Is count data for parametric analysis or non-parametric analysis?

#1
Hi,

In our paradigm, animals are presented with an stimulus and we count the number of calls they make.

I believe the number of calls is discrete variable and count data.

When we do an anlayis on these data, do we do parametric analysis or non-parametric analysis?

For example, we have brain imaging data from the same animals, these data are continuous variable and normally distributed.
When we correlate the brain imaging data with the call data, do we do Pearson's R or Spearman's rho?

Thank you, :)
 

trinker

ggplot2orBust
#2
Re: Is counr data for parametric analysis or non-parametric analysis?

I am just learning about handling count DV's myself and have found this article to be helpful:
Count Data Regression

Maybe Dason will weigh in on this, as he's often heard preaching the good news of Poisson Regression.
 

Link

Ninja say what!?!
#3
Re: Is counr data for parametric analysis or non-parametric analysis?

I'm just kind of curious.

NP analysis is a lot harder to carry out than parametric analysis. Seeing that you are unsure of which to use, would you be able and willing to put in the vast amount of time I think it'd take to understand and implement it?

That being said, it appears that Poisson Regression might fit your proposal well. However, I cannot be sure until I know more about your data and how it is set up. Basically, can you describe each observation in your dataset for me?
 
#4
Re: Is counr data for parametric analysis or non-parametric analysis?

Thank you for your reply.

So, there are different types of calls that the animal makes in the presence of a potential threat.
During the 5min test session, we count how many calls are made by the animal.
So, for animal 1, we get 20 type A call and 50 type B call; for animal 2, 0 type A call and 30 type B call and so on.

The types of call are the variables and the animals are the cases.

I wonder if we look at this data set as frequency data, then as long as they normaly are distributed, we can do parametric test on them.
 

CB

Super Moderator
#5
I would second the suggestion to use a model that treats the count variables as, well, count variables. Traditional parametric tests seem inappropriate - count data is discrete, whereas a parametric statistic like Pearson's R assumes continuous data. (The normal distribution is a continuous probability distribution, so non-normality of data and sampling distributions is essentially guaranteed here, I would've thought).

Poisson or negative binomial regression might be suitable and aren't really all that difficult to apply (both can be done in SPSS). You will need to do some reading on the methods though.
 

bugman

Super Moderator
#6
Re: Is counr data for parametric analysis or non-parametric analysis?

I wonder if we look at this data set as frequency data, then as long as they normaly are distributed, we can do parametric test on them.
Remember the equal variance assumption. This is often violated with count data. I would advocate looking at a possion regression / loglinear model for this.