Why Classification Matters
An NLME run can produce a numerical result without producing a trustworthy estimate. Conversely, a warning or a parameter at a limit does not automatically mean that the optimizer failed. These distinctions matter because the next step depends on the type of problem: an optimizer that stopped before meeting its convergence criteria is different from an optimizer that converged to a boundary, a covariance structure that is rank deficient, or a converged solution whose parameter values are incompatible with the model or data.
The classification should use several pieces of evidence together: the optimizer’s termination status and message, the value and behavior of the objective function, the estimates relative to their allowed ranges, and the reported covariance diagnostics. A single word such as “success” is not a complete assessment.
Different software may use different objective functions, stopping rules, and message wording. Before comparing objective values or interpreting a status label, confirm what was optimized, whether the runs use the same data and model, and whether the software says the convergence criteria were met. A finite objective value alone does not establish convergence.
Four Outcomes That Should Not Be Confused
| Classification | Typical evidence | What it means |
|---|---|---|
| Optimizer non-convergence | A message says the iteration limit was reached or the run stopped without satisfying its criteria; the objective or estimates may still be changing. | The run did not establish a numerical solution under the stated stopping rules. |
| Boundary estimate | An estimate equals or lies very near a specified lower or upper bound. | The fitted value is at the edge of the allowed parameter space. This can occur even when the optimizer reports convergence. |
| Singular covariance matrix | A reported random-effects covariance matrix has deficient rank, or a covariance calculation reports singularity. | At least one direction has zero variance in the matrix representation; some matrix operations requiring an inverse are unavailable. |
| Implausible converged solution | The status reports convergence, but one or more estimates or resulting predictions conflict with the model’s scientific meaning or the observed data. | The numerical stopping rule was met; that fact alone does not validate the solution. |
“Covariance matrix” can refer to more than one object. The random-effects covariance matrix, often denoted \(\Omega\), describes between-subject variation in the model. Software may also report a covariance matrix for the estimated parameters, commonly derived from local curvature of the objective function. Singularity in either object is a numerical or structural warning, but the interpretation depends on which matrix the software reports. State which matrix is involved rather than using “the covariance is singular” without qualification.
How to Read the Evidence
Start with the final status and diagnostic message. A message such as “maximum number of iterations reached” indicates that the run ended without meeting its convergence criteria; it is not equivalent to “converged.” Then inspect the objective-function history if available. A value that is still changing substantially at termination supports the concern that the search had not settled, although a nearly unchanged value by itself does not prove that all criteria were satisfied. Check the parameter estimates too: large changes between late iterations can be evidence that the search was still moving.
Next compare each estimate with its declared constraints. An estimate equal to a lower bound is a boundary estimate, not automatically an optimizer failure. For a variance constrained to be nonnegative, an estimate of zero is allowed if zero is included in the fitted parameter space. If software flags a boundary or singularity, retain that warning in the classification even when the termination status says “converged.”
For covariance matrices, inspect the reported matrix and its diagnostics. A square matrix is singular if it does not have an inverse; equivalently, its determinant is zero, or its rank is less than its dimension. In numerical computation, software may label a matrix “near singular” when it is sufficiently close to singular under its numerical tolerance. The tolerance is software-dependent, so the message and the matrix values should be reported together.
Worked Example: Identifying Optimizer Non-Convergence
Hypothetical fit. An NLME model is fit by minimizing \(F=-2\ell\). The run reaches its limit of 500 iterations and reports “maximum iterations reached; convergence criteria not satisfied.” Near the end, \(F\) changes from 1,280.4 to 1,262.7. The estimated clearance changes from 4.1 L/h to 5.3 L/h over the same final iterations.
Classification. This is optimizer non-convergence because the message explicitly says the criteria were not satisfied. The objective change is \(1,280.4-1,262.7=17.7\), a substantial decrease in the minimized objective across the reported interval. Clearance also changes by \(5.3-4.1=1.2\) L/h. These changes reinforce the message that the run had not settled.
The final \(F=1,262.7\) is a value produced by the run, but it does not turn the run into a converged fit. Nor can it be interpreted as evidence that this model fits better than another model unless the objectives are comparable and the relevant comparison is justified. The appropriate classification is “non-converged run; final objective and estimates are provisional,” not “successful fit with clearance 5.3 L/h.”
Boundary Estimates and Singular Covariance
A boundary estimate concerns the location of an estimate relative to the parameter space. Singularity concerns the rank of a matrix. They may occur together, but neither term is a substitute for the other. For example, an estimated random-effects variance of zero is on the boundary of the nonnegative variance space. If another random effect also has zero variance, the resulting random-effects covariance matrix may be singular. A covariance matrix can also be singular for other structural or numerical reasons.
Worked Example: A Converged Boundary Estimate and a Singular Matrix
Hypothetical fit. A model estimates random effects on two log-scale parameters. The software reports “converged,” estimates their covariance matrix as \[ \hat{\Omega}= \begin{pmatrix} 0.04 & 0.12\\ 0.12 & 0.36 \end{pmatrix}, \] and flags the matrix as singular. The entries are in squared log-scale units.
Check the matrix. For a two-by-two matrix with entries \(a,b,c,d\), the determinant is \(ad-bc\). Here, \[ \det(\hat{\Omega})=(0.04)(0.36)-(0.12)(0.12)=0.0144-0.0144=0. \] The determinant is zero, so \(\hat{\Omega}\) is singular and has no inverse. Its second row is three times its first row, so the rows are linearly dependent and its rank is one rather than two.
The estimates of the two random-effects variances are 0.04 and 0.36, both positive; thus this example is not a zero-variance boundary estimate. The matrix is singular because its two dimensions are perfectly dependent in the estimated covariance structure. The correct classification is “converged status, singular random-effects covariance matrix.” The convergence label does not remove the singularity warning, and the singularity does not, by itself, prove that the optimizer failed to meet its stopping criteria.
A parameter covariance matrix reported as singular should be named separately from \(\hat{\Omega}\). Such a warning can mean that the software cannot obtain its usual local uncertainty calculation from the fitted objective. Do not describe this as a singular random-effects matrix unless that is the matrix actually diagnosed.
Converged Does Not Mean Plausible
A converged solution can still be implausible. “Implausible” is a scientific and model-based judgment: an estimate may have an incoherent scale or imply predictions sharply at odds with the observations or with known features of the model. It is not a special optimizer status, and it should be justified with the parameter’s units, constraints, and consequences. A surprising estimate is a reason to investigate, not by itself proof of a software error.
Worked Example: A Converged but Implausible Clearance Estimate
Hypothetical data and model. Consider a one-compartment model after an intravenous bolus of 100 mg. In this model, concentration is \(C(t)=\frac{\text{Dose}}{V}\exp(-CLt/V)\), where \(V\) is volume in L and \(CL\) is clearance in L/h. Suppose the observed concentrations are about 5.0 mg/L at time zero and 0.50 mg/L at 8 h. A fitted NLME model reports convergence, \(V=20\) L, and typical clearance \(CL=0.050\) L/h.
Check the implied prediction. At time zero, the model predicts \(100/20=5.0\) mg/L. At 8 h, its elimination rate constant is \(CL/V=0.050/20=0.00250\) h\(^{-1}\), so \[ C(8)=5.0\exp[-(0.00250)(8)]=5.0\exp(-0.0200)\approx 4.90\text{ mg/L}. \] That prediction is far above the hypothetical observed concentration of about 0.50 mg/L. The optimizer’s convergence message is therefore not enough to treat the fitted values as credible: this solution is converged by its numerical status but implausible in light of the model-implied concentration.
The example does not establish the cause. The model, data handling, parameterization, or a numerical search issue could warrant investigation. Classification comes first: report a converged status and an implausible estimate with the prediction-based reason. Do not relabel it “non-convergence” unless the optimizer’s criteria were in fact not met.
A Practical Classification Sequence
Record whether the software says the convergence criteria were met, and note any iteration limit or other stopping condition.
Use reported objective values and parameter estimates near termination as supporting evidence. Do not treat a finite objective value as proof of convergence.
Identify estimates at or near bounds and specify whether a singularity warning concerns the random-effects covariance matrix or the parameter-estimate covariance calculation.
Use units and model-implied predictions to state why an estimate seems plausible or implausible. Keep this judgment distinct from the optimizer’s status.
These classifications can overlap. A run can fail to converge and end with an estimate at a bound; a converged run can have a singular random-effects covariance matrix; and either kind of run can produce estimates that raise scientific concerns. Report each supported finding rather than forcing the output into just one category.
Common Pitfalls and Limitations
- Calling every warning “non-convergence.” A boundary or singularity warning can accompany a converged status. State the warning and status separately.
- Treating the final objective as a success certificate. An optimizer can return an objective value even if it stopped before satisfying its convergence criteria.
- Calling every boundary estimate a failure. A boundary is a location in the allowed parameter space. The status, estimate, and scientific context determine what else can be concluded.
- Using “singular covariance” without naming the matrix. Identify whether the warning concerns \(\hat{\Omega}\) or the covariance calculation for estimated parameters.
- Equating convergence with model adequacy. Convergence describes numerical stopping; plausibility and fit to the data require separate evaluation.
A classification is limited by the diagnostics the software provides. Different algorithms may use different tolerances, and “near singular” is not a universal numerical threshold. Record the software’s message, the relevant estimates or matrix, and the objective convention so another analyst can understand what the classification is based on.
Check Your Understanding
For each item, identify the best-supported classification. Solutions follow the questions.
- A fit reaches its iteration limit, reports that convergence criteria were not met, and ends with an objective value of 850.2. What can you conclude from the objective value and status?
- A variance estimate is exactly zero, the software reports convergence, and the variance is constrained to be nonnegative. Is this necessarily optimizer non-convergence?
- A reported two-by-two random-effects covariance matrix is \(\begin{pmatrix}0.09&0\\0&0\end{pmatrix}\). Calculate its determinant and classify the matrix.
- A converged one-compartment fit has \(V=10\) L and \(CL=0.10\) L/h after a 50 mg intravenous bolus. What concentration does it predict at 10 h? If the hypothetical observed concentration is 0.50 mg/L, is the result automatically plausible because the run converged?
Solutions. (1) The fit is non-converged because the message says the criteria were not met. The finite objective \(850.2\) is simply the value at termination and does not establish successful convergence. (2) No. Zero is the lower boundary of the allowed variance space, so this is a boundary estimate with a converged status; it should not be mislabeled as optimizer non-convergence. (3) The determinant is \((0.09)(0)-(0)(0)=0\), so the matrix is singular and has rank one. (4) The initial concentration is \(50/10=5\) mg/L and \(CL/V=0.10/10=0.0100\) h\(^{-1}\). Thus \(C(10)=5\exp[-(0.0100)(10)]=5\exp(-0.100)\approx4.52\) mg/L. This is far above 0.50 mg/L, so the solution is not automatically plausible; convergence status and the prediction-based concern should be reported separately.