Why Degrees of Freedom Depend on the Model Comparison
A likelihood-ratio test (LRT) for covariate addition compares a reduced model with a nested full model. The test statistic measures how much the maximized likelihood improves when the covariate effect is allowed. To interpret that improvement against a chi-square distribution, the test also needs the right degrees of freedom. The key is not simply to count covariate names: it is to count the independent parameter restrictions imposed by the reduced model.
A single continuous covariate may contribute one coefficient, while a categorical factor with several levels usually contributes several. A covariate may also affect more than one PK parameter, with one coefficient per affected parameter or more, depending on the specified model. These cases can be handled with the same principle: compare the number of freely estimated parameters in the two models, provided the comparison satisfies the regularity conditions for the usual LRT reference distribution.
If \(\ell_{\text{full}}\) and \(\ell_{\text{reduced}}\) are the maximized log-likelihoods, the statistic is \(2(\ell_{\text{full}}-\ell_{\text{reduced}})\), equivalently the reduced-model objective function minus the full-model objective function when the objective function is \(-2\ell\). Under the null hypothesis and regularity conditions, this statistic is approximately distributed as \(\chi^2_q\). The earlier material on nested hypotheses and likelihood-ratio tests establishes the comparison; here the task is to determine \(q\) correctly.
Why the Parameter-Count Difference Gives the Reference Degrees of Freedom
The connection between restrictions and degrees of freedom comes from the local behavior of the likelihood. Suppose the full model has \(q\) additional independent coefficients, collected in a vector \(\boldsymbol{\beta}\), and the null hypothesis sets \(\boldsymbol{\beta}=\boldsymbol{0}\). Near the true parameter values, a regular log-likelihood is locally approximated by a quadratic function. After accounting for the nuisance parameters that are estimated in both models, the likelihood improvement from estimating \(\boldsymbol{\beta}\) has the form of a quadratic function of \(q\) approximately standard-normal quantities.
where the \(Z_j\) are asymptotically independent standard normal variables under the null. A sum of squares of \(q\) independent standard normal variables has a chi-square distribution with \(q\) degrees of freedom. Thus the relevant count is the dimension of the additional, independently estimable effect—not the number of observations, covariate values, or labels used to describe the effect.
This justification assumes the models are genuinely nested, fitted to the same observations using comparable likelihoods, and regular at the null. In particular, the added parameters must be identifiable, the null values must not lie on problematic parameter boundaries, and nuisance parameters must remain identifiable when the covariate effect is zero. The chi-square result is asymptotic; with sparse data or weak identification, its calibration can be poor even when the parameter count is clear.
Counting Coefficients in Common Covariate Tests
For one continuous covariate entering one PK parameter through one coefficient, the usual count is one. For example, a model may describe clearance as
Here \(x_i\) is the covariate for individual \(i\), and \(\beta\) is its fixed-effect coefficient. Testing no covariate effect imposes the single restriction \(\beta=0\), so \(q=1\). The fact that the effect is expressed on the log scale, and therefore corresponds to a multiplicative effect on clearance, does not change the number of coefficients.
A categorical factor requires a different count. With \(L\) observed levels, a standard reference-level parameterization uses \(L-1\) coefficients. The reference level is represented by the intercept, and each other level has a contrast against that reference. Testing the factor as a whole imposes \(L-1\) restrictions, so \(q=L-1\), if all those contrasts are independently estimable. The count is not \(L\): adding an intercept and a separate unrestricted coefficient for every level would be a redundant parameterization unless an identifying constraint were imposed.
For spline effects, count the coefficients that are added in the tested comparison. If the reduced model has a linear covariate effect and the full model adds two independent nonlinear basis terms, testing those nonlinear additions has \(q=2\). The number of knots alone does not determine the test degrees of freedom. Knot placement and basis construction determine the number of basis columns, while constraints or rank deficiency can reduce the number that is independently estimable.
Worked Example: One Continuous Covariate on Clearance
In a hypothetical population PK analysis, the reduced model has 12 freely estimated parameters. The full model adds one coefficient for body weight on clearance, with the effect written as \(\log(CL_i)=\theta_{CL}+\beta x_i+\eta_{CL,i}\). Assume both models are identifiable, use the same observations and likelihood, and the null value \(\beta=0\) is an interior value.
The full model has 13 parameters, so
Suppose the reduced and full objective function values are 842.6 and 838.8. The LRT statistic is \(842.6-838.8=3.8\), and its reference distribution under the null is \(\chi^2_1\). The upper-tail probability is approximately 0.0513. The degrees of freedom are one because the comparison adds one coefficient, not because the covariate is continuous or because clearance is the only outcome being recorded.
Worked Example: A Five-Level Categorical Factor
Consider a hypothetical factor describing one of five enzyme-activity groups, included as a covariate on clearance. Use one group as the reference and estimate a separate contrast for each of the other four groups. Suppose the full model improves the objective function by 9.50 units relative to the reduced model.
There are four independently estimated contrasts, so the null that the factor has no effect sets all four to zero. Therefore \(q=5-1=4\). The LRT statistic is 9.50, and its upper-tail probability is
The approximate p-value is 0.0497. The four degrees of freedom reflect the four restrictions tested jointly. Testing one selected contrast instead would be a different, one-degree-of-freedom hypothesis; it would not test whether the factor as a whole contributes to the model.
Worked Example: Testing Nonlinear Spline Terms
In a hypothetical model, the reduced model includes a linear effect of renal function on clearance. The full model retains that linear term and adds two independent spline basis coefficients to allow curvature. Suppose the objective function falls from 615.4 in the reduced model to 608.2 in the full model.
The linear coefficient is present in both models and is not an added restriction in this comparison. The two added spline coefficients are tested jointly as zero, giving \(q=2\). The LRT statistic is \(615.4-608.2=7.2\). For two degrees of freedom, the upper-tail probability is \(e^{-7.2/2}=e^{-3.6}\approx 0.02732\). This result concerns whether the two nonlinear terms improve fit beyond the linear effect. It is not a test of whether renal function has any effect at all, because the linear term remains in both models.
Joint Effects Across Multiple PK Parameters
A single covariate may be allowed to affect several PK parameters. Count the independently estimated coefficients added for the covariate across all those parameters. For instance, if age has one coefficient on clearance and a separate coefficient on central volume, testing both age effects jointly imposes two restrictions and ordinarily gives \(q=2\). The fact that both coefficients use the same measured covariate does not make them one parameter.
Conversely, a model may impose a shared effect. If a single coefficient is constrained to govern the covariate effect on both clearance and volume, adding that shared coefficient contributes one degree of freedom, not two. The model equations and constraints—not the number of PK parameters named in the scientific description—determine the count. If an effect on three parameters is represented by three coefficients but one is constrained to equal another, only two independent coefficients are added.
Worked Example: A Covariate Tested Jointly on Three PK Parameters
In a hypothetical model, a covariate is proposed to affect clearance, central volume, and absorption rate. The full model adds three distinct coefficients, \(\beta_{CL}\), \(\beta_V\), and \(\beta_{KA}\), and the reduced model sets all three to zero. Assume the three effects are identifiable and none is constrained to be a function of the others. The parameter-count difference is
If the reduced and full objective function values are 1,024.0 and 1,016.0, then the LRT statistic is \(1,024.0-1,016.0=8.0\), compared with \(\chi^2_3\). Its upper-tail probability is approximately 0.0460. This is a joint test: it assesses the null that all three coefficients are zero. It does not identify which parameter-specific effect is responsible for the improvement.
A Practical Counting Procedure and Diagnostic Checks
A reliable count begins with the null hypothesis as a set of explicit restrictions. Do not infer degrees of freedom from the label “one covariate test” or from the change in objective function. Write down which coefficients are removed, constrained, or added, then check whether those restrictions are independent in the fitted design.
State which covariate effects are in the full model and what the reduced model sets to zero or constrains. Confirm that the full model contains the reduced model as a special case.
Count coefficients by PK parameter and effect term. For factors, use the independent contrasts; for spline comparisons, count the additional basis coefficients, not the knots.
Look for empty factor levels, redundant contrasts, collinear covariates, or constraints that tie coefficients together. Such features can make the nominal coefficient count differ from the effective number of independent restrictions.
When regularity conditions hold, compare the LRT statistic with \(\chi^2_q\). Interpret that test as the particular joint null stated in the first step.
A practical audit is to compare the model specifications and fitted parameter reports. The reported number of freely estimated parameters should change by the expected amount. If it does not, investigate fixed parameters, aliases, dropped columns, or constraints before interpreting a p-value. A nominal four-contrast factor test is not a four-degree-of-freedom test if the data contain no observations in one level or if two contrasts are indistinguishable in the fitted design.
Common Pitfalls, Alternatives, and Limits
A frequent error is to count the number of covariates rather than the number of tested coefficients. Saying “this is one covariate, so the test has one degree of freedom” is wrong for a multi-level factor or for a covariate whose effects on multiple PK parameters are separately estimated. The correction is to express the null as restrictions on coefficients and count the independent restrictions.
Another error is to use the total number of parameters in the full model as the degrees of freedom. Parameters shared by both models are nuisance parameters estimated under each model; they do not contribute to the difference. Likewise, when comparing a linear effect with a linear-plus-spline model, only the added nonlinear coefficients determine the degrees of freedom for that comparison.
Wald and score tests are alternative large-sample tests of specified parameter restrictions. They do not change how many restrictions define the scientific null, although their statistics and finite-sample behavior differ from the LRT. The parameter-count rule described here is specifically the usual chi-square calibration for a regular LRT; it does not itself determine a decision threshold for a model search.
The rule also has limits. If a parameter is on a boundary under the null, if a nuisance parameter is undefined when the covariate effect is absent, or if parameters are weakly identified, the ordinary chi-square reference distribution may not apply. Sparse subgroups, severe covariate imbalance, or nearly redundant spline columns can also make the effective test unstable. In such cases, a nominal degrees-of-freedom count is not sufficient evidence that the asymptotic p-value is reliable; the model structure and data support need scrutiny.
Check Your Understanding
For each item, determine the degrees of freedom and identify the restrictions being tested.
- A binary covariate adds one coefficient on clearance. What are the LRT degrees of freedom?
- A six-level factor is represented by reference coding, with all levels represented in the data. The factor is added to one PK parameter. What are the degrees of freedom?
- A linear covariate effect is already in the reduced model; the full model adds three independent spline coefficients. What is the degrees-of-freedom count for testing the added nonlinear terms?
- A covariate has separate coefficients on clearance and volume, and both are tested jointly. What is the count? How would it change if one shared coefficient were constrained to govern both effects?
Solutions: (1) One, because the null imposes the single restriction that the clearance coefficient is zero. (2) Five, because reference coding for six observed levels gives \(6-1=5\) independent contrasts, all tested jointly. (3) Three, because the linear coefficient is common to both models and only the three added spline coefficients are restricted. (4) Two for separate coefficients, since both must be zero under the joint null; one for a shared coefficient, since only one free parameter is added and tested.