Methodology Question: IVs - org elvel, DV - individ level aggregated to org level


New Member
Hello everybody,

Please help me with the following issue. Consider the attached model.

All the IVs at both stages (Xs and Ys) are at organizational level. The DV (Z) was collected at individual level and then aggregated to organizational level in the form of "% of the respondents who A,B,C" (came like that from a data set).

I tried 2SLS, OLS, SEM estimation for systems of simultaneous equations, but they all yield rather low R squared of about 5% for Z. I believe there is a multilevel issue with the DV and that is exactly why I am getting low R squared.

My initial exploration of this problem directs me to "Mixed-effects linear regression".

Please let me know if my thoughts on all this are correct or not. Also, what would you recommend on how to estimate this model with such DV?


Smelly poop man with doo doo pants.
Nothing at all, huh?..
well, what do you expect? we're trying to squeeze out our last days of summer :D

in any case, i think we need lotsa more info to get an idea of what you're doing. like what are the X's and the Y's? it helps out a lot more if we know what they are and in which scale they were measured.

the path diagram that you're presenting is... peculiar. you have paths connecting all the X's to the Z through Y2 but what's going on with Y1 and Y3? are they covarying with something? predicting something? being predicted by something?

then there's also the idea of what's going on with this R-squared. if you're using a simultaneous-equations approach (assuming covariance modeling-SEM and not that weirdo PLS stuff you mentioned before), you should be getting more than one R-squared. are they all less than 5%? or if you're only getting one, do you know how it gets calculated? an overall R-squared measure for a SEM model is possible but usually actively discouraged because it's tricky to interpret and susceptible to all kinds of screw-ups. (or you could simply have a really, really crappy model or a good model with really crappy measures)

so linear mixed effects models could be something you would consider, but that's mostly contingent on what you're trying to test. if you're really only interested in getting the fixed effects (and their significance tests) right, you could take the econometrics approach and only adjust for the standard errors.

the estimation of your model's DV is contingent on what you're trying to test. so all of this is just one big rant to say "we need more info to help you out".

PS- i thought you had got ahold of an R-specialist/methodologist to help you out? it's always good to have someone reliable to bounce around ideas when you're working on this stuff.

so yeah... in conclusion: summer and we need more info.