Impact of covariate model building methods on their clinical relevance evaluation in population pharmacokinetic analyses: comparison of the full model, stepwise covariate model (SCM) and SCM+ approaches - further results based on more conventional practices.

Abstract

Covariate clinical relevance (CCR) is commonly assessed in population pharmacokinetics using forest plots visualizing parameter changes across covariate values. In our previous work (Philipp et al. 2024), CCR was evaluated using a [0.80-1.20] reference area and a 90% confidence interval for both relevance and significance assessment. However, more conventional thresholds include a broader reference area of [0.80-1.25] and the use of a 5% type I error to assess statistical significance. This commentary extends our previous analysis by evaluating CCR decisions under these more conventional thresholds, in order to assess whether the full model, the stepwise covariate modeling (SCM) and its enhanced version SCM+ remain robust. A comparison with the previous results is also provided. The revised CCR evaluation gave satisfactory results across all three approaches. For covariates with a simulated effect, the full model and SCM/SCM+ provided consistent conclusions with those of the true model. For covariates without a simulated effect, the full model mainly found them non-relevant (NR) non-significant or insufficient information (II) non-significant, while SCM/SCM+ mainly did not select them. These results align with our previous findings. Conclusions for covariates with a simulated effect were almost unchanged. For covariates without a simulated effect, the more conventional threshold allowed the full model to conclude more frequently to their NR instead of II, likely due to the broader reference area and stricter type I error control. Overall, the consistency of our results across different thresholds demonstrates their robustness and supports their generalizability.