04-28-2017 10:52 AM
First of all, I do not consider myself being a stats-guru and my "extended" SAS and stats-knowledge is mostly self-tought. My question deals with the SAS procedure GLIMMIX and, more precisely, the generated output table.
My research question is related to animal sciences and with my statistical analysis I am trying to identify how different feeds and ingredients may interact with one another affect animal performance (my response variable).
I thought of analyzing my data using a full factorial GLMM. I am having 2 class variables and I came up with the following model:
ods graphics on;
Data=Data2 plots =pearsonpanel;
class RefDiet TestIngr;
model ADCprotI=RefDiet TestIngr RefDiet*TestIngr / dist=beta
lsmeans RefDiet*TestIngr / adjust=bon ilink oddsratio;
ods graphics off;
The beta distribution was chosen according to the residual panel, as it seems to provide a better fit and be more appropriate than a gaussian-shaped distribution (response variables are expressed as proportions with values between 0 and 1). With one specific response variable I am now getting standard errors of 0 for all class variables, resulting in infinite t values. As a result, all class variables turn out to likely have a highly significant effect on my response variable. However, something seems wrong here, but I am currently stumped... Does anyone has a good suggestion on how to overcome this issue or am I doing something completely wrong here? Please let me know if you need more information + thanks a lot!
The result for e.g. fixed effects looks like this (lines should be labelled with: RefDiet TestIngr RefDiet*TestIngr)
And for the parameter estimates (just the upper lines, the shown line names are: RefDiet 1 RefDiet 2 TestIngr 10 TestIngr 20 TestIngr 30...)
04-28-2017 11:03 AM
You get infinity for a t-test value when there is zero variability in the residuals. In other words, the model has fit the data exactly.
04-28-2017 03:01 PM
Thank you very much for your quick reply, Paige. In other words: would you consider it legitimate to report this outcome "as is" (standard-error = 0 and F-value infinity)? Of course this makes running a multiple comparison test a bit difficult (impossible)...
04-28-2017 03:25 PM
I'm sure there are some real-world situations where this is a legitimate outcome, but there are also some real-world situations where this is not a legitimate outcome.
The zero error situation could arise from mis-specifiying (actually over-specifying) the model (i.e. statistical analysis error), it could also arise from degenerate data (perhaps some error in data collection or data transmission or error in entering the data into the database or computer coding error).
It could be caused by a many different things.
04-29-2017 01:11 PM
Your model is likely overspecified (too many terms). Also note (very important): the beta distribution is defined for 0 < y < 1. Thus, 0s and 1s are not allowed. All proportions equal to 0 or 1 become missing in the analysis. If you don't have 0s and 1s, then you are fine, but otherwise you are not using all your data. Be careful.
If your proportions are for a continuous variable, then binomial is not appropriate.
05-02-2017 02:34 PM
Apologies for the late reply and thanks for the notification. Indeed, binomial would not be appropriate and, likely, I should go with a beta (proportions are for a continuous variable). The model stated above seemed to work for a bunch of variables, but is failing for the one listed above...