Attention all psych and stats boffins!
I am running a study using 2 continuous IV's and 1 binary DV and 1 continuous DV. I have a few questions relating to the output I have already generated from SPSS. (N.B. The study is mainly exploratory):
1) Running a hierarchical linear regression for my 2 IV's and continuous DV I ended up with a situation where the order of the regression worked more effectively when I entered the smaller IV predictor 1st (i.e. both IV's were highly sig and accounted for decent proportions of the variance) than when entering the larger IV predictor first (i.e all of the variance being accounted for by this variable). Is it feasible and justifiable to run my hierarchical with the smaller predictor 1st?
2) Running the logistic regression using forward L-R the 2 IV's and an interaction with the binary DV produces all 3 variables as being sig in step 0 but one of the IV's is more sig than the interaction (2nd most sig) and I am mainly looking at interaction effect. In Step 1 the interaction effect is left out of the model all together and then becomes the least significant. Am I able to use theory to justify using the interaction effect in my equation?
3) My study would also like to look at the effect of any potential differences between countries of my 2 IV's against the continuous IV and my 2 IV's against the binary DV. It doesn't appear that I can run regression comparisons of the results of both countries against one another (logistic and linear). Is this the case? If so do I just run separate linear and logistic regressions for each country and compare them based on sig levels myself manually?
Appreciate any ideas or suggestions to help clear up these issues. Many thanks.
I am really not sure whether I understand your post correctly, but I will have a try:
1) What do you mean by hierarchical regression? It doesn't seem to be hierarchical model combining data from different levels of analysis. I guess you mean stepwise regression. Opinions are split about stepwise regression because of the importance of the order in which you introduce the regressors.
2) How can you have an interaction between an IV and a DV? It also seems that you seem to place a little bit too much emphasis on statistical significance (most sig, 2nd most sig, etc.). Besides, the justification for an interaction effect MUST be theory Brambor, Thomas, William R. Clark and Matt Golder (2006): Understanding Interaction Models: Improving Empirical Analyses. Political Analysis 14 (1): 63-82.
3) For some statistical reason, you should not compare estimates across logistic models. An alternative could be to estimate some interaction model with country dummies (but could make things too complex). In principle, in OLS you can compare estimates across models.
I hope it is not too late to reply. I have used SPSS 17.0 for all my multinomial logistic regression (MLR) models which are solved as an extension of the logistic model, ie each multi categorical level (k) of the DV is grouped with a reference level and converted into k-1 binary logistic models. I have obtained excellent results.
1) I do not see why the order of the IVs should make any difference. And if your DV is categorical, why are you considering it as a continuous DV???
2) Did you try running your model as a MLR - after all it will be converted into a logistic regression model.
The above views are based on my sketchy understanding of your problem from the above description - which may not be relevant to your actual results. In such a case, please ignore my responses.