The Covariate Search Stepwise (stepwise covariate search) run mode is only available for Population models. Covariates and covariate effects must be specified. This run mode performs an automatically parallelized stepwise forward or backward addition or deletion of covariates effects by adding one at a time to determine if they make a sufficient threshold improvement based on the specified criterion options. (Note that, during the first step, the baseline model is combined with the cases for the first addition step in order to avoid running the base model alone as the first step.)
With Covariate search stepwise selected from the Run Mode pulldown, select the OPTIONS tab.
Refer to the Simple run mode “Estimation” and “Advanced Options” sections for descriptions of most items available on the Covariate Search Stepwise OPTIONS tab.
Select the information Criterion on which to base the stepwise approach from the pulldown.
-2LL: If selected, enter p-values in the fields, such as 0.01 and 0.001, to use as the P-Value to Add a Covariate and P-Value to Remove a Covariate.
These p-values are used, in conjunction with the degrees of freedom for the particular effect, to determine the thresholds using the inverse cumulative distribution function of the chi-square distribution. The degrees of freedom is the number of fixed effects active under the actual current selection of covariate effects. Normally, this is the base set of fixed effects, plus one for each enabled covariate effect. However, if the covariate effect is for a categorical covariate having N categories (N > 1) the number of fixed effects (and thus degrees of freedom) for that covariate effect is N-1.
AIC: If selected, enter threshold values in the Threshold to Add a Covariate and Threshold to Remove a Covariate fields that serve as the threshold at which to add or remove a covariate effect, respectively.
BIC: If selected, enter threshold values in the Threshold to Add a Covariate and Threshold to Remove a Covariate fields that serve as the threshold at which to add or remove a covariate effect, respectively.
The stepwise covariate search method used is the forward addition, backwards elimination where the structural model is used as a baseline and the covariate model is made increasingly complex. After each model estimation, the covariates are evaluated to see which one has the greatest improvement in the goodness-of-fit statistic selected (-2LL, AIC, or BIC) greater than the user-specified threshold. That covariate is added to the regression model for the structural parameter and the model is estimated. This process is repeated until all significant effects are accounted for. Then the process works in reverse to eliminate covariates on parameters whose removal produces the smallest reduction in goodness-of-fit less than the user-specified threshold. This process can find a good set of covariate effects more quickly than the shotgun mode.
How to force some covariates to be parts of the model while others are evaluated via the stepwise covariate search
In some Population PK modeling studies, it may be necessary to “force” some covariates to be part of the base model (e.g., Cl=tvCl*(Wt/70)^0.75). However, when it comes to stepwise searching, the rest of the covariates (e.g., age, sex, etc.) need to be tested. This can be accomplished through the text model.
In the text editor, look for lines such as:
fixef(dVdwgt(enable=c(0))=c(, 0.949657, ))
fixef(dKedGender1(enable=c(1))=c(, 0.0586368, ))
Remove them as desired by deleting or commenting out the enable clause as shown below:
fixef(dVdwgt/*(enable=c(0))*/=c(, 0.949657, ))
fixef(dKedGender1(enable=c(1))=c(, 0.0586368, ))
The covariate search will ignore the covariate that was commented out (or deleted).
Another way to comment out the enable clause is as follows:
/* (enable=c(0)) /**/
The /*…*/ comments are non-nesting, so both opening and closing character sets would have to be removed in order to comment the clause back in. Using the /*…/**/ characters makes it easier to comment the clause back in:
(enable=c(0)) /**/