Count data

[mthorne]

Hello,

I have plant count data taken from 2x10 m areas in field plots. If I convert these numbers to a square meter basis, GLIMMIX doesn't often converge when applying a negative binomial distribution. Is there anything wrong with doing the analysis on the whole plot area, then converting to square meter densities in a publication, as long as I make it clear in the methods what I did?

Thanks,

Mark

[ballardw]

Suggestion: Provide some of the starting data, what you are doing to convert those numbers to a "square meter basis" and the Glimmix code you run. It may also help to provide the Log from the Proc Glimmix to see what notes the procedure may provide.

Provide data in the form of data step code pasted into a code box opened on the forum with the </> icon that appears above the main message window on the forum. Provide code and log results in a similar code box.

It may also help a bit to describe your conversion code is supposed to be doing in case that had any logic issues along with more detail of the collection. Did you have a single plot of 2mx10m and your count was a single number of that whole plot? Of were the collections for smaller sections such as 2mx2m?

If you are modeling the count as the dependent variable of model that directly would, if I understand, relate to a similar 2mx10m plots.

I would personally be a bid cautious of disagregating a count to any smaller area unless you could guarantee that each 2mX10m plot was pretty homogeneous. Do you believe that was the case in your collection? If the collection was with "designed to be homogenous" 2x10m plots then the average count per meter should be re-portable. If you have any reason to believe that the per square meter counts for some of the 2x10m plots varied significant then they are not homogeneous and probably should not do so. This may be indicated by the model fit statistics. If a negative binomial doesn't fit then it may be that is not a good model to start with or you data just does not support your model attempt.

[SteveDenham · Accepted solution]

@mthorne , what might be happening with the rescaling to plot area is that the data is no longer an integer count, and if you have a lot of values between zero and one, the log link gets a negative value which may be causing the failure to converge. Since all of the plots are identical in area, I think your plan to analyze the count data, and then post-process the back transformed lsmeans (ILINK mu values) to counts per square meter, is a good approach. Back transformed standard errors would also be post-processed to counts per square meter. 

SteveDenham

[mthorne]

Thank you, Steve! Yes, the data become small decimals when converted to a square meter basis. GLIMMIX has no problem, even with the zeros if the the numbers are left as integers. This helps me better understand GLIMMIX.  

Mark Thorne

[mthorne]

Good points! The problem with perennial weed species is that they don't often occur in uniform densities, and that is whey I counted the whole plots. I initially set up the blocks so that the densities would be as uniform as possible across all blocks, but within each plot the spatial variability was greater. The conversion was a simple division of the count per plot by the number of square meters in the plot. The count data was modeled as a dependent variable of the independent treatment variable. The only reason I had for converting to a square meter basis was that scale is common in the weed science journals.