X. Portilla posted on Thursday, December 18, 2014 - 9:58 pm
Dear Drs. Muthen,
I am interested in conducting an analysis with a sample that participated in an intervention with a 2x2 factorial design. However, I am interested in a developmental question, so would like to use the entire sample, yet I need to account for non-independence in the standard errors. My question is whether I should cluster by the RA unit (the practitioner who delivered the intervention -- practitioners, and thus families within, were randomized to one of the four groups) or whether I should cluster by intervention group. If I cluster by intervention group, I only have 4 clusters, whereas the RA unit provides me with 80 clusters. My sample size is ~1,300. Any insights are welcomed.
Dan Feaster posted on Friday, December 19, 2014 - 5:54 am
It would make the most sense to cluster (using two-level model) using the RA. You should probably control for the intervention using a fixed effect, even though your question is developmental.
X. Portilla posted on Tuesday, January 06, 2015 - 9:35 am
Thank you for your responses, Dan and Bengt. I have a couple of follow up questions:
1) In working with this sample to answer developmental questions, we are interested in running a path analysis with multiple mediators, predictors and outcomes. How does multi-level modeling work with multiple mediators?
2) When previously working with Stata to run the intervention analyses, we ran into a problem running a fixed effects model that uses cluster dummies to absorb all the between-cluster variation in the data: there was no variation in the practitioner assignment across RA groups. As such, intervention dummies have no within-cluster variation. In other words, each practitioner delivered only one intervention, so the practitioner and RA group are highly collinear. Since STATA will drop one variable when it detects high enough collinearity between two variables, the intervention variables would be omitted from the model. Would this issue apply in Mplus?
Thanks so much for your insights! Ximena
Dan Feaster posted on Wednesday, January 07, 2015 - 5:44 am
On Point 2, this is a limitation of the fixed effect approach to multilevel modeling--the fixed effect absorbs all variablity at the higher level and you can therefore not estimate the impact of a variable that does not vary at the lower level. This is a statistical/algorithmic issue not software so you would have the same issue in MPLUS or any program.
On Point 1, we would need a much more time/space to explore than this discussion or my schedule allow. I suggest checking out SEMNET and/or the book by MacKinnon.