Worked Diagnosis Of Weak Association Identification

When a Joint Model Cannot Pin Down the Association

A joint model can converge to a numerical estimate while still providing little evidence about the strength of the biomarker–survival association. The practical clue is often a profile likelihood that changes very little over a wide range of association values. In that situation, the reported point estimate may depend heavily on small changes in the data, model, or numerical settings.

This tutorial works through a hypothetical biomarker-survival analysis. The aim is not to decide whether a particular association is scientifically plausible, but to diagnose what the observed follow-up and measurement pattern can support. The earlier material on unstable-fit diagnostics provides the broader workflow; here the focus is the association parameter and the evidence specific to it.

Definition: In a shared-random-effects joint model, the longitudinal measurements are described through a subject-specific latent biomarker trajectory $m_i(t)$, and the event hazard depends on that trajectory. For a current-value association model, one common specification is $$ h_i(t\mid b_i)=h_0(t)\exp\{\gamma^\mathsf{T}w_i+\alpha m_i(t)\}. $$ Here $h_i(t\mid b_i)$ is the hazard for subject $i$ conditional on their random effects $b_i$, $h_0(t)$ is the baseline hazard, $w_i$ is a vector of survival covariates, $\gamma$ is its coefficient vector, and $\alpha$ is the association parameter. When the biomarker is measured in standard-deviation units, $\exp(\alpha)$ is the conditional hazard ratio for a one-unit increase in the latent biomarker value, holding the stated covariates and random effects structure fixed.

The longitudinal model and survival model are fitted together, so uncertainty in the latent trajectory affects inference about $\alpha$. A profile likelihood makes that dependence visible: fix $\alpha$ at a candidate value, maximize the joint log likelihood over the other parameters, and record the resulting maximized log likelihood. Write this as $\ell_p(\alpha)$. A broad, nearly level profile means that substantially different association values can fit almost equally well after the other parameters are re-optimized.

Conditions for Interpreting the Profile

The profile is meaningful only relative to a specified model and its assumptions. In the worked diagnosis, assume the following: the longitudinal trajectory model is an adequate description of the biomarker measurements; the hazard model has the stated current-value link; censoring is conditionally independent of the event process given the modeled information; and the event and measurement records are correctly aligned in time. The random-effects distribution and other likelihood assumptions must also be sufficiently appropriate for the fitted likelihood to be informative.

These assumptions are not established by a flat profile. They define what the profile assesses: how well the specified model distinguishes values of $\alpha$, conditional on its other assumptions. A poor trajectory model, informative dropout not represented in the model, or an incorrect hazard form can undermine interpretation even if the profile is sharply peaked. Conversely, a flat profile is evidence of weak information about $\alpha$, not proof that the true association is zero.

Under regular likelihood conditions, twice the difference between the best log likelihood and the profile log likelihood at a candidate value is approximately distributed as a chi-squared random variable with one degree of freedom:

\Lambda(\alpha)=2\{\ell_p(\hat{\alpha})-\ell_p(\alpha)\}\ \approx\ \chi^2_1.

Thus, an approximate 95% profile-likelihood confidence set contains values for which $\Lambda(\alpha)\leq 3.84$. This approximation relies on regularity conditions and can be unreliable in a very weakly identified or otherwise irregular problem. Treat it as a diagnostic summary, not as a guarantee that the model has enough information.

Worked Example: Reading a Nearly Flat Association Profile

Suppose a hypothetical analysis standardizes the biomarker to have a one-unit increase equal to one standard deviation. The maximized profile log likelihood is $-812.01$ at $\hat{\alpha}=1.0$. The profile log likelihood at $\alpha=0$ is $-812.18$, and at $\alpha=2.0$ it is $-812.10$.

At $\alpha=0$, the likelihood-ratio statistic is $$ 2\{(-812.01)-(-812.18)\}=2(0.17)=0.34. $$ At $\alpha=2.0$, it is $$ 2\{(-812.01)-(-812.10)\}=2(0.09)=0.18. $$ Both values are well below the approximate 95% cutoff of 3.84. The likelihood therefore does not distinguish either candidate sharply from the fitted maximum.

Suppose an extended profile grid crosses the 3.84 threshold at approximately $\alpha=-3.6$ and $\alpha=5.5$. The resulting approximate profile interval is extremely broad. On the hazard-ratio scale, its endpoints are $\exp(-3.6)\approx0.0273$ and $\exp(5.5)\approx245$, per one-standard-deviation increase in the latent biomarker. This range includes radically different association strengths and directions. The correct conclusion is weak identification in this fitted model, not that the hazard ratio is known to be near $\exp(1)\approx2.72$.

Why Follow-Up and Measurement Patterns Matter

The survival part of the likelihood informs $\alpha$ through event and censoring histories, while the longitudinal part informs the latent trajectory that enters the hazard. The association is especially difficult to learn when few events occur, when follow-up is short, or when biomarker measurements provide little information about the trajectory near the times subjects remain at risk. These conditions can occur together.

A measurement recorded after a subject's event cannot be used as a pre-event predictor of that event. For an event at time $T_i$, the hazard depends on the latent trajectory at times up to $T_i$; post-event measurements do not supply direct pre-event evidence about that risk. Sparse measurements before $T_i$ leave greater uncertainty about the relevant trajectory. The model may still estimate the trajectory using its assumptions, but that is not the same as having rich observed data near the event time.

The relationship between follow-up and measurements should therefore be inspected, not summarized only by the total number of biomarker records. Useful descriptions include the number of observed events, the distribution of follow-up times, how many pre-event measurements event cases contribute, and whether visits are concentrated early or late. These patterns do not prove that identification is weak; they provide a substantive explanation to compare with the likelihood evidence.

Worked Example: Checking the Event-Time Measurement Pattern

Consider a hypothetical dataset of 180 subjects followed for up to three years. There are 42 observed events and 138 right-censored subjects; median follow-up is 1.4 years. Of the 42 subjects with events, 35 have no more than one post-baseline biomarker measurement before their event. The remaining 7 have at least two such measurements.

The fraction of event cases with at most one post-baseline measurement before the event is $$ \frac{35}{42}=0.833, $$ or 83.3%. Only $7/42=0.167$, or 16.7%, of event cases contribute at least two post-baseline measurements before their event. If the planned schedule was every six months, this pattern indicates that many event cases contribute limited observed information about how their biomarker changed before the event.

This calculation does not establish that the model is wrong, and it does not quantify the association's information by itself. It does make a flat profile more understandable: only 42 events inform the survival association, and most event cases have sparse pre-event biomarker histories. If those histories also leave the latent trajectories uncertain, many values of $\alpha$ may be compensated by changes in trajectory or other model parameters.

A Short Justification: Flat Curvature Means Large Uncertainty

Near an interior maximum, a smooth profile log likelihood can be approximated locally by a quadratic:

\ell_p(\alpha)\approx \ell_p(\hat{\alpha})-\frac{1}{2}I_p(\hat{\alpha})(\alpha-\hat{\alpha})^2,

where $I_p(\hat{\alpha})$ denotes the local profile curvature, or profile information, in this approximation. Substitution into the likelihood-ratio statistic gives

2\{\ell_p(\hat{\alpha})-\ell_p(\alpha)\}\approx I_p(\hat{\alpha})(\alpha-\hat{\alpha})^2.

A small curvature means the likelihood-ratio statistic rises slowly as $\alpha$ moves away from its estimate. Under the same local quadratic approximation, the standard error is approximately $1/\sqrt{I_p(\hat{\alpha})}$, so small curvature corresponds to a large standard error. This explains why a point estimate can look precise to several decimal places while the likelihood supports a wide range of values: decimal precision is not information.

Worked Example: Relating Profile Curvature to a Standard Error

Suppose a local quadratic fit to a hypothetical profile gives estimated curvature $I_p=0.25$ per squared association unit. The approximate standard error is $$ SE\approx\frac{1}{\sqrt{0.25}}=\frac{1}{0.5}=2.0 $$ association units. At a displacement of one unit from the maximum, the approximate likelihood-ratio statistic is $0.25(1)^2=0.25$; at a displacement of two units, it is $0.25(2)^2=1.00$. Both are below 3.84. The local approximation therefore predicts a broad profile interval, consistent with weak information. This calculation is only a local diagnostic; the actual profile should be examined across the relevant range rather than inferred from curvature alone.

Choosing a Defensible Response

A remedy should address the source of weak information or make the target more limited and transparent. Different responses answer different questions. More follow-up and appropriately timed measurements can add information; simplifying the association structure can reduce the demands on the existing data; rescaling can make coefficients easier to interpret; and regularization can stabilize estimation by adding assumptions. None should be described as creating information that the observed data do not contain.

Response	What it can do	Key limitation
Collect longer follow-up or more pre-event measurements	Potentially add event and trajectory information relevant to the association	Requires new or additional data; more measurements alone do not guarantee more events or adequate model specification
Simplify the association structure	Reduce the number of association parameters the data must distinguish	Changes the scientific model and can omit relevant features of the trajectory
Rescale the biomarker	Put the coefficient on a more interpretable unit, such as one standard deviation	Changes coefficient units, not the underlying information or fitted predictions
Use penalization or a prior	Stabilize estimates by adding a penalty or prior information	Results can depend on that added information; this is not evidence that the likelihood alone identifies the association

Worked Example: Rescaling Without Claiming a Cure

Suppose the biomarker's standard deviation is 2.5 raw units, and the fitted association parameter is expressed per raw biomarker unit as $\alpha_{\mathrm{raw}}$. A one-standard-deviation increase is 2.5 raw units, so the corresponding coefficient is $$ \alpha_{\mathrm{SD}}=2.5\alpha_{\mathrm{raw}}. $$ If $\alpha_{\mathrm{raw}}=0.40$, then $\alpha_{\mathrm{SD}}=2.5(0.40)=1.00$. The hazard ratio per one-standard-deviation increase is $\exp(1.00)\approx2.72$, the same effect expressed on a different scale.

The log likelihood and predictions are unchanged by this re-expression when the model is transformed consistently. A flat profile remains flat after the corresponding change of units. Rescaling improves interpretability and can help numerical optimization, but it does not remedy weak identification.

If a simpler association model is scientifically defensible, it can be compared with a more complex one, but the comparison must be interpreted as a change in assumptions, not an automatic repair. For example, removing a poorly supported association component may narrow the question to a single current-value association. A narrower model may produce a more stable estimate because it asks less of the data; it can also be misleading if the omitted component matters scientifically. Report the rationale and show whether conclusions depend on that choice.

Penalized likelihood and Bayesian priors can likewise yield finite, stable estimates where the likelihood profile is broad. Their role is to combine the data with additional constraints or information. A responsible analysis identifies the penalty or prior, checks sensitivity to reasonable alternatives, and separates the resulting regularized or posterior summary from the information supplied by the likelihood.

Diagnostics, Failure Modes, and Limits

A flat profile is not interchangeable with a failed optimizer. Numerical convergence concerns whether the fitting algorithm has located and adequately characterized a solution. Weak identification concerns whether the observed data and specified model distinguish parameter values. An optimizer can converge reliably to the maximum of a shallow likelihood. Conversely, a poor numerical fit can make the profile appear unstable even when the underlying model would be more informative with accurate computation.

For a diagnosis, inspect a profile over a broad enough range to see its shape and whether it crosses a stated likelihood-ratio threshold. Check that nuisance parameters are re-optimized at each fixed $\alpha$; otherwise the curve is not the profile likelihood. Compare multiple starting values, gradients, Hessian-based uncertainty, and quadrature settings using the relevant diagnostics from earlier material. If the profile changes materially with numerical integration settings, computational error may be obscuring the statistical evidence.

Bootstrap instability and sensitivity to influential subjects can provide additional warning signs. These are not substitutes for the profile: they address related but distinct aspects of stability. In particular, a broad profile should not be reduced to “no association” just because the point estimate is modest or a test is not statistically significant. Nor should a large point estimate be presented as compelling evidence when the profile supports a wide range of effects.

Wrong reasoning: “The optimizer converged and returned $\hat{\alpha}=1.0$, so the association is identified.”
Correction: Convergence establishes a numerical result under the fitted procedure, not sharp information about $\alpha$. Examine the profile, uncertainty, and data pattern.
Wrong reasoning: “Standardizing the biomarker fixed the flat profile.”
Correction: Standardization changes the unit of $\alpha$; it does not add likelihood information.
Wrong reasoning: “A prior produced a narrow interval, so the data identify the effect.”
Correction: The narrow result may depend on prior information. Report that dependence and distinguish it from likelihood-based evidence.

The diagnostic remains conditional on the fitted joint model. A flat profile cannot, by itself, tell whether the cause is few events, sparse pre-event measurements, strong dependence among parameters, misspecification, or a combination. The follow-up summaries and model checks help explain the pattern, but cannot uniquely assign its cause. The central conclusion should therefore be calibrated: describe the weak identification shown by the profile, give the observed-data features that plausibly contribute, and state the limits of the proposed remedy.

Check Your Understanding

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

A profile log likelihood has maximum $-500.20$ and is $-500.35$ at $\alpha=0$. Calculate the likelihood-ratio statistic and state what it suggests relative to 3.84.
In a dataset, 24 of 30 event cases have at most one pre-event post-baseline measurement. What proportion is this, and what can it suggest?
Does converting a coefficient from per raw unit to per standard deviation create new information about the association? Explain.
Why is an optimizer's successful convergence not sufficient evidence that the association is well identified?

Solutions. (1) $2\{(-500.20)-(-500.35)\}=0.30$, below 3.84; the profile does not sharply distinguish zero from its maximum by this approximate criterion. (2) $24/30=0.80$, or 80%; it indicates that most event cases have sparse observed pre-event biomarker histories, which may leave their relevant latent trajectories uncertain. It is a descriptive clue, not proof of weak identification by itself. (3) No. Consistent rescaling changes the coefficient's units but leaves the underlying likelihood information unchanged. (4) Convergence indicates that the algorithm found a numerical solution; a shallow profile can still support many association values nearly as well as the maximum.