Identifying Shared Random-Effects Association Parameters

Why the Shared Association Parameter Needs Two Sources of Information

In a joint model, a subject’s repeated biomarker measurements and event history are connected through a latent subject-specific effect. The association parameter quantifies how that latent effect changes the event hazard. To understand what identifies this parameter, it helps to separate two questions: how well do the longitudinal observations locate each subject’s latent effect, and how strongly do event histories vary with that effect? Information about the parameter comes from the connection between those two sources, not from either one in isolation.

This tutorial focuses on a random-intercept model. The intercept represents a subject’s persistent departure from the population’s mean biomarker trajectory. The model is deliberately simple so that the roles of longitudinal measurement and survival outcomes are visible. It is not a claim that every biomarker trajectory is constant over time.

Definition: For subject $i$, let $Y_{ij}$ be the biomarker measured at visit $j$, and let $b_i$ be a latent random intercept. A random-intercept longitudinal model is $$ Y_{ij}=\beta_0+\boldsymbol{x}_{ij}^{\mathsf T}\boldsymbol{\beta}+b_i+\epsilon_{ij},\qquad b_i\sim N(0,\tau^2),\qquad \epsilon_{ij}\sim N(0,\sigma^2). $$ A shared-random-effects survival model links that same $b_i$ to the event hazard: $$ h_i(t\mid b_i)=h_0(t)\exp\{\boldsymbol{w}_i^{\mathsf T}\boldsymbol{\gamma}+\alpha b_i\}. $$ Here $h_0(t)$ is the baseline hazard, $\boldsymbol{x}_{ij}$ and $\boldsymbol{w}_i$ are covariate vectors, and $\alpha$ is the shared-random-effects association parameter.

For the basic formulation, assume the $b_i$ are independent across subjects and independent of the longitudinal errors. Conditional on $b_i$ and the covariates, the longitudinal errors are independent with mean zero and variance $\sigma^2$. Conditional on $b_i$ and survival covariates, the event process follows the specified hazard model. The longitudinal and event processes are conditionally independent given the shared effect and covariates. The random effects have the stated normal distribution, and censoring is noninformative conditional on the modeled quantities. These are modeling assumptions, not consequences of observing repeated measurements.

The parameter $\alpha$ is on the log-hazard scale per unit of $b_i$. If two subjects have latent intercepts differing by $\Delta b$, while their survival covariates are the same, their hazard ratio is $\exp(\alpha\Delta b)$. Thus $\alpha>0$ means a higher latent biomarker intercept is associated with a higher hazard, conditional on the model; it does not establish a causal effect. If the biomarker is measured in units such as milligrams per liter, then $\alpha$ has inverse-milligrams-per-liter units. Rescaling the biomarker changes the numerical value of $\alpha$.

What Each Part of the Data Contributes

Repeated biomarker measurements help locate $b_i$. When the fixed effects are known or have been adjusted for, define the residual $r_{ij}=Y_{ij}-\beta_0-\boldsymbol{x}_{ij}^{\mathsf T}\boldsymbol{\beta}$. A sequence of residuals consistently above zero points toward a positive intercept; one isolated high observation is less persuasive because it could be measurement error. The amount of repeated-measurement information depends on the number of visits and on the ratio of between-subject variation $\tau^2$ to within-subject variation $\sigma^2$.

Event histories provide information about $\alpha$ when their hazards differ systematically with $b_i$. A subject’s event time or censoring time contributes through the cumulative hazard as well as, for an observed event, the hazard at the event time. If every subject had the same known $b_i=0$, the hazard would not depend on $\alpha$ at all. More generally, event data cannot reveal a relationship with $b_i$ unless there is variation in $b_i$ and the event process contains information about how hazard changes across that variation.

The effects are estimated jointly, not by first treating estimated subject intercepts as error-free data. In the marginal likelihood, each subject’s latent effect is integrated out. Longitudinal observations influence which values of $b_i$ are plausible; the event history then contributes evidence about $\alpha$ across those plausible values. If longitudinal information about $b_i$ is diffuse, the event process may not clearly distinguish a weak association from uncertainty about the latent effect.

Worked Example: How Repeated Measurements Locate an Intercept

Suppose four hypothetical biomarker residuals for one subject are $0.4, 0.6, 0.5,$ and $0.7$ units. For this calculation, take the fixed effects as known, and assume $\tau=0.8$ units and $\sigma=0.6$ units. Under the normal model, the conditional distribution of $b_i$ given these residuals is normal. Its variance $V_i$ and mean $m_i$ are $$ V_i=\left(\frac{1}{\tau^2}+\frac{n_i}{\sigma^2}\right)^{-1}, \qquad m_i=V_i\frac{\sum_j r_{ij}}{\sigma^2}. $$

The four residuals sum to $2.2$ units. Substitution gives $$ V_i=\left(\frac{1}{0.8^2}+\frac{4}{0.6^2}\right)^{-1} =(1.5625+11.1111)^{-1}\approx 0.07890\text{ units}^2. $$ The posterior mean is $$ m_i=\left(\frac{1}{1/0.8^2+4/0.6^2}\right)\frac{2.2}{0.36} \approx 0.4822\text{ units}. $$

The four observations suggest a positive intercept, but the mean is smaller than the residual average $2.2/4=0.55$ units. This shrinkage toward the population mean of zero reflects the finite precision of the measurements and the assumed distribution of subject intercepts. The calculation describes how the longitudinal data inform $b_i$; by itself it says nothing about the survival association $\alpha$.

Likelihood Contribution and a Score Derivation

Let $T_i$ denote the observed follow-up time, $\delta_i$ indicate whether an event was observed ($\delta_i=1$ for an event and $0$ for censoring), and $H_0(T_i)=\int_0^{T_i}h_0(u)\,du$ be the cumulative baseline hazard. Conditional on $b_i$, the event-process log-likelihood contribution, up to terms not involving $\alpha$, is $$ \ell_{Ti}(\alpha\mid b_i)=\delta_i\{\boldsymbol{w}_i^{\mathsf T}\boldsymbol{\gamma}+\alpha b_i\} -H_0(T_i)\exp\{\boldsymbol{w}_i^{\mathsf T}\boldsymbol{\gamma}+\alpha b_i\}. $$

Differentiating with respect to $\alpha$ gives the conditional score:

U_i(\alpha\mid b_i) =\frac{\partial \ell_{Ti}}{\partial\alpha} =b_i\left[\delta_i-H_0(T_i)\exp\{\boldsymbol{w}_i^{\mathsf T}\boldsymbol{\gamma}+\alpha b_i\}\right].

This expression makes the two ingredients visible. The effect $b_i$ multiplies the contribution: if $b_i=0$, this subject’s conditional event likelihood has no sensitivity to $\alpha$. The bracket compares the observed event indicator with the cumulative-hazard term implied by the model. An observed event and a censored observation do not contribute in the same way, and follow-up time matters through $H_0(T_i)$.

The conditional score is not generally the score used to estimate $\alpha$, because $b_i$ is unobserved. To see the marginal version, write the subject’s likelihood as $$ L_i(\alpha)=\int L_{Yi}(b)L_{Ti}(b;\alpha)\phi(b)\,db, $$ where $L_{Yi}$ is the longitudinal likelihood and $\phi$ is the random-intercept density. If those two terms and the random-effects distribution have no direct dependence on $\alpha$, differentiating under the integral gives $$ \frac{\partial \log L_i(\alpha)}{\partial\alpha} =\int \frac{\partial\ell_{Ti}(\alpha\mid b)}{\partial\alpha} \,p(b\mid Y_i,T_i,\delta_i;\alpha)\,db =E\{U_i(\alpha\mid b_i)\mid Y_i,T_i,\delta_i;\alpha\}. $$ Thus, the marginal score averages the conditional event score over the latent effects supported by both data processes. This identity explains why better longitudinal measurement can sharpen information about the association: it can make the distribution used in that average more concentrated. It also explains why simply inserting a point estimate of $b_i$ and ignoring its uncertainty is not equivalent to joint likelihood estimation.

Worked Example: Event Contributions at a Trial Association Value

Consider two hypothetical subjects at a trial value $\alpha=0.7$, with no survival covariates and cumulative baseline-hazard exposures $H_0(T_A)=0.4$ and $H_0(T_B)=0.6$. Suppose their latent intercepts are $b_A=0.5$ and $b_B=-0.4$, in biomarker units. Subject A has an observed event ($\delta_A=1$); subject B is censored ($\delta_B=0$). Treat these latent values and exposures as specified for this conditional-score calculation.

For subject A, $\alpha b_A=0.35$. The score contribution is $$ b_A[\delta_A-H_0(T_A)\exp(\alpha b_A)] =0.5[1-0.4\exp(0.35)] \approx 0.2162\text{ per biomarker unit}. $$ For subject B, $\alpha b_B=-0.28$, so $$ b_B[\delta_B-H_0(T_B)\exp(\alpha b_B)] =-0.4[0-0.6\exp(-0.28)] \approx 0.1814\text{ per biomarker unit}. $$ The two conditional contributions sum to approximately $0.3976$ per biomarker unit. At this trial value, the conditional event log-likelihood for these specified latent effects has a positive local slope with respect to $\alpha$. That is not an estimate, standard error, or test result: actual joint estimation integrates over uncertainty in $b_i$, includes all subjects, and estimates nuisance parameters as appropriate.

A Small Information Calculation

A useful way to isolate the role of latent-effect variation is to consider a simplified setting with known $b_i$, no survival covariates, and a fixed censoring horizon whose baseline cumulative hazard is $q$ for each subject. For a proportional-hazards model with hazard multiplier $\exp(\alpha b_i)$, the probability of an event by that horizon is $1-\exp\{-q\exp(\alpha b_i)\}$. The expected information about $\alpha$ from that subject is $$ \mathcal{I}_i(\alpha\mid b_i) =b_i^2\left[1-\exp\{-q\exp(\alpha b_i)\}\right]. $$

One justification follows from the event-process score. The observed information is $b_i^2\exp(\alpha b_i)H_0(\min(T_i,C_i))$. Under the model and a fixed horizon, the expected cumulative baseline exposure until the event or horizon is $\{1-\exp(-q\exp(\alpha b_i))\}/\exp(\alpha b_i)$. Multiplying these expressions gives the stated expected information. This calculation assumes the hazard form is correct and the latent effects are known; in the joint model they are not known, so it is an illustration, not a formula for the full observed information.

Worked Example: Comparing Two Amounts of Latent-Effect Variation

Suppose two hypothetical subjects have known intercepts $b_1=0.5$ and $b_2=-0.5$ units, with $q=0.5$ and $\alpha=0.7$. The expected information under the simplified calculation is the sum of their contributions. For $b_1=0.5$, $$ 0.5^2[1-\exp\{-0.5\exp(0.35)\}] =0.25[1-\exp(-0.7095)] \approx 0.1270\text{ per squared association unit}. $$ For $b_2=-0.5$, $$ (-0.5)^2[1-\exp\{-0.5\exp(-0.35)\}] =0.25[1-\exp(-0.3523)] \approx 0.07424\text{ per squared association unit}. $$ Together they contribute approximately $0.2013$ in this simplified setting.

For comparison, if the two intercepts were $1$ and $-1$ units, keeping $q$ and $\alpha$ fixed, the sum would be $$ 1^2[1-\exp\{-0.5\exp(0.7)\}] +(-1)^2[1-\exp\{-0.5\exp(-0.7)\}] \approx 0.8545\text{ per squared association unit}. $$ The larger value illustrates how wider variation in known latent effects can increase information about an association. It does not imply that a fitted joint model will have this exact information: measurement error, uncertainty in the intercepts, baseline-hazard estimation, and other parameters alter the information available in practice.

Assumptions, Diagnostics, and Alternatives

The interpretation of $\alpha$ depends on the model being a reasonable description of both processes. In particular, the random-intercept structure must capture a meaningful persistent biomarker difference; the hazard must depend on that effect through the specified linear term $\alpha b_i$; and the conditional-independence and censoring assumptions must be defensible. The normal random-effects and Gaussian-error assumptions are also part of the model used in the posterior calculation above.

Useful checks therefore include examining the longitudinal residual pattern and whether the assumed random-intercept model leaves systematic subject-specific trends, and checking whether event predictions are adequate across relevant risk groups. In joint-model simulation or predictive checks, compare both the longitudinal measurements and event outcomes with replicated data, rather than relying on fit to only one process. For Bayesian fits, chain diagnostics and simulation-based calibration address computational behavior and calibration under the assumed model; they do not establish that the model assumptions are scientifically correct.

Warning signs for weak association information include very few observed events, short follow-up with little accumulated baseline hazard, little variation in the latent intercepts, or longitudinal measurements too noisy or sparse to distinguish subject effects. A nearly flat marginal likelihood over plausible values of $\alpha$, large uncertainty, or sensitivity to modeling assumptions can also indicate limited information. These symptoms do not by themselves identify the cause: for example, uncertainty about $b_i$ and limited event information can both make the association difficult to estimate. Diagnose the data support and model fit before interpreting a numerical estimate as a stable association.

The joint likelihood is the main alternative to a two-stage procedure that first estimates subject-specific intercepts and then uses those estimates as if observed in a survival model. The two-stage approach can be convenient, but it generally fails to propagate uncertainty in $b_i$ and can distort the estimated association, particularly when the intercept estimates are imprecise. Fitting a longitudinal model and a survival model separately does not estimate their shared association at all. Conversely, a joint model can estimate the relationship while accounting for latent-effect uncertainty, but it does not guarantee strong identification or protect against a misspecified link, baseline hazard, or random-effects distribution.

The introductory material on structural and practical identifiability distinguishes whether a parameter is uniquely determined by the model from whether the available data estimate it precisely. Here, the practical question is what variation supplies information: longitudinal replication helps resolve the latent intercepts, and event histories across those intercepts inform the hazard association. Neither source alone guarantees a precise estimate. The random-intercept formulation also has limits: it does not represent subject-specific biomarker slopes, and a real association may depend on a current biomarker level or a more complex trajectory rather than a persistent intercept.

Common Pitfall

Common mistake: “The longitudinal data estimate each subject’s random intercept, so those estimates alone identify the survival association.” The correction is that the longitudinal data inform the latent intercepts, while the event process supplies evidence about how hazard changes with those intercepts. Since the intercepts are estimated with uncertainty, the joint likelihood integrates over that uncertainty rather than treating a single estimated value as known.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions below.

If $\alpha=0.4$ per biomarker unit, what hazard ratio corresponds to a one-unit increase in the latent intercept?
Under the random-intercept model, what does the conditional event score imply when $b_i=0$?
For a hypothetical subject with $b_i=0.3$, $\delta_i=0$, and $H_0(T_i)\exp(\boldsymbol{w}_i^{\mathsf T}\boldsymbol{\gamma}+\alpha b_i)=0.5$, calculate the conditional score contribution.
Why can repeated, precise biomarker measurements help estimate $\alpha$ even though they are not event outcomes?

Solutions. (1) The hazard ratio is $\exp(0.4)\approx 1.492$ for a one-unit increase, conditional on the modeled covariates and shared-effect structure. (2) The score is zero, because $b_i$ multiplies the bracket; for that subject the conditional hazard does not depend on $\alpha$. (3) The score is $b_i(\delta_i-\text{cumulative-hazard term})=0.3(0-0.5)=-0.15$ per biomarker unit. (4) Repeated precise measurements concentrate the plausible values of $b_i$. The survival likelihood can then more clearly relate event histories to those latent values, while still accounting for their remaining uncertainty in the joint likelihood.