We have conducted a multilevel latent covariate linear regression model, and have some “paradoxical” results.
Study is two-wave full-panel with 2 years between T1 and T2. Respondents (N = 3000) distributed across 830 departments. Average dep. size; 4 (range 1-33).
Role conflict (predictor)and mental distress (outcome), adjusted for sex, age, skill level, and mental distress at baseline. Mental distress at T1 is strongest predictor for mental distress at T2 (ß = 0.7).
Variance 0-model; 0.165 (p = 0.000) at within level, 0.003 (p = 0.004) at between level.
Intercept+slope model show a stat. sig. effect at within level (ß = 0.09, p = 0.000), and at between level (ß = -0.08, p = 0.021). As you can see, ß values go in opposite directions. At department level, departments with a higher level of role conflict has a lower level of mental distress, contrary to what we would expect.
When plotting role conflict and mental distress at the between level it looks like the direction of the relation is as we would expect.
Could these “paradoxical” results could be a result of some “technical artefact”? For instance, could it be the adjustment for baseline distress? or; when adjusting for the relation at the individual level, is there so little variation left at the department level that this could cause us problems?
We have now conducted the multilevel regressions with aggregated mean scores and group-mean centering (we have also done grand-mean centering). We still get the same "paradoxical" results with the coeffisient at the between level going in the opposite direction of what we would expect (and opposite of the coeffisient at the within level).
Xu, Man posted on Wednesday, December 02, 2015 - 7:08 am
I´d also like to ask a question related to latent co-variate model. I have some groups of indicators representing distinct biological pathways. I would like to sum each of the pathways using latent aggregation and use these latent pathway means to predict the person' outcome. I think effectively this is a two level model, with person outcome at level 2 and biological indicators from different pathways at level 1.
The issue is that if I line up each pathway indicators in the long form, then I end up having missing data because there are different number of indicators in each pathways. Would this be a problem in statistical computation, or it is more of a substantive question in itself? Thank you.