gsem: Same code, no longer working

[Katelin Hudak]

Hello,

I am using gsem in Stata 15 for an LCA, and although I have been using Stata for some time, I am new to LCAs and sem/gsem. I would like to identify the underlying latent class variable (categorical) using 7 observed binary variables: pthink_par_qual_bin cthink_par_qual_bin know_datr felt_nlove talk_diff resp_hpe wor_depr using weighted survey data. I am working on identifying the number of classes that would be most appropriate for the data and from a theoretical standpoint, but I have 2 issues:

1. I am unable to use the command

 estat lcgof

I receive an error message that the command is unrecognized. I'm sorry I cannot provide additional details on the precise error code and message because . . .

2. I am no longer able to use gsem. I was running the following model earlier in the day:

svy: gsem (pthink_par_qual_bin cthink_par_qual_bin know_datr felt_nlove talk_diff resp_hpe wor_depr <-), logit lclass(C 2)

It ran beautifully, and I was able to compute some postestimation stats, e.g., estat lcmean, estat lcprob, etc. with no issues (but not estat lcgof). However, a few hours later I ran the same code and received the error message

����DB�ж�V���_�C�P��W�H� invalid name
invalid syntax
invalid syntax
an error occurred when svy executed gsem
r(198);

I tried shutting Stata and the computer down, restarting, installing available updates, but nothing has worked. It is the exact same code and exact same data, so I do not understand why it will not run. (Also, I tried running it without using the survey prefix, but this did not help at all.)

Does anyone have any ideas what may be going on and how I can fix it?

Thank you in advance!

[Andrew Musau]

You should start by running examples in the documentation to determine whether it is a problem with Stata or it is an issue with your code. Sometimes invisible characters in the code can lead to "invalid name" errors, and this can arise, e.g., from copying and pasting the code from a third party editor.

help gsem

[Katelin Hudak]

Thank you, Andrew, that was a great idea! I ran the example 50g in Stata's sem manual and it worked exactly as it should have. This is good news, as now I know that it is not an issue with Stata (or my specific version). However, I still do not understand why my code will not run, especially because I ran the exact same code earlier that day. I have gone over my code many times, e.g., checking spaces, as well as checking my variables to ensure there is nothing wrong there.

Any other advice would be greatly appreciated!!

[Richard Williams]

It is a relatively short command, so I would try retyping it from scratch. Like Andrew, i suspect your code got corrupted somehow and may have invisible characters you can’t see.

[Katelin Hudak]

Thank you for your advice, Richard. Following your suggestion, I have tried retyping it several times and in different ways (e.g., without the survey command, in the command window, in the do-file, with and without the " / / / " line connectors), and still no luck. I added (from the variable list window using the arrow) my own variables into the command from 50g in the sem manual (the example data and code in 50g worked just fine), and nothing has worked. I have no idea what else to try.

Thank you for your time and advice!

[Richard Williams]

Weird. Do you have another machine you can try? Also, are you sure your data set did not somehow get corrupted? My guess is that something got corrupted somewhere.

[Katelin Hudak]

Another good idea. I tried the virtual Stata on my company's VPN but with no more luck. Also, I should note that I checked all pertinent variables closely, including looking at their distribution, total n, and total missing and verified that it matched my original data that I was able to use previously and successfully with the command specified above. I will continue digging into the data and the issue.

Thank you again for your time and your advice!

[Hua Peng (StataCorp)]

log using gsem.log, text replace
set trace on
//!! run your commands here
log close
set trace off

post the gsem.log or the section before the error, make sure you delete any sensible information before post.

[Andrew Musau]

The one thing that I noticed after posting is that you have very long variable names, which in some contexts can cause problems. Try

preserve
rename (pthink_par_qual_bin cthink_par_qual_bin know_datr felt_nlove talk_diff resp_hpe wor_depr) (var#), addnumber(1)
svy: gsem (var1 var2 var3 var4 var5 var6 var7 <-), logit lclass(C 2)
restore

If this still results in the error, follow Hua Peng (StataCorp)'s advice and post the log.

[Katelin Hudak]

Thank you, Hua and Andrew, for your advice! I renamed the variables var1-7, but that did not work, indicating that it is something besides the variable names that is the issue.

Then I set the trace on but did not learn anything. I tried it without the survey command and got a little farther. The log is below (I was unable to upload it):

. set trace on

. svy: gsem (var1 var2 var3 var4 var5 var6 var7 <- ), logit lclass(C 2)
(running gsem on estimation sample)
¬ÉßØDB*жâVŠ‘»_ìC¥PøèWÛH invalid name
invalid syntax
invalid syntax
an error occurred when svy executed gsem
r(198);

end of do-file

r(198);

. do "C:\Users\HUDAK-~1\AppData\Local\Temp\STD1758_000000.tmp"

. *try without svy
. gsem (var1 var2 var3 var4 var5 var6 var7 <- ), logit lclass(C 2)
¬ÉßØDB*жâVŠ‘»_ìC¥PøèWÛH invalid name
invalid syntax

Fitting class model:

Iteration 0:   (class) log likelihood = -885.14856  
Iteration 1:   (class) log likelihood = -885.14856  

Fitting outcome model:

Iteration 0:   (outcome) log likelihood = -3853.6137  
Iteration 1:   (outcome) log likelihood = -3777.1893  
Iteration 2:   (outcome) log likelihood = -3772.8617  
Iteration 3:   (outcome) log likelihood = -3772.7736  
Iteration 4:   (outcome) log likelihood = -3772.7733  

Refining starting values:

Iteration 0:   (EM) log likelihood =  -4637.765
Iteration 1:   (EM) log likelihood = -4631.9045
Iteration 2:   (EM) log likelihood = -4629.7688
Iteration 3:   (EM) log likelihood = -4628.8446
Iteration 4:   (EM) log likelihood = -4628.4517
Iteration 5:   (EM) log likelihood = -4628.3138
Iteration 6:   (EM) log likelihood = -4628.2989
Iteration 7:   (EM) log likelihood = -4628.3407
Iteration 8:   (EM) log likelihood = -4628.4052
Iteration 9:   (EM) log likelihood =  -4628.475
Iteration 10:  (EM) log likelihood =  -4628.542
Iteration 11:  (EM) log likelihood = -4628.6023
Iteration 12:  (EM) log likelihood = -4628.6549
Iteration 13:  (EM) log likelihood = -4628.6996
Iteration 14:  (EM) log likelihood = -4628.7372
Iteration 15:  (EM) log likelihood = -4628.7688
Iteration 16:  (EM) log likelihood = -4628.7945
Iteration 17:  (EM) log likelihood = -4628.8157
Iteration 18:  (EM) log likelihood =  -4628.833
Iteration 19:  (EM) log likelihood = -4628.8471
Iteration 20:  (EM) log likelihood = -4628.8586
Note: EM algorithm reached maximum iterations.
invalid syntax
r(198);

end of do-file

r(198);

I would love to know what may be going on and how to move forward, especially because I know:
1) gsem works on my version of Stata
2) I was able to run the exact same code on the same data last week
3) the data does not seem to have changed because all my summary stats are the same

Thank you in advance!!

[Hua Peng (StataCorp)]

Unable to see your data, I think we need to try a more systematic approach. First let's try to eliminate other possibilities by:

1. fully update your Stata 15, type

update all

in Stata command window and follow the instruction

2. After fully update, download Stata installation qualification tool from https://www.stata.com/support/instal...lification/15/
install and run it to check if the Stata installation on you machine is good.

3. If problem persists, report back and we can move forward from there.

[Hua Peng (StataCorp)]

Also, can you show the output of

desc var1 var2 var3 var4 var5 var6 var7

[Katelin Hudak]

Update: I solved the issue of the gsem command not running. The dataset I was working with was too large. Once I cut it to the bare necessities, the gsem command ran without any issues.

However, I am still unable to use the

estat lcgof

to compare models to determine the optimal number of classes. The code I am using is:

svy: gsem (pthink_par_qual_bin cthink_par_qual_bin know_datr felt_nlove talk_diff ///
resp_hpe wor_depr <-), logit lclass(C 2)

estat lcgof

where all of my observed variables are binary indicators. The error message I get is

 estat lcgof
invalid subcommand lcgof
r(321);

end of do-file

r(321);

I have tried it in both Stata 15, which I have on my machine, and in Stata 16, using a VPN. However, the results and the error message are the same.

Thank you for your time and assistance!!

[Katelin Hudak]

I should note that I've also tried this with the estimates store command

svy: gsem (pthink_par_qual_bin cthink_par_qual_bin know_datr felt_nlove talk_diff ///
resp_hpe wor_depr <-), logit lclass(C 1)

estimates store oneclass

estat lcgof

with the same results and error message.

[Andrew Musau]

help svy_estat

does not indicate that estat lcgof is supported after svy: gsem. I do not know why this is the case, but what is allowed is

estat gof reports a goodness-of-fit test for binary response models using survey data. This command is available only after svy: logistic, svy: logit, and svy:
probit; see [R] logistic, [R] logit, and [R] probit.

Have you ever successfully gotten results previously? The following replicates your error:

webuse nhanes2f
svyset psuid [pweight=finalwgt], strata(stratid)
svy: gsem (highlead zinc age age2 weight <- ), logit lclass(C 1)
estat lcgof

Res.:

. estat lcgof
invalid subcommand lcgof
r(321);