Why Look Beyond the Best-Fitting Covariate Function?
A covariate exponent or nonlinear coefficient is usually reported as a single estimate. That estimate identifies the value giving the best fit under a specified model, but it does not show how much the fit would change for nearby values. A narrow, sharply curved objective-function minimum suggests that the data distinguish the estimate from alternatives. A broad or asymmetric minimum conveys a different message: many values may fit almost as well, or uncertainty may be greater on one side.
A likelihood profile makes this uncertainty visible. It is particularly useful when the effect parameter is nonlinear in the likelihood, when its likelihood is not close to a symmetric quadratic, or when a Wald standard error gives an unconvincing summary. The profile is conditional on the chosen covariate function and the rest of the model; it does not by itself determine which functional form should be selected.
Definition and Construction
Let \(\psi\) be the covariate functional-form parameter of interest, such as an exponent in a power relationship or a coefficient multiplying a nonlinear basis term. Let \(\lambda\) denote all remaining estimable parameters, including other fixed effects, variance parameters, and any nuisance parameters in the specified model. Write the log likelihood as \(\ell(\psi,\lambda)\). In a population PK analysis, the likelihood used for this purpose should be the same marginal or consistently approximated likelihood used to fit the model.
The profiling step matters. Holding \(\psi\) fixed while also holding every other parameter at its unrestricted estimate does not produce a profile: it produces a conditional slice that may overstate how much fit is lost. A profile instead allows other parameters to compensate as far as the model permits. The result is a curve of best achievable fit over values of \(\psi\).
For each grid value, the estimation procedure must converge to a suitable optimum. The curve should be based on the same model, data, likelihood approximation, and parameter constraints at every point. A grid that is too coarse can miss curvature or a crossing; a search that fails at some grid values can create apparent jumps that are optimization failures rather than likelihood features.
Interpreting the Profile Difference
Under regular likelihood conditions, the likelihood-ratio statistic for a fixed value \(\psi_0\) is \(2\{\ell(\hat{\psi},\hat{\lambda})-\ell_p(\psi_0)\}\). Since the OFV is minus twice the log likelihood, this statistic is exactly \(D(\psi_0)\). For a single freely varying scalar parameter, it is approximately distributed as chi-square with one degree of freedom when \(\psi_0\) is the true value and the usual large-sample regularity conditions hold.
Consequently, an approximate 95% likelihood-based confidence set contains values satisfying \(D(\psi)\leq \chi^2_{1,0.95}\), approximately \(3.84\). Its endpoints are where the profile difference crosses 3.84. This set need not be symmetric around the estimate. It can also be disconnected or unbounded, in which case the shape itself is important evidence about the information in the data. The cutoff is an asymptotic reference, not a guarantee of exact coverage in every population PK setting; boundary parameters, weak identification, small samples, or a problematic likelihood approximation can undermine it.
Worked Example: Profiling a Power-Effect Exponent
Suppose a hypothetical model describes clearance as \(\mathrm{CL}_i=\mathrm{CL}_{\mathrm{typ}}(C_i/C_{\mathrm{ref}})^\theta\exp(\eta_i)\), where \(C_i\) is a covariate, \(C_{\mathrm{ref}}\) is its reference value, and \(\theta\) is the exponent of interest. The unrestricted fit gives \(\hat{\theta}=0.72\) and \(\mathrm{OFV}_{\min}=812.40\). At every fixed grid value of \(\theta\), the other model parameters are re-estimated.
| \(\theta\) | Profile OFV | \(D(\theta)\) |
|---|---|---|
| 0.20 | 817.60 | 5.20 |
| 0.40 | 814.50 | 2.10 |
| 0.60 | 812.70 | 0.30 |
| 0.72 | 812.40 | 0.00 |
| 0.80 | 812.48 | 0.08 |
| 1.00 | 813.10 | 0.70 |
| 1.20 | 815.00 | 2.60 |
| 1.40 | 817.50 | 5.10 |
The approximate 95% cutoff is 3.84. The lower endpoint lies between 0.20, where \(D=5.20\), and 0.40, where \(D=2.10\). Linear interpolation gives
The upper endpoint lies between 1.20, where \(D=2.60\), and 1.40, where \(D=5.10\):
Thus the interpolated profile-based interval is approximately 0.288 to 1.30. It is an interval for the exponent, which is dimensionless in this model. The estimate is 0.72, but the data are also compatible with a substantial range of exponents under the stated model. The endpoints are approximate because the profile was evaluated on a grid and interpolated; additional fits near the crossings would refine them.
Asymmetry and Nonlinear Terms
A profile’s shape can be more informative than its width alone. If the curve is steep on one side of its minimum and shallow on the other, the corresponding confidence limits will be asymmetric. A symmetric estimate-plus-or-minus-standard-error summary may then conceal a long tail of plausible values. Nonlinear covariate terms can produce this pattern because changes in the term may be partly offset by changes in other parameters.
Worked Example: An Asymmetric Profile for Curvature
Consider a hypothetical covariate model whose log typical parameter includes a nonlinear term \(qg(C)\), where \(g(C)\) is a prespecified, dimensionless curved function and \(q\) is its coefficient. The best fit has \(\hat{q}=0.15\) and minimum OFV 904.00. The profile values below result from re-estimating nuisance parameters at each fixed \(q\).
| \(q\) | Profile OFV | \(D(q)\) |
|---|---|---|
| -0.40 | 909.60 | 5.60 |
| -0.20 | 906.20 | 2.20 |
| 0.00 | 904.60 | 0.60 |
| 0.15 | 904.00 | 0.00 |
| 0.30 | 904.70 | 0.70 |
| 0.50 | 906.00 | 2.00 |
| 0.80 | 907.50 | 3.50 |
| 1.00 | 908.70 | 4.70 |
On the lower side, the 3.84 crossing is between \(-0.40\) and \(-0.20\). Linear interpolation gives
On the upper side, the crossing is between 0.80 and 1.00:
The approximate profile interval is \(-0.296\) to \(0.857\). Its distances from the estimate are about 0.446 below and 0.707 above, so the uncertainty is visibly asymmetric. The profile does not establish that curvature is required; it describes uncertainty in \(q\) within this specified model. Assessing whether the nonlinear form improves on a simpler model is a separate model-comparison question.
Flat Profiles, Diagnostics, and Limits
A profile can remain below the likelihood-ratio cutoff across the entire range searched. That is not evidence that every value in the range is equally plausible, nor does it prove that the covariate has no effect. It means the data and model have not produced a crossing within the searched range, so a finite interval endpoint has not been located. The range may need to be expanded, or the parameter may be weakly identified by the available data.
Worked Example: A Profile That Does Not Close
In a hypothetical analysis, a nonlinear coefficient \(q\) has minimum OFV 1220.00 at \(q=0.40\). Profile differences are 1.20 at \(q=-1\), 0.35 at \(q=0\), 0.00 at \(q=0.40\), 0.50 at \(q=1\), and 1.40 at \(q=2\). All these values are below 3.84. For instance, at \(q=2\), the profile OFV is \(1220.00+1.40=1221.40\), still below the cutoff value \(1220.00+3.84=1223.84\).
The 95% profile set has no observed crossing on either side in this grid. Reporting \(-1\) to 2 as the confidence interval would be incorrect: those are only the searched limits, not likelihood-based endpoints. A denser, wider profile can determine whether crossings occur farther out. If the curve remains shallow, the appropriate conclusion is that \(q\) is poorly determined under this model and data set; a precise finite interval is not supported by the profile.
Several checks help distinguish a genuine profile feature from a numerical artifact:
- Check the minimum. The lowest profile value should agree with the unrestricted fit, within the numerical precision of the estimation procedure.
- Check convergence and continuity. Refit grid points near abrupt jumps or unexpected local peaks, using suitable starting values and more than one start when needed.
- Refine the grid near crossings. Endpoint interpolation is only an approximation; direct fits near the threshold make it more reliable.
- Inspect constraints and search limits. A profile that meets a parameter boundary or ends at the grid limit has not necessarily located a likelihood-based endpoint.
- Assess the likelihood assumptions. The chi-square cutoff may be unreliable with boundary estimates, weak identification, small samples, or a poorly behaved likelihood approximation. The plotted shape does not repair those limitations.
Profiling complements rather than replaces other summaries. A Wald interval uses local curvature near the estimate and is convenient when the likelihood is approximately quadratic; it can be misleading when the profile is asymmetric or flat. A profile likelihood ratio uses the fit across fixed parameter values and directly reveals that shape, but requires repeated optimization and relies on likelihood-ratio calibration. A numerical objective-function scan can show the same curve, but only if nuisance parameters are re-estimated consistently at each point. Finally, profiles for distinct functional-form families are not automatically one common curve: comparing a power form with a spline, for example, requires a model-comparison framework rather than treating their parameters as interchangeable.
Derivation of the Profile Cutoff
The cutoff follows from the likelihood-ratio statistic for a candidate value \(\psi_0\). The unrestricted maximum log likelihood is \(\ell(\hat{\psi},\hat{\lambda})\); when \(\psi\) is fixed at \(\psi_0\), the best available log likelihood is \(\ell_p(\psi_0)\). Their difference cannot be negative because the unrestricted fit includes all parameter values available to the restricted fit. Multiplication by two gives the likelihood-ratio statistic, and replacing each log likelihood by minus one-half the OFV gives the profile difference exactly:
For one scalar parameter under regularity conditions, the likelihood-ratio statistic is approximately chi-square with one degree of freedom. Inverting that test retains candidate values for which \(D(\psi)\leq 3.84\) at the approximate 95% level. This inversion explains why the confidence-set endpoints are read from threshold crossings rather than from a fixed distance around the estimate.
Common Pitfall
Wrong reasoning: “The estimate is 0.72, so I can report 0.72 plus or minus 1.96 times its standard error as the uncertainty, and a nearly flat profile means the estimate is precise because the OFV barely changes.” The first claim assumes a locally quadratic likelihood; the second reverses the interpretation. If the OFV barely changes over a broad range, the data distinguish those parameter values poorly. Use the profile’s threshold crossings where they exist, check numerical stability, and describe a flat or open profile as weak information or unresolved endpoint—not as precision.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- When constructing a profile at a fixed exponent, what happens to the nuisance parameters?
- A profile minimum is 500.0 and the OFV at a candidate value is 503.2. What is the profile difference, and is the candidate inside the approximate 95% set using a 3.84 cutoff?
- A profile crosses the 3.84 cutoff at 0.1 below the estimate and 0.6 above it. What does this indicate about symmetry?
- A profile stays below 3.84 throughout the searched grid. Can the grid endpoints be reported as the confidence limits? Explain.
- Why is a profile for a coefficient within one spline model not, by itself, a comparison of that spline model with a power model?
Solutions. (1) They are re-estimated at each fixed exponent to maximize the likelihood, which defines the profile. (2) \(D=503.2-500.0=3.2\); since \(3.2\leq3.84\), the candidate is inside the approximate 95% profile set. (3) The uncertainty is asymmetric: the upper-side distance is six times the lower-side distance. (4) No. The endpoints are only search limits; expand or refine the profile, and report that a finite crossing has not been found if it remains open. (5) The profile varies one parameter while holding the model family fixed; comparing different functional forms requires a separate model-comparison procedure.