Why a Likelihood Ridge Matters
A joint model can fit the observed longitudinal measurements and event times similarly well at quite different combinations of parameter values. For example, a larger random-effect variation may be offset by a smaller survival association parameter. If the likelihood changes little along such a path, the parameter pair is weakly distinguished by the data, even when an optimizer reports a precise-looking point estimate.
A two-parameter likelihood contour makes this compensation visible. Its shape can show whether the data support a compact region around an estimate or instead allow a long, narrow path of nearly equivalent fits. This is a diagnostic for the joint likelihood, not by itself a remedy for weak identification or a guarantee that an optimization has found the global maximum.
Model and Meaning of a Ridge
Consider a shared-random-intercept model. For subject \(i\), let \(Y_{ij}\) be the longitudinal outcome at time \(t_{ij}\), and let \(b_i\) be a subject-specific random intercept. A simple longitudinal submodel is
Suppose the event hazard is
Here \(h_0(t)\) is the baseline hazard, \(w_i\) is a vector of survival covariates, \(\gamma\) their coefficient vector, \(\alpha\) the association parameter, and \(\sigma_b\) the standard deviation of the shared random intercept. The model assumes the random effects are independent across subjects, that the stated conditional longitudinal and event models are appropriate, and that the two outcomes are conditionally independent given the shared random effect and included covariates. The marginal likelihood integrates each subject’s conditional contribution over the specified distribution of \(b_i\).
For the pair \((\alpha,\sigma_b)\), a useful intuition comes from the random part of the survival linear predictor, \(\alpha b_i\). Since \(b_i\) has variance \(\sigma_b^2\), the product \(\alpha b_i\) has variance \(\alpha^2\sigma_b^2\). When the event data are particularly influential and the longitudinal measurements provide limited information about \(\sigma_b\), combinations with similar \(|\alpha|\sigma_b\) can produce similar variation in the survival predictor. The longitudinal likelihood usually prevents this from being an exact equivalence, because it also depends on the distribution of \(b_i\) itself. The result can nevertheless be a long, shallow ridge.
How to Construct and Read a Two-Parameter Contour
Choose two parameters of interest, here \(\alpha\) and \(\sigma_b\), and let \(\lambda\) denote all remaining model parameters. The profile log likelihood for the pair is
Thus, at each plotted pair, the nuisance parameters are re-estimated to maximize the log likelihood. This differs from a fixed-nuisance slice, which holds \(\lambda\) at one chosen value. A slice can distort the apparent support for the pair if the nuisance parameters would adjust in response to changes in \(\alpha\) or \(\sigma_b\).
Let \((\hat\alpha,\hat\sigma_b)\) be the maximizer in the plotted region and let \(\ell_{\max}\) be the corresponding maximum profile log likelihood. A likelihood-ratio contour is based on
Under regularity conditions and sufficiently informative data, a contour \(D\leq 5.991\) is an approximate 95% joint likelihood region for two parameters, using the 95th percentile of a chi-squared distribution with 2 degrees of freedom. In weakly identified or boundary cases, this calibration may be unreliable. The contour’s geometry remains useful as a diagnostic, but its nominal coverage should not be treated as assured.
Worked Example: Seeing Compensation in the Model
Suppose a hypothetical joint model has candidate values \(\sigma_b=0.30\) outcome units and \(\alpha=1.20\) per outcome unit. Compare them with \(\sigma_b=0.45\) and \(\alpha=0.80\) per outcome unit. For this illustration, assume the model’s outcome and time units make \(\alpha b_i\) dimensionless.
At the first pair, the standard deviation of the random contribution to the survival linear predictor is
At the second pair it is
The two pairs therefore give the same marginal standard deviation for \(\alpha b_i\), even though the random-intercept standard deviation and association parameter differ. This calculation explains why the survival component may fit similarly at both pairs. It does not prove that their full joint likelihoods are equal: the longitudinal component also uses \(b_i\), and other fitted parameters may change. A contour plot is needed to assess how much the complete likelihood distinguishes these candidates.
Deriving the Approximate Ridge Direction
The compensation can be expressed by holding the scale of the survival random contribution approximately constant. For positive \(\alpha\), set \(c=\alpha\sigma_b\), where \(c\) is a fixed positive constant. Solving for \(\alpha\) gives
Taking logarithms yields \(\log(\alpha)+\log(\sigma_b)=\log(c)\). Differentiating along this curve gives
So a proportional increase in \(\sigma_b\) is accompanied approximately by an equal proportional decrease in \(\alpha\). In ordinary coordinates this relation is curved, not a straight line; on a log-scale plot it is a straight line with slope \(-1\). This is a justification for looking for a negatively sloped, elongated contour in the \((\alpha,\sigma_b)\) plane when both parameters are positive. It is only an approximation to the full likelihood geometry: longitudinal information, event information, nuisance-parameter adjustment, and the baseline hazard can all alter the contour.
Worked Example: Calculating a Likelihood-Ratio Contour
Consider a hypothetical profile log likelihood whose maximum is \(\ell_{\max}=-120.00\) at \((\hat\alpha,\hat\sigma_b)=(1.10,0.34)\). At another candidate pair, \((\alpha,\sigma_b)=(0.80,0.47)\), suppose nuisance parameters have been re-optimized and the profile log likelihood is \(-122.70\).
The likelihood-ratio statistic for this pair is
Since \(5.40<5.991\), this pair lies inside the nominal 95% joint contour under the regular two-parameter approximation. The values also show compensation: the product at the maximum is \(1.10(0.34)=0.374\), while at the candidate pair it is \(0.80(0.47)=0.376\). Their nearly equal products are consistent with the ridge intuition. The conclusion is not that there is a 95% probability that the fixed pair is correct. Rather, under the stated approximation, the joint region formed by the contour has nominal 95% repeated-sampling coverage.
If a second candidate had profile log likelihood \(-123.10\), its statistic would be
Because \(6.20>5.991\), it would fall outside that nominal contour. This comparison uses the re-optimized profile likelihood at each pair; using log likelihoods from a fixed-nuisance slice would answer a different question.
Practical Reading, Diagnostics, and Alternatives
A compact, roughly elliptical contour suggests that the selected pair is jointly well localized, subject to the model and calibration assumptions. A long, narrow contour indicates that many combinations fit similarly, often with one parameter increasing as the other decreases. If the contour is open at the edge of the plotted grid, the plot may simply be truncated; widen the grid before concluding that the likelihood is unbounded. A contour that touches \(\sigma_b=0\) or another parameter boundary deserves particular caution because regular likelihood-ratio approximations may fail there.
A useful check is whether the plotted contours are stable when the grid is refined, expanded, and the nuisance parameters are re-optimized at each point. Sparse or uneven optimization can create apparent gaps, kinks, or isolated islands that are numerical artifacts rather than likelihood structure. Conversely, a smooth elongated contour is not proof that the model is correctly specified; it only describes the fitted likelihood surface under the chosen model.
The main alternatives answer related but distinct questions. A one-parameter likelihood profile, as in likelihood profiles for weakly identified parameters, displays how likelihood changes for one parameter while other parameters are optimized. It can summarize uncertainty for that parameter but conceal which second parameter compensates for it. A Wald approximation summarizes a local quadratic shape near an estimate and can be misleading when the likelihood is strongly curved, asymmetric, or near a boundary. A two-parameter contour directly displays joint shape, but requires repeated nuisance-parameter optimization and is limited to the selected pair. These methods complement rather than replace one another.
The contour also depends on parameterization and scale. A ridge that looks curved in \((\alpha,\sigma_b)\) coordinates may look straighter in \((\alpha,\log\sigma_b)\) coordinates. The likelihood ordering is unchanged by a one-to-one transformation, but visual distances and apparent shape are not. Label axes with parameter names and units, state whether parameters were profiled or fixed, and distinguish an approximate likelihood region from a confidence statement that has been justified under the model’s conditions.
Worked Example: Interpreting an Elongated Contour
Suppose a hypothetical grid search profiles all nuisance parameters and finds the following pairs near the top of the likelihood surface. Assume \(\alpha\) is measured per outcome unit and \(\sigma_b\) in outcome units.
| \(\alpha\) | \(\sigma_b\) | \(\alpha\sigma_b\) | \(D\) |
|---|---|---|---|
| 1.40 | 0.28 | 0.392 | 1.1 |
| 1.10 | 0.36 | 0.396 | 0.4 |
| 0.85 | 0.47 | 0.400 | 1.3 |
| 0.65 | 0.62 | 0.403 | 3.8 |
The product is similar across the first three pairs: \(1.40(0.28)=0.392\), \(1.10(0.36)=0.396\), and \(0.85(0.47)=0.3995\), which rounds to \(0.400\). Their likelihood-ratio values are all below 5.991, and the set follows the predicted pattern of larger \(\sigma_b\) paired with smaller \(\alpha\). The last pair also has a similar product, \(0.65(0.62)=0.403\), but its \(D=3.8\) is still below 5.991; it remains inside the nominal contour while being less well supported relative to the maximum than the first three. The data therefore permit substantial movement along this direction, although the likelihood does not remain exactly constant.
The appropriate interpretation is joint: the contour supports a range of combinations, not a claim that all listed values are equally likely or that one parameter has been estimated precisely in isolation. A one-parameter profile may be useful for reporting uncertainty in \(\alpha\), but it should be read alongside this pairwise geometry if the random-effect scale is a plausible source of compensation.
Limits of the Diagnostic
A two-parameter contour diagnoses the fitted likelihood, not the truth of the joint-model assumptions. It cannot establish that the random-effects distribution, longitudinal mean structure, or hazard model is adequate. Nor can it resolve confounding involving parameters not shown on the axes: profiling nuisance parameters accounts for their best-fitting values, but a contour cannot display all dimensions at once. Interpretation also depends on adequate numerical optimization and integration when the marginal likelihood requires integration over random effects.
When a ridge is present, avoid presenting a single parameter estimate without its uncertainty and context. State which pair was plotted, whether other parameters were profiled, and whether the likelihood-ratio calibration is being used as an approximation. The central lesson is that a maximum can exist even when the data provide little separation among nearby parameter combinations.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- If \(\sigma_b\) increases from \(0.25\) to \(0.50\), what approximate change in positive \(\alpha\) preserves \(\alpha\sigma_b\)?
- A profiled pair has \(\ell_{\max}=-80.0\) and \(\ell_p=-83.2\). Calculate \(D\) and say whether it is inside the nominal two-parameter 95% contour.
- Why can a fixed-nuisance slice give a different contour from a profile likelihood contour?
- What does a long negatively sloped contour for positive \(\alpha\) and \(\sigma_b\) suggest, and what does it not establish?
Solutions. (1) Doubling \(\sigma_b\) requires approximately halving \(\alpha\), since preserving the product gives \(\alpha_{\text{new}}=\alpha_{\text{old}}(0.25/0.50)=0.5\alpha_{\text{old}}\). (2) \(D=2\{-80.0-(-83.2)\}=6.4\), which exceeds 5.991, so the pair is outside the nominal 95% contour under the regular approximation. (3) A fixed-nuisance slice holds the other parameters constant, while a profile contour re-optimizes them separately at every plotted pair. (4) It suggests that larger random-effect variation can be offset by a smaller association parameter, leaving the fitted likelihood similar. It does not prove exact equivalence, validate the model, or guarantee that the nominal contour has its stated coverage when identification is weak or a boundary is involved.