Why Candidate Covariates Can Compete
Candidate predictors often carry overlapping information. Body weight and body surface area, for example, may both describe body size. If both are related to a PK parameter, a model may fit better when either is added alone than when each is tested after the other. This does not necessarily mean that one covariate is biologically irrelevant. It can mean that the available data distinguish their separate contributions only weakly.
The practical question is not just whether a covariate is associated with an outcome on its own. In a multivariable covariate model, the question is whether it contributes information conditional on the other included terms. Correlation matrices help identify overlapping predictors before model comparisons; conditional model tests quantify the improvement attributable to a term given the rest of the specified model.
For a continuous covariate model on typical clearance, a simple illustration is
Here \(CL_{\mathrm{typ},i}\) is the typical clearance for individual \(i\), \(\theta_0\) is the intercept, and \(z_A,z_B\) are specified covariate values, perhaps centered and scaled. The coefficient \(\beta_A\) describes the change in log typical clearance for a one-unit increase in \(z_A\), at fixed \(z_B\). In a population PK analysis, these terms are estimated within the stated structural and variability model; the equation is an illustration of conditional interpretation, not a replacement for the population likelihood.
Correlation Matrices: A Screening Tool, Not a Verdict
A correlation matrix summarizes pairwise linear association among quantitative candidate covariates. Its entries range from \(-1\) to \(1\): values near either extreme indicate strong linear association, while values near zero indicate little linear association. A small pairwise correlation does not rule out a dependence involving several predictors, and a large correlation does not prove that two terms cannot both be useful.
Worked Example: Screening Correlated Size Covariates
In a hypothetical cohort of 180 individuals, investigators calculate correlations among centered body weight, body surface area (BSA), and age. The resulting matrix is:
| Covariate | Weight | BSA | Age |
|---|---|---|---|
| Weight | 1.00 | 0.94 | 0.18 |
| BSA | 0.94 | 1.00 | 0.22 |
| Age | 0.18 | 0.22 | 1.00 |
The weight–BSA correlation of \(0.94\) is strong. It warns that separate effects of weight and BSA may be difficult to estimate precisely if both are entered together. The correlations involving age are much smaller, so pairwise linear overlap with these size measures appears less pronounced.
In a simple linear regression with two predictors, the variance inflation factor for either predictor is \(1/(1-r^2)\), where \(r\) is their correlation. Using \(r=0.94\),
This value illustrates substantial inflation of coefficient variance relative to uncorrelated predictors in that simple setting. It is not a universal cutoff or a direct test of whether BSA should be excluded. With several predictors, the corresponding VIF is based on predicting one covariate from all the others, not merely on its largest pairwise correlation. The analysis should also consider the intended covariate parameterization and the model’s uncertainty estimates.
For categorical covariates, an ordinary correlation depends on how categories are coded and may be an unhelpful summary. Inspect category counts and cross-tabulations, and assess the model terms using their full joint tests where appropriate. A correlation matrix is most useful as an exploratory diagnostic for continuous candidates, not as an automatic selection algorithm.
Conditional Model Tests
A conditional test compares two nested models fitted to the same observations using the same likelihood and compatible estimation settings. The reduced model contains the terms being retained; the full model adds the covariate term or terms being tested. If the objective function is \(-2\) times the maximized log likelihood, the objective-function difference is
Under the null hypothesis that the added coefficients are zero, and under the usual regularity conditions, \(\Delta\mathrm{OFV}\) is approximately distributed as chi-square with degrees of freedom equal to the number of added parameters. These conditions include nested models, a common dataset and likelihood, adequate convergence to the relevant maxima, and parameters that are not on problematic boundaries. The approximation may be unreliable in boundary or other nonregular settings; numerical noise can also obscure small differences. A model comparison is meaningful only after checking that both fits are trustworthy.
Worked Example: Shared Signal and Conditional Tests
Suppose a hypothetical population PK analysis fits four models to the same data. The base model has neither age nor weight on typical clearance. Each covariate is represented by one coefficient. The OFVs are: base, 1240.0; age only, 1233.0; weight only, 1234.2; and both age and weight, 1231.8.
The age-only comparison against the base model gives \(\Delta\mathrm{OFV}=1240.0-1233.0=7.0\), with \(df=1\). Its chi-square tail probability is approximately \(0.0082\). Thus age improves fit when tested without weight, under the stated likelihood-ratio assumptions.
To test age conditional on weight, compare the weight-only model with the model containing both:
At a prespecified \(\alpha=0.05\), this conditional test is not statistically significant. Conversely, testing weight conditional on age compares the age-only model with the joint model: \(\Delta\mathrm{OFV}=1233.0-1231.8=1.2\), \(df=1\), and \(p\approx0.273\).
The joint model versus the base model gives \(\Delta\mathrm{OFV}=1240.0-1231.8=8.2\), with \(df=2\), for a tail probability of approximately \(0.0166\). Together, the two terms improve fit relative to the base model, even though neither one-term increment is significant at the 0.05 level after the other has entered. The results are consistent with overlapping information: the data support a joint contribution, but provide limited evidence for either predictor’s unique increment in this parameterization. They do not prove that both terms are biologically necessary.
Why an Apparent Effect Can Change
The effect estimate can change when a correlated predictor is added because a one-covariate model may attribute to its included covariate some association that is shared with an omitted covariate. A short derivation makes the mechanism explicit in a standardized linear-model analogy. Let \(Y,A,B\) each have mean zero and variance one, and consider the population regression of \(Y\) on both \(A\) and \(B\). The normal equations are
Solving these equations gives
The marginal slope for \(A\) alone is \(r_{AY}\). The conditional slope for \(A\) depends also on the relationship between \(A\) and \(B\), and on the relationship between \(B\) and \(Y\). When \(r_{AB}\) is near \(1\) or \(-1\), the denominator \(1-r_{AB}^2\) is small, so small changes in the correlations can produce large changes in the separate coefficients. This algebra explains the mechanism in a linear model; in a nonlinear mixed-effects PK model, estimates arise from the likelihood and need not obey these exact correlation formulas.
Worked Example: A Marginal Association Reverses Conditionally
Consider hypothetical standardized summaries with \(r_{AY}=0.60\), \(r_{BY}=0.70\), and \(r_{AB}=0.90\). In the model containing \(A\) alone, the standardized slope is \(0.60\), a positive marginal association.
After including \(B\), the conditional coefficient for \(A\) is
The conditional coefficient for \(B\) is
Thus the positive marginal association for \(A\) becomes a small negative conditional coefficient, while \(B\) has a larger positive conditional coefficient. The example is algebraically coherent: the determinant of the three-variable correlation matrix is \(1+2(0.60)(0.70)(0.90)-0.60^2-0.70^2-0.90^2=0.096>0\). The reversal is not proof that increasing \(A\) would reduce \(Y\). It describes the fitted conditional relationship for these hypothetical summaries, and highlights how shared information can make an isolated association a poor guide to a multivariable effect.
Diagnostics, Alternatives, and Limitations
When candidate predictors are correlated, examine more than a single pairwise correlation or p-value. Useful checks include the following:
- Inspect correlations and scatterplots for continuous candidates, looking for near-linear dependence, restricted ranges, and influential observations.
- Compare the estimates and standard errors for each term in single-covariate and joint models. Large standard errors, unstable signs, or substantial estimate changes are warnings that separate effects are weakly determined.
- Check that full and reduced models use the same observations, parameterization, likelihood, and converged fitting procedure before interpreting an OFV difference.
- Assess whether the functional forms and covariate scales are sensible. A correlation matrix describes linear association in the recorded variables; it does not diagnose every nonlinear relationship or coding problem.
A conditional likelihood-ratio test is the main model-based comparison when the question is incremental fit under a specified likelihood. A correlation matrix is a preliminary descriptive tool. Information criteria such as AIC or BIC compare fit with a complexity penalty rather than testing one nested increment at a chosen \(\alpha\). Biological knowledge and prespecified constraints can guide which parameterization is meaningful, but cannot make weakly distinguishable effects precise. These approaches answer related but different questions; none makes high collinearity disappear.
A common mistake is to say, “Age was significant by itself, so its effect is established, even though its conditional test is not significant.” The correction is that the single-covariate result answers a different question: it assesses age without weight in the model. The conditional test asks whether age adds detectable information after weight is already represented. Likewise, a nonsignificant conditional test does not prove no effect; correlated predictors can leave little information for separating their coefficients. If the scientific goal requires a stable individual effect, report the uncertainty and consider whether the available data and chosen covariates can support that distinction.
Collinearity is therefore a limitation on interpretation, not merely a technical inconvenience. A model may predict adequately while its individual covariate coefficients remain imprecise or sensitive to model specification. In population PK, keep the interpretation tied to the modeled parameter, covariate scale, and conditional set of terms. Do not treat the sign or significance of a coefficient as a standalone biological conclusion.
Check Your Understanding
Answer each question, then compare with the brief solutions.
- A pair of continuous covariates has correlation \(r=-0.92\). What does this suggest, and what does it not establish?
- A reduced model has OFV 810.0 and a nested model adding one coefficient has OFV 806.4. Calculate the test statistic and state the degrees of freedom under regularity conditions.
- Why can an effect estimate change when a correlated covariate is added?
- A covariate has \(p=0.02\) alone and \(p=0.18\) when another predictor is included. Does this demonstrate that the covariate has no effect?
Solutions. (1) The strong negative pairwise correlation warns of overlapping linear information and potentially imprecise separate effects; it does not prove redundancy or establish the conditional relationship among all predictors. (2) \(\Delta\mathrm{OFV}=810.0-806.4=3.6\), with \(df=1\) for one added coefficient; the chi-square approximation requires the stated nesting, common data and likelihood, reliable fits, and regularity conditions. (3) The single-predictor model can assign shared outcome association to the included predictor; conditioning on the additional covariate changes the comparison and the coefficient’s meaning. (4) No. The conditional test is not statistically significant at \(\alpha=0.05\), but that does not prove a zero effect; it indicates insufficient evidence for an incremental contribution under that model and dataset.