Why Profile a Bounded Parameter?
A parameter estimate at a bound is a warning to investigate, not a diagnosis by itself. For example, a fitted interindividual variance of zero might reflect a likelihood that genuinely favors the boundary, a broad region of nearly equivalent values, or an optimizer that stopped at the bound without adequately exploring the likelihood. A local curvature summary can be uninformative in these cases: near a boundary, the likelihood may be asymmetric, and a flat direction may make curvature estimates unstable.
A profile likelihood examines the fitted likelihood across values of one parameter while allowing the other parameters to adjust. Its shape shows how much support the data provide for values away from the reported estimate. It can reveal asymmetry, flatness, and whether the best fit is at a parameter bound. It does not, on its own, prove that a zero estimate represents a genuine absence of an effect or validate the model.
The profile is a function of the fixed parameter value, not just a line joining estimates from different fits. At each grid point, nuisance parameters are reoptimized subject to the fixed value. If software reports an objective function equal to minus twice the log likelihood, the corresponding difference in objective values is \(D(\theta)\), provided the same likelihood, data, and conventions are used for every fit. If the objective contains additional terms or is not a likelihood, the likelihood-ratio interpretation needs separate justification.
Constructing and Reading the Profile
Choose a range of admissible values for \(\theta\), including the reported estimate and the relevant bound. Fit the model with \(\theta\) fixed at each grid value, reoptimizing \(\eta\). Record the optimized log likelihood and calculate \(D(\theta)\). The global maximum should be at least as large as every constrained maximum, so \(D(\theta)\geq 0\). A negative value beyond numerical rounding, or a constrained fit with a better likelihood than the purported global fit, signals that the unconstrained optimization or the profile calculations need investigation.
A low value of \(D(\theta)\) means the best fit at that fixed value is close in log likelihood to the overall best fit. A rapidly increasing profile indicates that moving away from the optimum incurs a substantial likelihood loss. An asymmetric profile means the likelihood falls at different rates on either side; reporting a symmetric uncertainty based only on local curvature may then give a poor summary. A flat stretch indicates that a range of values fits almost equally well, subject to the model and data used.
State the parameter’s meaning and units, its permitted range, and the nuisance parameters that will be reoptimized.
Record the best log likelihood and verify that the fit is not merely an optimizer stopping point. Retain the estimates and diagnostics for later comparison.
Fix the parameter of interest at each chosen value and maximize over the remaining admissible parameters. Use adequate optimization starts and tolerances.
Compute \(D(\theta)\), check that it is nonnegative up to numerical tolerance, and refine the grid where the curve changes sharply or approaches a decision threshold.
The likelihood-ratio statistic has a familiar reference distribution only under conditions. For a regular, identifiable model with a smooth likelihood and a true parameter in the interior of its permitted range, the large-sample distribution of \(D(\theta)\), evaluated at the true value, is approximately chi-squared with one degree of freedom when one scalar parameter is tested. That result motivates the usual cutoff of about 3.84 for a nominal 95% likelihood-ratio interval in such settings. A bounded parameter whose optimum is at, or close to, a boundary does not automatically satisfy the interior condition. The usual cutoff may therefore be inappropriate; boundary-specific calibration, simulation, or another justified method may be needed.
Worked Profiles
Worked Example: A Variance Estimate at Zero
Suppose a hypothetical pharmacometric model estimates an interindividual variance \(\omega^2\), constrained by \(\omega^2\geq 0\). The unrestricted fit reports \(\hat{\omega}^2=0\) and \(\ell(\hat{\omega}^2,\hat{\eta})=-120.00\). Assume that each fixed-value fit below successfully reoptimizes all other parameters. The resulting profile values are:
| Fixed \(\omega^2\) | Profile log likelihood \(\ell_p(\omega^2)\) | \(D(\omega^2)\) |
|---|---|---|
| 0.00 | -120.00 | 0.00 |
| 0.01 | -120.18 | 0.36 |
| 0.04 | -121.10 | 2.20 |
| 0.09 | -123.20 | 6.40 |
At \(\omega^2=0.04\), for example,
The largest profile log likelihood in this grid occurs at the lower bound, zero. The increase in \(D\) shows that larger fixed variances fit less well in this hypothetical fit, but the grid alone does not establish that the underlying variance is truly zero. For orientation only, if one used the regular-model cutoff 3.84, the crossing would lie between 0.04 and 0.09. Linear interpolation of the displayed \(D\) values gives
This is a rough crossing estimate from a coarse grid, not a justified 95% upper confidence limit: the optimum is at a boundary, so the regular cutoff is not automatically valid. A finer profile and an appropriate boundary calibration would be needed for inferential reporting. The example’s diagnostic conclusion is narrower: the fitted maximum is at the bound, and the likelihood profile provides information about how quickly fit worsens as the variance is moved away from zero.
Worked Example: An Asymmetric Interior Profile
Consider a hypothetical positive clearance parameter \(CL\), measured in L/h, with an interior estimate \(\widehat{CL}=0.16\) L/h and maximum log likelihood \(-200.00\). The reoptimized profile has the following values:
| Fixed \(CL\) (L/h) | \(\ell_p(CL)\) | \(D(CL)\) |
|---|---|---|
| 0.02 | -202.70 | 5.40 |
| 0.04 | -201.80 | 3.60 |
| 0.08 | -200.70 | 1.40 |
| 0.16 | -200.00 | 0.00 |
| 0.24 | -200.40 | 0.80 |
| 0.32 | -201.35 | 2.70 |
| 0.40 | -202.60 | 5.20 |
For \(CL=0.08\) L/h, the profile statistic is
The profile is asymmetric: at 0.08 L/h, which is 0.08 L/h below the estimate, \(D=1.40\); at 0.24 L/h, the same distance above it, \(D=0.80\). Thus, the likelihood falls more sharply on the lower side across those particular points. If the regular chi-squared approximation were justified, interpolation at the 3.84 cutoff would give a rough lower crossing of
and a rough upper crossing of
These interpolated values suggest an asymmetric profile-based interval of about 0.0373 to 0.356 L/h under that cutoff. They are not final endpoints: a finer grid should locate the crossings, and the inferential cutoff must be appropriate for the model and parameter. In this example the estimate is interior; the table alone does not establish that all regularity conditions hold.
Worked Example: A Flat Profile at a Bound
Suppose a hypothetical fixed-effect slope \(\beta\), measured in concentration units per mg, is constrained to be nonnegative. The best fit is at \(\hat{\beta}=0\), with \(\ell_p(0)=-80.00\). Reoptimization gives:
| Fixed \(\beta\) (concentration units/mg) | \(\ell_p(\beta)\) | \(D(\beta)\) |
|---|---|---|
| 0.00 | -80.00 | 0.00 |
| 0.10 | -80.025 | 0.05 |
| 0.20 | -80.060 | 0.12 |
| 0.30 | -80.105 | 0.21 |
| 0.40 | -80.200 | 0.40 |
At \(\beta=0.30\), for example, \(D=2\{-80.00-(-80.105)\}=0.21\). The profile remains close to its maximum over the displayed range. This is evidence of a flat profile over that range, not strong evidence that the effect is exactly zero. The grid has not even reached a 3.84 crossing; one must extend it within scientifically and mathematically admissible bounds before deciding whether the profile closes on the upper side. If it does not, the data may provide only a weak upper constraint under the chosen model and calibration.
Why the Shape Matters
A useful way to understand profile asymmetry is to compare the profile with a local quadratic approximation. Near a well-behaved interior maximum, a second-order expansion of the log likelihood gives a locally quadratic loss. If the curvature is stable, that approximation makes \(D(\theta)\) roughly symmetric around its maximum. But the approximation is local: it need not describe a broad range of parameter values, and it is especially unreliable when a bound truncates one side of the parameter space.
where \(I_p(\hat{\theta})\) denotes local profile curvature at the maximum. This is a local justification for a symmetric, curvature-based approximation when its assumptions hold—not a guarantee that the observed profile is symmetric. If the curve is asymmetric or nearly flat, a single local curvature value can conceal important features. Near a boundary, only one side may be admissible, and the standard interior likelihood-ratio result may not apply.
For a simple scalar parameter at a regular boundary, a common asymptotic likelihood-ratio result under a boundary null is a mixture: probability one-half at zero and probability one-half on a chi-squared distribution with one degree of freedom. This result requires its own model and sampling assumptions. It is not a universal correction for every pharmacometric boundary problem; nuisance parameters, weak identification, small samples, or other irregularities can change the calibration. The profile curve should therefore be separated from the choice of inferential cutoff.
Diagnostics, Alternatives, and Limitations
A profile is only as reliable as the constrained fits used to construct it. At each fixed value, check convergence, gradients, objective values, and whether different starts lead to the same optimized likelihood. The profile should not show unexplained jumps caused by failed optimization or by a switch between local solutions. Increase grid density near sharp bends, apparent plateaus, or likelihood-ratio crossings. When a constrained likelihood exceeds the claimed unrestricted maximum, refit the unrestricted model and inspect the fitting procedure before interpreting the curve.
A profile is most informative when compared with other diagnostics rather than used in isolation. A Hessian or covariance-based standard error summarizes local behavior near the estimate and is computationally economical, but it relies on local curvature and can be misleading for a boundary estimate or an asymmetric profile. An optimizer trace helps show the path taken by a fit but does not describe the likelihood over a parameter range. A profile explicitly explores that range, at the cost of many constrained fits. These approaches answer related but different questions.
A flat profile can signal limited information about the parameter, but it does not by itself distinguish structural from practical nonidentifiability or identify the scientific source of the weakness. A profile can also look flat because the grid is too coarse, numerical tolerances are loose, or the nuisance-parameter optimization is unreliable. Conversely, a steep profile does not guarantee that the model is correctly specified. Report the parameter’s units and bounds, the fitting and profiling procedure, the profile scale, and how any interval cutoff was chosen.
The practical value of profiling is diagnostic as well as inferential. A boundary peak with a steep rise away from the bound, a broad plateau, and an asymmetric interior peak are visibly different likelihood geometries. Describing which geometry is present is more informative than reporting only the optimizer’s boundary estimate or a standard error that may not adequately summarize the likelihood.
Check Your Understanding
Answer each question, then compare with the solutions.
- In defining \(\ell_p(\theta)\), what is held fixed and what is maximized?
- A fitted model has maximum log likelihood \(-50.00\). At a fixed parameter value the reoptimized profile log likelihood is \(-51.25\). Calculate \(D(\theta)\).
- Why can a flat profile near a boundary not establish that the parameter is exactly zero?
- A profile gives a constrained log likelihood larger than the reported unrestricted maximum. What should be checked first?
- Why might the usual 3.84 likelihood-ratio cutoff fail when the optimum is at a parameter boundary?
Solutions. (1) Hold \(\theta\) at its selected value and maximize the log likelihood over all admissible nuisance parameters \(\eta\). (2) \(D=2\{-50.00-(-51.25)\}=2(1.25)=2.50\). (3) A flat profile means the likelihood changes little across the examined values, so the data may not distinguish them well; it does not establish an exact value. (4) Check whether the unrestricted fit reached the global maximum and whether the constrained fit and likelihood evaluations used consistent settings; then refit as needed. (5) The usual cutoff relies on regular interior-parameter asymptotics. A boundary violates the interior condition, so a boundary-appropriate calibration may be required.