A Diagnostic Workflow For Unstable Fits

Why an Unstable Fit Needs a Workflow

A joint-model fit can produce a warning, an implausibly large coefficient, or a standard error that changes dramatically when the optimizer is rerun. These symptoms have different possible causes: poor numerical scaling, sensitivity to starting values, inaccurate integration, or a likelihood that contains little information about one or more parameters. Repeating the fit with a different optimizer setting may help, but it does not by itself reveal which problem occurred.

A useful diagnostic workflow changes one source of numerical difficulty at a time and records what changes. Start with the data and parameter scales; then investigate optimization starts and local derivatives; assess curvature and integration accuracy; and finally inspect profiles for parameters that remain unstable. This sequence does not guarantee that a model is scientifically appropriate. It helps determine whether the reported fit is a reliable numerical representation of the likelihood under the specified model.

For subject \(i\), a shared-random-effects joint model commonly expresses the contribution to the observed-data likelihood as an integral over the subject-specific random effects \(b_i\). In a schematic notation,

Here \(y_i\) denotes the observed longitudinal measurements, \(T_i\) the observed event or censoring time, \(\delta_i\) the event indicator, and \(\theta\) the model parameters. The schematic assumes independent subjects, a specified random-effects distribution, and the stated conditional longitudinal and event-time models. In practice, numerical integration approximates \(L_i\), and an optimizer searches for a maximum of the resulting log likelihood. The workflow below checks both the statistical information in that likelihood and the numerical procedures used to evaluate and maximize it.

Stage 1: Check Data and Parameter Scales

First check units, ranges, missingness, and coding for both the longitudinal and event-time data. A time variable measured in days rather than years, or a biomarker recorded in very large units, can create parameters on sharply different numerical scales. Optimizers often work less reliably when a small parameter change in one coordinate has a vastly different effect on the objective than a comparable change in another.

Rescaling a predictor can improve numerical conditioning, but all coefficient interpretations must be translated back to the chosen units. For example, if \(x^*=x/100\), then \(\beta x=\beta^*x^*\) requires \(\beta^*=100\beta\). This is a change of parameter units, not new information in the data. Centering can also make an intercept more interpretable and reduce avoidable dependence between an intercept and slope. Do not rescale or recode silently: record the transformation and check that fitted predictions and likelihood values agree after translating parameters.

Stage 2: Separate Starting-Value Problems from a Reliable Optimum

Fit from multiple substantively plausible starting values, using the same data, model, integration settings, and stopping tolerances. Compare the final log likelihoods, parameter estimates, convergence status, and diagnostics—not only whether the software prints “converged.” Competing solutions with similar likelihoods but substantially different parameters may indicate multiple optima or a likelihood that is poorly informative in some direction. A single successful run cannot rule this out.

When starts lead to different solutions, inspect whether the optimizer stopped at a boundary, encountered a non-finite objective, or terminated before the gradient was small. Reparameterizations that enforce constraints, such as a log scale for a positive variance, can prevent invalid trial values. They do not create information or resolve every weakly identified parameter.

Stage 3: Check Gradients and Hessian Curvature

At a smooth interior maximum, the gradient of the log likelihood should be close to zero. Because parameters may have different units, inspect scaled or componentwise gradients as well as a single unscaled norm. A finite-difference gradient check compares numerical derivatives computed with more than one step size; a large discrepancy can reflect numerical noise, an implementation problem, or an unsuitable step size. The gradient is necessary evidence of local stationarity, not proof of a maximum.

The Hessian is the matrix of second derivatives of the log likelihood. At a well-behaved interior maximum, its curvature should be negative in every identifiable direction; equivalently, the observed information, often defined as the negative Hessian, should be positive definite. Near-zero curvature directions produce large uncertainty and sensitivity to small perturbations. Strong correlations among parameter estimates can also make the matrix poorly conditioned. If a parameter lies on a boundary, ordinary interior curvature interpretations and standard approximations may not apply.

Stage 4: Test Integration Accuracy Before Trusting Fine Curvature

The likelihood integral over random effects is generally evaluated approximately. A fit can appear converged for a coarse approximation while the approximation itself shifts the objective, derivatives, or estimates. Refit with progressively more accurate integration settings appropriate to the method, holding other choices fixed. Compare estimates, log likelihoods, gradients, and uncertainty—not just whether each run converged. The comparison is informative only if the settings are genuinely more accurate for the model and the numerical method is otherwise comparable.

A useful check is whether the changes shrink as accuracy increases. If point estimates shift materially, or a Hessian warning appears only under one setting, conclusions based on that setting are not robust. Agreement across a finite number of settings supports numerical stability but cannot prove exact integration; especially difficult integrals may require additional checks.

Stage 5: Profile Parameters That Remain Unstable

A profile likelihood examines how well the model can fit when one parameter is fixed at a sequence of values and the other parameters are reoptimized. For parameter \(\psi\), define \(\ell_p(\psi)=\max_{\lambda}\ell(\psi,\lambda)\), where \(\lambda\) denotes the remaining parameters. A narrow, smooth maximum suggests that the likelihood localizes \(\psi\); a flat, asymmetric, or irregular profile suggests that its estimate may be weakly determined, near a boundary, or affected by numerical problems.

Use the same sufficiently accurate likelihood evaluation at every profile point, and check that nuisance parameters are actually reoptimized. Failed optimization at individual points can make a profile look artificially jagged or too low. For a regular interior parameter, likelihood-ratio differences are sometimes compared with a chi-squared reference distribution with one degree of freedom; boundary cases or other violations require different care. A profile is a diagnostic, not a cure for poor fit or model misspecification.

Putting the Workflow Together

The stages complement rather than replace one another. Multiple starts test the search over the likelihood; gradient and Hessian checks assess the local solution; integration comparisons assess the objective’s numerical approximation; and profiles show how the likelihood behaves as a parameter varies. Simulation-based recovery and bootstrap stability answer related but different questions: they examine behavior under generated or resampled data, not whether a particular observed-data fit was numerically optimized accurately.

Common Pitfalls and Limits

• Equating a convergence message with a trustworthy fit. A termination rule can be met at a saddle point, boundary, or poorly determined solution. Check gradients, curvature, and sensitivity as well.
• Changing many settings at once. If the estimate changes after altering scaling, starts, tolerances, and quadrature together, the source is unclear. Change one factor at a time where practical.
• Using a small gradient as a precision measure. A gradient near zero says the objective is locally stationary; it does not say the likelihood is sharply curved or the estimate is precise.
• Ignoring units in a scale comparison. Raw coefficients can differ greatly after rescaling even when fitted values are identical. Compare transformed parameters and predictions.
• Treating instability as proof of model failure. Numerical instability may arise from optimization or integration. Conversely, a numerically stable fit can still be based on a misspecified model.

No single diagnostic resolves every case. Near-boundary parameters, complex random-effects structures, limited event information, or a highly irregular likelihood may defeat standard local approximations. If the checks disagree, preserve the evidence, simplify or reparameterize only with scientific justification, and avoid presenting a single unstable estimate as definitive.