1
$\begingroup$

I have 5 groups of very different sizes. I want to know if various attributes are the same for the groups when I have corrected for the differences in size, e.g.

> Nordjylland_CG #smallest group

     A     B     C     D     E     
     0     0     96    9     29

> Hovedstaden_CG #largest group

     A     B     C     D     E     
     0     7     457   93    158

My idea is to sum across the categories and divide by the sum so that I get normalised or scaled numbers in stead of raw counts.

The counts can be assumed to be poisson distributed, but what is the distribution of the scaled counts (I'm using R with glm to do anova analysis of them)?

Thanks!

$\endgroup$

1 Answer 1

1
$\begingroup$

If you want to compare the proportions, you could do that with a GLM or a log-linear model (loglin in R) -- or, for that matter, a chi-squared test or Fisher test, but you shouldn't just scale the counts to proportions, since the variance of those proportions depends on the number you scaled by.

If you're using a GLM you probably want to leave the numbers unscaled but write the model in such a way that the comparisons of interest correspond to a simple test of equality over a set of parameters.

If the categories are ordered you might want to consider an analysis that takes some account of that, perhaps ordered logit models

$\endgroup$
1
  • $\begingroup$ Can you elaborate a bit? Which family should I choose in the GLM? Choosing "poisson" or "quasipoisson" and doing anova with Chi-sq or F-test returns a significant difference between groups (no surprise, the largest is several times larger than the smallest), but doing the same with the scaled counts only returns a significant difference between categories A,B,C,D,E as expected. What should I do instead? $\endgroup$
    – SiKiHe
    Jul 1, 2015 at 6:37

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

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