Separating Weak Identification From Optimizer Failure

Why an Unstable Estimate Needs Diagnosis

An NLME fit can return a parameter estimate that changes substantially when the data, model, or initial values change. That instability has at least two distinct explanations. The likelihood may contain little information about the parameter, so many values fit almost equally well. Alternatively, the likelihood may contain a clear optimum that the numerical algorithm has failed to find or characterize. These situations call for different responses: weak identification calls for interpreting uncertainty and possibly reconsidering the information in the design, whereas optimizer failure calls for resolving the numerical problem before interpreting the estimate.

A convergence message alone does not settle the question. The useful evidence comes from examining how the objective behaves near and away from the reported estimate, whether nuisance parameters can be optimized reliably, and whether controlled changes to starting values lead to the same fit.

For an NLME model, the likelihood commonly refers to the likelihood for the observed data after accounting for the random effects, often by integrating them out. In practice, a software package may use a numerical approximation to that likelihood. Profile and curvature conclusions therefore concern the objective actually fitted; they may also reflect approximation error or an inadequately optimized inner calculation.

What the Profile and Curvature Show

A well-defined, sharply curved profile has a distinct minimum of \(Q_p\), with the objective increasing as the fixed parameter moves away from its optimum. A relatively flat profile means that a broad range of values can be offset by changes in other parameters with little loss of fit. This is evidence of weak identification in that direction. A long, narrow ridge in the full objective is a common multivariate version of the same problem: individual parameters may vary considerably while a combination of them remains well determined.

Local curvature summarizes behavior close to an optimum. The Hessian is the matrix of second derivatives of \(Q\) with respect to the parameters. At an interior minimum, positive curvature in a direction means the objective rises when moving in that direction; small curvature means it rises slowly. Because parameter units and scales affect numerical curvature, comparisons should be made on a meaningful scale, such as log-transformed positive parameters or parameters standardized by a stated scale. A small eigenvalue of the appropriately scaled Hessian indicates a locally shallow direction, but does not by itself identify its cause.

The cutoff 3.84 is the 95th percentile of a chi-squared distribution with one degree of freedom. It is a large-sample likelihood-ratio approximation, not a guarantee of good optimization. If a profile point is computed inaccurately, the resulting interval can be misleading. A profile that does not cross the cutoff on one or both sides may indicate very weak information, a parameter constraint, or a profile search that has not been carried out far enough; the cause must be investigated.

Separating a Flat Likelihood From a Failed Search

The key distinction is whether the objective is genuinely shallow after the nuisance parameters have been adequately optimized. To assess that, inspect the profile over a range of fixed values, not merely a single standard error or the optimizer’s final message. At each point, check that the nuisance-parameter optimization is stable enough to make the profile comparison meaningful. A profile made from inconsistent or incomplete nuisance fits can look jagged, artificially high, or artificially flat.

Starting-value sensitivity is informative when tested in a controlled way. Hold the data, model, objective settings, parameter bounds, and numerical tolerances fixed; change only the initial values, then compare final objective values and parameter estimates. If multiple starts reach essentially the same objective and similar parameters, that supports a reproducible solution. If they reach materially different objective values, at least one run may have failed to find the best fit. If they reach essentially the same objective but very different parameter values, a flat ridge or weak identification is plausible. Neither pattern proves its explanation: distinct local optima can have similar objectives, and weak identification can coexist with optimization difficulty.

Deriving the Curvature Connection

The local link between objective curvature and uncertainty follows from a second-order expansion. For one parameter near an interior maximum \(\hat{\theta}\), the profile negative log likelihood is approximately \(Q_p(\theta)\approx Q_p(\hat{\theta})+\tfrac{1}{2}H(\theta-\hat{\theta})^2\), where \(H\) is its second derivative at the optimum and is positive at a local minimum. Multiplying the increase by two gives the likelihood-ratio deviance:

Equating this approximation to the one-degree-of-freedom cutoff 3.84 yields \(|\theta-\hat{\theta}|\approx\sqrt{3.84/H}\). Thus larger curvature corresponds to a narrower local interval, while smaller curvature corresponds to a wider one. This justification depends on a smooth, approximately quadratic profile near an interior optimum. It does not establish that the objective is quadratic over a broad range, that the likelihood approximation is adequate, or that a numerically reported Hessian is reliable.

Diagnostics, Alternatives, and Limitations

Several patterns are useful but should be interpreted together. A stable, broad profile with similar objective values across starts suggests weak information. A jagged profile, profile minima that fail to match the unrestricted fit, or substantial variation in objective values across starts suggests an optimization or numerical-evaluation problem. A narrow profile combined with different estimates at similar objective values may indicate multiple optima or a parameterization issue rather than ordinary weak identification. Strong correlations among parameter estimates or a very small eigenvalue of a properly scaled Hessian can reveal a shallow joint direction that one-dimensional standard errors obscure.

Profile likelihood is more informative than relying only on a local standard error because it can reveal asymmetry, long tails, and departures from a quadratic shape. Local Hessian-based uncertainty is faster and summarizes behavior near the optimum, but is fragile when curvature is nearly singular, parameters are poorly scaled, or the optimum is not genuine. Multiple starting values are a practical check for search sensitivity, not a proof of global optimality. These methods complement rather than replace one another.

Important limitations remain. A flat profile can result from a genuinely weakly identified parameter, a parameter constraint, a misspecified model, or a poor likelihood approximation. Profile likelihood intervals can have unreliable coverage in small samples, at boundaries, or under nonidentifiability; the usual chi-squared cutoff should not be treated as universally valid. Numerical tolerances and the accuracy of integration or approximation can also affect small objective differences. When the evidence is mixed, the responsible conclusion is that identification and optimization have not yet been cleanly separated.