26th July 2011 at 4:25 pm #3326Katy PearceMember
I’m working with 2 years of data (not panel, but same items, equal samples) and looking at the effect of a continuous variable (scale 0-5) on a 5 level categorical variable.
In the highest level of the continuous variable, there is a 30% change from category 1 to category 5 betweent the 2 years and I think that I know why (and there was no such change in any of the other levels of the continuous variable).
(I almost never work with categorical data, so this is tough for me)…
I THINK that I could do a regression model.
What I want to know:
– the relationship between the IV and the categories of the DV
– if there was a statistically significant difference between year 1 and year 2
(Assuming that there were no demographic changes between year 1 and year 2 for the IV).
Tools at hand: SPSS, MPlus.
Thanks!28th July 2011 at 5:02 pm #3332
As Dave suggests, clarifying whether you have repeated measures is important. I would be cautious of collapsing the five levels in order to run a particular statistical procedure, especially if the five levels are conceptually important (I am guessing they are, otherwise you wouldn’t have five different levels). Let’s assume you have cross sectional data. If the categorical DV is ordered categorical, then you could examine the distribution of the data and select an appropriate regression model (e.g., ordinal, Poisson, negative binomial). If the data are categorical (unordered), then you should probably look at a multinomial regression. The multinomial regression can be a major pain for interpretation, but if you have resources to guide you with Stata or R, you can implement the “clarify” or “zelig” procedures for post-processing your analysis to greatly improve interpretation. Definitely check out these implementations by Gary King, along with an tour de force paper, “Making the Most of Statistical Analyses: Improving Interpretation and Presentation.”
Brian28th July 2011 at 5:27 pm #3331Katy PearceMember
Thanks. It isn’t panel data, but same items, equivlent sample.28th July 2011 at 5:34 pm #3330
Since it isn’t panel data, then I suggest selecting the regression model that works with your distribution of data, as I described.28th July 2011 at 6:40 pm #3329Mads Meier JægerMember
Ordered logit/probit would be appropriate for your setup. This can be done easily in SPSS.
Mads28th July 2011 at 7:42 pm #3328
Ordered probit / logit would not really be appropriate if the data were at the nominal level, right? From the description, the categorical data were not described as having ordering. If the DV is, indeed categorical without an ordering of values, then I stand by my recommendation that nominal logistic regression is the proper method.
Brian28th July 2011 at 8:16 pm #3327
I meant multinomial logistic regression….sorry
- The forum ‘Default Forum’ is closed to new topics and replies.