Efficiency Gains From Common Transition Effects

Why a Common Effect Can Be More Precise

In a multi-state model, different transitions generally have different eligible individuals, event times, and risk sets. Estimating a separate covariate coefficient for every transition respects those differences, but it also divides the available event information among several estimates. If the covariate has the same effect on each transition, a common coefficient can use information from all of those risk sets to estimate one shared effect. The potential gain is precision, not a guarantee that the model is appropriate.

The central condition is substantive: the coefficient must genuinely be common across transitions, on the same covariate scale and for the same effect specification. The baseline hazards need not be common. Pooling the coefficient does not mean pooling transition risk sets, treating all events as the same event, or assuming that transitions occur at the same rate.

The model concerns the intensity of a particular transition among individuals eligible to make it. Its risk set is transition-specific. A person ineligible for transition \(r\) contributes no risk-set information for that transition at that time, even if that person is observed in the study. The separate baseline hazards \(\lambda_{0r}(t)\) allow the underlying rates to differ freely across transitions.

How the Information Adds

For a single covariate, the transition-specific Cox partial likelihood compares the covariate value for each event with values among individuals in that transition’s risk set. At an event time \(t\) for transition \(r\), define the risk-set weighted mean and variance by \( \bar X_r(\beta,t)=S_r^{(1)}(\beta,t)/S_r^{(0)}(\beta,t)\) and \(V_r(\beta,t)=S_r^{(2)}(\beta,t)/S_r^{(0)}(\beta,t)-\{\bar X_r(\beta,t)\}^2\), where \(S_r^{(j)}(\beta,t)=\sum_iY_{ir}(t)\exp\{\beta X_i(t)\}X_i(t)^j\). The risk-set variance measures how much covariate contrast is available to distinguish the event individual from those still at risk.

For a vector covariate, \(V_r\) is the corresponding weighted covariance matrix. The observed information for a transition is obtained by summing these risk-set variances or covariance matrices over that transition’s event times. The common-coefficient model combines the transition-specific score contributions; its information is the sum of the contributions from all transitions.

This addition explains the precision gain. Under the common-effect model and regularity conditions, the large-sample covariance matrix of the common estimator is approximately \(I_{\mathrm{common}}^{-1}\). Information contributed by one transition can therefore help estimate the same coefficient using events from other transitions. With separate coefficients, the approximate covariance for \(\hat\beta_r\) instead depends on \(I_r^{-1}\) alone.

A short justification follows from the partial-likelihood score. For transition \(r\), each event contributes the event individual’s covariate minus its risk-set weighted mean. The score for a common coefficient is the sum of those transition-specific contributions, since every contribution depends on the same \(\beta\). Differentiating the negative score gives the sum of the transition-specific weighted covariance contributions. Thus the combination is a consequence of imposing one shared parameter, not of combining the risk sets themselves.

This information calculation supports the usual model-based large-sample standard error when the intensity model is correctly specified and subjects are independent, or when dependence is handled by an appropriate variance estimator. In a multi-state process, transitions from the same person are not separate independent people. The counting-process likelihood accounts for the transition contributions, while uncertainty calculations must respect the sampling and dependence structure. The simple inverse-information expression should not be used blindly when its assumptions fail.

Worked Information and Precision Calculations

Assumptions, Checks, and Failure Modes

The precision argument depends on the model and data structure, not merely on fitting a shared coefficient. Before interpreting a smaller standard error as a gain, check the following conditions:

• Equal coefficient meaning. The covariates must be measured and coded consistently, and the log-hazard coefficient for each covariate must be shared across the transitions being pooled. Equality is an assumption, not a consequence of using one parameter.
• Correct transition models. Within each transition, the specified covariate form and proportional-hazards structure must be adequate. A common coefficient does not repair a wrong functional form or time-varying effect.
• Correct risk sets and observation process. Eligibility for each transition must be represented correctly. Censoring must be non-informative conditional on the modeled information, or its dependence must be handled appropriately.
• Usable covariate variation. The relevant covariate must vary in the risk sets that contribute events. If a covariate is constant within a transition’s risk sets, that transition supplies no information about its coefficient, regardless of its event count.
• Appropriate uncertainty calculation. Repeated transitions from one individual do not create independent subjects. When the sampling design or dependence requires it, use a variance estimator that accounts for within-subject clustering rather than relying only on inverse model-based information.

Diagnostics should be transition-specific. Inspect event counts and risk-set covariate distributions, and look for separation or very limited overlap. Compare fitted effects and uncertainty across transitions as a descriptive check, while recognizing that noisy transition-specific estimates can conceal real differences. Examine proportional-hazards and covariate-form diagnostics within each transition; an aggregate check can hide a problem confined to one risk process. A small standard error alone is not evidence that the equality constraint is plausible.

The main alternative is to estimate transition-specific coefficients. This avoids imposing equality but can produce unstable estimates when some transitions have few events or little covariate variation. A common coefficient uses fewer parameters and can be substantially more precise when equality holds. Its cost is bias for transition-specific targets when effects differ. Neither model is universally preferable: the choice depends on the scientific target and the credibility of the common-effect assumption.

The gain also has limits. Information matrices add, but the amount added depends on risk-set covariate variation, not just the total number of observed transitions. Information may be weak for a particular covariate even in a large study. Further, a correctly estimated common coefficient is a shared conditional log-hazard effect; it does not by itself summarize absolute transition probabilities or overall prognosis.