Re: Stepwise covariate modeling

Hi Sumeet, If you have rich sampling (and rich information on all parameters of interest) then one would not expect much difference between the individual parameter estimates with/without covariates in the model. This does not make the covariate model meaningless, since future patients may be sparsely sampled, or the model may be used to identify subpopulations, or for predictions of future patients, etc. When you say that pop predictions do not change, exactly what do you mean by that? The population typical value is not expected to change much (it may for categorical covariates with high impact) - the interpretation of the population parameter value has shifted from e.g. "median parameter value in population" (base model) to "median parameter value for a subject with typical covariate values". This is because the covariate equations are generally centered around typical covariate values: We do not want the population parameter to represent CL for a subject with zero kilo body weight - had it been coded that way the population parameters would have changed dramatically. So the question is rather if the typical parameter values for two subjects with different covariate values are different to a degree that it is important to account for (i.e. clinically relevant). If we assume that you have body weight on CL, you can calculate e.g. the 2.5th and 97.5th percentiles of body weight in your population (or in another population or relevance). And then you can calculate TVCL for these two different weights and compare to the typical body weight (e.g. 70 kg). You may have for example this equation*: TVCL=THETA(1)*(WT/70)**THETA(2) Based on the point estimate and SE of THETA 2, you can then calculate percent change (from the typical 70 kg body weight) with point estimate and 95% CI, for each of the two extreme body weights. And you can illustrate this in a so-called Forest plot (or tornado plot), for all covariate coefficients. If the CI is wide, the data does not contain enough information to rule out clinical relevance (if you think the parameter in question is important - maybe abs rate is in some cases not, for examples). But given that it has been selected by SCM, if the SE agrees (with LRT) CIs should not overlap with zero percent change. If the CI is tight and with small change in the parameter, then that covariate relation can be concluded to be clinically irrelevant, despite being statistically significant. This may happen if you have many subjects in your data. (Or if your limit for what is a relevant change is very wide) In this case it may be justified leaving that covariate relation out of the final model. Then of course, the fact that something was not statistically significant does not mean that the covariate effect is clinically irrelevant - it may just be that you do not have enough information. To assess that you would need to use FREM or FFEM (instead of SCM) - but this is out of scope for your original question. Best wishes Jakob *actually, for this example, THETA 2 may be fixed according to allometric principles, but let’s assume this is a large molecule and that allometry was not deemed suitable in this case, and therefore the covariate was tested in SCM, or otherwise estimated.