PROC GLIMMIX & Missing values

[buder]

Getting a lot of observations that are dropped when using PROC GLIMMIX procedure. The note reads: 

Some observations are not used in the analysis because of: missing response values (n=52508), missing fixed effects

(n=53078), zero or negative weight (n=5972), not a proportion (n=36597).

NOTE: Convergence criterion (GCONV=1E-8) satisfied.

When comparing the number of observations for the PROC GLIMMIX procedure and running an OLS model, these are the comparisons: 

Number of observations read: 132032

Number of observations used Glimmix procedure: 41428

Number of observations used OLS procedure: 77209

Number of observations for dependent variable: 79524

Any suggestions as to why there is such a discrepancy in number of observations read for Glimmix would be appreciated. 

[Rick_SAS]

It would be helpful to know what options you are using for PROC GLIMMIX and the other procedure.

For example, the "not a proportion" note might indicate that you are using the DIST=BINOMINAL option instead of the DIST=BINARY option. There was an post last year in which the OP had a similar note.

[buder]

Doing a quasi-likelihood regression, so here are the options: 

PROC GLIMMIX DATA = TEST;
MODEL QALY =
INCOME

AGE

....ETC

 /LINK = LOGIT S DIST = BETA;
output out=fracout pred(ilink)=pred lcl(ilink)=lower ucl(ilink)=upper;
RANDOM _RESIDUAL_;
WEIGHT WTFA;
RUN;

[SteveDenham]

With link=logit and dist=beta, I would guess that the values falling into the "not a proportion" are greater than 1 or less than 0.  That would lead to this.  As far as missing response/missing fixed effects part, it looks like that a large part of these overlap.  OLS would not detect that a large part of the values did not fit the specified distribution.

Steve Denham

[lvm]

You seem to have several observations where your response is larger than 1 or smaller than 0. That is why these are "not proportions". Probably you made an error in calculating the proprotion. And more seriously: the beta conditional distribution is defined between 0 and 1 (0 < prop < 1). That means that all 0s and 1s are converted to missing values. You cannot use the beta distribution if you have 0 and 1 unless you want to throw away data. Many references fail to make this clear. To get around this, you only have ad hoc solutions. For instance: Convert all 0s to a very small number (smaller than the smallest nonzero real value that you could observe). Same idea for the 1s. Of course, this is creating artificial data. Would be OK for occassional 0s and 1s, but it appears that you have many.