Flat Profiles as Evidence of Practical Nonidentifiability

Why a Flat Profile Matters

An NLME fit can return a precise-looking estimate even when the observed dosing and sampling design contains little information about that parameter. The estimate may be a numerical maximizer, but many substantially different values may fit almost as well once the remaining parameters are re-estimated. A profile likelihood makes this lack of information visible: it asks how much the best achievable likelihood deteriorates when one parameter is fixed at different values.

A broad, nearly flat profile is evidence of practical nonidentifiability: under the specified model and observed design, the data do not distinguish a wide range of values well. It is not, by itself, proof that the model is structurally nonidentifiable, nor proof that the optimizer failed. Those are different explanations and require different checks.

Definition and Conditions for Interpretation

Let $\psi$ be the parameter of interest and let $\lambda$ collect all other model parameters, including nuisance fixed effects and random-effects or residual parameters as appropriate. Write $\ell(\psi,\lambda)$ for the log likelihood of the observed data under the specified NLME model. The likelihood used here should be the model likelihood for the observed data, with random effects handled consistently with the model, rather than a likelihood that changes definition across profile points.

Definition: The profile log likelihood for $\psi$ is $$ \ell_p(\psi)=\max_{\lambda\in\Lambda(\psi)}\ell(\psi,\lambda), $$ where $\Lambda(\psi)$ is the admissible set of nuisance-parameter values when $\psi$ is fixed. The maximized log likelihood is $\hat{\ell}=\max_{\psi,\lambda}\ell(\psi,\lambda)$. The profile likelihood-ratio deviance is $$ D(\psi)=2\{\hat{\ell}-\ell_p(\psi)\}. $$

A useful profile requires that nuisance parameters be reoptimized at each fixed value of $\psi$. Holding them at their full-model estimates instead produces a conditional slice, not the profile defined above. The maximization must also be adequate: a poorly optimized point can make $\ell_p(\psi)$ appear artificially low and the profile appear more curved than it is.

The usual approximate 95% likelihood-ratio reference for one interior parameter is $D(\psi)=3.84$, the 95th percentile of a chi-squared distribution with one degree of freedom. This approximation relies on regularity conditions, including a correctly specified likelihood, an interior true parameter value, and sufficiently informative data for the large-sample approximation. At a boundary, or under severe nonregularity, this cutoff may not apply; boundary behavior is addressed in the earlier material on profile intervals for variance components at zero. Here the cutoff is a scale for interpreting profiles, not a guarantee that every NLME profile has exact chi-squared calibration.

Why Flatness Indicates Weak Information

Near a well-behaved maximum $\hat{\psi}$, a profile may be approximated locally by a quadratic:

\ell_p(\psi)\approx\hat{\ell}-\frac{1}{2}I_{\mathrm{prof}}(\psi-\hat{\psi})^2, \qquad D(\psi)\approx I_{\mathrm{prof}}(\psi-\hat{\psi})^2,

where $I_{\mathrm{prof}}$ is the local curvature after nuisance parameters have been reoptimized. The second relation follows by subtracting the first approximation from $\hat{\ell}$ and multiplying by two. If $I_{\mathrm{prof}}$ is small, moving $\psi$ a substantial distance changes the log likelihood little. Consequently, the deviance rises slowly and a likelihood-based uncertainty region is broad. This is a local justification, not a claim that every profile is quadratic across its full range.

In an NLME model, weak information can arise when different parameter combinations produce similar predicted concentration-time patterns. For example, a change in clearance may be offset by a change in volume or bioavailability over the observed sampling window. The profile can then remain high as $\psi$ changes because the nuisance parameters compensate. This trade-off is one reason to inspect the nuisance estimates along the profile, not only the plotted likelihood.

Worked Example: A Broad Clearance Profile

Suppose a hypothetical NLME analysis estimates typical clearance $CL$, measured in L/h. The maximized log likelihood is $\hat{\ell}=-120.00$ at $\widehat{CL}=1.20$ L/h. After reoptimizing all nuisance parameters, the profile log likelihoods are as follows.

Fixed $CL$ (L/h)	$\ell_p(CL)$	$D(CL)=2\{\hat{\ell}-\ell_p(CL)\}$
0.80	-120.08	0.16
1.00	-120.01	0.02
1.20	-120.00	0.00
1.40	-120.03	0.06
1.60	-120.10	0.20

At $CL=0.80$ L/h, for example, $D=2\{-120.00-(-120.08)\}=2(0.08)=0.16$. At $CL=1.60$ L/h, $D=2\{-120.00-(-120.10)\}=0.20$. Both values are far below the approximate 95% reference value of 3.84, and the entire displayed range is similarly close to the maximum.

The profile therefore provides little evidence, within the displayed range, for distinguishing these clearance values by likelihood. It does not establish that every value from 0.80 to 1.60 L/h is scientifically plausible, nor does this finite grid identify the endpoints of a confidence interval. It does show that this design and model yield weak likelihood information about clearance over that range, assuming the profile maximizations are reliable.

Distinguishing Weak Information from Optimizer Failure

A profile point is computed by constrained optimization. If that optimization fails, the reported value can be lower than the true profile value. A sequence of such failures may create a false bowl or a false impression of strong information. Conversely, a poor full-model fit can make all profile points misleading. A visually flat curve should therefore be interpreted only after checking the numerical construction of the profile.

Check the full-model fit.
Confirm that the unconstrained fit is converged and that its log likelihood is at least as high as every constrained profile value. If a profile value exceeds the purported full-model maximum, the full fit or profile computation is inconsistent.

Repeat constrained fits.
At selected fixed values, use the full-model estimates, nearby profile solutions, and additional reasonable starting values. Reoptimize nuisance parameters each time and compare the resulting log likelihoods.

Inspect convergence, not only software status.
Review convergence messages, gradients or other stopping criteria, boundary hits, and parameter estimates. A nominal success flag alone does not show that the constrained maximum was found.

Look for a coherent profile.
Successful points should vary smoothly in a way compatible with the fitted model. Abrupt jumps, isolated dips, or large changes after a new start are warnings to investigate rather than evidence about parameter precision.

Worked Example: A False Bowl from Incomplete Optimization

Consider a hypothetical profile for clearance $CL$, measured in L/h, with a verified full-model maximum $\hat{\ell}=-100.0$. At each fixed value, the first optimization uses one starting point; a second set of fits uses additional starts and checks convergence.

Fixed $CL$ (L/h)	First fit $\ell_p$	Checked fit $\ell_p$	Deviance from checked fit
8	-102.0	-101.9	3.8
10	-101.0	-100.4	0.8
12	-103.0	-100.0	0.0
14	-102.0	-101.8	3.6

For the checked fit at $CL=10$ L/h, $D=2\{-100.0-(-100.4)\}=0.8$. At $CL=12$ L/h, the checked fit reaches the full-model maximum, so $D=0$. The initial fit at $CL=12$ had log likelihood $-103.0$, which is 3.0 log-likelihood units below the checked result at the same fixed value. Its apparent dip was caused by the unsuccessful optimization, not by the profile.

The checked deviance values, 3.8, 0.8, 0.0, and 3.6, show a substantially sharper local pattern than the first fits suggest. The endpoint values are near the approximate 95% reference of 3.84, so this small grid alone does not locate precise interval endpoints. The lesson is that repeat optimization can change the diagnosis: a curve should not be called flat or steep based on unreliable constrained fits.

Design Dependence and Parameter Trade-Offs

Practical identifiability is conditional on the model, the parameterization, the observed data, and the design. A parameter may be weakly informed because observations miss the part of the response curve that is most sensitive to it, because measurements are noisy, or because another parameter can compensate for its effect. The profile is therefore a diagnostic of information under the analysis at hand, not a permanent label attached to a parameter.

For example, an absorption-rate parameter may be difficult to estimate from a design with few early samples. Additional early observations may make the predicted curves for different absorption rates more distinguishable. A flat profile can also be asymmetric: values on one side of the estimate may be weakly constrained while values on the other side are more strongly constrained. A single standard error or a visual impression of curvature near the maximum can hide this behavior.

Worked Example: Sampling Design Changes Information

Suppose a hypothetical model estimates an absorption rate $k_a$, measured in h$^{-1}$. For a sparse design with few early samples, assume the checked profile has maximum log likelihood $-50.00$ at $k_a=2.0$ h$^{-1}$. The deviance values at $k_a=0.5, 1.0, 2.0,$ and $4.0$ h$^{-1}$ are respectively 0.20, 0.05, 0.00, and 0.20. For instance, $D(0.5)=0.20$ corresponds to $\ell_p(0.5)=-50.00-0.20/2=-50.10$.

Now consider a hypothetical enriched design that adds informative early samples. Its own maximum log likelihood is $-60.00$, and its deviance values at the same four fixed $k_a$ values are 5.0, 1.5, 0.0, and 4.2. The value at $k_a=4.0$ corresponds to $\ell_p(4.0)=-60.00-4.2/2=-62.10$.

The absolute log likelihoods should not be compared across these different data sets. The within-design deviances show the relevant contrast: the sparse design changes very little across this range, whereas the enriched design has deviance above 3.84 at both $0.5$ and $4.0$ h$^{-1}$. Under the approximate one-parameter reference, this suggests appreciably more information about $k_a$ in the enriched design. It does not prove that the model is correct or that the enriched design will solve every estimation problem; it illustrates how sampling can change practical identifiability.

Comparisons, Limitations, and Common Mistakes

A local covariance estimate or Wald interval summarizes uncertainty using curvature near the fitted maximum, often relying on an approximately quadratic log likelihood. It can be informative when the profile is smooth, approximately symmetric, and well behaved near its maximum. A profile instead displays how likelihood changes over a range of fixed parameter values while nuisance parameters are reoptimized, so it can reveal asymmetry, plateaus, and parameter trade-offs that a local summary obscures. Bootstrap distributions provide another view of estimator variability across resampled data sets, but they can themselves be affected by failed fits and by the resampling design. These approaches answer related but not identical questions; agreement is reassuring, while disagreement calls for investigation rather than automatic preference for one number.

A flat profile is evidence of weak practical information, not a unique diagnosis of its source. It does not alone prove structural nonidentifiability, which concerns whether distinct parameter values can produce the same data distribution under the model. A finite observed design can yield weak information even when the model is structurally identifiable. Nor does a flat profile alone justify removing a parameter, fixing it to a convenient value, or changing the model; those choices require scientific justification and further evaluation.

Common mistake: “The optimizer returned an estimate, so the parameter is identified precisely.” An optimizer returning a maximum only establishes that a numerical solution was reported. Correction: examine the checked profile: if a broad range has small deviance after nuisance reoptimization, the data provide limited likelihood information over that range, even if the point estimate is numerically stable.

Other warning signs include a profile that changes materially with starting values, frequent boundary estimates, sharp isolated jumps, or a full-model likelihood lower than a constrained fit. These patterns indicate numerical or modeling problems to resolve before interpreting flatness. If the profile is reliably broad, report that limitation in the parameter’s units and context, describe the range examined, and avoid implying precision unsupported by the data.

Key takeaway: A broad, reliably computed profile means that substantial changes in the fixed parameter cause little loss of likelihood after nuisance parameters are reoptimized. Treat it as evidence of practical nonidentifiability under the specified model and design, then check optimization, boundaries, and parameter trade-offs before drawing conclusions.

Check Your Understanding

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

At a fixed parameter value, the profile deviance is 0.6. What does this say, and what does it not say?
Why must nuisance parameters be reoptimized when constructing a profile likelihood?
A profile point has a lower log likelihood than expected, and a second starting value produces a much higher one. What is the appropriate next interpretation?
Why can an informative change to sampling times make a parameter’s profile less flat?

Solutions. (1) The fixed value loses only 0.6 deviance units relative to the maximum, indicating little likelihood deterioration at that point; it does not by itself establish a confidence interval or prove structural nonidentifiability. (2) Profiling asks for the best fit possible at each fixed value. Holding nuisance parameters fixed gives a conditional slice and can exaggerate the likelihood loss. (3) The constrained optimization may have failed to find the nuisance-parameter maximum. Check convergence, gradients, and repeated starts before interpreting the point as evidence of strong information. (4) New sampling times can make predicted responses under different parameter values more distinguishable, increasing profile curvature and the deviance away from the maximum.