[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: st: ratios generated from count data

From   Maarten buis <>
Subject   Re: st: ratios generated from count data
Date   Mon, 22 Dec 2008 10:12:00 +0000 (GMT)

--- Nikolaos Pandis <> wrote:
> I have counted the number of articles that show non significant and 
> significant results in 5 different journals, over a five year period.
> The count data is further divided in 4 categories depending on the
> subject of the article. 
> The strategy I followed is:
> I collapsed my data to avoid the zero values that were generated per
> monthly journal issue.
> I calculated the ratios for nonsig/sign.
> I transformed the ratio variable into log scale and I got reasonably
> normally distributed data.
> 1. My concern relates to the count nature of my data. I was not sure
> that creating ratios of count data and then treating the ratio
> outcome as a continuous variable and perform linear regression is
> correct. Is this valid or would you recommend another approach?

A more common approach is to use count models (like -poisson-) with the
-exposure()- option. There are also models out there (e.g. -zip-) that
model zeros differently from the other counts, which may make
substantive sense in your case, as journals tend to be hesitent to
publish non-significant results.
> 2. If I use an interaction term (journalnumb*subject-not in the
> analysis shown above), I get a significant result. I am not sure how
> to interpret that in conjunction with the rest of the analysis. 

If you try enough things, you will find significant results (remember
that significant means that you expect to make a mistake in 1 out of 20
times that you declare something to be significant).

> 3. How would you perform pairwise comparisons of the subgroups
> created?

I think that that may not be good idea, for the reasons stated under
point 2 and for improving the clarity of your paper (too many numbers
for the reader to digest). 

-- Maarten

Maarten L. Buis
Department of Social Research Methodology
Vrije Universiteit Amsterdam
Boelelaan 1081
1081 HV Amsterdam
The Netherlands

visiting address:
Buitenveldertselaan 3 (Metropolitan), room N515

+31 20 5986715

*   For searches and help try:

© Copyright 1996–2015 StataCorp LP   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index