Updating Random-Effects Weights With Biomarker History

From a Population Distribution to a Patient-Specific Distribution

A joint model represents a patient’s latent biomarker trajectory through shared random effects. Before observing that patient’s biomarker values, the random effects have a population distribution. After observing a history of values at particular times, the distribution should generally change: histories that are more compatible with some random-effect values than others shift probability toward those values. This update is the bridge between a population model and an individualized trajectory.

The task here is to update using the longitudinal observations and their measurement times. The update is conditional on the longitudinal model and its parameters. It is not obtained by simply substituting the observed measurements for latent trajectory values, nor by treating an estimated random effect as known without uncertainty.

Bayes’ Rule for a Biomarker History

Suppose the longitudinal model implies conditional densities \(f(y_{ij}\mid b_i,t_{ij},x_i,\theta)\), and the measurements are conditionally independent given \(b_i\), their times, and the model parameters. Conditioning on the observed schedule \(t_i=(t_{i1},\ldots,t_{im_i})\), the history likelihood is

The posterior density is this likelihood times the prior density, divided by a normalizing constant. The integral in the denominator makes the posterior integrate to one over the possible values of \(b_i\).

The times appear in the likelihood because the longitudinal model may assign different expected values, variances, or response probabilities at different times. If the schedule is fixed by design, or is ignorable under the analysis assumptions, conditioning on the observed times is appropriate. If visit timing itself carries information about \(b_i\) after conditioning on the model variables, this likelihood alone is not enough: a model for the visit process may also be needed.

This update is conditional on \(\theta\), which commonly includes population parameters estimated from the full data. Treating \(\theta\) as fixed gives a conditional posterior for \(b_i\); it does not automatically include uncertainty in estimating those parameters. The history likelihood also uses only the biomarker observations specified in \(H_i(s)\). Other information should enter only if it is explicitly part of the conditioning set or model.

Gaussian Linear Mixed Model: A Closed-Form Update

For a Gaussian linear mixed model, write the patient’s observations in vector form as

Assume \(b_i\) and \(\varepsilon_i\) are independent, \(D\) and \(R_i\) are positive definite, and the errors are conditionally Gaussian with covariance \(R_i\). The matrices \(X_i\) and \(Z_i\) are formed using the patient’s covariates and measurement times. Given fixed model parameters, the posterior is normal, with covariance \(V_i\) and mean \(\mu_i\):

The covariance reflects both sources of information: \(D\), the population variation in random effects, and the precision of the patient’s longitudinal data. More precise observations generally contribute more to the update, while a small number of noisy observations usually leave substantial posterior uncertainty.

Why the Gaussian Formula Holds

Let \(r_i=y_i-X_i\beta\). The prior density contributes a quadratic term in \(b_i\) involving \(D^{-1}\); the Gaussian observation likelihood contributes a quadratic term involving \(Z_i^\mathsf{T}R_i^{-1}Z_i\). Their product has exponent, up to a constant independent of \(b_i\),

Expanding the second quadratic gives a coefficient of \(b_i^\mathsf{T}b_i\) equal to \(D^{-1}+Z_i^\mathsf{T}R_i^{-1}Z_i\), and a linear term in \(b_i\) equal to \(b_i^\mathsf{T}Z_i^\mathsf{T}R_i^{-1}r_i\). Completing the square therefore yields

where \(V_i^{-1}=D^{-1}+Z_i^\mathsf{T}R_i^{-1}Z_i\) and \(\mu_i=V_iZ_i^\mathsf{T}R_i^{-1}r_i\). This is the exponent of a multivariate normal density, establishing the stated posterior. Positive definiteness ensures the precision matrix is invertible and the covariance is well defined.

Non-Gaussian Biomarkers and Posterior Weights

For binary, count, ordinal, or other non-Gaussian biomarkers, the posterior update has the same Bayes-rule form, but it will not generally be normal or available in closed form. When random effects have a discrete distribution or are represented by quadrature nodes, the update can be implemented by reweighting. For node \(b^{(k)}\) with prior probability \(w_k\), calculate its history likelihood \(L_i(b^{(k)};H_i)\), multiply by \(w_k\), and normalize:

For a continuous random-effects distribution, quadrature or Monte Carlo methods approximate the corresponding integral. The likelihood must use the correct conditional response model; for example, a binary response contributes a Bernoulli probability, not a Gaussian density.

Computation, Diagnostics, and Alternatives

A posterior update should be checked as a calculation and as a model-based inference. In a discrete computation, confirm that all likelihoods and prior weights are nonnegative, the normalizing constant is positive, and the normalized weights sum to one, subject only to rounding. For a continuous computation, check numerical stability and whether quadrature or simulation adequately covers regions with appreciable posterior density. In high dimensions, likelihood products can underflow; summing log-likelihoods and using a log-sum-exp normalization avoids this problem.

The model assumptions also need scrutiny. Conditional independence may fail if residual serial correlation remains after including the random effects. A Gaussian posterior calculation can be misleading if the response distribution or error variance is badly misspecified. A visit schedule that depends on latent biomarker values may require a model for the visit process rather than conditioning on times as though they were noninformative. Sparse or noisy histories may leave a posterior close to the prior; an unexpectedly extreme posterior based on little information can signal a coding, scaling, or model problem.

The posterior distribution is especially useful when a later calculation must average over uncertainty in the patient’s latent trajectory. A point estimate can be a convenient summary, but it is not generally a substitute for that distribution. The update also has limits: it is only as reliable as the fitted longitudinal model, the random-effects distribution, and the assumptions about the observation process. If population parameters are estimated with substantial uncertainty, a plug-in posterior can understate the full uncertainty unless parameter uncertainty is propagated.

Common Mistake

A common mistake is to set the patient’s random effect equal to the difference between one observed biomarker and its expected value, then use that value as if it were known. This ignores measurement error and the population distribution. For example, a large residual could reflect a large random effect, a noisy measurement, or both. The posterior update balances these explanations through the likelihood and prior; it also retains uncertainty about the random effect. In the Gaussian case, the posterior covariance \(V_i\) makes that remaining uncertainty explicit.