Why the Level of Frailty Matters
Frailty represents unobserved multiplicative variation in event risk. When observations are grouped, a key modeling decision is which units share that unobserved variation. Relatives in a family may share background risk, patients at the same hospital may share institutional conditions, and repeated events from one person may share that person’s underlying susceptibility. These structures are not interchangeable: each assigns latent risk to a different level of the data.
This tutorial focuses on specifying that level. It does not require choosing a particular frailty distribution or estimating its parameters. The central question is whether the latent multiplier is common to a whole cluster, varies among members within a cluster, or includes both components.
Basic Model Structures
Let \(i\) index clusters and \(j\) index members within cluster \(i\). Let \(U_i\) denote a shared frailty and \(V_{ij}\) an individual frailty. A general combined structure for a member’s event hazard is
Here \(h_0(t)\) is a baseline hazard, \(X_{ij}(t)\) is the member’s covariate vector, and \(\beta\) is the associated coefficient vector. The hazard is specified conditional on the covariates and the latent frailties. For recurrent-event outcomes, the same structural idea can be used for the member’s event intensity while that member is under observation and at risk. Event-order effects or other history terms, when needed, belong in the event-process specification separately from the frailty structure.
A shared-only model sets \(V_{ij}=1\); an individual-only model sets \(U_i=1\). A combined model includes both. For a model to separate these latent components on a common scale, a usual convention is that each positive frailty has mean one. When both terms are included, a basic specification also assumes they are independent of one another and of the modeled covariates. Those are modeling assumptions, not consequences of the data structure; they should be reconsidered if there is substantive reason to expect dependence.
The mean-one convention makes the baseline scale interpretable as the average frailty multiplier under the assumed model. It does not say that every cluster or individual has multiplier one. The particular distributions and estimation methods for these latent terms are separate choices.
Why the Combined Structure Has Two Levels
The product form makes the levels explicit. Taking logarithms of the frailty contribution gives
Thus a combined multiplicative specification is equivalent, on the log-hazard scale, to adding a cluster-level latent term and a member-level latent term. Every member of cluster \(i\) receives the same \(\log(U_i)\), but each receives their own \(\log(V_{ij})\). This algebra explains why a shared-only model cannot represent member-to-member latent variation within a cluster, and why an individual-only model cannot explicitly represent a common cluster component.
This decomposition is meaningful only when the data contain the relevant levels. With just one member per cluster, the observed multiplier for that member is a product of the two terms, so the data structure alone does not provide multiple members with which to distinguish a cluster component from a member component. More generally, very small or few clusters can make the distinction difficult in practice.
Worked Example: Families With Several Members
Suppose a hypothetical study follows three members in each family for a first event. A family-level latent factor is considered plausible because members share an environment, while individual susceptibility may also differ. For one family, suppose the specified shared multiplier is \(U_i=1.20\); for one of its members, suppose \(V_{ij}=0.80\). At a particular time and covariate value, let the baseline-and-covariate component be \(0.05\) events per person-year.
Under a combined structure, that member’s conditional hazard is
The other family members share the factor \(1.20\), but they do not share this member’s factor \(0.80\). If another member has \(V_{ik}=1.10\), the corresponding calculation is \(1.20(1.10)(0.05)=0.0660\) events per person-year. These numbers illustrate the model’s arithmetic; they are hypothetical specified multipliers, not estimates from a real study.
A shared-only structure would use \(1.20(0.05)=0.0600\) for both members at the same covariate value. It therefore cannot express the stipulated individual difference. An individual-only structure could represent different member multipliers, but would omit the common family factor. The combined form matches both sources of variation only if the scientific rationale and data support both levels.
Choosing a Structure for the Data
The cluster label is determined by the scientific grouping, not simply by the presence of repeated rows. A family, hospital, or other site can define a cluster if its members plausibly share latent risk. The member is the unit to which individual frailty is attached. In recurrent-event data, repeated events from a person are observations on the same member, not new members with independently redrawn individual frailties.
| Data structure | Possible shared level | Possible individual level | Structural question |
|---|---|---|---|
| Relatives followed for an event | Family | Each relative | Could family background and relative-specific susceptibility both vary? |
| Patients treated at hospitals | Hospital | Each patient | Could institutional conditions and patient susceptibility both vary? |
| Repeated events per patient | A higher-level group, if present | Each patient across recurrences | Does one patient-level latent propensity persist across events? |
Worked Example: Patients Nested Within Hospitals
Consider a hypothetical study of patients treated at four hospitals. Suppose the scientific model allows both hospital-level latent variation and patient-level latent variation. At a specified time and covariate value, let the baseline-and-covariate hazard be \(0.020\) per patient-year. For a patient at Hospital A, take \(U_A=1.30\) and \(V_{A1}=0.90\). Then
For a second patient at the same hospital with \(V_{A2}=1.20\), the shared factor remains \(1.30\), and the calculation is \(1.30(1.20)(0.020)=0.0312\) events per patient-year. For a patient at Hospital B, the hospital-level factor would instead be \(U_B\). The values are hypothetical and serve only to show where each multiplier enters.
A hospital-only model would impose the same latent multiplier on all patients at Hospital A, apart from their modeled covariates. A patient-only model would omit the common hospital-level factor. Choosing between these structures should follow the plausible sources of unmeasured variation and the number and size of hospitals and patients represented in the data—not the numerical example.
Worked Example: Persistent Frailty Across Recurrent Events
Suppose a hypothetical patient experiences recurrent events and is followed over time. The model assigns that patient an individual frailty \(V_j=1.40\), which persists across the patient’s observed recurrences. Assume the baseline-and-covariate component is \(0.030\) events per person-month at one event order and \(0.025\) at another, perhaps because the model specifies different event-order components.
At the first event order, the frailty-scaled intensity is \(1.40(0.030)=0.0420\) events per person-month. At the other event order it is \(1.40(0.025)=0.0350\) events per person-month. The event-order component changes, while the patient-level frailty remains the same.
If patients are also grouped into clinics and a clinic-level term is included, the intensity has both multipliers, \(U_iV_{ij}\). By contrast, assigning a fresh independent frailty to every recurrence would define an event-specific latent structure, not the persistent individual-frailty structure described here. That alternative would answer a different scientific question about how latent risk varies over events.
Assumptions, Diagnostics, and Failure Modes
A frailty structure is a substantive model choice. Before using one, specify the cluster and member indices, state which frailty terms are included, and clarify whether each term persists over the observation period. The usual conditional model also assumes that the specified hazard or intensity form is appropriate given the covariates, latent terms, and relevant event history. If members’ event processes are assumed independent conditional on all frailties and covariates, that conditional-independence assumption should be stated rather than left implicit.
Several checks help reveal a mismatch between the intended structure and the available data or model:
- Verify the hierarchy. Confirm that each member is assigned to the intended cluster and that repeated events remain attached to the same person.
- Check replication at each level. A shared term is informed by multiple members per cluster; a member term requires multiple members, and recurrent-event variation requires suitable repeated follow-up. Sparse replication can leave the levels difficult to distinguish.
- Assess whether the structure has a substantive basis. A cluster label in the data does not by itself establish a shared latent risk factor. Consider whether members plausibly share unmeasured influences.
- Inspect the fit and sensitivity to structure. Compare scientifically plausible shared-only, individual-only, and combined specifications. Large changes in fitted results or unstable estimates across reasonable choices signal that conclusions may depend on the structural assumption.
- Separate latent variation from measured predictors and event history. A frailty term should not be used as a catch-all substitute for omitted measured covariates or a needed event-order or time-varying effect.
These checks do not prove that a particular latent structure is correct. Frailties are unobserved, so diagnostics can reveal incompatibility, weak support, or sensitivity, but cannot directly verify the latent values or guarantee that all relevant heterogeneity has been modeled.
How the Main Structures Differ
The alternatives differ in what is shared, not merely in the name given to a random effect. A shared-only model is most direct when the target latent variation is common to cluster members. An individual-only model is appropriate when latent variation is member-specific and there is no modeled cluster component. A combined model represents both, at the cost of additional assumptions and a greater demand on the data’s hierarchical structure. A model without frailty may be appropriate when the scientific question and data do not justify a latent multiplier; it should not be replaced automatically by a frailty model just because observations are clustered.
In all cases, a frailty term is distinct from an observed covariate: it represents latent multiplicative risk variation. It is also distinct from a fixed event-order effect, which describes a systematic change associated with recurrence number. Model components should be chosen to reflect these different roles.
Check Your Understanding
Answer each question, then compare with the brief solutions below.
- A study follows siblings from multiple families. Which term would be common to siblings in one family in a shared-plus-individual model?
- In a combined model, calculate the frailty-scaled hazard if the baseline-and-covariate hazard is \(0.04\) per year, \(U_i=0.75\), and \(V_{ij}=1.50\).
- A patient has six recurrences. In the usual persistent individual-frailty structure, how many distinct patient-level frailty terms are assigned to that patient?
- Why is a combined hospital-and-patient frailty structure difficult to distinguish from one product multiplier when every hospital contributes only one patient?
Solutions. (1) The family-level \(U_i\) is common to the siblings; each sibling has their own \(V_{ij}\). (2) \(0.75(1.50)(0.04)=0.045\) events per year. (3) One, shared across that patient’s observed recurrences; event-order effects, if included, are separate. (4) With one patient per hospital, the data provide no multiple patients within a hospital to distinguish a hospital multiplier from that patient’s multiplier; the observed product does not by itself identify which level generated it.