proc glimmix, abnormally high predicted probabilities

[grateful]

Hello!  

I am using proc glimmix (SAS 9.4) for a two level model (health care providers and organizations, which are both categorical) that regresses provider type (0/1) on imaging (0/1).  health care provider is specified as the random intercept.

There are 92,100 cases nested in 7,133 providers, nested in 2,370 organizations. All providers and organizations have at least 5 cases and there are no multi-organization providers.

The model results are consistent with what I would expect, until I add organization as a fixed effect.  The fixed effect intercept increases considerably, which triples the predicted probability (model 3).  The syntax and models are below.  Any ideas as to what the problem could be?

PROC GLIMMIX DATA=please_work METHOD=LAPLACE NOCLPRINT;

CLASS  provider org;

MODEL  X_RAY   (EVENT=LAST) = provider_type org / CL DIST=BINARY LINK=LOGIT SOLUTION ODDSRATIO (DIFF=FIRST LABEL);

random intercept / type=vc subject=provider;

COVTEST / WALD;

run;

	Model 1	model 1b	model 1c	Model 2	Model 3
changes to model	naïve, provider as class, random effects for the provider intercept	naïve, organization as class, random effects for the organization intercept	org and provider as class, org as fixed effect, random provider intercept	Model 1A + added provider type to fixed effects	model 2 + org as class and fixed effect
Fixed Effects (Solution for Fixed Effects)
Intercept (log odds of receiving a xray with a typical provider)	-1.204	-1.2581	0.9022	-1.23	0.5816
	(<.0001)	ns	(<.0001)	(<.0001)	NS
type				0.151	0.188
				(<.0001)	(<.0001)
Odds Ratios
type				1.16	1.21
Predicted Probability	0.231	0.221	0.711	0.226	0.641
Error Variance ( Covariance Parameter Estimates)
Intercept(provider)	0.445	0.383	0.078	0.452	0.073
	(<.0001)	(<.0001)	(<.0001)	(<.0001)	(std error 0)
Model Fit
(-2) Log Likelihood (lower is better)	100,237.00	100,847.70	95,864.76	100,135.00	95,729.98
NPI ICC	11.91%	10.43%	2.32%	12.08%	2.17%

[PaigeMiller]

Adding a class variable into a model will change the intercept. This is a known fact, based upon the way SAS handles categorical variables. I don't see anything wrong here.

Predicted probability of what?