Why Structural Model Choice Matters for Covariate Selection
A covariate test is conditional on the population PK model used to describe concentration-time data. For an oral drug, a misspecified absorption model may leave systematic residual patterns that a covariate effect partly absorbs. An overly simple compartment model may likewise attribute distribution differences to a covariate on clearance or volume. A candidate can therefore appear strongly supported under one structural model and weakly supported under another, even when the same individuals and covariate values are analyzed.
Sensitivity analysis asks a focused question: does the candidate covariate effect remain supported across plausible structural descriptions of the PK data? It does not establish that every tested structure is true, nor does it replace careful structural model development. Instead, it reveals whether the covariate conclusion depends on assumptions about absorption, distribution compartments, or elimination.
A Matched Test Within Each Structural Model
Let \(S_j\) denote one defensible structural model, such as a one-compartment model with first-order absorption. Within each \(S_j\), define a reduced model without the candidate covariate effect and a full model with that effect. Keep the dataset, likelihood, covariate coding and functional form, and other analysis choices matched within the pair. The reduced model must be nested in the full model for the usual likelihood-ratio test.
For a candidate continuous covariate \(z_i\) on clearance, one possible specification is \(CL_i=CL_{\mathrm{pop}}(z_i/z_{\mathrm{ref}})^\beta\exp(\eta_{CL,i})\). Here \(CL_i\) is individual clearance, \(CL_{\mathrm{pop}}\) is typical clearance at the reference value, \(\beta\) is the covariate coefficient, and \(\eta_{CL,i}\) is the individual random effect. In the reduced model, \(\beta=0\); the full model estimates \(\beta\). The same covariate definition and reference value should be used across structures.
If the models are fitted by maximum likelihood, the likelihood-ratio statistic is the difference in objective function values (OFV), with the usual convention that a smaller OFV indicates a higher maximized likelihood. For one freely estimated covariate coefficient, the nominal degrees of freedom are one. More generally, the degrees of freedom equal the number of additional parameters when the nested-model regularity conditions hold. Those conditions include a correctly specified likelihood, identifiable parameters, an interior null value, and adequate maximum-likelihood optimization. Boundary parameters or nonstandard testing conditions may require a different reference distribution.
Under the null hypothesis that the added covariate coefficients are zero, and under the stated regularity conditions, \(\Delta OFV_j\) is approximately distributed as a chi-square random variable with the corresponding degrees of freedom. A large positive value is evidence against the reduced model within that structural model. This test does not compare the structural models with one another.
Worked Examples Across Structural Alternatives
Worked Example: Renal Function on Clearance Under Three Structures
Suppose a hypothetical oral-drug analysis tests a prespecified eGFR effect on clearance. The same patients, concentration data, likelihood, and power covariate form are used in each paired test. The candidate coefficient is \(\beta\) in \(CL_i=CL_{\mathrm{pop}}(eGFR_i/90)^\beta\exp(\eta_{CL,i})\). Consider one-compartment first-order absorption with linear elimination, a two-compartment model with first-order absorption and linear elimination, and a one-compartment model with transit absorption and linear elimination. Assume all six fits converged, parameter estimates were identifiable, and each full model differs from its reduced counterpart only by estimating \(\beta\).
| Structural model | Reduced OFV | Full OFV | Estimated \(\beta\) |
|---|---|---|---|
| One compartment, first-order absorption | 1200.0 | 1191.8 | 0.42 |
| Two compartments, first-order absorption | 1178.0 | 1174.9 | 0.31 |
| One compartment, transit absorption | 1192.0 | 1186.0 | 0.38 |
For the first structure, \(\Delta OFV=1200.0-1191.8=8.2\). With one added coefficient, this is compared with a chi-square distribution with \(df=1\), giving \(p\approx0.0042\). For the two-compartment structure, \(\Delta OFV=1178.0-1174.9=3.1\), giving \(p\approx0.078\). For transit absorption, \(\Delta OFV=1192.0-1186.0=6.0\), giving \(p\approx0.014\).
At a prespecified \(\alpha=0.05\), the eGFR effect meets the nominal test threshold in the first and third structures, but not in the second. The estimated coefficient remains positive and of similar magnitude, from 0.31 to 0.42, across the models. This is evidence that the estimated direction and magnitude are relatively stable, while the test's strength is sensitive to the compartment assumption. It is not evidence that the covariate effect is absent in the two-compartment model.
Worked Example: An Absorption-Dependent Covariate Signal
In a second hypothetical oral-drug analysis, age is tested on the absorption rate parameter \(k_a\), using the same log-linear covariate form in each structure. A one-compartment first-order absorption model gives reduced and full OFVs of 845.0 and 837.1, respectively, with an estimated age coefficient of \(-0.55\). A two-compartment model with transit absorption gives corresponding OFVs of 812.0 and 811.4, with an estimated coefficient of \(-0.08\). Assume the fits converge and the added age coefficient is the only difference within each pair.
For first-order absorption, \(\Delta OFV=845.0-837.1=7.9\), which has \(df=1\) and \(p\approx0.0049\). For transit absorption, \(\Delta OFV=812.0-811.4=0.6\), giving \(p\approx0.44\). The age effect appears supported under the first absorption model but is small and weakly supported under the transit model. Both estimates are negative, but their magnitudes differ substantially.
The appropriate conclusion is that the evidence for an age effect on absorption is structurally sensitive in this analysis. Before attributing the difference to biology, inspect whether the absorption structures fit the observed concentration-time patterns, whether \(k_a\) or transit parameters are identifiable, and whether the age effect is compensating for absorption misspecification. The two p-values do not by themselves identify which absorption model is correct.
Worked Example: Why Cross-Structure OFV Differences Are Not Covariate Tests
Consider hypothetical baseline structures fitted to the same dataset: a one-compartment model has OFV 1200.0 and a two-compartment model has OFV 1178.0. It is tempting to call the difference \(1200.0-1178.0=22.0\) a covariate statistic or to use it as the test for adding a covariate. That subtraction does not compare a reduced and full covariate model within a matched structure, so it is not the required covariate likelihood-ratio test.
Suppose the two-compartment model has a reduced OFV of 1178.0 and a covariate full-model OFV of 1174.9. The valid within-structure statistic is \(\Delta OFV=1178.0-1174.9=3.1\), with \(df=1\) if the added effect is one regular parameter. Its nominal p-value is about 0.078. The baseline OFV difference of 22.0 instead reflects a comparison of distinct structural specifications; it cannot substitute for the paired covariate contrast.
If structures are non-nested, their likelihood difference does not have the ordinary nested-model chi-square reference distribution. Structural comparison requires its own justification and criteria. The covariate sensitivity question is answered by the matched tests, alongside checks of structural adequacy and covariate estimates.
How to Read Persistence and Sensitivity
A useful summary records more than whether each p-value falls on one side of \(\alpha\). For every structure, report the covariate estimate, its uncertainty measure, the reduced and full OFVs, the test degrees of freedom, and the test result. Then assess three features together:
- Direction: does the estimated effect keep the same sign across structures?
- Magnitude: are the estimates similar enough to support a consistent interpretation, given their uncertainty?
- Statistical support: does the within-structure test show similar evidence, or does it change materially?
An effect that keeps its direction and approximate magnitude but has different p-values is not identical to an effect whose direction reverses or whose magnitude changes sharply. The latter pattern suggests stronger dependence on structural assumptions and warrants closer examination of identifiability, residual patterns, and parameter correlations. A coefficient estimate can also be unstable while p-values happen to fall on the same side of a threshold. Thus, threshold persistence alone is a limited summary.
Use the same dataset and compatible likelihood definitions wherever possible. If records, likelihood contributions, estimation methods, or covariate coding differ, changes in test results cannot be attributed cleanly to structural choice. Fit each reduced-full pair separately within every structure, confirm that the optimizer found comparable valid solutions, and check that the intended covariate-free model is genuinely nested in the full model. A failed or non-identifiable fit is not evidence for or against the covariate.
Diagnostics, Alternatives, and Limitations
Structural sensitivity is informative only if the candidate structures are defensible. Examine goodness-of-fit plots and residual behavior for systematic time patterns, particularly around absorption and distribution phases. Review parameter precision, correlations, and boundary estimates. Check whether the covariate coefficient is strongly correlated with structural parameters such as \(k_a\), intercompartmental clearance, or elimination parameters. If a structure cannot adequately describe the data or cannot identify its parameters, its covariate test should not be treated as an equal-quality vote.
For linear versus saturable elimination, for example, confirm that the observed dose range and concentration data contain information to distinguish the alternatives. A change in a clearance covariate estimate may reflect differences in parameter meaning or identifiability rather than a genuine change in the biological relationship. Covariate parameterization should therefore target a comparable PK quantity, and its interpretation should be stated in the context of each structure.
| Approach | Question answered | Key caution |
|---|---|---|
| Matched covariate tests across structural models | Does covariate evidence persist across specified plausible structures? | Each test is conditional on its structural model. |
| Select one structural model, then test covariates | What is the covariate evidence conditional on the chosen structure? | May conceal uncertainty about structure. |
| Model averaging after selection | How can estimates or predictions account for uncertainty across selected models? | Does not remove the need to define and fit plausible models carefully. |
The earlier material on model averaging after covariate selection addresses a different step: combining information across selected models. Here, the primary task is to diagnose whether a covariate finding changes across a planned set of structural alternatives. Sensitivity analysis does not automatically select the best structure, correct structural misspecification, or prove a causal covariate effect. Results also depend on which alternatives are considered; an unexamined plausible structure can still alter the conclusion.
Common Pitfall and Practical Workflow
Identify defensible choices for absorption, compartment structure, or elimination before examining the candidate covariate results.
For every structure, fit a covariate-free reduced model and a full model that adds only the prespecified covariate effect, using the same data and covariate form.
Verify convergence, parameter identifiability, likelihood comparability within each pair, and adequacy of the structural model.
Compare effect direction and magnitude, uncertainty, and within-structure test evidence. State explicitly when the conclusion depends on structural assumptions.
Check Your Understanding
Answer each question, then compare with the brief solutions below.
- For one structural model, the reduced and full OFVs are 560.0 and 555.5. The full model adds one covariate coefficient. Calculate the test statistic and state whether it exceeds the approximate 0.05 critical value for one degree of freedom.
- Why is subtracting the OFV of a two-compartment model from the OFV of a one-compartment model not ordinarily a covariate likelihood-ratio test?
- A covariate coefficient is positive under three structures, with similar estimates, but its p-value is below 0.05 in only two. What is a suitable interpretation?
- Name two diagnostics that matter before interpreting a covariate test under an alternative absorption model.
Solutions. (1) \(\Delta OFV=560.0-555.5=4.5\). For \(df=1\), the approximate 0.05 critical value is 3.84, so 4.5 exceeds it. (2) The structures are not the matched reduced and full covariate models; if they are non-nested, the difference does not have the ordinary nested chi-square reference distribution. (3) The effect's estimated direction and magnitude appear relatively stable, while its statistical support is not identical across structures. Report the estimates and uncertainty and describe the evidence as structurally sensitive rather than declaring the effect absent in one model. (4) Examples include checking residual patterns over time, convergence, precision and identifiability of absorption parameters, and correlations between the absorption parameter and the covariate coefficient.