My dependent variable is not normally distributed. I chose the Gamma log link, and I hope this is correct. I wasn't entirely sure with this because the dependent variable is actually a whole lot of percentages. I know this isn't technically continuous, but if I choose Poisson, it excludes most of my data. so question 1. if this is the wrong distribution, what is going wrong with the others?

Next, when I run the model, I have a few factors which are sex, stage (5 categories) and number of offspring (1 or 2 only). My covariates are weight and the other dependent variable, which are blood cell counts. I then selected all of these for the model, with an interaction between sex and stage, sex stage and offspring, sex and offspring. I have also played around by taking these in and out, and not having interactions at all. However, I have done single analysis on this before, and am interested in sex and stage related to my dependent variable. I also checked the box for pairwise comparisons, and Bonferroni.

So these are the questions I have:

1) if my interactions are either approaching significance or not significant, can I still look at the pairwise Bonferroni figures and quote the significant values from that?

2) If I take all my interactions out, but want to look at how my dependent variable is affected by stage in females only, can I exclude male cases, and run two separate models (then exclude females). If I explain this in a paper, saying that I ran three different (looking at the original also) models and use the results for all, is this valid? Or would this open up the possibility of errors.

3) I have been told by one person that if I have an interaction (significant or not), I can no longer look at the single effects. Is this true?

4) If my interaction between sex/stage and my dependent variable is not significant, should I exclude it from the model? The AIC doesn't change very much between them. If I am allowed to do 1 from above, there are a couple of interesting things happening.

5) why are my tests of model effects always significant but my parameter estimates aren't? is it because of the various levels in some of my factors?

  • 1
    $\begingroup$ The apologies are well meant, but not needed. CV exists to answer beginners' questions too. If the question is out of order, we will tell you. $\endgroup$ – Nick Cox Nov 11 '13 at 16:55

Most of the questions are so tied up with what makes sense for your problem and your dataset, which clearly we can't see, that remote advice is problematic. So, slow right down, please!

More crucially, it is vital to get the overall model right in terms of model family and link long before you start worrying about interactions and multiple comparisons.

So, what is your dependent variable (response or outcome)? (The reference to my "other dependent variable" also needs clarification.)

What is its definition? What are its actual and possible minimum and maximum values?

Gamma is a family; log is a link.

"a whole lot of percentages": if your response variable is bounded by 0 and 100, the most appropriate GLM is arguably binomial family and logit link with Huber-White-sandwich standard errors, after division by 100. I can't comment on whether SPSS supports that.

"if I choose Poisson, it excludes most of my data": what is SPSS objecting to? non-integers? That's a poor implementation of GLM if so.

| cite | improve this answer | |
  • $\begingroup$ My dependent variable is a response. It is bacterial killing - how well blood kills bacteria. We compare the number killed to a control, so you end up with the percentage killed or the killing rate, so technically it should fall between 0 and 100. This data is positively skewed.I do have 2 dependent variables. My other is white blood cell count. It is not a percentage, it is calculated per ml of blood. It doesn't need to fall between values. This was not normally distributed either. $\endgroup$ – Jess Nov 11 '13 at 17:22
  • $\begingroup$ So I suggest binomial. See e.g. stata-journal.com/sjpdf.html?articlenum=st0147 and its references. $\endgroup$ – Nick Cox Nov 11 '13 at 17:25

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.