 |
 |
Jon - you might look into using negative binomial regression for these scores: I have found this model to fit well for very skew measures like the GHQ12. If you have the good fortune to be a user of Stata, you can fit the models very easily to your repeated measures data using 'xtnbreg'. Cheers, Keith Dear At 08:54 30/11/00 +0000, you wrote: > Dear Allstatians, > > > I have a depression score measured at three separate time points. > Each score can take values from 0 to 30 and all 3 are similarly negatively >skewed. > > I want to assess whether the change in score is greater between time points >2 and 3 > compared with between time points 1 and 2. > > My first though was to calculate two change scores (2 minus 1 and 3 minus >2) and then use a paired t-test > to compare them. However, I am not sure this is appropriate. Due to the >scores > being restricted to [0,30] there is a strong dependency between the two >changes - the range > of possible values for the second change is dependent on what the first >change was. > As a result it is often impossible for the two changes to be equal (and >this would surely be my null). > > > The correlations between the three scores are 0.633 (1 and 2), 0.571 (2 and >3) and 0.522 (1 and 3) > with n = 9028. Are these similar enough to use some kind of repeated >measures technique or > am I violating homogeneity of covariance assumptions? > > > Any help would be gratefully recieved. > > > Cheers > > > > Jon > > >=========================================================================== > Jon Heron > Research Statistician > Dept of Child Health > University of Bristol ---- Dr Keith B.G. Dear, Senior Fellow The Australian National University Canberra, ACT 0200, Australia National Centre for Epidemiology and Population Health (NCEPH) http://nceph.anu.edu.au +61(0)2 6249 4865, fax 6249 0740 Centre for Mental Health Research (CMHR) http://www.anu.edu.au/cmhr +61(0)2 6279 8412, fax 6249 0733