bayesian method in sas

[Ameurgen]

Hi everyone,

I am working on a linear mixed model using PROC HPMIXED, but I have encountered an issue with my dependent variable, which is not normally distributed as it is a count trait. Since I am dealing with count data, I considered using a zero-inflated model with PROC NLMIXED. However, this is not feasible due to the large size of my dataset (over 100k observations).

I am exploring other approaches, such as Gibbs sampling or Bayesian methods, for resampling. My question is: Does SAS have a procedure that can handle this approach, as I do not have time to explore R or other programs?

Thank you for your suggestions.

Best regards,

[StatDave]

For a count response, you can fit appropriate models (Poisson, negative binomial, or zero-inflated versions of either) in PROC GENMOD. They can also be fit PROC HPGENSELECT as well as in PROC COUNTREG and PROC HPCOUNTREG in SAS/ETS. Zero-inflated models can also be fit using PROC FMM. 

[Ksharp]

@SteveDave is right.
GEE model(PROC GEE or PROC GENMOD) is good(high efficient) for your BIG table.
But GEE model is a little different with Mixed model.You can use both of them.

[Ameurgen]

Hi everyone,

I am working on a linear mixed model using PROC HPMIXED, but I have encountered an issue with my dependent variable, which is not normally distributed as it is a count trait. Since I am dealing with count data, I considered using a zero-inflated model with PROC NLMIXED. However, this is not feasible due to the large size of my dataset (over 100k observations).

I am exploring other approaches, such as Gibbs sampling or Bayesian methods, for resampling. My question is: Does SAS have a procedure that can handle this approach, as I do not have time to explore R or other programs?

Thank you for your suggestions.

Best regards,

[ballardw]

Did you actually run the zero-inflated and encounter problems? If so what problems?

[Ameurgen]

yes,  i used zero inflate, but as i mentioned i musing quiet large data so proc NLmixed can handle the large matrice x dimension, so now im looking for another method, so i said may be the bayesian method can do the job,

My data is around 200 000 records

Thank you for your response

regards

[Ksharp]

1) You can use PROC GLIMMIX to fit Possion Distribution of Mixed model. But you have a very large table.
2)If you want use Bayesian method of Mixed model ,Check PROC MCMC .

And better post your statistic question at Stat Forum:
https://communities.sas.com/t5/Statistical-Procedures/bd-p/statistical_procedures

[Ameurgen]

thank you for your response  Ksharp, ill ckeck mcmc procedure , its quiet complex 

what do you think about Proc BGLIMM (Bayesian Generalized Linear Mixed model, can do the job.? what is the differences?

regards

[Ksharp]

Yes. You can do Bayesian method of Mixed model by PROC BGLIMM.
I almost forgot this PROC when @SteveDenham mentioned it before.

[Ksharp]

"what is the differences?"
No different.
PROC BGLIMM is for rookie of Bayesian method.
PROC MCMC is for Bayes expert to do some customize method.

[Ameurgen]

Thank you Sir Ksharp,  follow-up  my questions, 

I do like proc bglimm, it is straight forward application for LMM using bayesian method, but in my case i asking if any one cant run it in sas because , i have error ' proc bglimm not found ' im using 9.4 version of sas but this procedure is not available.

Thank you 

Regards

[Ksharp]

I can ruing proc bglimm without problem.

49   data MultiCenter;
50   input Center Group$ N SideEffect @@;
51   datalines;

NOTE: INPUT 语句到达一行的末尾时，SAS 转到新的一行。
NOTE: 数据集 WORK.MULTICENTER 有 30 个观测和 4 个变量。
NOTE: “DATA 语句”所用时间（总处理时间）:
      实际时间          0.01 秒
      CPU 时间          0.00 秒


67   ;
68
69   proc bglimm data=MultiCenter nmc=10000 thin=2 seed=976352
NOTE: 正在写入 HTML Body（主体）文件: sashtml.htm
70   plots=all;
71   class Center Group;
72   model SideEffect/N = Group / noint;
73   random int / subject = Center;
74   run;

NOTE: Generating the burn-in samples.
NOTE: Beginning sample generation.
NOTE: Beginning calculation of summary and diagnostics statistics.
NOTE: Generating diagnostic plots.
NOTE: “PROCEDURE BGLIMM”所用时间（总处理时间）:
      实际时间          2.59 秒
      CPU 时间          0.31 秒

What version of sas are you using? Mine is SAS9.4M7 .

78   %put &=sysvlong. ;
SYSVLONG=9.04.01M7P080520

If your sas version is too low to run the proc bglimm. you could try @StatDave  's suggestion GEE model by PROC GEE or PROC GENMOD .

[Ameurgen]

Hi;

Thank for the response, 

My version sas is : SYSVLONG=9.04.01M5P091317

I think i am far way back from the recent one, do have any idea how can i upgrade mine,  i think this is the reason why i cant run the proc bglimm

Thank you so much

[Ksharp]

You could freely use SAS by SAS OnDemand for Academic :

https://welcome.oda.sas.com/login