About Garnett

Garnett · ‎05-25-2023

This solution from Bill at SAS Tech Support fixed the issue: --------------------------------------------------------------------------------------------------------- For this one program, add PROC PRINTTO code to write the log to disk to see if that avoids the issue for this particular case. Example code: proc printto log=" within quotes specify the directory and filename to write the log to disk on the SAS server"; run; Your proc iml code goes next...

Garnett · ‎05-22-2023

I get this message while running SAS EG. No subsequent action is possible until SAS restarts. Other program files work as expected.

Garnett · ‎08-26-2020

Greetings! My goal is a panel plot with distinct y-axis scaling for each panel. -I can do this with LAYOUT LATTICE and a macro to create each panel, but this seems inefficient. -I can create a LAYOUT DATAPANEL with one column and several rows, then COLUMNDATARANGE=, but this is unwieldy for the viewer with more than three panels. Can I create this plot in conjunction with LAYOUT DATAPANEL? Is there a non-GTL solution? Thanks in advance!

Garnett · ‎11-29-2019

Greetings! I get the following message when calling PROC MCMC from PROC IML: "YHQSRC/XZPWRIT failure" SAS usage note 7275 describes the error message and states that To solve the problem, copy the data to the local hard drive. I'm not sure how to do this when calling PROC MCMC from PROC IML. Any ideas?

Garnett · ‎07-26-2019

Hi, I've lost track of the problem in this thread, but here a couple of points: -The rate difference can be negative, of course, but your original code had the rate difference as the binomial mean parameter. By definition, the binomial parameter has to be between 0 and 1. -Given data, the posterior on beta0 or beta1 will be different from the prior unless the prior has zero variance. I'm not sure why it is important to show that the posterior and the prior are the same.... -If the "Sex_Bin" variable is (0,1), then the rate difference is given by Rate_Difference=logistic(beta0 + beta1*1) - logistic(beta0+beta1*0); I don't know if this is what you are looking for....

Garnett · ‎07-23-2019

I don't think you want "Rate_difference" as the binomial parameter. In theory, it could be negative. Instead, replace the model code (from "Parms" to "Model") with parms (beta0 beta1); prior beta0 ~ normal(0.10, var=0.025); prior beta1 ~ normal(0.80, var=0.025); *Rate_Difference = logistic(beta0 + beta1*age_bin) - logistic(beta0 - beta1*age_bin); prob=logistic(beta0 + beta1*age_bin); model Y ~ binomial(1,prob); /*n=number of parameters from input dataset, (e.g., 1)*/ Rate_Difference=logistic(beta0 + beta1*1) - logistic(beta0 + beta1*-1); This assumes that "Age_bin" is a binary indicator taking values of 1 and -1.

Garnett · ‎05-17-2019

A HA!!! I'm working from the documentation for 14.1, which defines the interior knot placement in terms of k. My guess is that this is a typo in the 14.1 documentation. Thanks!

Garnett · ‎05-17-2019

Greetings! Has anyone noticed a possible typo in the documentation for the BSPLINE function in SAS/IML? Under the argument definitions, "i is an optional argument that specifies the number of interior knots when n=1 and k contains a missing value." Then goes on to to define the interior knot placement as a function of k, which is a missing value. What am I missing, here? Thanks in advance! Garnett

Garnett · ‎01-15-2019

Hi, An easy way is to assume the posterior on each effect is Gaussian: 1) Monitor the posterior distributions of the random effects. 2) Write ODS output postsumint=Px; 3) In a subsequent datastep, sample from the mean and Standard deviation given in Px (or just use the posterior info in Px.) Let me know if this isn't what you are thinking... G

Garnett · ‎01-02-2019

Hello, I missed the datastep specification of the priors in GENMOD. sorry about that. In any event, the prior information about the event probabilities are different because the link functions are different. As long as you use the logistic link in PROC MCMC and log link in GENMOD, you can expect different inferences about the regression coefficients. I wouldn't use a log link for probabilities. If you want to stick with it, just replace the "logistic" function in MCMC with "exp". You'll get errors as this link function allows binomial probabilities outside (0,1). Personally, I wouldn't bother with that approach, and just stick with the logit link. Then you can compute the rate difference within the MCMC iterations. Garnett

Garnett · ‎12-28-2018

Hi there, You are using a log link function in the GENMOD syntax and a logistic link function in the MCMC syntax. It isn't clear what priors you are using in GENMOD, but b/c the link functions are different, priors on the parameters can't be identical if you want to get the same result. One HUGE benefit of MCMC in Bayesian analysis is that it's trivial to make inferences about any function of the parameter. If you want the posterior on the difference in probabilities between "teaching_service" = 1 and "teaching_service" = -1, then add the following to your MCMC syntax: Rate_Difference = logistic(beta0 + beta1) - logistic(beta0 - beta1); Also, somewhere in the MCMC statement add "monitor=(_PARMS_ Rate_Difference)" and that'll get you what you want. Good luck! Garnett

Garnett · ‎10-05-2018

I assume we are talking about PROC MCMC: "Is there any limit on the maximum number of subjects?" -Not that I'm aware of. I've run models with 1000s of subjects. "How do the subject effects affect the convergence?" -It doesn't, but a lot depends on the choice of priors and the amount of data you have on the individual subjects. "Moreover, is there a limit on the number of fixed effects and random effects on the model building process?" -Not that I've ever encountered. "The computer is taking too long to finish, most of the cases, it does not finish and ends up with error messages." -My guess is that the model is specified incorrectly for the data structure. What do you mean "it does not finish"? What error messages are you getting? You must, must, must test a much reduced version of the model for stability!

Garnett · ‎09-28-2018

Hi there! Thanks for your reply. The challenge is italicizing theta without italicizing "P(" and ")". Apparently it cannot be done. I received a reply from tech support, pasted below: --------------------------------------------------------------------------------------------------------------------------- Garnett, There is not an option to italicize only a portion of an axis label. The text style applies to the entire label. You can use a DRAWTEXT statement to place the text on the graph to simulate the label. Below is a simple sample that will suppress the default Y axis label and use the DRAWTEXT statement to place text to the left of the graph. The PAD= option on the BEGINGRAPH statement adds space to the left of the graph where we want the text to be placed. You might need to adjust the PAD= value, depending on how the text looks on your graph. proc template; define statgraph plot; begingraph / pad=( left=5%); layout overlay / yaxisopts=(display=(ticks tickvalues line)); scatterplot x=age y=height; drawtext textattrs=(style=normal) "P" textattrs=(style=italic) "(*ESC*){unicode theta}" / xspace=graphpercent yspace=wallpercent x=2 y=50 rotate=90 anchor=center width=50; endlayout; endgraph; end; run; proc sgrender data=sashelp.class template=plot; run;

Garnett · ‎09-28-2018

"How can I specify the same model like proc glimmix with seven random slopes and a random intercept for the Subject = region and the four random slopes and a random intercept for the subject = V11" You need to have a RANDOM statement for each component, i.e. Random V01~Normal(0, var=VarV01) subject=Region; Random V02~Normal(0, var=VarV02) subject=Region; etc. Same idea for subject=V11. Note the different variances for each "random effect". However, you can specify any kind of covariance structure you like.

Garnett · ‎09-27-2018

Hi there, The answer is yes, if I understand the data structure correctly. It's easy enough to test the model in GLIMMIX and MCMC without covariates! Garnett

Online Status	Offline
Date Last Visited	‎04-19-2024 03:45 PM

Re: How to clean log is SAS EG rather than close and open the program ...

Re: How to clean log is SAS EG rather than close and open the program ...

LAYOUT DATAPANEL and distinct y-axis scales

YHQSRC/XZPWRIT failure when calling PROC MCMC from IML

Re: Convert PROC GENMOD BAYES model to PROC MCMC

Re: Convert PROC GENMOD BAYES model to PROC MCMC

Re: documentation of the BSPLINE function

documentation of the BSPLINE function

Re: How to output 200K parameters in PROC MCMC?

Re: Convert PROC GENMOD BAYES model to PROC MCMC

Re: Convert PROC GENMOD BAYES model to PROC MCMC

Re: Convert PROC GENMOD BAYES model to PROC MCMC

Re: How to clean log is SAS EG rather than close and open the program ...

Re: How to clean log is SAS EG rather than close and open the program ...

LAYOUT DATAPANEL and distinct y-axis scales

YHQSRC/XZPWRIT failure when calling PROC MCMC from IML

Re: Convert PROC GENMOD BAYES model to PROC MCMC

Re: Convert PROC GENMOD BAYES model to PROC MCMC

Re: documentation of the BSPLINE function

documentation of the BSPLINE function

Re: How to output 200K parameters in PROC MCMC?

Re: Convert PROC GENMOD BAYES model to PROC MCMC

Re: Convert PROC GENMOD BAYES model to PROC MCMC

Re: How to fit Mixed Effect Model in Proc MCMC?

Re: How to create a formatted unicode character within a GTL label

Re: How to fit Mixed Effect Model in Proc MCMC?

Re: How to fit Mixed Effect Model in Proc MCMC?