Hi,
I have three groups that I compare with a different count as outcome.
I would like to calculate the power for this comparison.
I was thinking of proc glmpower and using Poisson as link function.
However this is not supported.
Is there a macro written and available for use that would be able to support my question?
Thanks, Katrien
That is one of the assumptions behind the link function--that it normalizes the data. For years before good maximum likelihood algorithms existed, data were transformed to meet normality assumptions. Actually, the preferred transformation for count data was a square root transformation (Sokal and Rohlf, Biometry, 1969). However, since you indicate that you will probably be analyzing the data using one of SAS's generalized linear model procedures (dist=poisson), and the Poisson distribution defaults to a log link, I suggested transforming using the log. Warning: If you have zeroes in your dependent variable, you probably ought to transform as log(y + 1). You may want to do some data exploration to check the assumption of having a Poisson distribution (e.g. are the mean and variance approximately equal?, are the data over-represented with zeroes?). You could find yourself wanting to use a negative binomial distribution (variance>mean) or a zero-inflated Poisson or negative binomial.
An alternative to the log transform would be to use the square root transformation prior to GLMPOWER, and analyze with LINK=POWER(0.5).
Good luck.
Steve Denham
You could try transforming your data prior to using glmpower. Since the analysis for the Poisson defaults to a log link, that seems like a logical first choice.
Steve Denham
Hi Steve,
Suppose I log transform the data, do you then suggest that I could treat my data as if normal?
Then proc glmpower would work.
Katrien
That is one of the assumptions behind the link function--that it normalizes the data. For years before good maximum likelihood algorithms existed, data were transformed to meet normality assumptions. Actually, the preferred transformation for count data was a square root transformation (Sokal and Rohlf, Biometry, 1969). However, since you indicate that you will probably be analyzing the data using one of SAS's generalized linear model procedures (dist=poisson), and the Poisson distribution defaults to a log link, I suggested transforming using the log. Warning: If you have zeroes in your dependent variable, you probably ought to transform as log(y + 1). You may want to do some data exploration to check the assumption of having a Poisson distribution (e.g. are the mean and variance approximately equal?, are the data over-represented with zeroes?). You could find yourself wanting to use a negative binomial distribution (variance>mean) or a zero-inflated Poisson or negative binomial.
An alternative to the log transform would be to use the square root transformation prior to GLMPOWER, and analyze with LINK=POWER(0.5).
Good luck.
Steve Denham
Thanks a lot Steve. I will proceed as you suggest.
Nice day, Katrien
Are you ready for the spotlight? We're accepting content ideas for SAS Innovate 2025 to be held May 6-9 in Orlando, FL. The call is open until September 16. Read more here about why you should contribute and what is in it for you!
ANOVA, or Analysis Of Variance, is used to compare the averages or means of two or more populations to better understand how they differ. Watch this tutorial for more.
Find more tutorials on the SAS Users YouTube channel.