I am building a LCMM model, and trying to understand how to select the covariates for the model. Is it feasible to first use LMER with Anova (add critical parameter, measure loglikelihood between models, retain if significant) to select the covariates, and then use the same covariates for constructing the LCMM model?
$\begingroup$
$\endgroup$
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
4
The best approach to selecting covariates is to be guided by the underlying domain theory, and expert knowledge, rather than from statistical procedures involving p values.
answered Jul 14, 2020 at 6:28
-
$\begingroup$ Thank you @robert-long. We do have a list of covariates from the domain expert, but trying to validate their significance, and trying to get the minimal set of covariates that suffice for modeling. $\endgroup$– nikarjCommented Jul 14, 2020 at 7:34
-
1$\begingroup$ Draw a DAG with guidance from the expert and this will inform the minimal sufficient set of variables. The free online tool Dagitty can do this for you. $\endgroup$ Commented Jul 14, 2020 at 7:45
-
$\begingroup$ Thanks again, will incorporate that. But from a statistical perspective, would the approach that I am suggesting be feasible? $\endgroup$– nikarjCommented Jul 14, 2020 at 8:18
-
1$\begingroup$ It's a feasible approach, but it is bad practice. You are talking about a stepwise procedure and you can read on this site and many others why is it bad. Don't do it. $\endgroup$ Commented Jul 14, 2020 at 8:53