About multivariate response model

Welcome to the forum for MLwiN users. Feel free to post your question about MLwiN software here. The Centre for Multilevel Modelling take no responsibility for the accuracy of these posts, we are unable to monitor them closely. Do go ahead and post your question and thank you in advance if you find the time to post any answers!

Remember to check out our extensive software FAQs which may answer your question: http://www.bristol.ac.uk/cmm/software/s ... port-faqs/
Post Reply
Posts: 1
Joined: Thu Feb 28, 2019 4:57 pm

About multivariate response model

Post by chenr392 »

Hi everyone

I'm analysing some effect sizes collected from a number of empirical studies, so the data structure is multilevel, with ES nested within studies. ES is my response variable, which comprises of response type A and B. So I'm trying to fit a multivariate response model. I have encountered some difficulties therefore I'm here to seek help.

1) I don't have ES level explanatory variables but only study level explanatory variables. I want to test the influence of these study level variables on A and B. However, for some variables I propose they have similar effects on A and B, and others have differential effects on the A and B. For the variables I proposed to have similar effects, it possible to estimate only one coefficient for each explanatory variable on both A and B, because I'm not interested how different these effects are on A and B. I found a function of "add common coefficient", which will produce an h equation. Can it be used for my purpose? When should the h equation be used?

2) Within each study I have multiple ES reporting both A and B, not only one ES for each type of response. Will this model be useful for such a situation?

3) How to calculate the ICC for this model?

Thank you very much for your help

Posts: 140
Joined: Fri May 21, 2010 1:21 pm

Re: About multivariate response model

Post by billb »

Hi Chenr392,
That all sounds quite complicated. Perhaps a starting point would be to fit separate univariate models for the type A and B responses you could then look at the estimates produced.
I would then be tempted to fit a univariate model but with indicator functions to identify A and B so that you can test for differences of effects on A and B so model something like
y_ij = b0 + b1* B + b2* x1 + b3 * x1*B +... + u_j + e_ij
then b1 identifies differences between types A and B, b2 gives effect of x1 on type A and b3 gives differential effect (and if zero can be removed and then you have common effect). If you expert different variability then you can make random part u_0j + u_1j*B for example

In terms of other questions I think 2 is easier in this univariate formulation as you are not matching each ES for A with ES for B which sounds like it would be difficult here and thus make multivariate hard.

You can work out ICC fairly straightforwardly - in simplest model it is simply sigma^2_u / (sigma^2_u + sigma^2_e). If you decide to make random part depend on A/B then it should still be straightforward.

Hope that helps
Post Reply