Frailty Selection and Marginal Hazard Attenuation

Why Marginal Hazards Change Over Time

A frailty model can have a simple proportional hazards structure for individuals with the same frailty, while its population-level hazard changes in a more complicated way. The reason is selection: among individuals who remain event-free, those with higher frailty become progressively less common because they tend to experience the event earlier. Thus the composition of the risk set changes over time, even when no individual's conditional hazard ratio changes.

The distinction is between a conditional hazard, defined for a given frailty value, and a marginal hazard, averaged over the frailty distribution of those still event-free. This tutorial derives that change in composition and its consequences. The main selection argument concerns time to a first event, or another setting in which an event removes an individual from the risk set. It should not be applied without modification to recurrent-event data in which individuals remain under observation and can experience further events.

Model and assumptions: Let $Z$ be a positive, time-constant frailty, independent of covariates at baseline, with $E(Z)=1$. For a covariate value $x$, suppose the hazard conditional on $Z$ is $$ \lambda(t\mid x,Z)=Zq(x)\lambda_0(t), \qquad q(x)=\exp(\beta x), $$ where $\lambda_0(t)$ is the baseline hazard and $H_0(t)=\int_0^t\lambda_0(u)\,du$ is the baseline cumulative hazard. Conditional on $Z$ and $x$, the event-free survival function is $\exp\{-Zq(x)H_0(t)\}$. The frailty is fixed over time, and censoring or entry into the risk set must not selectively alter the frailty distribution in a way omitted from the model.

The conditional hazard ratio comparing $x_1$ with $x_0$, for the same frailty value, is $q(x_1)/q(x_0)$ at every time. But a marginal hazard compares people who are still event-free in each covariate group. Their frailty distributions need not be the same as at baseline, or as each other. We first derive this survivor selection explicitly.

The Frailty Distribution Among Survivors

Write $s=q(x)H_0(t)$, the cumulative exposure for covariate value $x$, and let $L(s)=E\{\exp(-sZ)\}$ be the Laplace transform of the baseline frailty distribution. Conditional survival is $\exp(-sZ)$. By Bayes' rule, the frailty distribution among individuals event-free at time $t$ is therefore the baseline distribution reweighted in proportion to $\exp(-sZ)$.

Definition: For any function $g$ for which the expectation exists, define its survivor-weighted expectation at exposure $s$ by $$ E_s\{g(Z)\}=\frac{E\{g(Z)\exp(-sZ)\}}{E\{\exp(-sZ)\}}. $$ The mean frailty among survivors is $A(s)=E_s(Z)$. Frailties with larger values receive less weight as $s$ increases, because their event-free probability $\exp(-sZ)$ decreases more rapidly.

Marginal survival is obtained by averaging conditional survival over the baseline frailty distribution. Differentiating its logarithm gives the marginal hazard:

\bar S(t\mid x)=L\{q(x)H_0(t)\}, \qquad \bar\lambda(t\mid x)=q(x)\lambda_0(t)A\{q(x)H_0(t)\}.

Thus the marginal hazard is the conditional multiplier $q(x)\lambda_0(t)$ times the mean frailty in the current survivor group. The multiplier $A(s)$ equals one at baseline because $E(Z)=1$, but generally declines as survivor selection proceeds.

Why the Survivor-Weighted Mean Decreases

The decline can be shown directly, rather than inferred only from the reweighting intuition. Assume the expectations needed below are finite so that differentiation under the expectation is valid. Let $M_k(s)=E\{Z^k\exp(-sZ)\}$. Then $A(s)=M_1(s)/M_0(s)$, $M_0'(s)=-M_1(s)$, and $M_1'(s)=-M_2(s)$. The quotient rule gives:

\begin{aligned} A'(s) &=\frac{-M_2(s)M_0(s)+M_1(s)^2}{M_0(s)^2}\\ &=-\frac{E\{Z^2\exp(-sZ)\}}{E\{\exp(-sZ)\}} +\left[\frac{E\{Z\exp(-sZ)\}}{E\{\exp(-sZ)\}}\right]^2\\ &=-\operatorname{Var}_s(Z)\leq 0. \end{aligned}

The first ratio in the second line is the survivor-weighted second moment, and the squared ratio is the square of the survivor-weighted mean. Their difference, with the negative sign shown, is minus the survivor-weighted variance. Consequently, the mean frailty among survivors cannot increase with cumulative exposure. It decreases strictly wherever the survivor-weighted frailty distribution has positive variance; if frailty is constant, there is no selection effect.

Worked Example: Selection in a Two-Point Frailty Distribution

Suppose $Z$ equals $0.5$ or $1.5$, each with probability $0.5$. This hypothetical distribution has mean one. At cumulative exposure $s=1$, calculate the probability that a survivor has frailty $1.5$, the mean frailty among survivors, and the derivative of that mean with respect to exposure.

The survivor weights are proportional to the baseline probabilities times $\exp(-sZ)$. The common baseline probability $0.5$ cancels in the ratio, so the probability of frailty $1.5$ among survivors is $$ \frac{\exp(-1.5)}{\exp(-0.5)+\exp(-1.5)} =\frac{1}{1+\exp(1)} \approx 0.2689. $$

The survivor-weighted mean is the lower frailty plus the difference between the two frailty values times the probability of the higher value: $$ A(1)=0.5+(1.5-0.5)(0.2689)=0.7689. $$

For a two-point distribution separated by one unit, the variance is $p(1-p)$, where $p=0.2689$ is the survivor probability of the higher value. Hence $$ A'(1)=-\operatorname{Var}_1(Z) =-0.2689(1-0.2689)\approx -0.1966. $$

The mean frailty among those still event-free is about $0.7689$, below its baseline value of $1$, and it is declining at exposure $s=1$. The derivative is with respect to cumulative exposure, not calendar time; the rate of change over time also depends on how quickly $H_0(t)$ increases.

Marginal Hazard Ratios and Attenuation

For two covariate values, the marginal hazard ratio follows from the marginal hazard formula:

\frac{\bar\lambda(t\mid x_1)}{\bar\lambda(t\mid x_0)} = \frac{q(x_1)A\{q(x_1)H_0(t)\}} {q(x_0)A\{q(x_0)H_0(t)\}}.

The conditional hazard ratio is the fixed quantity $q(x_1)/q(x_0)$; the marginal ratio also includes the two survivor-weighted means. When $q(x_1)>q(x_0)$, the higher-risk group accumulates exposure faster and generally experiences stronger depletion of high-frailty individuals. Its marginal hazard can therefore become closer to the lower group's marginal hazard. The precise pattern depends on the frailty distribution, so a general frailty distribution does not imply that every marginal hazard ratio must decrease monotonically or converge to one. Gamma frailty gives a particularly clear, exact example.

Gamma frailty result: Suppose $Z$ has a gamma distribution with mean $1$ and variance $\theta>0$. Its Laplace transform is $L(s)=(1+\theta s)^{-1/\theta}$, so $$ A(s)=(1+\theta s)^{-1}, \qquad \bar\lambda(t\mid x)=\frac{q(x)\lambda_0(t)}{1+\theta q(x)H_0(t)}. $$ For $q(x_1)>q(x_0)>0$, the marginal hazard ratio is $$ \frac{q(x_1)}{q(x_0)} \frac{1+\theta q(x_0)H_0(t)} {1+\theta q(x_1)H_0(t)}. $$ It starts at the conditional ratio when $H_0(t)=0$, decreases as $H_0(t)$ increases, and tends to $1$ as $H_0(t)$ tends to infinity.

Worked Example: Gamma Frailty Attenuates a Hazard Ratio

Consider a hypothetical gamma frailty model with variance $\theta=0.5$. Compare a reference group with $q(x_0)=1$ and a higher-risk group with $q(x_1)=2$. At a time when $H_0(t)=1$, calculate both marginal hazards as multiples of $\lambda_0(t)$ and their ratio.

For the reference group, $$ \bar\lambda(t\mid x_0) =\frac{1\lambda_0(t)}{1+0.5(1)(1)} =\frac{1}{1.5}\lambda_0(t) \approx 0.6667\lambda_0(t). $$ For the higher-risk group, $$ \bar\lambda(t\mid x_1) =\frac{2\lambda_0(t)}{1+0.5(2)(1)} =\frac{2}{2}\lambda_0(t) =1.000\lambda_0(t). $$ Thus the marginal hazard ratio is $1.000/0.6667=1.500$, below the conditional hazard ratio $2/1=2$.

At $H_0(t)=2$, the marginal ratio is $$ 2\frac{1+0.5(1)(2)}{1+0.5(2)(2)} =2\frac{2}{3} \approx 1.333. $$ The ratio has moved closer to one as cumulative baseline hazard increased. These are hazard ratios at the specified cumulative hazards; they are not ratios of cumulative event probabilities.

Worked Example: The Marginal Survival and Risk-Set Frailty

Use the same gamma frailty variance $\theta=0.5$, but consider $q(x)=1$ and $H_0(t)=1$. Compute marginal event-free survival, the mean frailty among survivors, and the marginal hazard relative to the baseline hazard.

The gamma Laplace transform gives $$ \bar S(t\mid x)=\{1+0.5(1)(1)\}^{-1/0.5} =(1.5)^{-2} =\frac{4}{9}\approx 0.4444. $$

The mean frailty among survivors is $$ A(1)=\frac{1}{1+0.5(1)}=\frac{2}{3}\approx 0.6667. $$ Consequently, the marginal hazard is $$ \bar\lambda(t\mid x) =1\lambda_0(t)\frac{2}{3} \approx 0.6667\lambda_0(t). $$

The event-free probability $0.4444$ is an average over frailty, not the survival probability for an individual with $Z=1$, which would be $\exp(-1)\approx0.3679$. The marginal hazard is lower than the baseline conditional hazard because the survivors have mean frailty below one.

Assumptions, Diagnostics, and Comparisons

The derivation relies on a well-defined multiplicative frailty model: $Z$ is nonnegative, fixed over the follow-up considered, and independent of the covariates at baseline; conditional on $Z$ and the covariates, the stated hazard determines conditional survival. The risk set must represent event-free individuals under the modeled process. Informative censoring, selective delayed entry, frailty that changes over time, or dependence between frailty and covariates can alter the survivor distribution and invalidate the simple formula unless explicitly modeled.

A marginal hazard ratio that changes over time is not, by itself, proof of frailty. A changing baseline or covariate effect, nonproportional hazards, unmodeled subgroups, or selection into observation can also produce time variation. Useful checks include comparing the model-implied marginal survival and hazard patterns with observed time patterns, checking whether the assumed frailty distribution gives plausible survivor selection, and assessing sensitivity to alternative frailty distributions. A close match does not establish that latent frailty is the unique explanation: different mechanisms can produce similar marginal patterns.

Quantity	What it conditions or averages over	Key implication
Conditional hazard	Individuals with a specified frailty $Z$	Proportional hazards holds by the model specification.
Marginal hazard	The frailty distribution among those still event-free	Can vary relative to another group's hazard as survivor composition changes.
Cluster-robust standard errors	Sampling uncertainty with within-cluster dependence	They do not, by themselves, model the survivor-weighted frailty distribution.

These approaches answer different questions. A conditional frailty model describes risk at a given latent frailty, while its marginal hazard describes the event rate in the evolving population of survivors. Cluster-robust standard errors can address uncertainty under dependence for an estimating procedure, but do not create a frailty-based model of risk-set selection. Neither conditional nor marginal hazard ratios are automatically causal effects; causal interpretation requires additional design and identification assumptions.

For recurrent events, the event-free survivor calculation applies directly only when the event of interest removes a person from the risk set, such as a first event or terminal event. If people remain at risk after each recurrence, a prior event can provide information about their frailty, while remaining under observation without an event can also affect its distribution. The relevant risk-set distribution then depends on the observation and event history. Treating every recurrence as if it permanently removed the individual would give the wrong selection argument.

Common Mistakes and Practical Interpretation

Confusing the two hazards. The conditional hazard ratio is constant in the specified model; the marginal hazard ratio need not be. State which population and conditioning set a reported ratio refers to.
Giving the derivative the wrong sign. The derivative is $-\operatorname{Var}_s(Z)$, not $+\operatorname{Var}_s(Z)$. Survivor weighting lowers the mean frailty as exposure increases.
Claiming universal convergence to one. The gamma model has this limiting behavior under the conditions shown. It is not a universal theorem for every frailty distribution and every covariate comparison.
Calling attenuation evidence of a changing individual effect. The marginal change can occur even though the conditional coefficient $\beta$ is constant; it reflects the changing mix of survivors.
Applying first-event selection to all recurrences. If individuals remain at risk after an event, event history and subsequent risk-set membership must be represented appropriately.

Key takeaway: Frailty selection is a change in who remains event-free. Survivor weighting makes the mean frailty $A(s)$ nonincreasing, with $A'(s)=-\operatorname{Var}_s(Z)$. The marginal hazard is the conditional hazard multiplied by this survivor-specific mean, so conditional proportional hazards need not imply marginal proportional hazards.

Check Your Understanding

Answer each question, then compare with the solutions below.

In one sentence, explain why the frailty distribution among event-free individuals differs from its baseline distribution.
For gamma frailty with variance $\theta$, write the survivor-weighted mean frailty and state its value at $s=0$.
With $\theta=1$, $q(x_0)=1$, $q(x_1)=3$, and $H_0(t)=1$, calculate the marginal hazard ratio and compare it with the conditional ratio.
Why does the selection formula for event-free survivors not automatically describe a recurrent-event risk set where individuals can experience multiple events?

Solutions. (1) Event-free probability is proportional to $\exp(-sZ)$, which weights higher frailties less as exposure increases. (2) $A(s)=1/(1+\theta s)$, so $A(0)=1$. (3) The marginal ratio is $3\{1+1(1)(1)\}/\{1+1(3)(1)\}=6/4=1.500$, compared with the conditional ratio $3$. (4) A recurrence does not necessarily remove an individual from observation; the risk-set frailty distribution can depend on both remaining event-free over intervals and the individual's event history.

Quantity	What it conditions or averages over	Key implication
Conditional hazard	Individuals with a specified frailty \(Z\)	Proportional hazards holds by the model specification.
Marginal hazard	The frailty distribution among those still event-free	Can vary relative to another group's hazard as survivor composition changes.
Cluster-robust standard errors	Sampling uncertainty with within-cluster dependence	They do not, by themselves, model the survivor-weighted frailty distribution.