Multilevel regression problem

#1
I am new to multilevel models and I am basically wondering whether I can use a multilevel model at all, given the structure of my data.

I have collected data on membership of lobby groups in committees used by the Directorates-General (DGs) of the European Commission (the DGs are a bit like ministries in a national context). I have data on the number of committee seats occupied by each group for all DGs. Level 2 of the analysis would be lobby groups, level 1 would be the DGs.

Now, my problem is that many lobby groups are present in more than one DG. It is my understanding that multilevel regression should normally only be used if the cases on the lower level are nested in those of the higher level (i.e. students are parts of a class, classes are part of a school, etc.). In effect, the "students" (lobby groups) in my data are members of several "classes" (DGs) each. Is it still possible to do a multilevel model in this case?

I am thinking of "forcing" the lobby groups to be nested in the DGs by adding data rows, so that one lobby group present in 5 DGs effectively becomes 5 identical lobby groups present in 1 DG each. Is this advisable?

Many thanks in advance!
 
Last edited:

Jake

Cookie Scientist
#2
Now, my problem is that many lobby groups are present in more than one DG. It is my understanding that multilevel regression should normally only be used if the cases on the lower level are nested in those of the higher level (i.e. students are parts of a class, classes are part of a school, etc.). In effect, the "students" (lobby groups) in my data are members of several "classes" (DGs) each. Is it still possible to do a multilevel model in this case?
Yes, it is still possible. You are dealing with crossed random effects (as opposed to hierarchical/nested random effects, which is the somewhat more traditional or classic case).

A perfectly hierarchical dataset is one in which every DG is in one and only one lobby group.

A perfectly crossed dataset is one in which every DG is observed with every lobby group (so you could imagine, for example, a matrix where rows = DGs and columns = lobby groups, and there is at least 1 observation in every cell).

Your data is in between. This is sometimes called a partially crossed dataset. Most modern multilevel model packages can handle this kind of data (although the popular package "HLM" cannot).