Convergence and Identifiability · Tutorial 346 of 697

Near-Zero Random-Effect Variance And Association Weakness

Learn why little random-effect variation leaves the survival association weakly informed, and how to interpret profile-likelihood evidence near a variance boundary.

Advanced 12 min read

What You'll Learn

Explain why the survival association is unidentifiable at zero random-effect variance.
Relate the association parameter to the random-effect scale and the induced hazard variation.
Construct and interpret a profile likelihood for the association parameter.
Recognize boundary and flat-profile features that make standard likelihood tests unreliable.
Distinguish weak association information from evidence that the association is absent.
Identify practical diagnostics and alternative ways to assess uncertainty.

Why a Small Variance Changes the Association Problem

In a shared-random-effects joint model, the survival submodel links an individual’s event hazard to a latent quantity that also appears in the longitudinal submodel. The association parameter describes how that latent quantity shifts the hazard. If the corresponding random effect varies substantially between individuals, the data can potentially reveal whether individuals with different latent values have different hazards. If its variance is close to zero, there is very little between-person variation through which that association can be learned.

This is a specific identifiability issue, not merely a convergence inconvenience. When the random-effect variance is exactly zero, the random effect is zero for every individual under the model, so its coefficient in the hazard has no effect on the likelihood. Near zero, the coefficient may change substantially while producing almost no change in fitted survival probabilities. The association estimate can therefore be highly uncertain even when an optimizer returns a precise-looking number.

Definition: Consider a shared random-intercept model with $b_i\sim N(0,\sigma_b^2)$, where $b_i$ is the latent intercept for individual $i$. A conditional survival model may be written $$ h_i(t\mid b_i)=h_0(t)\exp\{z_i^\top\gamma+\alpha b_i\}. $$ Here $h_0(t)$ is the baseline hazard, $z_i$ is a vector of survival covariates, $\gamma$ is its coefficient vector, and $\alpha$ is the association parameter. If $b_i$ is measured in longitudinal-outcome units, $\alpha$ is measured in inverse outcome units. The random-effect variance $\sigma_b^2$ has squared outcome units.

The longitudinal model provides information about $\sigma_b^2$, while the event data provide information about how the shared effect relates to hazard. They contribute jointly, but the key scale issue is visible by writing $b_i=\sigma_b u_i$, where $u_i\sim N(0,1)$ and $\sigma_b=\sqrt{\sigma_b^2}$. The random part of the log hazard is then $\alpha\sigma_bu_i$. Thus the standard deviation of the random contribution to log hazard is $|\alpha|\sigma_b$. If $\sigma_b$ is tiny, a moderate change in $\alpha$ can correspond to a tiny change in the survival effect.

This product is a useful way to reason about the survival scale, but it does not mean the entire joint likelihood depends only on $\alpha\sigma_b$. The longitudinal likelihood also depends on the random-effect scale, and the survival and longitudinal observations are linked through the same latent $b_i$. The concern is that survival data may have little leverage on $\alpha$ when the shared effect scarcely varies.

A Short Likelihood Justification

For a subject with event or censoring time $T_i$ and event indicator $\delta_i$, the conditional event contribution, given $b_i$, has log likelihood $$ \ell_i(b_i)=\delta_i\log h_i(T_i\mid b_i)-\int_0^{T_i}h_i(s\mid b_i)\,ds. $$ The derivative of the conditional log hazard with respect to $\alpha$ is $b_i$. More explicitly, differentiating the event contribution gives $$ \frac{\partial\ell_i(b_i)}{\partial\alpha} =b_i\left\{\delta_i-\int_0^{T_i}h_i(s\mid b_i)\,ds\right\}. $$

This equation gives the basic justification: when $b_i$ is close to zero, the event contribution changes little as $\alpha$ changes. Under bounded follow-up and regularity conditions that keep the hazard contributions controlled, the conditional information about $\alpha$ is proportional in scale to $b_i^2$; averaging over $b_i$ brings in $E(b_i^2)=\sigma_b^2$. This is a scale argument, not a universal exact formula for the observed information. In the actual joint likelihood, $b_i$ is latent and integrated out, and information also depends on event frequency, follow-up, longitudinal precision, covariates, and the baseline hazard.

At $\sigma_b^2=0$, $b_i=0$ with probability one. The hazard becomes $h_i(t)=h_0(t)\exp(z_i^\top\gamma)$, regardless of $\alpha$. Consequently, at that boundary the likelihood is constant in $\alpha$, after holding the other parameters fixed. This is why the association is not identified at the zero-variance boundary. Near the boundary, a broad or nearly flat likelihood profile for $\alpha$ is a natural warning sign.

Worked Example: Comparing Association Scales

Suppose the random intercept is measured in assay units, and consider two hypothetical fitted parameter combinations: (A) $\sigma_b=0.05$ assay units and $\alpha=2$ per assay unit; (B) $\sigma_b=0.025$ assay units and $\alpha=4$ per assay unit. Compare the implied standard deviation of the random contribution to log hazard.

For (A), $|\alpha|\sigma_b=2(0.05)=0.10$. For (B), $|\alpha|\sigma_b=4(0.025)=0.10$. Each combination therefore implies a one-standard-deviation random log-hazard shift of 0.10, corresponding to a hazard ratio of $\exp(0.10)=1.105$ for a one-standard-deviation increase in the random effect.

The two combinations need not have equal joint likelihood: their longitudinal random-effect variances differ, and the longitudinal observations can distinguish them. But if the data estimate that variance imprecisely, survival evidence that mainly tracks the product $\alpha\sigma_b$ may not sharply distinguish the two association values. A large $\alpha$ alone does not establish a strong survival association.

Profiling the Association Parameter

A profile likelihood makes the weakness visible by asking how well each fixed candidate value of $\alpha$ can fit the data after the other parameters are re-estimated. Let $\psi$ denote all nuisance parameters, including $\sigma_b^2$, longitudinal parameters, survival covariate effects, and baseline-hazard parameters. Define

\ell_p(\alpha)=\max_{\psi:\,\sigma_b^2\geq 0}\ell(\alpha,\psi).

The profile likelihood is the likelihood at each $\alpha$ after maximizing over $\psi$. Its maximum occurs at the joint maximum-likelihood estimate of $\alpha$, subject to any constraints in the model. A narrow, curved profile indicates that values away from the maximum fit noticeably worse. A broad, shallow profile indicates that a wide range of association values fits nearly as well. A plateau or a profile that continues nearly unchanged toward the edge of the evaluated range is especially concerning.

To examine a profile, fit the model at a grid of fixed $\alpha$ values, optimizing all nuisance parameters at each grid point. Record the maximized log likelihood and the nuisance estimates, particularly $\sigma_b^2$. A profile can be misleading if nuisance optimization fails at some grid values, so check that the optimizer has actually found a comparable maximum throughout the grid. Plotting the fitted variance against $\alpha$ can reveal whether the model compensates for a changing association by driving the variance toward its boundary.

Worked Example: Reading a Shallow Profile

Suppose a hypothetical profile calculation gives the following maximized log likelihoods. The largest value is $-120.00$, at $\hat\alpha=1.0$. The table also reports the nuisance-optimized random-intercept variance.

Fixed $\alpha$	$\ell_p(\alpha)$	Optimized $\sigma_b^2$
0	-120.70	0.0005
1	-120.00	0.0004
3	-120.12	0.0002
6	-120.45	0.0001

The log-likelihood loss at $\alpha=3$ is $-120.00-(-120.12)=0.12$; the likelihood-ratio statistic comparing that fixed value with the maximum is $2(0.12)=0.24$. The loss at $\alpha=6$ is 0.45, giving $2(0.45)=0.90$. These small changes show that the profile is shallow over the displayed range. The declining variance also suggests a trade-off: as $\alpha$ increases, the fit reduces the amount of random-effect variation.

This table is evidence of weak profile information over the range shown, not proof that every value of $\alpha$ is plausible. Nor should the displayed likelihood-ratio statistics automatically be compared with a standard chi-square cutoff: the variance boundary and nonidentification at zero can invalidate the usual regular approximation. The profile should be extended far enough to determine whether it eventually falls, and uncertainty should be calibrated with methods appropriate to the boundary problem.

Boundary Evidence, Interpretation, and Alternatives

A profile is most informative when it is examined alongside the estimate and uncertainty of $\sigma_b^2$, the longitudinal evidence about heterogeneity, and the actual scale of the outcome. For instance, a numerically nonzero variance estimate may still be so small relative to the measurement scale that the survival association is practically weakly informed. Conversely, a variance estimate near zero does not by itself prove that the association parameter equals zero; it says the model has little estimated latent variation through which that parameter can operate.

Worked Example: Translating a Near-Zero Variance

Suppose a hypothetical assay model estimates $\hat\sigma_b^2=0.0004$ assay-units squared and $\hat\alpha=3$ per assay unit. The estimated random-effect standard deviation is $\hat\sigma_b=\sqrt{0.0004}=0.02$ assay units. The standard deviation of the random contribution to log hazard is therefore $3(0.02)=0.06$.

A one-standard-deviation increase in the random intercept corresponds to a conditional hazard ratio of $\exp(0.06)=1.062$. This describes the fitted scale of the random hazard shift; it does not show that $\alpha$ is precisely estimated. If a profile includes, for example, values of $\alpha$ from 0 to 10 with little likelihood loss while the optimized variance remains near zero, the data do not distinguish those association values well. The estimate 3 should then not be reported as strong evidence of a well-determined association.

There are important differences between this profile approach and simpler alternatives. A Wald interval uses the estimated curvature near the maximum and can be unreliable when the profile is flat, asymmetric, or near a boundary. A likelihood-ratio interval uses profile likelihood but its usual cutoff relies on regular large-sample conditions; here, nonidentification at zero variance complicates those conditions. A parametric bootstrap can calibrate uncertainty or tests under a specified model by simulating data and refitting, but it must represent the boundary and the fitting procedure accurately. Bayesian inference can use a prior to regularize weak information, but the resulting association inference may be sensitive to prior choices when the likelihood is flat. None of these approaches creates information absent from the data.

A profile is a diagnostic, not a repair. It can fail to reveal the true shape if the grid is too narrow, if nuisance optimization converges to inferior local maxima, or if numerical integration is too inaccurate. Estimates may also depend strongly on how the longitudinal outcome is scaled, because $\alpha$ is expressed per outcome unit. Scaling choices should preserve the model’s meaning and be reported clearly; changing units rescales $\alpha$ but does not create latent variation. The central limitation remains: if the corresponding random-effect variance is essentially zero, the data offer little basis for estimating its survival association.

Practical Checks and a Common Pitfall

Check the estimated variance and its scale.
Report $\hat\sigma_b^2$ and $\hat\sigma_b$ in the longitudinal outcome’s units. Consider whether the estimated variation is substantively negligible.

Inspect the association profile.
Fix $\alpha$ over a sufficiently broad grid and re-optimize nuisance parameters. Inspect the profile shape and the variance estimates along it.

Verify the numerical fit.
Check that each profile point is adequately optimized and that integration accuracy is sufficient for likelihood differences of the size being interpreted.

Choose uncertainty statements cautiously.
Do not treat a routine Wald interval or chi-square likelihood-ratio cutoff as automatically valid at a variance boundary. Consider suitable calibration and disclose weak identification.

Common mistake: “The estimated association is large, so the survival association must be strong.” This reasoning ignores the random-effect scale. If $\hat\sigma_b$ is tiny, then $\hat\alpha\hat\sigma_b$—the standard deviation of the random log-hazard contribution—may also be small. In addition, a large point estimate can coexist with a nearly flat profile and substantial uncertainty. Assess the profile and variance together, and interpret the association in its stated units.

Key takeaway: As the variance of a shared random effect approaches zero, the hazard has less latent variation on which its association parameter can act. The association becomes weakly informed and is completely absent from the likelihood at zero variance. A shallow profile, especially with variance estimates near the boundary, is evidence to report uncertainty rather than overinterpret a point estimate.

Check Your Understanding

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

In the shared random-intercept hazard model, what happens to the hazard when $\sigma_b^2=0$?
If $\sigma_b=0.04$ outcome units and $\alpha=2.5$ per outcome unit, calculate the standard deviation of the random contribution to log hazard and its one-standard-deviation hazard ratio.
A profile log likelihood is $-50.0$ at its maximum and $-50.3$ at a candidate association value. Calculate the likelihood-ratio statistic and explain what it says descriptively.
Why can a large estimated $\alpha$ still indicate a weakly informed survival association?
Name one reason a routine likelihood-ratio cutoff may be unreliable in this setting.

Solutions. (1) Since $b_i=0$ with probability one, $h_i(t)=h_0(t)\exp(z_i^\top\gamma)$; the hazard no longer depends on $\alpha$. (2) The log-hazard standard deviation is $2.5(0.04)=0.10$. The corresponding hazard ratio is $\exp(0.10)=1.105$. (3) The statistic is $2\{-50.0-(-50.3)\}=0.60$. The candidate value has a small likelihood loss on this scale, suggesting a shallow profile locally; this alone does not establish a calibrated p-value or interval. (4) The random-effect variation may be tiny, making the product $|\alpha|\sigma_b$ small, and the profile may remain nearly flat across association values. (5) At zero variance, $\alpha$ is not identified, and the variance is constrained to be nonnegative; these features violate regular conditions underlying the usual chi-square approximation.

Back to Tutorials