About RyanSimmons

RyanSimmons · ‎05-02-2016

Hardy quadrature appears to converge better than the default, but it's not perfect (i.e. some rather large gradients). However, at least it is able to give me sensible estimates when the offset term isn't in the model. Interestingly, I've noticed rather consistently that the variable for age seems to always have the largest gradient value, and that some of the models without age converge better. I'm not entirely sure why; the range for age in my dataset is not very large (~21-30 years) and doesn't appear to have any meaningful relationship with the outcome, but it's possible that something about the mixture of categorical and continuous covariates is giving the standard integration method some problems. It is especially interesting that it is only the models without the offset that have these problems (which is too bad, because for theoretical reasons the non-offset model is more appropriate for our analysis, given how we have defined the risk sets). I am guessing the problem lies somewhere in the difference in the outcome's distribution when it is offset (in which case it is essentially a proportion) versus as a straight-up count (which is wildly overdispersed, with values ranging from 0 to over 200, and that's AFTER screening out some of the worst outliers). Some combination of the covariate values must result in a small cell sizes with wildly disparate ranges that force the estimation procedure into local optima, though I have not been able to find this combination through standard exploration of the data. In any case, the Hardy method appears adequate (and a couple of covariates will end up being dropped for non-statistical reasons, simplifying the model).

RyanSimmons · ‎04-29-2016

I agree with you that it doesn't seem possible. That's why I posted here, because I legitimately cannot understand how or what happened with those files. I don't know what external changes could have been made to either my computer or the servers in question. None that have been clearly communicated to me, at least. It looks like I am just going to have to file a ticket with SAS Support and talk to our IT department.

RyanSimmons · ‎04-29-2016

The cascading effect was not harmless, however. It wasn't just the display that was different, the FILE EXTENSION WAS CHANGED. They were changed from .sas7bdat files to .sd2 files! And, as I said in my last post, the only way I could fix this was by deleting them and redownloading the raw data! It went far beyond just changing the icon called by Windows Explorer, it completely eliminated my ability to even ACCESS those files, period (even declaring V6 option in the LIBNAME statement). And it wouldn't let me change the file type association back; in fact, right-clicking these .sd2 files didn't even give me the "Open with" option anymore! As for your last sentence: "In the case of SAS datasets, the association with SAS 9.4 does not seem to make sense: It appears that SAS then tries to "execute" the datasets as if they were SAS programs (see the error messages you posted)." What do you mean by "the association with SAS 9.4 does not seem to make sense"? What part of the error message indicates it is trying to run them as a program? (I am just trying to understand the error log; I'm not an expert on this part of SAS)

RyanSimmons · ‎04-29-2016

Whatever happened with those SAS datasets I will probably never know. I remained unable to open them (even with explicitly using the v6 option in the LIBNAME statement). The only way I was able to get around this was by redownloading the raw (version 6) SAS datasets form the original source and reconvert them into v9 datasets. But this still offers me no clues as to how this whole corruption thing happened in the first place, or why it seemed to have a cascading effect on unrelated files in different directories.

RyanSimmons · ‎04-29-2016

I am aware of the issues with V9 vs V6 libraries. It was a major issue for me when I was first converting these files. However, as I said, I solved that issue MONTHS ago, and haven't had any issues with it in the ensuing time period. That directory has not been modified in any way since 3/11/2016 according to my logs.

RyanSimmons · ‎04-29-2016

Yes, I do. And no, all of these datasets were written with SAS 9.4. As I said, I ran this exact code on these exact datasets only two weeks ago and didn't encounter any of these problems. Interestingly, the ORIGINAL versions of these datasets (which were downloaded from dhsprogram.org) were SAS 6 files, and I had to convert them into SAS 9 files so I could use them. But I did that conversion MONTHS ago, and have used the SAS 9 versions of the datasets multiple times (literally hundreds, actually) over that span without any issues or errors or corruption until today. EDIT: And what the $%*&$*(%&*$ I just noticed that whatever is going on here HAS EFFECTED SAS DATASETS LOCATED ON A DIFFERENT SERVER! How is this even possible?!

RyanSimmons · ‎04-29-2016

My covariates are all binary variables, with the exception of age, which is continuous. So correlations aren't particularly meaningful. Regardless, I have not been able to find any strong relationships between any of these covariates, no matter how I try to break them down graphically or numerically.

RyanSimmons · ‎04-29-2016

I have a directory of files that I am trying to read into SAS. I am using a macro that has always worked in the past without error on this same directory. Now, inexplicably, it has stopped working properly. The relevant part of the code is, simply, this: libname infiles 'H:\DGHI\DHS\Data\'; DATA _indiv; set infiles.&dind; keep caseid bidx v024 v002 v005 v006 v007 v008 v009 v010 v011 v012 v025 v113 v116 v136 v137 v149 v190 v201 v501 bord b0 b1 b2 b3 b4 b5 b6 b7 b8 b11 b12 b13; run; Where "&dind" is the macro variable for the name of the input dataset. There are 14 pre-existing SAS datasets in the directory, all of which need to be read into SAS. 13 of them are working normally, and 1 of them has stopped working (inexplicably ... two weeks ago I ran this exact code on these files with no problems). When I run the above code on this file (which is called KEBR70FL), it gives me: ERROR: File INFILES.KEBR70FL.DATA does not exist. Further, if I run PROC DATASETS on INFILES, the file KEBR70FL does not show up! It only shows 13 of the 14 SAS datasets. However, if I navigate to the directory specified by INFILES in Windows Explorer, I can see that the KEBR70FL dataset DOES in fact, exist in that directory. Now, here is where things get even stranger. The 13 other datasets in that directory (the 13 that are recognized by SAS as existing) I can all right-click and open in either SAS 9.4 OR Sas Enterprise Guide 7.11, just as normal. However, the 14th dataset doesn't even have an option for opening it in SAS 9.4 when I right-click the file! It ONLY lets me open the dataset in Enterprise Guide (and it opens there succesfully, without error). So, I right-clicked this file and maneuevered to "Open with..." and selected SAS 9.4 in the window that pops up when you use that option. However, it doesn't open the file at all. Nothing happens in SAS 9.4, but when I look back in Windows Explorer after having done this, it has now changed the file type from a SAS dataset to an SD2 file, and output a text document full of error messages (I've copied those at the below). Even STRANGER is the fact that doing this changed a COMPLETELY DIFFERENT DATASET to an SD2 file as well! One of the 13 files that had previously been working is now suddenly changed to an SD2 file, simply by trying to open this other file in SAS 9.4 (despite it working in Enterprise Guide, as I mentioned)! Not only that, but it changed THE BACKUP COPIES OF THESE DATASETS THAT WERE LOCATED IN A DIFFERENT DIRECTORY! After the file got changed to SD2, I went to the backup directory to retrieve the original SAS dataset version and somehow it had ALSO been changed to an SD2, despite not being connected to any of the above operations in anyway that I can tell. EDIT: Even worse, the backup file that was overwritten into an SD2 was in a compressed (zip) folder in a different directory, and it was still somehow altered by just trying to open this one file in SAS. It also altered a completely different backup file in a different compressed folder in a different directory. Does anybody have the slightest idea what is going on, here? I don't understand at all how this dataset got corrupted so that SAS 9.4 wouldn't recognize it anymore (yet it was still recognized by Enterprise Guide), and I CERTAINLY don't understand how any of the above was able to independently corrupt other files located in completely different directories... (Here are the error messages in the text file that is created along with the SD2 file: NOTE: Unable to open SASUSER.REGSTRY. WORK.REGSTRY will be opened instead. NOTE: All registry changes will be lost at the end of the session. WARNING: Unable to copy SASUSER registry to WORK registry. Because of this, you will not see registry customizations during this session. NOTE: Unable to open SASUSER.PROFILE. WORK.PROFILE will be opened instead. NOTE: All profile changes will be lost at the end of the session. NOTE: Copyright (c) 2002-2012 by SAS Institute Inc., Cary, NC, USA. NOTE: SAS (r) Proprietary Software 9.4 (TS1M3) Licensed to DUKE UNIVERSITY - T&R - SFA - NCICU GRANT, Site 70082794. NOTE: This session is executing on the X64_7PRO platform. NOTE: Updated analytical products: SAS/STAT 14.1 SAS/ETS 14.1 SAS/OR 14.1 SAS/IML 14.1 SAS/QC 14.1 NOTE: Additional host information: X64_7PRO WIN 6.1.7601 Service Pack 1 Workstation NOTE: SAS initialization used: real time 0.21 seconds cpu time 0.18 seconds WARNING: A line of input contains one or more null (hexadecimal 00) characters, which might cause the SAS System to ignore part of the line. The line, with null characters printed as question marks, is: SAS 6.08.04 WIN 3.10 STATTRAN 1 SAS 6.08.04 WIN 3.10 STATTRAN ÀL; ___ 180 ERROR 180-322: Statement is not valid or it is used out of proper order. 1 ÑA•c‘T2Æ __ 180 1 ! #8Û 8 ¬Q ÀL; ERROR 180-322: Statement is not valid or it is used out of proper order. 1 8Û 8 ¬Q ÀL; _ 180 2 The SAS System 11:59 Friday, April 29, 2016 ERROR 180-322: Statement is not valid or it is used out of proper order. 1 ÑAwm%w©\·!êh ____ 180 1 ! s `7 è6 x x6 1 ! p 6 p ˜5 p (5 p ¸4 p H4 p Ø3 p h3 p 1 ! ø2 p ˆ2 p 2 p ¨1 p 81 p È0 p X0 p 1 ! è/ p x/ p / p ˜. p (. p ¸- p H- p Ø 1 ! , p h, p ø+ p ˆ+ p + p ¨* p 8* p È) 1 ! p X) p è( p x( p ( p ˜' p (' p ¸& p 1 ! H& p Ø% p h% p ø$ p ˆ$ p $ p ¨# p 1 ! 8# p È" p X" p è! p x! p ! p ˜ p ( 1 ! p ¸ p H p Ø p h p ø p ˆ p 1 ! p ¨ p 8 p È p X p è ERROR 180-322: Statement is not valid or it is used out of proper order. ERROR: Errors printed on pages 1,2. NOTE: SAS Institute Inc., SAS Campus Drive, Cary, NC USA 27513-2414 NOTE: The SAS System used: real time 0.31 seconds cpu time 0.18 seconds )

RyanSimmons · ‎04-14-2016

I am running a mixed-effects zero-inflated Poisson model using PROC NLMIXED. The following code fits my model without any problems: PROC NLMIXED data=_model method=gauss; parms a0=-1.1433 b0=-0.8371 b1=0.07413 b2=0.08834 b3=-0.2387 b4=0 b5=0 b6=0 b7=0 b8=0 sigma=0.3030 ; zeromodel = a0; phi = 1/(1+exp(-zeromodel)); mu = exp(b0 + subj + b1*Time + b2*Intervention + b3*Interaction + b4*phone + b5*positive + b6*education + b7*income + b8*age + logTI); if UAVI=0 then logLik = log(phi + (1-phi)*exp(-mu)); else logLik = log(1-phi) + UAVI*log(mu) - mu - lgamma(UAVI+1); model UAVI ~ general(logLik); random subj ~ normal(0,sigma) subject=Study_ID; run; This gives me reasonable parameter estimates and no convergence errors. I wanted to compare this model to the same model run without the offset term (logTI). However, as soon as I remove the offset term, I can no longer get the model to converge. If I run the above code AS IS (removing only the "+ logTI") I get the following error: ERROR: Quadrature accuracy of 0.000100 could not be achieved with 31 points. The achieved accuracy was 1.000000. I figured the error had to do with the starting values for the parameters. So I removed the random intercept term (subj) and ran the model without offset using only fixed-effect terms; this converges, but gives me large gradient values, as well as a warning about the final Hessian matrix not being positive definite. I have tried several dozen different combinations of starting values, but it always gives me the same end result. I have also tried changing the minimization technique, the update method, the linesearch procedure, and the linesearch precision, but all to no avail (regardless of what I specify, it always seems to converge to some sort of local extremum with unreasonable values for the parameters and large gradients). This only seems to happen when the offset term isn't in the model; with the offset term things seem to work better (though, some of the gradients are still relatively big). My sample size is ~400, and each of the covariates is binary, except for the continuous "age". Examining the data itself, I don't see any problems that would throw off the model (and the data is sorted by the subject variable). Does anybody have any ideas for anything I can try, or some other options I can run that might help? EDIT: If I simplify the model down a bit, and run the following, it converges just fine. As soon as I start adding other covariates in, however, I start getting issues with the gradient or the Hessian. I can't figure out exactly why this is going on. My sample size isn't huge, but it isn't tiny and should be able to support more than three binary indicator variables ... PROC NLMIXED data=_model method=gauss; parms a0=-0.5732 b0=0.9552 b1=0.1298 b2=0.2687 b3=-0.6070 sigma=2.1479 ; zeromodel = a0; phi = 1/(1+exp(-zeromodel)); mu = exp(b0 + subj + b1*Time + b2*Intervention + b3*Interaction); if UAVI=0 then logLik = log(phi + (1-phi)*exp(-mu)); else logLik = log(1-phi) + UAVI*log(mu) - mu - lgamma(UAVI+1); model UAVI ~ general(logLik); random subj ~ normal(0,sigma) subject=Study_ID; run;

RyanSimmons · ‎03-29-2016

Aha! I found the mistake. I mentally mixed up the log-likelihood expressions for the ZIP and ZINB models. The code that reads as: if UAVI=0 then logLik = log(p0 + (1-p0)*exp(p**theta)); Should read as: if UAVI=0 then logLik = log(p0 + (1-p0)*(p**theta)); There is no exponentiation in this part of the log-likelihood (in contrast with the ZIP, which would have exp(-mu) in this term. It's always the silly mistakes that undermine you!

RyanSimmons · ‎03-29-2016

Incidentally, I also modified the last line of code in NLMIXED before the model statement to read: logLik = log(1-p0)+logpdf('NEGBINOMIAL',UAVI,p,theta) This gives me the same parameter estimates as the code using lgamma() in place of the log(gamma()) calls. So I am confident that my log-likelihood is correctly specified. However, I don't understand why NLMIXED and GENMOD are giving different parameter estimates.

RyanSimmons · ‎03-29-2016

I didn't even realize lgamma() was a function in SAS! Always full of surprises ... In any case, that removed the error I was experiencing. However, the parameter estimates are still different from those produced by the GENMOD call. The intercept for the zero-inflated model in particular is way off, with a value of -1.1399 given by GENMOD and a value of -18.9018 given by NLMIXED. Any idea what would account for this discrepency (incidentally, all the other parameter values are also different between GENMOD and NLMIXED, but not by much. The zero-inflation parameter is the only one that is radically off.)

RyanSimmons · ‎03-28-2016

I fit a zero-inflated negative binomial model in PROC GENMOD, using the following code (both Time and Treatment are binary indicator variables): PROC GENMOD data=analysis_dataset; class Time(ref="Baseline") Treatment(ref="Control") / param=ref; model UAVI = Time|Treatment / dist=zinb offset=logTI; zeromodel; run; This model converges and produces no errors. If I try to fit the same model in PROC NLMIXED (note that "Interaction" here is an indicator variable I created in a datastep to signifiy the Time X Treatment interaction, since NLMIXED can't create interaction terms on its own): PROC NLMIXED data=analysis_dataset method=gauss qpoints=25; /* Set initial values for parameters */ parms a0=-1.1365 alpha=1 b0=-0.8222 b1=0.06963 b2=0.09920 b3=-0.2334 sigma=0.3131 ; /* Specificy zero model */ zeromodel = a0; p0 = 1/(1+exp(-zeromodel)); /* Specify Negative Binomial model */ mu = exp(b0 + b1*Time + b2*Treatment+ b3*Interaction + logTI); theta = 1/alpha; p = 1/(1+alpha*mu); /* Now build ZINB likelihood */ if UAVI=0 then logLik = log(p0 + (1-p0)*exp(p**theta)); else logLik = log(1-p0) + log(gamma(theta + UAVI)) - log(gamma(UAVI + 1)) - log(gamma(theta)) + theta*log(p) + UAVI*log(1-p); model UAVI ~ general(logLik); run; So far as I can tell, this is the correct specification of the log-likelihood. Now, this NLMIXED code produces errors. Specifically, it gives me 6 of the following errors: NOTE: Invalid argument to function GAMMA(#) at line # column # I used the NLMIXED code in a DATA step to identify the 6 problematic observations. It turns out that these six observations are all individuals whose counts of the outcome (UAVI) are all exceptionally high (>200). After fiddling around with some datasteps, I've determined that the GAMMA function will return this error if given as an argument any number greater than 171. Since, for positive integers, Gamma is calculated as (Integer-1)!, I'm guessing the error is because the numbers are too big. That is, Gamma(171)=7.257416e306! Now, likely, these 6 observations will end up being excluded from my analysis because they are huge outliers, and I have some rather convincing evidence that the reasons they are outliers have more to do with data quality than any intrinsic effect of the treatment. But, the question I have is; how is PROC GENMOD able to calculating values of the log-likelihood for these observations, if decomposing the likelihood function in PROC NLMIXED returns an error? Clearly, something about PROC GENMOD's method of calculating the likelihood is different, because it is able to fit the model to those 6 datapoints with an error. This is troublesome for me because ultimately I want to include random intercepts in my model, thus why I am fooling around with PROC NLMIXED instead of PROC GENMOD. However, it concerns me that in this situation GENMOD and NLMIXED are treating the data so differently. Unless I have made a major mistake in my PROC NLMIXED code above, what accounts for GENMOD's ability to fit those datapoints?

RyanSimmons · ‎01-29-2016

Thank you everyone for your help! All the answers are great. Wish I could accept them all!

RyanSimmons · ‎01-28-2016

Have you done diagnostics on the degree of zero-inflation to verify whether or not you need zero-inflated model to begin with? For example, in many cases a negative binomial model (which can be easily fit in GENMOD with a repeated statement) will fit zero-inflated data quite well (in fact, anecdotally speaking, in my experience a ZIP or ZINB model rarely offers any practical advantages over a negative binomial model in the case where you have zero-inflation and overdispersion).

Online Status	Offline
Date Last Visited	‎09-04-2020 06:50 PM

Re: Identifying if a set of variables meets a condition

Identifying if a set of variables meets a condition

Re: Adding TEXT to SGPLOT changes x-axis

Adding TEXT to SGPLOT changes x-axis

Re: Import SAS dataset from .zip file

Re: Import SAS dataset from .zip file

Import SAS dataset from .zip file

Re: Correctly calculating hours between two times

Re: Correctly calculating hours between two times

Correctly calculating hours between two times

Re: Adding TEXT to SGPLOT changes x-axis

Re: Adding TEXT to SGPLOT changes x-axis

Re: Adding TEXT to SGPLOT changes x-axis

Re: Correctly calculating hours between two times

Re: Correctly calculating hours between two times

Re: Annotating an SGPANEL plot; "xspace invalid" error

Re: Saving PROC GPLOT graph to folder

Adding TEXT to SGPLOT changes x-axis

Re: Adding TEXT to SGPLOT changes x-axis

Re: Correctly calculating hours between two times

Re: NLMIXED optimization problems

Re: File (NAME) does not exist

Re: File (NAME) does not exist

Re: File (NAME) does not exist

Re: File (NAME) does not exist

Re: File (NAME) does not exist

Re: NLMIXED optimization problems

File (NAME) does not exist

NLMIXED optimization problems

Re: ZINB in GENMOD vs NLMIXED

Re: ZINB in GENMOD vs NLMIXED

Re: ZINB in GENMOD vs NLMIXED

ZINB in GENMOD vs NLMIXED

Re: Count number of changes in variables over time

Re: Zero-inflated and correlated count data