Using Event-Free Follow-Up to Refine Latent Risk

Why Event-Free Follow-Up Contains Information

A patient’s biomarker history is not the only information that can distinguish plausible latent trajectories. If two trajectories imply different event hazards, then remaining event-free for a period changes how plausible those trajectories are. All else equal, a trajectory predicting a higher hazard assigns a lower probability to observing no event over that period. Event-free follow-up therefore acts as evidence about latent risk; it is not merely an absence of data.

The update is probabilistic. It does not establish which trajectory a patient follows, and it does not make event-free status a cause of a particular trajectory. Rather, the joint model combines the observed biomarker history and the observed survival status to revise the posterior distribution over latent trajectories. The resulting distribution can then be used in a prediction that averages over trajectory uncertainty.

Definition: Let \(b\) denote a patient’s latent random effects, which determine the patient-specific longitudinal trajectory and, through the joint model, the event hazard. Let \(Y(s)\) denote the biomarker measurements observed through time \(s\). The event-free survival likelihood contribution through \(s\), conditional on \(b\), is \(S(s\mid b)=P(T>s\mid b)\), where \(T\) is the event time. The posterior distribution of \(b\) given both \(Y(s)\) and event-free follow-up through \(s\) is proportional to the prior density of \(b\), the longitudinal-data likelihood, and \(S(s\mid b)\).

Model and Assumptions for the Update

Suppose the event hazard conditional on the latent effects is \(h(t\mid b)\). The corresponding cumulative hazard and survival function are

H(s\mid b)=\int_0^s h(u\mid b)\,du, \qquad S(s\mid b)=\exp\{-H(s\mid b)\}.

Thus, the probability of remaining event-free through \(s\) is smaller for a trajectory with a larger cumulative hazard over that interval. For a patient whose biomarker history has likelihood \(f\{Y(s)\mid b\}\) and whose latent effects have prior density \(p(b)\), Bayes’ rule gives

p\{b\mid Y(s),T>s\} = \frac{p(b)f\{Y(s)\mid b\}S(s\mid b)} {\int p(v)f\{Y(s)\mid v\}S(s\mid v)\,dv}.

This expression relies on the joint-model assumption that, conditional on \(b\), the longitudinal measurements and event time are independent. The biomarker likelihood and survival likelihood may therefore be multiplied conditional on \(b\); their association in the observed data is represented through their shared dependence on \(b\). The calculation also presumes that the longitudinal and survival models are adequately specified, that the patient is known to be event-free through \(s\), and that the measurement process is ignorable under the model. If the observation or censoring process carries additional information not represented in the model, it may require separate modeling.

A useful equivalent form starts with the posterior based on biomarker history alone, \(p\{b\mid Y(s)\}\). The event-free update multiplies that density by \(S(s\mid b)\) and normalizes:

p\{b\mid Y(s),T>s\} = \frac{p\{b\mid Y(s)\}S(s\mid b)} {\int p\{v\mid Y(s)\}S(s\mid v)\,dv}.

The survival term is a likelihood for the observed event-free status, not a penalty chosen by the analyst. Trajectories that make the observed survival status more probable receive more posterior weight.

Deriving the Posterior Weight Update

For clarity, first suppose the patient may follow one of \(K\) latent trajectory classes, indexed by \(k\). Let \(w_k\) be the class probability after accounting for the biomarker history but before using the event-free follow-up. Let \(S_k(s)\) be the probability of being event-free through \(s\) for class \(k\). The joint probability of class \(k\) and event-free status, conditional on the biomarker history, is \(w_kS_k(s)\). Summing over classes gives the probability of event-free status under that history, \(\sum_{j=1}^K w_jS_j(s)\). Dividing the joint probability by this sum yields

w_k^{\,\text{updated}} = \frac{w_kS_k(s)} {\sum_{j=1}^K w_jS_j(s)}.

The same logic applies to continuous random effects, with integrals replacing sums. In practice, the continuous posterior is often evaluated by numerical integration or approximated using weighted draws. The essential operation remains multiplication by each draw’s survival likelihood followed by normalization.

Worked Example: Updating Two Trajectory Weights From Survival

Hypothetical data. After observing a patient’s biomarker history, suppose the posterior weights for two latent trajectories are \(0.40\) and \(0.60\). The trajectories imply constant hazards of \(0.02\) and \(0.06\) per year, respectively. The patient is known to be event-free through 4 years. Assume the constant-hazard model is appropriate over this period and that the weights already account for the biomarker history.

For a constant hazard \(\lambda\), survival through time \(s\) is \(S(s)=e^{-\lambda s}\). The two survival likelihoods are therefore \(e^{-0.02(4)}=e^{-0.08}\) and \(e^{-0.06(4)}=e^{-0.24}\). The unnormalized updated weights are

0.40e^{-0.08}\approx 0.3692465, \qquad 0.60e^{-0.24}\approx 0.4719767.

Their sum is approximately \(0.8412233\). After normalization, the posterior weights are

\frac{0.3692464}{0.8412232}=0.4389, \qquad \frac{0.4719768}{0.8412232}=0.5611.

The first, lower-hazard trajectory gains posterior weight, from \(0.40\) to approximately \(0.4389\), because event-free survival for 4 years is more probable under that trajectory. The higher-hazard trajectory still has the larger posterior weight, \(0.5611\), since its starting weight was larger; an update need not make the lower-risk trajectory the most probable one.

Updating Over a Follow-Up Interval

An update must use the survival information that has accrued since the distribution being updated was formed. Suppose the current weights already condition on survival to an earlier landmark \(s_0\). If the patient remains event-free through a later time \(s_1\), the relevant conditional survival likelihood for trajectory \(k\) is

P(T>s_1\mid T>s_0,k) = \frac{S_k(s_1)}{S_k(s_0)}, \qquad s_1>s_0.

This ratio follows from the definition of conditional probability: \(P(T>s_1\mid T>s_0,k)=P(T>s_1,T>s_0\mid k)/P(T>s_0\mid k)\). Since \(T>s_1\) implies \(T>s_0\), the numerator is \(S_k(s_1)\), and the denominator is \(S_k(s_0)\). Multiplying by \(S_k(s_1)\) instead would count survival to \(s_0\) again if it was already incorporated into the current weights.

Worked Example: Using Only New Event-Free Follow-Up

Hypothetical data. At an earlier landmark, after accounting for the available marker history and survival to that landmark, a patient’s trajectory weights are \(0.70\) for class A and \(0.30\) for class B. Over the next 2 years, assume constant class-specific hazards of \(0.03\) and \(0.09\) per year. The patient remains event-free throughout that interval. These weights are already conditioned on being event-free at the start of the interval.

The conditional survival likelihoods over the 2-year interval are \(e^{-0.03(2)}=e^{-0.06}=0.9417645\) and \(e^{-0.09(2)}=e^{-0.18}=0.8352702\). Multiplying by the starting weights gives

0.70(0.9417645)=0.6592352, \qquad 0.30(0.8352702)=0.2505811.

The normalizing constant is approximately \(0.9098162\). Therefore, the updated weights are

\frac{0.6592352}{0.9098163}=0.7246, \qquad \frac{0.2505811}{0.9098163}=0.2754.

Class A’s weight rises from \(0.70\) to \(0.7246\), because it assigns a higher probability to remaining event-free during the interval. The calculation uses survival over the new 2-year interval only; it does not reapply the likelihood for survival up to the earlier landmark.

From Updated Weights to Conditional Risk

The updated weights are not themselves an event-risk prediction. They describe uncertainty about the patient’s latent trajectory after considering the observed information. To obtain a conditional risk over a future horizon, average the trajectory-specific risks using those weights. If the patient is event-free at time \(s\), then under trajectory \(k\), the probability of an event by \(s+\Delta\) is \(1-S_k(s+\Delta)/S_k(s)\). Consequently,

P(T\leq s+\Delta\mid T>s,\text{observed history}) = \sum_{k=1}^{K} w_k^{\,\text{updated}} \left\{1-\frac{S_k(s+\Delta)}{S_k(s)}\right\}.

Worked Example: Predicting Risk After the Weight Update

Hypothetical data. Use the posterior weights \(0.4389\) and \(0.5611\) from the first example, which incorporate the patient’s biomarker history and event-free follow-up through 4 years. Suppose the patient-specific hazards remain \(0.02\) and \(0.06\) per year, and consider the risk of an event during the next 3 years. Under a constant hazard, the class-specific conditional risks are \(1-e^{-0.02(3)}=1-e^{-0.06}=0.0582355\) and \(1-e^{-0.06(3)}=1-e^{-0.18}=0.1647298\).

Weighting those risks gives

0.4389(0.0582355)+0.5611(0.1647298) \approx 0.02556+0.09243=0.1180.

The estimated event risk over the next 3 years, conditional on being event-free at 4 years and on the information used to form the weights, is approximately \(11.80\%\). For comparison, using the original weights \(0.40\) and \(0.60\) would give \(0.40(0.0582355)+0.60(0.1647298)=0.1221\), or \(12.21\%\). The lower risk after updating reflects the greater weight placed on the lower-hazard trajectory, not a change in either class-specific hazard.

Diagnostics, Alternatives, and Limitations

A weight update can be computed correctly while still being untrustworthy if the model or the inputs are wrong. First check the time origin and interval: the survival likelihood must correspond to the period represented by the observed event-free status. Next verify that the patient’s event status and censoring time are coded consistently with the definition of \(S(s\mid b)\). For a continuous random-effects calculation, check that integration or posterior sampling adequately represents the regions where both the biomarker likelihood and survival likelihood are appreciable.

The magnitude of the update is also informative. If trajectories imply nearly identical survival probabilities over the observed period, their relative weights should change little. If the posterior changes dramatically, inspect the hazard scale, time units, cumulative-hazard calculation, and starting weights. A posterior concentrated on extreme latent values can reflect genuine information, but it can also signal a poorly fitting trajectory or survival model. Comparing predicted and observed survival patterns across relevant patient histories and checking sensitivity to reasonable model specifications can help identify such problems.

Approach	What it uses	Main implication
Biomarker-history weights only	Longitudinal likelihood and prior for the latent effects	Does not use event-free status to distinguish trajectories.
Survival-updated weights	Biomarker history plus the survival likelihood for the observed event-free period	Uses both data sources while retaining uncertainty over trajectories.
Single plug-in trajectory	One estimated or most likely trajectory	Can omit uncertainty and may give a different risk than averaging over trajectories.

The survival-updated approach is appropriate only to the extent that the joint model captures the relevant dependence and the hazard is adequately specified. Event-free follow-up may add little information when the candidate trajectories have similar hazards, or when the interval is short. Conversely, a long interval can strongly favor trajectories with lower cumulative hazards. Such selection among latent trajectories is a consequence of conditioning on survival; it does not show that surviving changed the patient’s underlying trajectory.

A common mistake is to treat the patient’s event-free status as irrelevant because no event time was recorded. The correction is that event-free status through \(s\) has likelihood \(S(s\mid b)\), which can vary with \(b\) and therefore change posterior weights. A related mistake is to multiply by survival from the study origin after starting with weights already conditioned on survival to a landmark. In that case, use the conditional survival ratio \(S(s_1\mid b)/S(s_0\mid b)\) for the additional interval. These updates are model-based associations, not causal effects of remaining event-free.

Key takeaway: Event-free follow-up refines latent-risk uncertainty through a survival likelihood. Multiply each trajectory’s current posterior weight by its probability of surviving the newly observed interval, then normalize; use the updated mixture, rather than a single selected trajectory, for risk calculations.

Check Your Understanding

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

A patient’s current trajectory weights are \(0.50\) and \(0.50\). Over an observed interval, the two trajectories have survival probabilities \(0.90\) and \(0.60\). What are the updated weights?
If weights already condition on survival through year 3, what survival likelihood should be used to update them using event-free follow-up through year 5?
Why can event-free follow-up change posterior trajectory weights even when no new biomarker value is observed?
Under what circumstance should event-free follow-up have little effect on relative trajectory weights?

Solutions. 1. The unnormalized weights are \(0.50(0.90)=0.45\) and \(0.50(0.60)=0.30\), with sum \(0.75\). The updated weights are \(0.45/0.75=0.60\) and \(0.30/0.75=0.40\). 2. Use the conditional survival likelihood \(S(5\mid b)/S(3\mid b)\), not \(S(5\mid b)\), because survival through year 3 is already represented in the current weights. 3. The event-free observation has likelihood \(S(s\mid b)\), which differs across trajectories when their hazards differ; Bayes’ rule therefore changes their relative posterior weights. 4. The update should be small when the candidate trajectories assign similar probabilities to surviving the observed interval.