Covariate Effects in Panel-Observed Transition Models

Why Covariates Need Transition-Specific Models

In panel-observed multi-state data, a subject’s state is recorded at visits, while changes between visits may be unobserved. A covariate model must therefore describe the probability of the state at the next visit conditional on the state at the current visit. The effect of a covariate can differ according to the starting state and the destination state; a single coefficient for “progression” may conceal these differences.

This tutorial formulates transition-specific models for a fixed observation interval. The modeled quantities are endpoint transition probabilities: probabilities of being in each possible state at the end of the interval, given the state at its start and covariates. They need not equal probabilities of making exactly one transition, because an unobserved path may include several changes before the next visit.

The formulation below treats \(x\) as a covariate measured at the start of the interval and holds the interval length \(\Delta\) fixed. Its coefficients therefore describe differences in endpoint probabilities over that particular interval, not effects per unit of elapsed time.

Baseline-Category Logit Model for a Transition Row

For each starting state \(i\), select one possible destination \(r_i\) as a reference category. For every other allowed destination \(j\), model the log odds of ending in \(j\) rather than \(r_i\) as a linear function of covariates:

Writing \(\eta_{ij}(x)=\alpha_{ij}+\beta_{ij}^{\mathsf T}x\), and setting \(\eta_{i r_i}(x)=0\), gives the complete row of probabilities:

where \(A_i\) is the set of destinations allowed from state \(i\), including the reference destination. The probabilities in each row are nonnegative and sum to one. If a destination is structurally impossible from \(i\), it should be excluded from \(A_i\), rather than assigned an ordinary finite logit coefficient.

The coefficient \(\beta_{ij,q}\) for covariate \(x_q\) is a change in the log odds of destination \(j\) versus the reference destination, per one-unit increase in \(x_q\), holding other modeled covariates fixed. Thus \(\exp(\beta_{ij,q})\) is the corresponding odds ratio. It is not, in general, a ratio of probabilities, and it is not automatically a causal effect.

Assumptions and Interpretation

The regression equation is meaningful only alongside a clear data and sampling model. The following assumptions define one common setting:

• Well-defined states and visits. State categories are mutually exclusive and collectively exhaustive at each scheduled visit. The interval being modeled is fixed at \(\Delta\), or the analysis is restricted to intervals of that length.
• Conditional transition description. Given the state at the start of the interval and the included covariates, the modeled next-visit probabilities do not depend on earlier observed history. This is a first-order Markov assumption at the visit times. If history has additional predictive value, it must be represented by covariates or the assumption is inadequate.
• Correct covariate specification. The chosen linear predictors, reference categories, interactions, and functional forms adequately describe the log odds. Effects may vary by starting state; imposing the same coefficient across rows is an additional restriction, not a default fact.
• Observed endpoint and appropriate missingness handling. The endpoint state and relevant covariates are observed, or missingness is handled under a stated model. Unobserved transitions between visits are allowed; the endpoint model does not require their path to be observed.
• Independent subjects for the usual likelihood analysis. Subjects contribute independent histories, unless clustering or other dependence is explicitly modeled. Repeated intervals from one subject are not independent merely because they are separate rows.

The model does not identify the timing or order of unobserved changes within the interval. For example, observing state 0 at one visit and state 2 at the next establishes an endpoint transition probability from 0 to 2, but alone does not show whether state 1 was visited in between. This distinction is central when interpreting covariates.

A Short Derivation: Why One Odds Ratio Does Not Determine One Probability Effect

Consider a starting-state row with destinations \(j\), \(k\), and reference \(r_i\). Let a scalar covariate be \(x\), and write the slope for destination \(k\) as \(\beta_{ik}\), with reference slope zero. Differentiating the softmax probability gives:

To see this, write \(P_{ij}=\exp(\eta_{ij})/D\), where \(D=\sum_k\exp(\eta_{ik})\). Then \(\partial\exp(\eta_{ij})/\partial x=\beta_{ij}\exp(\eta_{ij})\), and \(\partial D/\partial x=\sum_k\beta_{ik}\exp(\eta_{ik})\). The quotient rule yields \(\partial P_{ij}/\partial x=P_{ij}\{\beta_{ij}-\sum_kP_{ik}\beta_{ik}\}\).

Consequently, a positive slope for destination \(j\) versus the reference does not guarantee that \(P_{ij}\) rises: competing destinations’ probabilities and slopes also matter. Probabilities should be computed from the full row before making statements about absolute probability changes.

Estimation and a Binary Transition Row

With one observed interval per independent subject and fully observed endpoints, the contribution to the likelihood for a subject starting in \(i\) and ending in \(j\) is \(P_{ij}(\Delta\mid x)\). Multiplying these contributions over subjects gives the likelihood. With repeated visits, a subject’s likelihood contribution is the product of the relevant conditional row probabilities only under the stated conditional transition structure; uncertainty calculations must still account for the subject as the sampling unit. Estimates are commonly obtained by maximum likelihood, and standard errors describe sampling uncertainty under the model.

When a starting-state row has only two possible destinations, the baseline-category model reduces to ordinary logistic regression for the probability of one destination. The coefficient then gives a covariate effect on the odds of that endpoint versus the other endpoint. This is often easier to interpret, but it remains specific to the chosen starting state and interval length.

State-Specific Effects and Probability Contrasts

A model can use different coefficients in each starting-state row. This is often scientifically important: a covariate could be associated with recovery among ill subjects but have a different association with illness onset among well subjects. A shared coefficient across rows should be imposed only when it is a defensible restriction. Likewise, destination-specific coefficients allow a covariate to shift the balance among competing endpoints in different directions.

A useful reporting sequence is to state the modeled interval and starting state, identify the reference destination, report the log-odds coefficient or odds ratio with its units, and then show predicted probabilities for meaningful covariate values. Predicted probabilities make the competing destinations visible. If uncertainty is reported, intervals for probabilities should reflect the joint uncertainty in the row’s estimated coefficients, rather than treating each destination probability as an unrelated estimate.

Diagnostics, Alternatives, and Limits

Warnings that the model may be failing include fitted probabilities near zero or one with very large coefficient estimates, sparse starting-state/destination combinations, and poor agreement between observed and predicted endpoint proportions across covariate groups. Complete or quasi-complete separation can make maximum-likelihood estimates unstable or effectively unbounded. A nearly empty destination category may require combining categories only if scientifically justified, simplifying the predictor structure, or using an appropriate penalized method. It should not be “fixed” by silently treating an allowed destination as impossible.

Check that each predicted row sums to one, that no impossible endpoint has been included, and that covariate effects are not being interpreted outside their observed range. Compare predicted and observed frequencies by starting state and relevant covariate strata, and examine whether earlier history still predicts the next observed state after conditioning on the model variables. Such checks can reveal inadequate functional forms or a failure of the visit-time Markov assumption. Repeated observations also require appropriate subject-level uncertainty calculations.

A multinomial row model respects the requirement that destination probabilities sum to one. Fitting separate binary regressions for every destination can yield probabilities that do not sum to one and can produce mutually incompatible predictions; it is appropriate only when those separate outcomes are explicitly part of a different model. A proportional hazards regression for interval-censored times instead models event-time hazard structure and is not interchangeable with a regression for a multi-state endpoint row. A marginal model for whether a subject is in a particular state, without conditioning on the starting state, also answers a different question.

Finally, fixed-interval transition probabilities should not be compared as though they were rates per unit time when interval lengths differ. The model does not recover transition timing, distinguish alternative paths with the same observed endpoints, or guarantee good predictions beyond the covariate range used to fit it. The introductory material on panel-observed transition likelihoods establishes how endpoint transitions contribute to likelihoods; here the additional task is to model how covariates alter those endpoint probabilities.