I have a model with a continuous dependent variable (DV) and 9 independent variables (IVs). My model is a mixed-model, because 7 of the 9 IVs are nested within the two levels of another IV (lets call it X1) which is applied in a repeated-measure design to the same group of subjects (zero at first trial and 1 at the second trial). So I have done the nesting according to the SubjectID. X1 is either 0 or 1 for each subject, but other IVs except another IV are similar for each subject. As I stated right now, a remaining IV (lets call it X2) is not nested. As a better explanation, X2 does not follow a repeated-measures pattern; and thus it is not necessarily the same at the first and second trials in each subject (it might be the same for some subjects, and there is some correlation, but it is not intentionally repeated like the other 8 IVs).
Now I should build a mixed-model in which the nesting is done according to subjects/trials/X1 (SubjectID). The effect of X1 is being already added as a random factor. Now, all the variables except X2 are properly modeled.
However, I am not quite sure about X2. I know I can (and perhaps should) add X2 as a fixed-effect to the mixed-model. But that way, I will lose some of good information because the model will treat the X2 as a nested variable, while it is not actually nested. I also can run a fixed-effect regression with only X1 and X2 to assess the role of X2 regardless of the nesting. That is somehow good, but that way I cannot add the nested variables (because they will cause atomistic fallacy). So I want something in between. I don't know how to build a mixed model that has the advantages of a mixed-model (has random factor) and also when it comes to X2, does not lose some power (does not consider the nesting according to SubjectID for X2). So I need a model that in most of parts considers SubjectID to nest the data (all the IVs except X2) and at some part (in the case of X2), does not consider SubjectID.
Is it possible?
Do you have any other suggestions?
...On a second thought, it appears to me that perhaps X2 can be considered repeated-measures as well. It is true that this variable is not like the other ones, exactly repeated within each subject, but there is still some correlation. So perhaps it is good to run a routine mixed-model regression without worrying about losing some power. What do you think?
