Cause-Specific Loadings on Shared Random Effects

Why Use Cause-Specific Loadings?

A longitudinal biomarker may carry information about more than one possible event, but that information need not point in the same direction for every event type. For example, a latent tendency toward higher biomarker values could be associated with a higher hazard of one event and a lower hazard of another. A model that forces the same association for all causes cannot represent that pattern.

Cause-specific loadings provide one way to express these differences. The longitudinal model supplies a shared set of subject-level random effects, and each cause-specific hazard has its own coefficients multiplying those effects. “Shared” means that the same latent subject effects enter all cause-specific hazards; it does not mean that their coefficients must be equal.

Model and Interpretation

Suppose subject \(i\) has longitudinal measurements \(Y_{ij}\) at times \(t_{ij}\), and let \(m_i(t)\) denote the latent trajectory for the marker. A general longitudinal submodel is

Here \(X_i(t)\) and \(Z_i(t)\) are known design vectors, \(\beta\) is a vector of fixed effects, and \(b_i\) is a vector of subject-specific random effects. A common assumption is \(b_i\sim N(0,D)\), where \(D\) is a covariance matrix. The measurement errors \(\epsilon_{ij}\) are commonly assumed to have mean zero and to be independent across measurements conditional on the random effects and covariates, with a specified variance model.

Let \(T_i\) be the time to the first event and \(K_i\) its cause, taking values in \(1,\ldots,K\). A cause-specific hazard conditional on the random effects can be written

In this expression, \(h_{0k}(t)\) is the baseline hazard for cause \(k\), \(W_i(t)\) is a vector of covariates, \(\gamma_k\) is its cause-specific coefficient vector, and \(\alpha_k\) is the cause-specific loading vector. Given the random effects and covariates, the hazard for cause \(k\) describes the instantaneous rate of that cause among subjects who have not yet experienced any event.

The loading interpretation is conditional: holding the other random-effect components and the covariates fixed, a one-unit increase in component \(r\) of \(b_i\) multiplies the cause-\(k\) hazard by \(\exp(\alpha_{kr})\). A positive loading means a larger value of that latent component is associated with a higher cause-specific hazard, conditional on the model; a negative loading means a lower hazard. The units matter: random effects are defined on the scale of the longitudinal model, so a loading is not interpretable without knowing that scale.

What the Cause-Specific Coefficients Allow

For a fixed subject and time, compare two hazard predictors that differ only in the random-effect vector, from \(b_i\) to \(b_i+\Delta b\). The cause-\(k\) hazard ratio is

This follows by substituting each vector into the model: the baseline hazard and covariate term are identical in numerator and denominator, and cancel; the difference between the two random-effect terms is \(\alpha_k^\top\Delta b\). Thus the loading vector determines how a change in the latent subject effects shifts each cause-specific hazard. It does not, by itself, give an unconditional hazard ratio for a population in which random effects are unobserved.

A shared-loading model imposes \(\alpha_1=\cdots=\alpha_K\). It is more parsimonious and can be suitable when a common association is scientifically justified. Cause-specific loadings relax that constraint while retaining a shared latent trajectory. Using entirely separate random effects for each cause is a different and more extensive specification: it no longer makes the causes respond to the same latent subject-level effects in the stated way. The choice should reflect the scientific question and the information available, not merely a preference for more parameters.

Comparing Loadings Across Causes

A cause-specific loading can be compared with the corresponding loading for another cause when both refer to the same random-effect component on the same scale. For example, \(\alpha_{1r}-\alpha_{2r}\) is a contrast in log-hazard change per one-unit increase in component \(r\). Its exponential is the ratio of the two cause-specific hazard ratios for that one-unit change, with other terms held fixed. Estimating each coefficient separately is not enough to test their difference: the covariance between the estimates is needed.

Assumptions, Diagnostics, and Limits

The model is useful only if its shared-effects structure is credible. The random effects must capture the relevant subject-level features linking repeated marker measurements to event risk. The assumed distribution of those effects, the longitudinal mean and error structure, and the baseline hazards must be reasonable enough for the intended inference. In particular, if the marker's latent variation is poorly estimated, the loadings can reflect misspecification in the longitudinal model rather than a stable association with event risk.

Practical checks include examining longitudinal residual patterns and whether the fitted model reproduces the spread and correlation of repeated measurements. For the event component, inspect cause-specific event counts and estimated uncertainty, and check whether conclusions are sensitive to plausible baseline-hazard and random-effects specifications. Convergence warnings, very large standard errors, estimates near numerical boundaries, or strong changes under modest specification changes are reasons to qualify interpretation and investigate further. A sparse event cause may provide little information for its loading vector, even when the longitudinal data are extensive.

A loading is conditional on the other random effects in its vector. If random intercept and slope are correlated, for instance, the coefficient of the intercept describes a change in that component while holding the slope component fixed. It is not automatically the effect of a typical subject whose intercept changes in a way that also changes the slope. Moreover, marginal hazards can change over time as the mix of surviving subjects changes; the conditional hazard ratio derived above should not be reported as a population-marginal hazard ratio.

The model describes an association within a specified joint distribution. It does not prove a causal effect of the biomarker, nor does a positive loading mean that most individuals experience the event. The competing event processes also remain distinct: a loading concerns one cause-specific hazard, not the probability of a particular event by a chosen time. Interpretation should therefore name the cause, random-effect component, scale, and conditional nature of the comparison.

A Common Modeling Pitfall

A frequent mistake is to interpret \(\alpha_{kr}\) as the effect of a one-unit increase in the observed biomarker at time \(t\). That is not what the coefficient means in this model: it multiplies the \(r\)th latent random effect, not necessarily the observed value or a particular summary of the trajectory. The correction is to state the coefficient on the random-effect scale and interpret its hazard ratio conditionally. The way to associate a hazard directly with a specified feature of the latent trajectory requires a different predictor specification.

Another error is to infer that two causes have different associations because one loading estimate is statistically significant and the other is not. Significance in one separate test and nonsignificance in another do not test whether the coefficients differ. Form the contrast, use its estimated covariance, and test or interval-estimate that contrast directly.