I'm having hard time choosing the right statistical method for my study. The data that I have is summarized below.
- Waterbird count data (absolute counts). Non-standard, varying effort but standard effort is assumed.
- The counts are done yearly and the years are categorized into two groups: high water level years and low water level years. Sample sizes are different for these groups (e.g. there are 8 high level years and 6 low water level years)
- I want to see if there is any statistical difference between high water level years and low water level years in terms of mean/median number of birds counted.
Because the count data is not normally distributed, I directly used Mann-Whitney U-test for the purpose. But I wanna know what other or better methods can I use for the same purpose? I also want to compare the two groups in terms of different biodiversity measures like species richness.
Hi Kaan, a useful approach would be to run a GLM using counts as response variable and groups as independent variable. You should also run some goodness of fit tests to see what error-distribution model (Poisson, negative binomial, etc) best fit to your data.
Search for "count data and GLM" in google and you will find some more answers.
Good luck!
Hi Kaan, a useful approach would be to run a GLM using counts as response variable and groups as independent variable. You should also run some goodness of fit tests to see what error-distribution model (Poisson, negative binomial, etc) best fit to your data.
Search for "count data and GLM" in google and you will find some more answers.
Good luck!
Dear Kaan,
I would first and foremost support the answer provided by Nicolas. If you have the correct specification of the counts it will be provably more powerful as an approach.
However, there are two issues with your question that do have an influence on this. First is whether the counts are being conducted at multiple sites - which may or may not all be represented each year. GLM can handle this but MW cannot. Second, you state that you are interested in a mean difference but MW may be more sensitive to medians. Accordingly the tests are not necessarily equivalent. Depending on your beliefs about the counts MW may in fact be more appropriate. For example. if the distribution may be contaminated with some extreme values then the higher breakdown of MW might be what you want.
If this is the case, then consider a permutation test on medians as being potentially more efficient/powerful then MW; otherwise GLM is very likely what you want.
Dear Nicolas and Bruce, thank you very much for your answers. I am already trying GLM which was also my advisor's choice. Bruce, I understand what your concerns are. From what you said, it seems like GLM would be better. Any software advices to run GLM on? Or should I just go with R.
Dear Kaan,
Nothing wrong with R in this case. Personally I use Stata and/or Mathematica most often, but most major packages can handle this job and the price is right for R.
- Badges
- Science topic