SAS Support Communities

Thanks.  Even better: I bought YOUR book 🙂 ... View more

You are too kind.  Much obliged!  Thanks!    It's a challenge to transition from decades of just columns, rows and cells to vectors and arrays. ... View more

And I am a total IML idiot.  Why is this code not working? proc iml; use SAShelp.class; read all var {Age}; PP = Age/100; Eta = log(PP / (1-PP)); start Integrand(u) global(eta); L = logistic(u + eta); fu = pdf("Normal", u, 0, 0.2); return( L * fu ); finish; call quad(iPP, "Integrand", {.M,.P}) scale=.0001; print iPP; ... View more

Oh, and one more thing, I think the notation in the formula is what's throwing you off.  "RE" subscripts for alpha and beta signify that it's a Random Effects model rather than that alpha and beta are random effects coefficients.  Alpha and Beta are fixed.  U is the random intercept.   ... View more

Thanks so much, Rick.   Extra background detail.  There is a problem with marginal observation-level Predicted Probabilities from logistic regression when random effects are involved.  Straight up linear predictor expressed as probability is systematically different from the 'correct' value.  See this for more detail (https://bmcmedresmethodol.biomedcentral.com/articles/10.1186/s12874-015-0046-6).   The correct way to calculated marginal probabilities according to the reference and the formula I pasted, is to integrate Predicted Probability for each row over the distribution of all random intervals in the model.  I wish SAS had this correction built in into all PROCs that do multi-level regression.  I requested that feature.  In the absence of it, my starting point is a marginal NOBLUP Predicted Probability value for each row in the model from GLIMMIX, for example.  From there, I back-transform it to linear predictor so that the LP could be integrated over the distribution of random effects or Random Intervals (the u term).  For each row in the model (patients), the Predicted Probability needs to be corrected using the distribution of random intercepts (hospitals). Normal integral QUAD() gives me AUC for the transformation without any density weights.  What you shared seems to do the trick.  I tested a couple of values and it seems to correctly increase the input for values below 0.5 and decreases them for values above .5.  I modified the code slightly and now just need to figure out how to apply it to an entire column of Predicted Probability values.  I think I can simplify this a little bit by outputing Linear Predictors from the model rather than Predicted Probabilities -- skip the `iLink` option. proc IML; PredProb = .73; /* Marginal Predicted Probability from RE GLMM */ Eta = log(PP / (1-PP)); /* Eta = linear predictor */ start Integrand(u) global(eta); L = logistic(u + eta); fu = pdf("Normal", u, 0, 0.2); /* specify f(u) here. eg, PDF for N(0,0.20) */ return( L * fu ); finish; call quad(iPP, "Integrand", {.M,.P}); print PP Eta iPP; quit; Thanks again.  Happy to discuss further.  Hope you can put in a good word with SAS development teams to add this correction to the SAS/STAT core PROCS.   Would love to hear if you disagree with the diagnosis for marginal PPs. ... View more

Thank you Steve. I was not looking to deliberate between PL and Quadrature approaches but this is certainly a relevant topic. Thanks for sharing! ... View more

Perfect! Thanks. Found the same solution. Any idea why one needs to request partitioning to get these basic model fit statistics to output? Does not make sense to me that they won't show up without typing extra code for the PARTITION option and specifying this contrived (TEST=0) scenario to, basically, negate the PARTITION request. ... View more

The equivalent in the PROC LOGISTIC is SCORE FITSTAT option. That does not work in the HP version. ... View more

Thanks for the recommendation regardgin WORK.  I am looking at Windows Resource Monitor and don't see anything of note on CPU, RAM, or Disk I/O.  Maybe up to 10-15% utilization.  Anything else you can recommend? ... View more

Sorry, should have been more specific: I am running SAS on Windows 11 desktop. The ten programms running GLIMMIX models are submitted by right-clicking and "Batch Submit with SAS 9.4". The code of GLIMMIX is probably not relevant but here it is just in case: proc glimmix data=BootData Method=Quad; where SampleID LE 50; by SampleID; freq NumberHits; class Stratum; * pre-sorted ; class ageGrp / reference = last ; model OpMort (event="1 = Yes") = &ModelVars / dist=binary link=logit cl; random intercept / subject= Stratum; output out=out.B0 pred(iLink BLUP)=PredProb_boot pred( noBLUP)=PredProb_boot_noBLUP_xBeta ; run; ... View more

I want a single column PROC in the SAS dataset that contains value PROC = "[ 2800, 2980, 4090, 4170 ]". ... View more