Likelihood-Ratio Tests for Covariate Addition

Why Compare Nested Models?

A covariate can be added to a population pharmacokinetic model to represent a possible systematic difference in a parameter, such as clearance. The central statistical question is whether the data support the added flexibility: does the model with the covariate fit sufficiently better than the model without it, given that the larger model has more parameters? A likelihood-ratio test (LRT) formalizes this comparison for nested models.

For example, a base model might describe typical clearance as the same across body weights, while an expanded model allows typical clearance to vary with body weight. The models are nested if the base model is obtained from the expanded model by setting its covariate-effect parameter to zero. The test compares their maximized likelihoods using the same observations and likelihood specification. It does not, by itself, establish that the covariate effect is clinically important or causal.

Definition: Let $M_0$ be a reduced model and $M_1$ an expanded model that contains $M_0$ as a special case. Write $\ell_0$ and $\ell_1$ for their maximized log likelihoods, respectively. The likelihood-ratio statistic is $$ D = 2(\ell_1-\ell_0). $$ Because the expanded model includes the reduced model, its maximized likelihood cannot be lower when both are fitted correctly to the same data. Thus $D$ should be nonnegative, apart from small numerical discrepancies.

In population PK analyses, software often reports an objective function value (OFV) equal to $-2$ times the maximized log likelihood, up to a constant common to the fitted models. If $\mathrm{OFV}_0$ and $\mathrm{OFV}_1$ are the reduced- and expanded-model values, then

D = \mathrm{OFV}_0-\mathrm{OFV}_1.

A decrease in OFV therefore gives a positive LRT statistic. The subtraction order matters: reduced-model OFV minus expanded-model OFV. The two fits must use the same data and comparable likelihood calculations; otherwise, the difference does not represent the intended nested-model comparison.

Conditions for the Chi-Square Reference Distribution

Under the null hypothesis, the added covariate effect is absent and the reduced model is adequate within the expanded model. Under appropriate regularity conditions, the LRT statistic has an approximate chi-square distribution under that null hypothesis. If the expanded model adds $q$ freely estimated parameters, the reference distribution is $\chi^2_q$, where $q$ is the difference in the number of free parameters between models.

Theorem: For properly nested models fitted by maximum likelihood to the same data, if the null parameter value is an interior point of the expanded model's parameter space, the parameters are identifiable, and the likelihood is sufficiently smooth with regular information, then under $H_0$, $D$ is asymptotically distributed as $\chi^2_q$. The value $q$ is the number of independent parameter restrictions imposed by $M_0$ relative to $M_1$.

“Asymptotically” matters: the chi-square result is a large-sample approximation, not an exact finite-sample identity. In a population PK model, the relevant information comes from the observed concentration data and the structure of the likelihood, not simply from counting concentration records. Sparse sampling, few individuals in a covariate subgroup, weak covariate variation, or parameters that are difficult to distinguish can make the approximation unreliable.

Other conditions also deserve attention. The added effect must be represented by a regular, identifiable parameterization; the likelihood must be correctly specified for the comparison; and both fits must reach the relevant maximum rather than different local or failed solutions. A null value on the boundary of the parameter space—for example, a variance constrained to be nonnegative and tested at zero—does not generally have the ordinary $\chi^2_q$ reference distribution. Covariate-effect tests are often regular when the added coefficient can be positive or negative and zero is an interior value, but this should be checked for the specific model.

Why the Statistic Has This Form

The LRT begins with the ratio of the largest likelihood allowed under the null model to the largest likelihood allowed under the expanded model. Since the expanded model includes all null-model possibilities, this ratio is no greater than one. Taking the negative twice the logarithm changes the ratio into a nonnegative measure of the loss in fit imposed by the null restrictions.

\Lambda = \frac{L_0}{L_1} \qquad\Longrightarrow\qquad -2\log(\Lambda) = -2\{\ell_0-\ell_1\} = 2(\ell_1-\ell_0) = D.

The chi-square approximation follows from the large-sample behavior of the likelihood near its maximum. In brief, a smooth log likelihood can be approximated locally by a quadratic function of the parameter deviations. Estimating the $q$ additional directions improves the fit by a quantity that, under the null and regularity conditions, behaves like the sum of squares of $q$ approximately independent standard normal variables. That sum has a $\chi^2_q$ distribution. This justification explains both the degrees of freedom and why boundary constraints, weak identification, or poor likelihood approximation can invalidate the usual reference distribution.

Worked Example: Adding a Body-Weight Effect on Clearance

Suppose a hypothetical population PK analysis compares two models using the same concentration records and the same residual-error model. The reduced model has an OFV of 1240.60. The expanded model adds one body-weight coefficient on typical clearance and has an OFV of 1235.58. Assume both fits converged to the relevant maxima, and the coefficient is identifiable with zero in the interior of its allowed range.

Calculate the statistic. The expanded model adds one parameter, so $q=1$. The OFV difference is $$ D = \mathrm{OFV}_0-\mathrm{OFV}_1 =1240.60-1235.58 =5.02. $$

Under $H_0$, use a $\chi^2_1$ reference distribution. Its upper-tail probability at 5.02 is approximately 0.02506, or 0.0251 to four decimal places. At a prespecified $\alpha=0.05$, $p<\alpha$, so the test rejects the null hypothesis that the added body-weight coefficient is zero, under the stated model and approximation conditions.

This result is evidence that the covariate improves fit beyond what would typically be expected from adding one parameter when the null is true. It does not show the direction or magnitude of the body-weight effect; those require examining the estimated coefficient and its uncertainty. Nor does statistical evidence alone establish clinical importance.

Interpreting Tests with More Than One Added Parameter

An addition can require multiple parameters. For instance, a categorical covariate with three levels may require two independent contrasts relative to a reference level. If both contrasts are added together, the joint null hypothesis sets both coefficients to zero, and the LRT compares the models using the corresponding multi-parameter reference distribution. The test is a joint assessment; it is not equivalent to conducting separate tests for each coefficient.

Worked Example: A Three-Level Genotype Covariate

In a hypothetical analysis, a reduced model has an OFV of 811.40. An expanded model adds two genotype contrasts to describe typical clearance across three genotype groups and has an OFV of 804.02. Assume the two models are nested, use the same observations and likelihood, and both fit successfully. Further assume the contrast parameters are identifiable and the null values are interior.

The expanded model adds $q=2$ parameters. Therefore, $$ D=811.40-804.02=7.38. $$ For a $\chi^2_2$ random variable, the upper-tail probability is $P(\chi^2_2\geq 7.38)=\exp(-7.38/2)=\exp(-3.69)\approx 0.0250$. At $\alpha=0.05$, this joint test rejects the null hypothesis that both genotype contrasts are zero.

The conclusion concerns the genotype terms as a set: at least one contrast contributes to the improved fit, subject to the model and test assumptions. This test alone does not identify which group differs, quantify a clinically meaningful contrast, or establish a biological mechanism.

Checking the Calculation and the Model Fits

A test statistic and p-value are meaningful only if the model comparison is valid. Before interpreting the result, verify that the expanded model truly contains the reduced model, that both use the same data and likelihood conventions, and that the difference in free parameters matches the restrictions being tested. A change in the dataset, residual model, or other structural assumptions can make an OFV difference unsuitable as a test of covariate addition.

Then inspect the fitting evidence. Confirm that both minimizations converged, parameter estimates are finite and plausible, and the expanded fit is not an inferior local solution. Review gradients, covariance or uncertainty diagnostics where available, and whether the added coefficient is estimable rather than effectively unbounded or confounded with another model component. A negative OFV difference larger than numerical tolerance is a warning: it conflicts with the nesting argument and often indicates a failed optimization, inconsistent likelihood calculation, or models that are not actually nested.

Worked Example: A Small Fit Improvement That Is Not Strong Evidence

Consider a hypothetical comparison for a continuous renal-function covariate. The reduced model has an OFV of 960.00, and the expanded model adds one covariate coefficient and has an OFV of 958.70. Both models use the same records and error structure, and assume the regularity and convergence checks are satisfactory.

The statistic is $$ D=960.00-958.70=1.30. $$ There is one added parameter, so the reference distribution is $\chi^2_1$. The upper-tail probability is approximately 0.254. Since $0.254>0.05$, this comparison does not provide sufficient evidence at the 5% level to reject the zero-effect null hypothesis. It does not prove the covariate has no effect; it says that this test, with these data and model assumptions, did not find a statistically significant improvement.

The estimated coefficient may still be scientifically relevant, especially if its uncertainty is large. The LRT is a test of fit improvement under a specified null, not a replacement for examining effect size, uncertainty, predictive consequences, and scientific plausibility.

How the LRT Differs from Other Tests

The LRT, Wald test, and score test address related hypotheses but use different information. The LRT compares the maximized fits of the reduced and expanded models. A Wald test uses the estimated coefficient and its estimated standard error, assessing how far the estimate lies from its null value. A score test evaluates the likelihood slope at the null fit, without requiring full maximization of the expanded model. Under regular large-sample conditions, these tests often give similar results, but in finite samples or with weakly identified parameters their results can differ.

The LRT is especially direct when both nested models can be fit reliably and their likelihoods are comparable. It can be less trustworthy when optimization is unstable or the chi-square conditions fail. Conversely, a Wald result can be sensitive to imprecise standard errors or parameterization. No test removes the need to diagnose the model or assess whether the proposed covariate relationship is scientifically reasonable.

Common Mistakes and Limitations

Reversing the OFV subtraction. The correct difference is reduced-model OFV minus expanded-model OFV. Subtracting in the opposite order can produce a negative number and does not represent the LRT statistic.
Using the wrong reference degrees of freedom. The reference distribution depends on the number of independent restrictions, not automatically on the number of covariate names. For a specific comparison, count the added free parameters represented by the null restrictions.
Interpreting a p-value as the probability that the null is true. A p-value is the probability, under the null model and the reference-distribution assumptions, of observing a statistic at least as large as the one calculated.
Treating a significant test as proof of clinical importance. A small p-value indicates evidence against the specified null, not the size, practical importance, or causal meaning of the effect.
Applying the ordinary chi-square approximation at a boundary without justification. If a tested value lies on a parameter-space boundary, the usual reference distribution may not apply; a suitable alternative calibration may be needed.

Finally, an LRT is conditional on the models being compared. A non-significant result can reflect a genuinely small effect, limited information, or model misspecification; a significant result can reflect a detectable but clinically trivial change. Covariate selection from repeated searches also raises issues beyond a single prespecified comparison. The statistic should therefore be reported with the models, parameter restrictions, reference degrees of freedom, p-value, and relevant estimation diagnostics.

Key takeaway: For nested population PK models fitted by maximum likelihood to the same data, calculate $D=2(\ell_1-\ell_0)$, or equivalently subtract the expanded-model OFV from the reduced-model OFV. Compare $D$ with a chi-square distribution only when the nesting, estimation, identifiability, interior-point, and large-sample conditions make that approximation reasonable.

Check Your Understanding

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

A reduced model has OFV 500.0 and an expanded nested model has OFV 496.4. What is the LRT statistic?
If a covariate addition contributes two independent parameters, what chi-square reference distribution is used under the regularity conditions?
In words, what does a p-value of 0.03 mean for a valid LRT?
Why might a negative OFV difference indicate that a comparison should not be interpreted as a valid LRT?
Does rejecting the null hypothesis establish that a covariate has a clinically important or causal effect? Explain briefly.

Solutions. (1) $D=500.0-496.4=3.6$. (2) Use $\chi^2_2$, because two independent parameters are added. (3) If the null hypothesis and reference-distribution assumptions hold, the probability of a statistic at least as large as the observed one is 0.03. (4) For correctly optimized nested models using the same likelihood and data, the expanded model cannot have a worse maximized likelihood, so its OFV should not be higher; a negative difference can signal failed optimization, inconsistent calculations, or non-nested models. (5) No. Rejection is evidence against the specified zero-effect null under the model; effect magnitude, clinical relevance, and causal interpretation require separate assessment.