Event Follow-Up And Information For Association Parameters

Why Follow-Up Is Not the Same as Information

A survival association parameter describes how a shared latent quantity connects the longitudinal and event processes. But the number of participants enrolled, the length of follow-up, and even the number of observed events do not by themselves tell us how precisely that association can be estimated. Participants must be observed while at risk, and their latent marker values or random effects must vary enough for the survival process to distinguish different association strengths.

This tutorial examines that connection in a shared-random-effects model. The focus is not a universal events-per-parameter rule. Instead, we derive a simple conditional information calculation, use it to compare follow-up patterns, and then explain why latent random effects and nuisance parameters make real joint-model information more complicated.

A Shared-Random-Effects Survival Model

Definition: Let $T_i$ be the event time for subject $i$, $C_i$ the censoring time, $Y_i=\min(T_i,C_i)$ the observed follow-up time, and $\delta_i=I(T_i\leq C_i)$ the event indicator. A shared-random-effects model uses subject-specific latent effects $b_i$ in both the longitudinal and event processes. One common survival specification is $$ h_i(t\mid b_i)=h_0(t)\exp\{\gamma^\top w_i+\alpha m_i(t;b_i)\}, $$ where $h_i(t\mid b_i)$ is the event hazard, $h_0(t)$ is the baseline hazard, $w_i$ denotes observed survival covariates, $m_i(t;b_i)$ is a longitudinal latent quantity such as the current underlying marker value, and $\alpha$ is the survival association parameter.

The parameter $\alpha$ measures how the specified latent quantity is associated with the instantaneous event hazard, conditional on the model’s other terms. For example, when $m_i(t;b_i)$ is a marker level, a one-unit increase in that level multiplies the hazard by $\exp(\alpha)$, holding the other model terms fixed. This is a conditional association, not a causal effect.

To isolate the role of follow-up, first consider a deliberately simplified conditional model with a known, time-constant random effect $b_i$, no survival covariates, and a constant baseline hazard $\lambda$:

h_i(t\mid b_i)=\lambda\exp(\alpha b_i).

Assume censoring is administrative and fixed at $c_i$, or more generally is independent of the event process conditional on the modeled covariates and random effects. The simplified calculation treats $b_i$ as known. In a joint model, $b_i$ is latent and must be integrated out using the longitudinal and survival data; therefore, this calculation isolates a source of information rather than giving the exact information in every fitted joint model.

How Event Observation and Follow-Up Enter Information

For the simplified model, the conditional event-time likelihood contribution, up to factors not involving $\alpha$, is

L_i(\alpha\mid b_i)= \{\lambda\exp(\alpha b_i)\}^{\delta_i} \exp\{-\lambda\exp(\alpha b_i)Y_i\}.

Writing the log likelihood and differentiating gives the conditional score and observed information:

\ell_i(\alpha)=\delta_i\{\log(\lambda)+\alpha b_i\} -\lambda\exp(\alpha b_i)Y_i,

U_i(\alpha)=\frac{\partial\ell_i}{\partial\alpha} =b_i\{\delta_i-\lambda\exp(\alpha b_i)Y_i\}, \qquad J_i(\alpha)=-\frac{\partial^2\ell_i}{\partial\alpha^2} =\lambda\exp(\alpha b_i)b_i^2Y_i.

An event contributes the term $\delta_i b_i$ to the score. The exposure term also matters: a subject who remains event-free for a long time contributes evidence through the accumulated hazard. The observed information depends on the actual time at risk $Y_i$, not simply on the planned censoring time $c_i$.

For a fixed censoring time $c_i$, the event time has an exponential distribution with rate $r_i=\lambda\exp(\alpha b_i)$. Thus $Y_i=\min(T_i,c_i)$, and

E(Y_i\mid b_i) =\int_0^{c_i}P(T_i>t\mid b_i)\,dt =\int_0^{c_i}\exp(-r_it)\,dt =\frac{1-\exp(-r_ic_i)}{r_i}.

Taking the expectation of $J_i(\alpha)$ yields the expected conditional information:

\mathcal I_i(\alpha\mid b_i) =E\{J_i(\alpha)\mid b_i\} =b_i^2\{1-\exp[-\lambda\exp(\alpha b_i)c_i]\}.

This is also $b_i^2$ times the conditional probability of an event by $c_i$. In this one-event-per-subject model, longer follow-up increases the probability of observing an event, but that probability approaches one rather than increasing without bound. When the cumulative hazard is small, $1-\exp(-x)\approx x$, so the information is approximately $b_i^2\lambda\exp(\alpha b_i)c_i$. For longer follow-up, this approximation can be poor.

With a time-varying latent quantity $m_i(t;b_i)$, the corresponding idealized expected conditional information, with the latent trajectory treated as known, is

\mathcal I_i(\alpha\mid b_i) =\int_0^{c_i}S_i(t\mid b_i)\, h_i(t\mid b_i)\,m_i(t;b_i)^2\,dt,

where $S_i(t\mid b_i)=P(T_i\geq t\mid b_i)$. This expression shows the roles of the risk set, hazard, and magnitude of the latent quantity at each time. It also makes clear why both the timing of events and the timing of longitudinal measurements can matter: they affect how well the relevant latent trajectory is known. The earlier material on longitudinal visit timing and slope identifiability addresses the measurement side of that issue.

Worked Information Calculations

Worked Example: Expected Events and Information With Fixed Follow-Up

Hypothetical data. Suppose 80 subjects have the same known $b_i=0.5$, the baseline hazard is $\lambda=0.12$ per year, and each is followed administratively for 6 years. Evaluate expected events and conditional information at $\alpha=0$.

At $\alpha=0$, each subject’s cumulative hazard over 6 years is $0.12(6)=0.72$. Therefore, the event probability by 6 years is

1-\exp(-0.72)\approx 0.5132477.

The expected event count is

80\{1-\exp[-0.12(6)]\} =80(1-\exp(-0.72)) \approx 80(0.5132477) =41.06\text{ events}.

Expected conditional information per subject is $b_i^2$ times the event probability:

$$ (0.5)^2(0.5132477)=0.1283119. $$

For all 80 subjects, the expected information is

$$ 80(0.1283119)=10.26495. $$

Thus the cohort is expected to have about 41.06 events, but its expected conditional information is about 10.26 in the model’s information scale. The event count and information are different quantities; here the squared effect value, $b_i^2$, scales the latter.

Worked Example: Unequal Follow-Up Changes Expected Information

Hypothetical data. Consider 80 subjects with known $b_i=0.5$, $\lambda=0.10$ per year, and $\alpha=0$. Forty subjects are followed for 2 years and 40 for 8 years. Compare the expected event total and information.

For 2 years, the event probability per subject is

1-\exp[-0.10(2)]=1-\exp(-0.2)\approx 0.1812692.

For 8 years, it is

1-\exp[-0.10(8)]=1-\exp(-0.8)\approx 0.5506710.

The expected total event count is

40(0.1812692)+40(0.5506710) =7.250768+22.02684 =29.27761\text{ events}.

Since $b_i^2=0.25$ for every subject, expected information is $0.25$ times the expected event count:

$$ 0.25(29.27761)=7.319403. $$

The longer-followed group contributes more expected information in this setup because it has a higher probability of an observed event. But a person followed for 8 years does not automatically contribute four times as much information as a person followed for 2 years: event-free survival and the one-event-per-subject structure cause the contribution to level off.

Worked Example: The Same Expected Event Count Can Mean Different Information

Hypothetical data. Compare two cohorts of 100 subjects, all followed for 5 years, with $\lambda=0.20$ per year and $\alpha=0$. In both cohorts, the known random effects are balanced around zero. In Cohort A, each $b_i$ is either $-0.2$ or $0.2$; in Cohort B, each is either $-1$ or $1$.

At $\alpha=0$, the event probability is the same for both cohorts:

1-\exp[-0.20(5)]=1-\exp(-1)\approx 0.6321206.

Each cohort therefore has the same expected event count:

100(0.6321206)=63.21206\text{ events}.

For Cohort A, $b_i^2=0.04$, so expected information is

$$ 100(0.04)(0.6321206)=2.528482. $$

For Cohort B, $b_i^2=1$, giving

$$ 100(1)(0.6321206)=63.21206. $$

The event totals are equal, while the conditional information differs greatly because the association is expressed through $b_i$. This idealized calculation treats the effects as known. In a fitted joint model they are latent, so their uncertainty and the estimation of other parameters also affect the information available for $\alpha$.

Assumptions, Diagnostics, and Comparisons

The calculations above rely on a specified hazard model, correctly measured follow-up and event status, and censoring that is independent of the event time conditional on the modeled information. They also treat the random effects or latent trajectory as known. These conditions are useful for understanding the calculation; they are not a claim that real joint-model random effects are observed.

In a shared-random-effects model, the observed likelihood integrates over the distribution of $b_i$. Longitudinal data help estimate each subject’s latent trajectory, but sparse or noisy measurements can leave that trajectory uncertain. Consequently, the conditional information formula is an idealized guide: it does not by itself equal the information for $\alpha$ after latent effects and nuisance parameters have been estimated. Correlation between the latent effects and other model components can further reduce information about the association parameter.

Useful checks therefore go beyond a single event total:

Tabulate events and censoring across ranges of the longitudinal marker or predicted latent values. If almost all events occur in a narrow range, the data may provide little contrast for estimating association.
Summarize the number at risk over time, event times, and follow-up distributions. A long nominal study duration is not informative for every subject if many are censored early.
Check whether follow-up or censoring patterns differ systematically with observed covariates or longitudinal history. If censoring remains related to event risk after conditioning on the model, the independent-censoring assumption may be implausible.
Use the fitted model or a planning model to calculate predicted event probabilities and information across plausible follow-up designs, while treating those calculations as conditional on the assumed model.

A common alternative is to summarize exposure using total person-time or to compare studies by the number of observed events. These are helpful descriptive summaries, but neither accounts by itself for the association signal carried by the latent values. A model-based calculation incorporates the hazard and the magnitude of the latent quantity; a design simulation can additionally reflect uncertainty in latent effects and nuisance parameters. An events-per-parameter rule is at most a rough planning heuristic, not a guarantee of precise association estimation.

Summary or approach	What it captures	What it can miss
Event count	How many events were observed	Latent-value variation, timing, and uncertainty in random effects
Total person-time	Aggregate time observed at risk	Which subjects contributed that time and the latent values during it
Conditional information calculation	Hazard, follow-up, and known latent values under a stated model	Uncertainty from integrating over latent effects and estimating other parameters
Design simulation	Expected behavior under a complete assumed data-generating model	Misspecification of the assumed model or censoring mechanism

The central limitation is that follow-up duration is not a substitute for variation in the shared effect or accurate measurement of it. Longer follow-up can increase the chance of observing events, but it cannot create information about a parameter if the latent quantity is effectively constant at zero, nor can it repair severe uncertainty about the latent trajectory. Conversely, a smaller event count may still be useful if events occur across a broad range of well-measured latent values.

Common Pitfall

Key takeaway: Do not infer that equal event counts imply equal information about a survival association parameter. Information also depends on when subjects were at risk, the latent values associated with their risk, and how well those values are estimated. The conditional calculations in this tutorial clarify these factors but do not replace assessment of the full joint likelihood.

A frequent mistake is to multiply the planned follow-up time by the baseline hazard and call the result each subject’s information. That ignores the fact that subjects who experience an event stop contributing to the risk set. In the constant-hazard example, expected information is $b_i^2\{1-\exp(-\lambda\exp(\alpha b_i)c_i)\}$, not $b_i^2\lambda c_i$ in general. The latter is only a small-cumulative-hazard approximation.

Check Your Understanding

Answer each question using the conditional model and assumptions stated above. Solutions follow the questions.

At $\alpha=0$, $b=0.4$, $\lambda=0.05$ per year, and fixed follow-up $c=4$ years, calculate the event probability and expected information for one subject.
Why does expected information approach $b^2$, rather than grow without bound, as fixed follow-up becomes very long in the simplified model?
Two cohorts have equal expected event counts, but one has random effects closer to zero. Which cohort has less conditional information at $\alpha=0$, and why?
Name two reasons the known-random-effect information calculation may overstate the information available for $\alpha$ in an actual joint model.

Solutions. (1) The cumulative hazard is $0.05(4)=0.20$, so the event probability is $1-\exp(-0.20)\approx0.1812692$. Since $b^2=0.16$, information is $0.16(0.1812692)\approx0.0290031$. (2) The event probability $1-\exp(-\lambda c)$ approaches 1 as $c$ increases, so expected information approaches $b^2$. Each subject can experience at most one event in this setup. (3) The cohort with effects closer to zero has less information, because at $\alpha=0$ each subject’s information is $b_i^2$ times the event probability. (4) In a joint model, random effects are latent and uncertain, and other model parameters must also be estimated; either can reduce the information specifically available for $\alpha$ compared with the known-effect calculation.