Multilevel Multigroup Approach?

是 否 +2 论坛币

k人 参与回答

经管之家送您一份

应届毕业生专属福利!

求职就业群

赵安豆老师微信：zhaoandou666

经管之家联合CDA

送您一个全额奖学金名额~ !

立即领取

感谢您参与论坛问题回答

经管之家送您两个论坛币！

+2 论坛币

I am measuring depression changes over time, and I am working with a multi-site dataset, which includes 5 sites and about 100-200 people at each site, N=800, and repeated measures across three time points.

To control for site variation, non-equivalence of interventions,  and adjust for within-person correlation over time, I am proposing hierarchical linear models for repeated measurements to assess changes in severity of depression.

Using SPSS version 19, hierarchical linear modeling with repeated measures will be conducted to test for site location interaction with depression since the intervention at each site was non-equivalent. Data will be nested in two levels (client data are at level 1 and site data are at level 2) to account for site variation. Since there is possible clustering within the sites, site location will be set to random effects. The predictor will be defined as the intervention (i.e., integrated care or enhanced referral). The dependent variable is defined as scores on the CES-D.

However, I do not think I have a large enough sample at level two (n=5). Is it possible to set the site location as a covariate and random effect in the model? By doing so, have I truly adjusted for site differences? I am not interested in the sites, just the people.

Does anyone have any recommendations to justify the level 1 and level 2 sample size?

扫码加我 拉你入群

请注明：姓名-公司-职位

以便审核进群资格，未注明则拒绝

分享0 收藏0 回帖

关键词：Multilevel Approach iGroup multi Roach variation includes measures repeated version

I don't think you need a random effect for site, and, as you suggest, by most people's rules of thumb, you have too few anyway.

I think you should include dummies for each site (minus one reference category), and interact them with a dummy for treatment (rather than control). If desired, you could also interact site with other covariates.

Keep the random effects for person... so you would have just a two-level, repeated measures model.

You don't really need to control for possible clustering at the site level, since from the sounds of it you don't have any site-level covariates. The dummies for site will control for differences in the mean, and the interactions with treatment will allow for the possibility that the intervention had different impacts at different sites (and in fact will tell you whether it did).

My $0.02,

Malcolm Fairbrother

You could also use a multilevel multigroup approach; where client is level 2, its measure occasions are level 1, and sites are clusters groups). By fitting this multilevel model, in 5 different sites, you could test if this are similar or not.

If you only want to show how much variation its attributable for each factor (as an initial step for modelling), and by factor I'm meaning levels, I'm not sure you could use just the null model of a 3 level specification for your outcome (occasion nested in client, and clients nested in sites). But maybe if its just to estimate variance would be enough (anova like approach) in terms of observations per level, you would need to check that (maybe Malcolm would know this).
Good Luck!

Diego A. Carrasco Ogaz