multinomial logistic regression - assumptions violated?

N

~Narayan~

Guest
#1
Hi everybody :)

I ended up asking a few questions in another thread that however didn't have a matching title, which probably explains why nobody replies. Since it's really important for me to understand, I am reposting my questions in this new thread with a matching title... and hope someone can explain things to me. :D

I want to do a multinomial logistic regression (for the first time ever) in SPSS 17.0 to test interactions between reproductive success in a bird species and several variables (e.g. age/origin/experience of females, same things for their males), my factors and covariates. The reproductive success, my dependent variable, is measured as a categorical variable with the 3 possible states NH no hatch, H hatch, Y surviving young. The important thing is that this success is measured for individuals over multiple decades, so each individual has multiple records for NH (and possibly H and Y) throughout its life. For example, some lines of my dataset would look something like this
outcome1: no hatch, female: A, male: B, year: 2010, age female: 10, age male: 15, [...]
outcome2: hatch, female: A, male: B, year: 2012, age female: 12, age male: 17, [...]
outcome3: ...

I know multinomial logistic regression assumes independent measures, so I am concerned about this. Do I violate this assumption?

As far as I understand the greatest risk arising from correlated non independent observations in multinomial logistic regression would be overdispersion. I have non significant values of Pearson and Deviance in the chi-square "Goodness-of-Fit" test in SPSS and the dispersion parameters are close to 1 and far away from 2: chi-square-Pearson/df = 1.03, chi-square-Deviance/df = 0.98. From what I have read, I can conclude overdispersion not to confound my regression in that case. So is my regression valid then?
Are there other things to worry resulting from non-independent observations in a multinomial logistic regression besides overdispersion?

Thanks for reading,
looking forward to your comments,
N