Model with PROC GENMOD + BAYES with pre-specified priors not being recognized in output

[lsandell]

Hi there, I am trying to run a model in the Bayesian framework using PROC GENMOD with a BAYES statement. I am specifying my priors using a data step to create a reference data set with the values of interest for my priors. However, when I call that dataset into my proc genmod + bayes using the coeffprior=normal(input=dataset), the values shown in the output are still the defaults. 

The output shows the "Independent Normal Prior for Regression Coefficients" table with mean=0 and precision=1e^-6, which is not equivalent to my specified variance=1. I have tried changing the input to be precision=1 in the dataset, mean to a non-zero value, and still no changes in the output. 

Can anyone assist me on how to get SAS to process the desired priors? I am running multiple models (both simple and multiple linear and logistic regressions). Thank you.

CODE: 

* Normal prior (0,1) PROC GENMOD USING BAYES;
data Prior; 
	input _TYPE_ $ group ; 
datalines; 
Mean 1 
Var 1 
; 
run;

title1 "Bayesian Unadjusted Model";
title2 "Normal (0, 1) Priors";
proc genmod data=dataproc.trial2 DESCENDING ;
	class groupn(ref="Fentanyl");
	model blockn = groupn / dist = bin 
							link = logit ;
	format groupn groupn_f. blockn yesno.;
	bayes seed=123 nbi=0 DIAG=ALL coeffprior=normal(input=Prior);
run;

LOG:

26   * Normal prior (0,1) PROC GENMOD USING BAYES; run;
27   data Prior;
28       input _TYPE_ $ group ;
29   datalines;

NOTE: The data set WORK.PRIOR has 2 observations and 2 variables.
NOTE: DATA statement used (Total process time):
      real time           0.01 seconds
      cpu time            0.00 seconds


32   ;
33   run;
34
35   title1 "Bayesian Unadjusted Model of Block = Beta0+Beta1(Group)";
36   title2 "Normal (0, 1) Priors";
37   proc genmod data=dataproc.trial2 DESCENDING ;
38       class groupn(ref="Fentanyl");
39       model blockn = groupn / dist = bin
40                               link = logit ;
41       format groupn groupn_f. blockn yesno.;
42       bayes seed=123 nbi=0 DIAG=ALL coeffprior=normal(input=Prior);
43   run;

NOTE: The default sampling algorithm is the Gamerman algorithm, which is different from the
      default in SAS/STAT 9.3 and earlier releases. To revert to the previous behavior, specify
      the SAMPLING=ARMS option in the BAYES statement.
NOTE: PROC GENMOD is modeling the probability that blockn='Yes'.
NOTE: Algorithm converged.
NOTE: The scale parameter was held fixed.
NOTE: Algorithm converged.
NOTE: PROCEDURE GENMOD used (Total process time):
      real time           0.62 seconds
      cpu time            0.37 seconds

[SAS_Rob · Accepted solution]

The problem is the format of your Prior data set.  It should look something like the example in the documentation.

SAS Help Center: Bayesian Analysis of a Poisson Regression Model

Using an artificial data set, similar to your code, it would need to look something like this:

data trial2;
seed=2534565;
do groupn=1 to 3;
do i=1 to 250;
logit=-2 + .05*groupn;
p=exp(-logit)/(1+exp(-logit));
if ranuni(seed)>p then blockn=1; else blockn=0;
output;
end;
end;

run;

title1 "Bayesian Unadjusted Model";
title2 "Normal (0, 1) Priors";
data NormalPrior;
input _type_ $ Intercept groupn2 groupn3;
datalines;
Var 1e6 0.0005 1e6
Mean 0.0 0.1385 0.0
;
proc genmod data=trial2 DESCENDING ;
class groupn(ref="1");
model blockn = groupn / dist = bin
link = logit ;

bayes seed=123 nbi=0 DIAG=ALL coeffprior=normal(input=NormalPrior);
run;

[lsandell]

Thank you! So I must have a column for every level and variable in my model, within the input priors dataset, from what I gathered. I appreciate it.