Profile Likelihood for Vmax and Km

Why Profile Vmax and Km?

In a Michaelis-Menten elimination model, $V_{\max}$ and $K_m$ jointly determine the concentration-time curve. A fitted value for each parameter is only part of the result: the shape of the likelihood around that fit shows how strongly the data distinguish nearby values. This matters especially when the observed concentrations cover only a limited range. In that setting, multiple combinations of $V_{\max}$ and $K_m$ may fit almost equally well.

A standard error summarizes local curvature near the maximum likelihood estimate. That summary can be misleading when the likelihood is asymmetric, flat, or shaped like a ridge. A profile likelihood examines the fit across a wider range of one parameter while allowing the other parameters to adjust. It can therefore reveal uncertainty that a single local standard error does not describe well.

Definition and Assumptions

Definition: Let $\ell(\theta)$ be the log likelihood for the specified nonlinear mixed-effects model, and let $\psi$ denote the parameter of interest, either $V_{\max}$ or $K_m$. The profile log likelihood is $$ \ell_p(\psi)=\max_{\lambda}\ell(\psi,\lambda), $$ where $\lambda$ collects all remaining model parameters. At each fixed $\psi$, the nuisance parameters $\lambda$ are re-estimated. The profile likelihood is $L_p(\psi)=\exp\{\ell_p(\psi)\}$, and a convenient relative-likelihood statistic is $$ D(\psi)=-2\{\ell_p(\psi)-\ell_p(\hat{\psi})\}. $$ Here $\hat{\psi}$ is the unrestricted maximum likelihood estimate. Thus $D(\hat{\psi})=0$, and smaller values indicate profile fits closer to the best-fitting model.

For a standard marginal NLME likelihood, the likelihood for each subject is obtained by integrating over that subject’s random effects under the specified random-effects distribution. The profile calculation then maximizes this marginal likelihood over nuisance parameters at each fixed value of $\psi$. Nuisance parameters may include the other Michaelis-Menten parameter, other fixed effects, random-effects covariance parameters, and residual-error parameters. The fitted structural model, residual-error distribution, random-effects distribution, and treatment of observations must remain the same throughout the profile.

The usual approximate 95% likelihood-ratio interval for one parameter includes values satisfying $D(\psi)\leq 3.84$, the 95th percentile of a chi-square distribution with one degree of freedom. This approximation relies on a correctly specified likelihood, an interior and identifiable parameter value, and enough information for the likelihood-ratio statistic to have its usual large-sample distribution. It can be inaccurate with very small samples, parameters near boundaries, severe weak identification, or substantial model misspecification. A profile is a way to inspect the likelihood; it does not remove these assumptions.

How to Construct and Read a Profile

To profile $V_{\max}$, choose a sequence of plausible fixed values for $V_{\max}$. At each value, refit $K_m$ and all other nuisance parameters, recording the maximized log likelihood. To profile $K_m$, reverse their roles. The unrestricted fit should be included as a check: its parameter value should be at or near the minimum of $D$. Because $V_{\max}$ and $K_m$ are positive, a grid on a logarithmic scale can be useful for covering a wide range without spending most of the grid at implausible large values.

A profile interval is found from the points where $D(\psi)$ reaches 3.84. If the profile is tabulated on a grid, a crossing can be approximated by interpolation between neighboring points. The interval is not necessarily symmetric around the estimate. If the profile remains below the cutoff across the full explored range, the grid does not establish a finite endpoint: extend the range or report that the endpoint was not found. Do not treat the edge of the grid as a confidence limit.

Worked Example: A Profile Interval for Vmax

Suppose a hypothetical one-compartment concentration model uses Michaelis-Menten elimination, with $V_{\max}$ measured in mg/L/h and $K_m$ measured in mg/L. The observed concentrations span 2 to 18 mg/L. The residual-error and random-effects structures are held fixed in form, and at each fixed $V_{\max}$, $K_m$ and the other nuisance parameters are re-estimated by the same marginal likelihood method. Assume the fit converges to the same relevant maximum at each grid point. The unrestricted fit gives $\hat V_{\max}=10$ mg/L/h and a minimum deviance of 100. The following values are hypothetical profile results.

Fixed Vmax (mg/L/h)	Profile deviance	D(Vmax)
8	105.0	5.00
9	101.5	1.50
10	100.0	0.00
12	101.0	1.00
14	104.0	4.00

For example, at $V_{\max}=8$ mg/L/h, the likelihood-ratio difference is $105.0-100.0=5.00$. The approximate 95% interval includes values with $D\leq 3.84$. On the lower side, the crossing lies between 8 and 9 mg/L/h. Linear interpolation gives $$ 8+\frac{5.00-3.84}{5.00-1.50}(9-8) =8.331\ \text{mg/L/h}. $$ On the upper side, the crossing lies between 12 and 14 mg/L/h: $$ 12+\frac{3.84-1.00}{4.00-1.00}(14-12) =13.89\ \text{mg/L/h}. $$ The approximate profile-likelihood interval is therefore 8.33 to 13.89 mg/L/h. Under the stated likelihood and regularity assumptions, it is a procedure that would cover the true $V_{\max}$ in about 95% of repeated samples. It is not a statement that the parameter has a 95% probability of lying in this particular interval.

Asymmetry and Weak Identification

The profile’s shape contains information beyond its cutoff. A profile that rises at different rates on the two sides of its minimum gives an asymmetric interval. A broad, shallow region means that substantially different parameter values produce nearly as good a fit after nuisance parameters adjust. For Michaelis-Menten elimination, this can happen when the data do not span concentrations informative about both the low-concentration slope and the high-concentration elimination limit.

Worked Example: An Asymmetric Profile for Km

Consider a hypothetical fit to concentration observations measured in mg/L. At every fixed $K_m$, the model re-estimates $V_{\max}$, random-effects parameters, and residual-error parameters using the same specified marginal likelihood. The unrestricted fit has $\hat K_m=4$ mg/L and minimum deviance 80. The profile values below are hypothetical.

Fixed Km (mg/L)	Profile deviance	D(Km)
1	88.0	8.0
2	83.0	3.0
4	80.0	0.0
6	80.8	0.8
8	82.0	2.0
10	84.5	4.5

The lower cutoff is between 1 and 2 mg/L. Interpolating to $D=3.84$ gives $$ 1+\frac{8.0-3.84}{8.0-3.0}(2-1) =1.832\ \text{mg/L}. $$ The upper cutoff is between 8 and 10 mg/L: $$ 8+\frac{3.84-2.0}{4.5-2.0}(10-8) =9.472\ \text{mg/L}. $$ Thus, the approximate profile interval is 1.83 to 9.47 mg/L. Its lower and upper distances from the estimate are different: $4-1.832=2.168$ mg/L below and $9.472-4=5.472$ mg/L above. Reporting only a symmetric estimate-plus-or-minus-standard-error summary could hide this unequal uncertainty.

A flat profile can indicate that the data weakly identify a parameter, not that the parameter is unimportant. At concentrations much smaller than $K_m$, $$ \frac{V_{\max}C}{K_m+C}\approx \frac{V_{\max}}{K_m}C. $$ In this range, the data may primarily inform the ratio $V_{\max}/K_m$, rather than the two parameters separately. This approximation follows because $K_m+C$ is close to $K_m$ when $C/K_m$ is small. At concentrations much larger than $K_m$, the rate is approximately $V_{\max}$, so the data may inform the elimination limit but provide less direct information about $K_m$. These are explanations for possible profile shapes, not proof that a particular fit is correctly specified.

Worked Example: A Shallow Profile Along a Parameter Ridge

Suppose a hypothetical study records concentrations only between 0.2 and 0.9 mg/L, while fitted $K_m$ values are much larger. The best fit is at $V_{\max}=40$ mg/L/h and $K_m=20$ mg/L, with minimum deviance 50. The following fitted pairs are hypothetical; each pair is obtained by fixing $V_{\max}$ and re-estimating $K_m$ and other nuisance parameters.

Fixed Vmax (mg/L/h)	Re-estimated Km (mg/L)	Profile deviance	D(Vmax)
20	10	50.2	0.2
40	20	50.0	0.0
80	40	50.8	0.8

The ratios are $20/10=2$, $40/20=2$, and $80/40=2$ h$^{-1}$. Because the observed concentrations are small relative to these $K_m$ values, these parameter pairs imply similar low-concentration slopes. Each profile difference, $0.2$, $0$, and $0.8$, is below 3.84. The data therefore do not distinguish these $V_{\max}$ values well by this profile. This three-point grid does not determine the full interval or establish whether an endpoint exists; the profile should be extended and refined before reporting endpoints. The appropriate conclusion is weak information about $V_{\max}$ separately, not a precise estimate merely because the optimizer returned 40 mg/L/h.

Justification, Diagnostics, and Alternatives

The likelihood-ratio profile statistic compares the best fit with $\psi$ fixed to the unrestricted best fit. The unrestricted fit is at least as good, because it can choose the same $\psi$ and nuisance values available to the constrained fit. Therefore $D(\psi)\geq 0$. Under regular large-sample conditions, the likelihood-ratio statistic for one fixed parameter value has approximately a chi-square distribution with one degree of freedom. Comparing it with 3.84 yields the usual approximate 95% one-parameter cutoff. This explains why a profile interval is based on likelihood differences, rather than on a fixed distance from the estimate.

Several checks help establish whether a profile is trustworthy:

Check the minimum. The profile should reach its lowest deviance near the unrestricted estimate. A mismatch can signal inconsistent optimization, an insufficient grid, or convergence to different local maxima.
Check convergence and refitting. At each fixed value, nuisance parameters must be re-estimated. Reusing their unrestricted estimates without refitting is not the profile likelihood. Use multiple starting values when optimization appears sensitive, and verify that the reported solution is the best converged fit found.
Check grid coverage and resolution. Add points near the minimum and near each cutoff crossing. Extend a side that remains below the cutoff; do not infer a finite limit from a grid boundary.
Check parameter constraints. $V_{\max}$ and $K_m$ are positive. A profile near zero or another imposed boundary may violate the regularity conditions behind the chi-square cutoff, so the usual interval may need additional justification.
Check the model assumptions. A precise-looking profile cannot correct a misspecified elimination model, residual-error distribution, or random-effects distribution. Inspect whether fitting and likelihood calculations used the intended model consistently.

A Wald interval uses the estimate and its standard error, typically from local curvature around the likelihood maximum. It is computationally convenient and can be adequate when the likelihood is approximately quadratic and symmetric. A profile interval instead allows nuisance parameters to adapt and makes non-quadratic shape visible, though it costs more computation and still relies on likelihood assumptions. A joint likelihood contour for $(V_{\max},K_m)$ shows their combined uncertainty and any ridge directly; its cutoff differs from the one-parameter profile cutoff. A parametric bootstrap is another option for assessing sampling behavior, particularly when asymptotic approximations are doubtful, but it depends on the fitted model and can also be computationally demanding.

Profiles have limitations. They summarize uncertainty conditional on the chosen model and the likelihood used. They do not identify the cause of a flat or unusual profile, guarantee that the model is scientifically adequate, or make an uninformative design informative. Nor should the 3.84 cutoff be treated as exact for a small sample or a boundary case. Report the profile interval together with its approximation, the parameter units, relevant fitting choices, and any endpoint that could not be found.

Key takeaway: A profile likelihood fixes $V_{\max}$ or $K_m$, re-estimates the nuisance parameters, and measures the resulting loss of fit. Its asymmetry and flatness reveal features of uncertainty that a local standard error may miss.

Check Your Understanding

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

When profiling $K_m$, which parameters must be re-estimated, and why is holding $V_{\max}$ fixed at its unrestricted estimate generally incorrect?
A profile deviance is 103.2 at a fixed parameter value and 100.0 at the unrestricted maximum. Calculate $D$ and state whether this value is inside the usual approximate 95% one-parameter profile interval.
For a $V_{\max}$ profile, $D=2.0$ at 12 mg/L/h and $D=5.0$ at 14 mg/L/h. Use linear interpolation to estimate the upper 95% cutoff crossing.
Why can observations only at concentrations far below $K_m$ leave $V_{\max}$ and $K_m$ weakly distinguished?
A profile stays below 3.84 at every point up to the largest tested $K_m$. What can and cannot be concluded?

Solutions. 1. Fix $K_m$ at each grid value and re-estimate $V_{\max}$ and all other nuisance parameters. Otherwise the likelihood is evaluated with nuisance values that may no longer fit best, so it is not the profile likelihood. 2. $D=103.2-100.0=3.2$. Since $3.2<3.84$, this value is inside the usual approximate 95% profile interval, subject to the likelihood and regularity assumptions. 3. Interpolate between the two points: $12+\{(3.84-2.0)/(5.0-2.0)\}(14-12)=13.23$ mg/L/h. 4. When $C\ll K_m$, the rate is approximately $(V_{\max}/K_m)C$, so many pairs with a similar ratio can produce similar predictions. 5. The explored values have not reached the usual cutoff, so the upper endpoint has not been found on that grid. Extend the profile to determine whether and where it crosses; do not report the grid maximum as a confidence limit.