Testing Saturable Against Linear Elimination

Why the comparison needs care

A Michaelis–Menten elimination model can describe concentration-dependent clearance, while a linear model describes elimination proportional to concentration. Comparing the models is not simply a matter of fitting both and applying a standard likelihood-ratio test. That test requires a particular relationship between the models and regular parameter behavior. Here, both can fail: a pure linear model is a limiting approximation to a pure Michaelis–Menten model, not generally a finite-parameter special case, and a saturable pathway added to a linear model has parameters on a boundary under the linear null.

A sound comparison therefore begins by stating exactly which models are being compared. Information criteria can compare models that are not nested, subject to their assumptions and limitations. A likelihood-ratio test is appropriate only when its nesting and reference-distribution conditions are justified.

What is nested, and what is not?

At concentrations small relative to \(K_m\), the Michaelis–Menten rate is approximately linear:

Thus, at low concentrations the ratio \(V_{\max}/K_m\) behaves like a clearance. This explains why the pure Michaelis–Menten model may resemble linear elimination over a restricted concentration range. It does not make the linear model a regular, finite-parameter special case of the pure Michaelis–Menten model: approaching the linear limit while keeping \(V_{\max}/K_m\) fixed requires \(K_m\) and \(V_{\max}\) to grow together. A conventional likelihood-ratio test between these two pure models is therefore not justified by ordinary nested-model theory.

The parallel model provides a different comparison. The linear-only model is obtained by setting \(V_{\max}=0\). But \(V_{\max}\) is constrained to be nonnegative, so the null value is on the boundary. Moreover, when \(V_{\max}=0\), the rate no longer depends on \(K_m\): \(K_m\) is unidentified under the null. These features prevent the usual chi-square likelihood-ratio reference distribution from being assumed automatically.

Requirements for a fair model comparison

For likelihood-based comparisons, the models should use the same observations and the same likelihood definition. In an NLME analysis, that means keeping the observation records, residual-error specification, random-effects structure, handling of below-quantification observations, and other modeling choices aligned unless their change is part of the stated comparison. Likelihoods from different data or incompatible likelihood calculations cannot be compared as though they came from the same experiment.

For comparisons of fixed effects or structural models, use maximum likelihood rather than restricted maximum likelihood. The random-effects distribution and the residual model must be specified well enough for the likelihood to be meaningful; convergence to a numerical optimum should also be checked. A likelihood value from a run that failed to converge, or from a solution at an unintended parameter constraint, is not reliable evidence for either model.

Information criteria are useful for comparing nonnested models. For maximized log likelihood \(\ell\), \(k\) estimated parameters, and \(n\) independent sampling units, the standard forms are:

Smaller values indicate a better trade-off between fit and the criterion’s parameter penalty among the candidate models being compared; they do not establish that a model is true. In clustered NLME data, the appropriate BIC sample size depends on the asymptotic convention and software implementation. State whether \(n\) denotes subjects or records, and use the same convention for every candidate. Count all estimated parameters consistently, including residual and random-effects parameters.

Boundary tests and their limits

When a single nonnegative parameter is tested at zero and all other parameters remain identifiable and interior, a common asymptotic reference distribution is a mixture: one-half probability at zero and one-half probability on a chi-square distribution with one degree of freedom. This is not the usual chi-square distribution. The mixture accounts for the fact that, under the null, random score fluctuations in the direction of a constrained parameter can improve the alternative fit only when they point into the allowed region.

That simple mixture does not automatically apply to the parallel linear-plus-saturable model when \(K_m\) is estimated. Under the null \(V_{\max}=0\), \(K_m\) is unidentified, so the alternative contains a nuisance parameter that has no defined null value. The asymptotic distribution can then depend on the search over \(K_m\) and on the design and likelihood. Simply counting “one extra parameter” and comparing \(D\) to \(\chi^2_1\), or applying the one-boundary-parameter mixture without checking the unidentified parameter, can produce an unjustified p-value.

Parametric bootstrap when the reference law is uncertain

A parametric bootstrap can calibrate the likelihood-ratio statistic under a specified null model when standard asymptotic reference distributions are doubtful. Fit the null model, simulate new data under it while retaining the study design and the assumed random-effects and residual-error structure, then refit both null and alternative models to each simulated data set. The simulated statistics approximate the null distribution, including the fitting procedure and parameter constraints. This approach requires credible null-model simulation and reliable refitting; it does not repair a misspecified model.

Diagnostics, alternatives, and limitations

Before interpreting any comparison, check that both fits reached stable solutions and that their likelihoods are comparable. Examine parameter estimates, uncertainty, boundary hits, and sensitivity to starting values. A nominally better likelihood from a poorly converged alternative is not persuasive. If the estimated \(K_m\) is extremely large relative to observed concentrations, the data may provide little separation from an approximately linear rate; a large or unstable estimate can also signal weak identification rather than strong evidence for a particular saturation scale.

The main alternatives answer different questions. AIC and BIC compare relative fit with different complexity penalties and can compare nonnested models, but they do not provide a conventional null-hypothesis p-value. A standard likelihood-ratio test is attractive for regular nested models, but is inappropriate for pure linear versus pure Michaelis–Menten models and may be miscalibrated for a boundary model with unidentified \(K_m\). A boundary-mixture test is useful for a single constrained parameter only when the nuisance parameters remain identifiable under the null. A parametric bootstrap can reflect a more complicated null distribution, at the cost of computation and dependence on the assumed data-generating model.

No one criterion resolves poor design, weak concentration coverage, or misspecification. A comparison summarizes evidence within the candidate models and likelihood assumptions; it does not by itself establish the mechanism of elimination. Do not select a saturable model solely because it has a larger maximized likelihood: the larger alternative is expected to fit at least as well. Consider the penalty, uncertainty, parameter behavior, and whether the observed data can distinguish the models.