Why Identifiability Matters
A joint model can be fitted by a sophisticated algorithm and still fail to answer a basic statistical question: do the observed data distinguish the parameters being estimated? If two different parameter values imply exactly the same distribution for all observable data, no increase in sample size can separate them under that model and design. If they imply different distributions that are nevertheless very similar, finite data may provide too little information to estimate them precisely.
These are different problems. The first is structural identifiability, a property of the model and observation design. The second is practical identifiability, a matter of how informative the available finite sample is. Neither is the same as a numerical fitting failure, although weak information can make computation more difficult.
Model and Definitions
For subject \(i\), let \(Y_i=(Y_{i1},\ldots,Y_{in_i})\) denote the longitudinal measurements, \(T_i\) the event time, \(C_i\) the censoring time, \(U_i=\min(T_i,C_i)\) the observed follow-up time, and \(\Delta_i=I(T_i\le C_i)\) the event indicator. Let \(b_i\) be a subject-specific random effect. A common shared-random-effects model specifies a random-effects distribution, a longitudinal distribution conditional on \(b_i\), and an event hazard conditional on \(b_i\). For example,
Here \(x_i\) denotes observed covariates, \(\lambda_0(t)\) is a baseline hazard, and \(g_i\) describes how the latent effects enter the event hazard. A standard assumption for this formulation is conditional independence of the longitudinal and event processes given the random effects and covariates. The particular form of \(g_i\) is model-specific; the identifiability concepts below do not require choosing one.
The observed-data distribution is obtained by integrating over the unobserved \(b_i\). With right censoring, a contribution based on the longitudinal and event processes alone has the form below only when the censoring mechanism is ignorable for the parameters of interest.
A sufficient censoring condition for omitting the censoring-process contribution is that censoring be independent of the event and longitudinal processes and of \(b_i\), conditional on observed covariates, with censoring parameters distinct from the parameters of interest. If censoring depends on \(b_i\), it can reveal information about that latent effect even when censoring and event times are independent conditional on \(b_i\). In that case its contribution belongs inside the integral. For example, under conditional independence of event and censoring times given \(b_i\), the event-censoring factor is \(f_T(u_i\mid b)S_C(u_i\mid b)\) when \(\Delta_i=1\), and \(S_T(u_i\mid b)f_C(u_i\mid b)\) when \(\Delta_i=0\). The observation scheme is part of the distribution whose identifiability is being assessed.
Structural identifiability concerns the population distribution implied by the model, not whether one realized sample happens to look informative. Practical identifiability concerns the latter: with the available sample size, event count, visit schedule, and covariate variation, distinct parameter values may imply distributions that are so similar that estimates remain highly uncertain. It is therefore relative to both the model and the study design.
Structural Identifiability: Exact Equivalence
A direct way to investigate structural identifiability is to ask whether a change in parameters leaves the distribution of all observable quantities unchanged. If it does, the parameters cannot be uniquely recovered from those data. The following example illustrates an exact redundancy in a survival component.
Worked Example: An Intercept Confounded With an Unrestricted Baseline Hazard
Suppose a joint model has event hazard \(\lambda_i(t\mid b_i)=\lambda_0(t)\exp\{\gamma_0+\gamma x_i+g_i(t,b_i;\alpha)\}\), where \(\lambda_0(t)\) is unrestricted and \(\gamma_0\) is a separate intercept. For any constant \(c\), define \(\gamma_0'=\gamma_0+c\) and \(\lambda_0'(t)=\lambda_0(t)e^{-c}\).
Substitution gives
The conditional hazard is unchanged for every \(t\), covariate value, and random effect. Consequently the conditional event-time distribution is unchanged, as is the marginal distribution after integrating over \(b_i\). The longitudinal model is unchanged too, because this transformation alters no longitudinal parameter. Thus the observed-data distribution cannot distinguish \((\lambda_0,\gamma_0)\) from \((\lambda_0',\gamma_0')\). This parameterization is structurally nonidentifiable for \(\gamma_0\) and the baseline scale. A constraint on the baseline hazard or a parameterization without a redundant intercept can remove this particular equivalence; it does not automatically establish identifiability of every other parameter.
This argument illustrates a useful distinction: identifiability is a property of the complete parameterization, not of a coefficient considered in isolation. One might still identify a combination such as the product \(\lambda_0(t)e^{\gamma_0}\), even though its two factors are not separately identified.
Identifiability From Repeated Measurements
Not every latent-effect model contains an exact equivalence. Consider a longitudinal random-intercept component with repeated observations. Its population covariance structure can distinguish between variation across subjects and measurement error, provided the assumptions and design supply enough repeated measurements.
Worked Example: Separating Random-Intercept and Residual Variances
Assume independent subjects each have two measurements following \(Y_{ij}=\beta_0+b_i+\epsilon_{ij}\), where \(b_i\sim N(0,\tau^2)\), the errors \(\epsilon_{ij}\sim N(0,\sigma^2)\) are independent across visits and subjects, and the errors are independent of \(b_i\). For this calculation, take the measurements to be at the same modeled mean and assume no additional time or covariate terms.
The marginal mean of either observation is \(\beta_0\). Since the same random intercept contributes to both measurements, the marginal covariance matrix is
Suppose the population covariance matrix has diagonal entries \(1.00\) measurement-units squared and off-diagonal entries \(0.64\) measurement-units squared. Then \(\tau^2=0.64\), because the covariance between the two visits is due to the shared random intercept. The diagonal equation gives \(\tau^2+\sigma^2=1.00\), so \(\sigma^2=1.00-0.64=0.36\) measurement-units squared. The mean identifies \(\beta_0\), and these covariance entries uniquely recover both variance components under the stated model.
This is a statement about the implied population distribution. In a real finite sample, empirical covariance entries are uncertain, so estimates of \(0.64\) and \(0.36\) need not be precise. Also, the conclusion depends on the assumptions: correlated residual errors or a different random-effects structure would change the covariance equations and therefore the identifiability question.
In a joint model, the event process may contribute additional information about latent effects, but that does not guarantee that every parameter becomes identifiable. The observed-data distribution must be examined under the full model, including the actual visit and follow-up design.
Practical Identifiability: Different but Hard to Distinguish
Structural identifiability does not guarantee useful precision. Even when two parameter settings produce different distributions, their difference may be too small for the available sample to reveal. The next calculation gives a measure of how little information a small sample may provide about two distinct random-intercept variance settings.
Worked Example: Two Identifiable Variance Settings With Little Separation
Continue with the two-visit Gaussian model above, with mean zero and independent subjects. Compare \((\tau^2,\sigma^2)=(0.64,0.36)\) with \((0.70,0.30)\), in squared measurement units. Both settings have marginal variance \(1\), but their covariance matrices differ:
The settings are structurally distinguishable because their covariances differ. For one subject, the Kullback–Leibler divergence from the first bivariate normal distribution to the second is
Here \(\det(\Sigma_0)=1-0.64^2=0.5904\) and \(\det(\Sigma_1)=1-0.70^2=0.5100\). Also,
Therefore \(D_{\mathrm{KL}}(0\Vert1)\approx\frac12\{2.1647-2+\log(0.5100/0.5904)\}\approx0.0092\) nats per subject. For ten independent subjects, the divergence is approximately \(10(0.0092)=0.092\) nats. The distributions are not identical, but this small divergence indicates limited expected separation in such a sample. It does not give the probability that either setting is true, nor does it prove that an estimate will be inaccurate in every sample. It quantifies why finite-sample uncertainty can remain substantial despite structural identifiability.
In a joint model, limited information can arise when there are few events, sparse longitudinal visits, little covariate variation, or weak variation in the latent effects. These features can make different parameter values yield similar observed-data distributions. Practical identifiability should therefore be assessed for the actual design, not inferred solely from whether the model has a unique theoretical parameterization.
Censoring, Diagnostics, and Common Confusions
The observed-data distribution depends on what is observed and how observation ends. Suppose, in a simplified example, each subject has a latent class \(b\in\{0,1\}\) with prior probabilities \(P(b=0)=P(b=1)=0.5\). Let \(R=1\) denote one particular censoring pattern, with \(P(R=1\mid b=0)=0.8\) and \(P(R=1\mid b=1)=0.2\). Then
Observing this censoring pattern changes the latent-class probability from \(0.5\) to \(0.2\). A likelihood that drops \(R\) would discard information about \(b\); a likelihood that integrates over \(b\) must include the censoring factor \(P(R\mid b)\). In a continuous-time model, the corresponding censoring density or survival contribution must be handled in the same way. Conditional independence of censoring and the event time given \(b\) does not justify omitting a censoring mechanism that depends on \(b\).
Worked Example: Checking Whether an Identifiability Claim Is Justified
A hypothetical analysis reports a precise estimate of a parameter from a joint model. The analyst argues that this precision establishes that the parameter is identifiable. What additional checks are needed?
First, inspect the model parameterization for exact equivalences. For example, if an unrestricted baseline hazard and a separate event intercept enter only through their product, the calculation above shows that the intercept is not structurally identified, regardless of the reported standard error. Second, verify the assumptions behind the observed-data likelihood, especially whether censoring can inform the latent effects. Third, examine whether the design contains the repeated measurements, follow-up, and covariate variation needed to distinguish the relevant distributions.
Finally, distinguish information in the data from information supplied by modeling choices or priors. A prior can regularize a weakly informed direction and produce a concentrated posterior, but that concentration alone does not show that the likelihood contains strong information. A precise estimate is interpretable only in light of the model, likelihood, prior, and observation scheme.
Useful warning signs include estimates that change markedly under small changes to the parameterization or prior, wide uncertainty despite apparently stable fitting, and strong dependence among parameter estimates. These signs motivate further investigation; none alone proves structural nonidentifiability. A numerical algorithm can fail to converge even when a model is identifiable, and it can converge to a numerical solution even when the data do not identify all parameters.
| Situation | What it means | What it does not establish |
|---|---|---|
| Exact parameter transformation leaves the observed distribution unchanged | Structural nonidentifiability for at least part of the parameterization | That every function of the parameters is unidentifiable |
| Distinct parameter values imply different but similar distributions | Possible practical weakness for the available design and sample | Exact structural nonidentifiability |
| Algorithm fails to stabilize | A computational or numerical problem may be present | That the model is structurally nonidentifiable |
| Posterior is concentrated | The posterior is concentrated under the likelihood and prior used | That the data alone strongly identify the parameters |
A disciplined assessment proceeds in layers:
State the visit schedule, event and censoring information, covariates, and any assumptions that permit the censoring contribution to be omitted.
Ask whether distinct parameter settings can be transformed into one another while leaving the full observed-data distribution unchanged.
Even if the distribution map is one-to-one, consider whether the finite design and sample can practically distinguish nearby settings.
Interpret uncertainty and prior dependence as information questions, and treat numerical instability as a separate diagnostic problem.
The central limitation is that identifiability is conditional on a specified model and observation process. A conclusion under one random-effects distribution, censoring assumption, or visit design need not hold under another. Establishing identifiability is necessary for unique interpretation of parameters, but it does not guarantee low bias, adequate precision, or a correct model.
Check Your Understanding
Answer each question, then compare with the solutions below.
- In precise terms, what makes a parameter structurally identifiable?
- In the hazard model with an unrestricted \(\lambda_0(t)\) and intercept \(\gamma_0\), give a transformation that leaves the hazard unchanged.
- For the two-visit random-intercept model, a population covariance matrix has diagonal \(1.00\) and off-diagonal \(0.25\), in squared measurement units. Find \(\tau^2\) and \(\sigma^2\) under the stated assumptions.
- Can two parameter settings be structurally identifiable but practically difficult to distinguish? Explain.
- Why might censoring need to be included inside the integral over random effects?
Solutions. (1) Structural identifiability means that equality of observed-data distributions, \(P_\theta=P_{\theta'}\), implies \(\theta=\theta'\). (2) For any constant \(c\), set \(\gamma_0'=\gamma_0+c\) and \(\lambda_0'(t)=\lambda_0(t)e^{-c}\); their product in the hazard is unchanged. (3) The off-diagonal covariance gives \(\tau^2=0.25\); the diagonal equation gives \(\sigma^2=1.00-0.25=0.75\) squared measurement units. (4) Yes. Distinct parameters may imply different distributions, while a small finite sample provides too little information to separate those distributions precisely. (5) If censoring depends on the latent random effect, the observed censoring pattern updates its distribution. Its contribution must then be included in the integral rather than discarded.