampling or decision process does not undermine the intended interpretation.

That conditional model does not automatically provide the distribution of future covariates, the joint probability of the observed trajectories, or predictions under a new dosing strategy. Nor does a coefficient for a post-treatment covariate automatically represent a causal effect. If treatment changes a covariate that in turn predicts later response and treatment, conditioning on that covariate can remove part of a treatment pathway or induce selection. Causal comparisons of strategies require appropriate causal assumptions and methods, such as sequential exchangeability and positivity where applicable; a joint descriptive model alone does not establish them.

A joint dynamic model is attractive when the scientific target includes the evolution of both processes, when covariates are needed for future predictions, or when treatment decisions adapt to evolving states. Its additional complexity is useful only if the transition forms are plausible and the data contain enough information to estimate them. A simpler conditional model may be more stable for a narrowly defined conditional prediction, but should not be presented as a model of the complete feedback system.

Estimation Checks and Failure Modes

A feedback model can fit the response at observed visits and still describe the dynamics poorly. Check the covariate and response transitions separately as well as jointly. Compare observed and simulated trajectories, transition residuals over time, distributions of changes, and predictive behavior across relevant prior histories. If treatment adapts to history, check whether the model reproduces the observed treatment probabilities across those histories when treatment is modeled.

• Incorrect time ordering. A covariate measured after a dose decision cannot be treated as information available for that decision unless the records support that order. Reconstruct the timeline before specifying the transition.
• Missing feedback path. If prior response predicts later covariate values but the covariate equation omits it, simulated covariate trajectories may systematically diverge across response histories.
• Overly strong independence assumptions. Separate equations do not imply independent innovations. Residual dependence may indicate shared unmodeled drivers or a need for a joint transition distribution.
• Conditioning on adaptive future doses. A full dose sequence can select earlier states. Do not assume that independent transition errors make this conditioning harmless.
• Extrapolating beyond observed histories. Joint models can generate covariate and treatment combinations rarely represented in the data. Predictions there depend heavily on structural assumptions.

Useful alternatives depend on the estimand. Conditional response modeling targets response given recorded histories. Joint transition modeling targets the evolving state distribution and can support forecasts under a specified decision process. For causal effects of adaptive treatment, methods that explicitly address time-varying confounding may be more appropriate than interpreting conditional regression coefficients as intervention effects. These approaches answer different questions and should be compared on the target, assumptions, and available data—not only on fit statistics.

Deriving the Sequential Factorization

The likelihood factorization is a direct application of the chain rule, not an independence claim. Let the observed sequence be \(S_0,A_0,S_1,\ldots,A_{K-1},S_K\). The chain rule writes its density as the baseline density multiplied by each next conditional density given everything preceding it. Group each decision with the state transition that follows it:

The second line uses the stated history definition and includes exposure when it is part of the transition information. The identity holds as a sequential factorization when those conditional distributions are defined using the preceding history. It does not license replacing them with distributions conditional on all future doses. Future treatment factors can reweight earlier histories when treatment depends on those histories. This is why the treatment mechanism is essential to the joint distribution even when the scientific response equation is the main object of interest.