Computer for Mplus4 (Bootstrapped LRT... PreviousNext
Mplus Discussion > Latent Variable Mixture Modeling >
Message/Author
 Michael J. Zyphur posted on Sunday, May 21, 2006 - 2:00 pm
Hi Bengt/Linda,
I was wondering if you had any special recomendations regarding the fastest computer to use for Mplus4? I am going to write a grant for an expensive system soon; if money was no option, is there anything special out there for time-intensive processes (e.g., Bootstrapping the LRT for class enumeration, numerical integration with complex models with random slopes, etc.)? Should a dual-processor computer be purchased in case Mplus gets compiled to take advantage of multi-threading, etc?

Sorry for the somewhat off-topic post, but bootstrapping the LRT is taking a lot of time with complex models.

Thanks!
mike
 Linda K. Muthen posted on Sunday, May 21, 2006 - 3:24 pm
If you plan on keeping the computer for a long time, I would get a dual process computer with the fastest processor and a lot of RAM.

Have you downloaded Version 4.1. There have been speed improvements related to TECH14 and new options to use with TECH14. Following are recommendations for using TECH14 which can be found in the users' guide on the website.

1. Run without TECH14 using the STARTS option of the ANALYSIS command to find a stable solution if the default starts are not sufficient.
2. Run with TECH14 using the OPTSEED option of the ANALYSIS command to specify the seed of the stable solution from Step 1.
3. Run with LRTSTARTS = 0 0 40 10; to check if the results are sensitive to the number of random starts for the k class model.
 Daniel Rodriguez posted on Monday, June 19, 2006 - 11:52 am
I have a question about step 2. You say to run Tech14 using the OPTSEED option "...to specify the seed of the stable solution from step 1." Which is the seed from the stable solution? Is it the seed associated with the best -2log likelihood? or is it somewhere else?
 Linda K. Muthen posted on Monday, June 19, 2006 - 12:19 pm
It is the seed associated with the best loglikelihood from Step 1. This loglikelihood should be replicated in Step 1 for a stable solution to be obtained. This is described more fully in the most recent user's guide which is on the website. See TECH14.
 Daniel Rodriguez posted on Monday, June 19, 2006 - 12:37 pm
I checked the manual but it wasn't clear to me. I think I got it right. I got a new warning that says to increase the number of bootstrap LRT random starts. I don't understand from the manual how to do that. Is it in the command lrtstarts=0 0 40 10. Tha I make the change?
 Bengt O. Muthen posted on Monday, June 19, 2006 - 2:54 pm
Yes. It is particularly important to increase the last 2 numbers which are for the k-class model. This is described in the new Version 4.1 User's Guide on the website.
 Daniel Rodriguez posted on Tuesday, June 20, 2006 - 4:50 am
Thanks
 Daniel Rodriguez posted on Tuesday, June 20, 2006 - 6:47 am
Hi Bengt and Linda,
I appologize for bothering you all so much with the material regarding tech 14, but I had a hard time finding the info. I re-read the paper describing the monte carlo analyses to test the efficacy of the BLRT. I also read through the manual again and now feel more comfortable. Thank you very much for your patience. I am doing my best to learn this method well so that I may do better analyses.
DR
 Linda K. Muthen posted on Tuesday, June 20, 2006 - 8:34 am
You may also find the new section in the Version 4 Mplus User's Guide on Multiple Solutions For Mixture Models useful. This is in Chapter 13.
 Christian Geiser posted on Monday, July 10, 2006 - 7:01 am
I am currently "testing" the new TECH14 option in Mplus (I am doing classical LCA). I followed your 3-step recommendations in the user's guide and found that Mplus crashed very often when I increased the number of starts (using the "lrtstarts = 0 0 40 10" option). That is, Mplus performed a number of bootstrap draws (sometimes up to 400) without obvious problems and then suddenly stopped without providing any error message in the output. The same also happened sometimes in step 2 (when using the optseed option without additional lrtstarts), but somewhat less frequently.

Is this a known problem? What can one do about it? (With a small number of bootstrap draws, say 50-100, the problem occured less frequently, but I feel somewhat uncomfortable relying on less than about 500 draws...)
 Thuy Nguyen posted on Monday, July 10, 2006 - 9:44 am
There are currently no known problems with TECH14. Please send input and data to support@statmodel.com so we can look into this problem. Thanks.
 Scott Grey posted on Wednesday, November 01, 2006 - 10:28 am
I'm trying to use TECH14 with TYPE IS MIXTURE COMPLEX but I get an error message indicating that "TECH 14 IS NOT AVAILABLE FOR THIS MODEL." This is not noted in the 4.1 user's guide, is it correct?

Thanks, Scott
 Linda K. Muthen posted on Wednesday, November 01, 2006 - 1:10 pm
It is not disallowed as far as I know. If you send the input, data, output, and your license number to support@statmodel.com, I will take a look at it and see why it is disallowed.
 aprile benner posted on Friday, March 09, 2007 - 2:57 pm
Good afternoon -

I am running a GMM with tech14 (using the optseed command and LRTSTARTS = 2 1 50 15). the model estimation terminates normally, but i get these warnings for tech14:

WARNING: THE BEST LOGLIKELIHOOD VALUE WAS NOT REPLICATED IN 12 OUT OF 12 BOOTSTRAP DRAWS. THE P-VALUE MAY NOT BE TRUSTWORTHY DUE TO LOCAL MAXIMA. INCREASE THE NUMBER OF RANDOM STARTS USING THE LRTSTARTS OPTION.

WARNING: 2 OUT OF 12 BOOTSTRAP DRAWS DID NOT CONVERGE. INCREASE THE NUMBER OF RANDOM STARTS USING THE LRTSTARTS OPTION.

What should i do?

Thanks so much,

aprile
 Linda K. Muthen posted on Saturday, March 10, 2007 - 7:43 am
Was the loglikelihood replicated in the analysis for the OPTSEED you are using? If not, I would get a seed for which the loglikelihood is replicated. If it was, you can try increasing the last two numbers of the LRTSTARTS option.
 Cameron McPhee posted on Monday, March 19, 2007 - 12:23 pm
Hi, I am running a very preliminary LCA (6 categorical indicators) and no covariates yet. I am using Type=mixture missing; and I am using the weight = option, but not type=complex. When I include the weight I am getting the message that says "Tech 14 is not available for this model." I do not get the error if I take out the weight. Does M-Plus not produce tech 14 when the data is weighted?
 Linda K. Muthen posted on Monday, March 19, 2007 - 1:44 pm
TECH14 is not available with sampling weights.
 aprile benner posted on Monday, April 02, 2007 - 3:33 pm
Hi -

I have been running GMM models using TECH14 and the optseed command, and I have been getting warnings regarding replication of my loglikelihood value and convergence of my bootstrap draws. the logliklihood value was replicated for my optseed value.

i have been increasing the last 2 LRTSTARTS values, but i continue to get these warnings. how high i should increase the last 2 LRTSTARTS values (i've gone as high as 2 1 100 50, which is taking forever!)?

thanks so much!

aprile
 Linda K. Muthen posted on Monday, April 02, 2007 - 4:24 pm
Please send your input, data, output, and license number to support@statmodel.com.
 Kurt Beron posted on Thursday, July 12, 2007 - 2:24 pm
Hi,
I am running a censored mixture model with integration and, when I ask for tech14, I see a quick view of iterations and then no output appears. The only difference in runs is adding tech14. Page 519 of manual suggests this should work. Below is a code snippet. Is this a known issue? Thanks.

variable:
CLASSES = c (3);
Censored = x3-x7 (b);
analysis:
type= mixture missing;
algorithm=integration;

optseed = 887676;
lrtstarts = 0 0 40 10;

model:
%OVERALL%
tsocint tsocslp tsocslp2| x3@0 x4@1 x5@2 x6@3 x7@4;
output:
tech14;
 Linda K. Muthen posted on Thursday, July 12, 2007 - 3:04 pm
Please send your input, data, output, and license number to support@statmodel.com.
 Stata posted on Saturday, March 17, 2012 - 11:25 pm
Linda,

I am running categorical latent class analysis with missing data and have increased LRTSTARTS values; however, the output kept showing the following warning:

ANALYSIS: TYPE = Mixture;
estimator = MLR;
LRTSTARTS = 2 1 300 20;
LRTBOOTSTRAP = 100;


WARNING: THE BEST LOGLIKELIHOOD VALUE WAS NOT REPLICATED IN 76 OUT OF 100
BOOTSTRAP DRAWS. THE P-VALUE MAY NOT BE TRUSTWORTHY DUE TO LOCAL
MAXIMA. INCREASE THE NUMBER OF RANDOM STARTS USING THE LRTSTARTS OPTION.

Any suggestions/comments are highly appreciated.

Thank you.
 Linda K. Muthen posted on Sunday, March 18, 2012 - 9:44 am
I would run the k and k-1 analyses without TECH11 or TECH14 increasing the random starts using the STARTS option until the best loglikelihood is replicated at least five times. I would use the number of starts needed to achieve this in the LRTSTARTS option.
 Deryl Hatch posted on Thursday, July 19, 2012 - 2:59 pm
Hello,

I am likewise receiving warnings that best loglikelihood value was not replicated in many of the bootstrap draws, even though I get a duplicated best loglikelihood value (in fact, more than 30 duplicated best loglikelihoods).

To improve the bootstrap draws, do I still
increase the STARTS option? Or do I
increase the LRTSTARTS option? Or do I increase the LRTBOOTSTRAP option?

For example, in my calculation of 4 class, 5 factor model, I get 30+ repeated best loglikelihoods, but in TECH14, I get warning:

WARNING: THE BEST LOGLIKELIHOOD VALUE WAS NOT REPLICATED IN 11 OUT OF 15
BOOTSTRAP DRAWS. THE P-VALUE MAY NOT BE TRUSTWORTHY DUE TO LOCAL
MAXIMA. INCREASE THE NUMBER OF RANDOM STARTS USING THE LRTSTARTS OPTION.


WARNING: 10 OUT OF 15 BOOTSTRAP DRAWS DID NOT CONVERGE.
THE P-VALUE MAY NOT BE TRUSTWORTHY.
INCREASE THE NUMBER OF RANDOM STARTS USING THE LRTSTARTS OPTION.


Type = mixture ;
Starts = 500 50;
LRTSTARTS ; ! (Using the default 0 0 20 5)
! (I've not yet used the LRTBOOTSTRAP option)
 Bengt O. Muthen posted on Thursday, July 19, 2012 - 3:27 pm
You might find the following web note on our web site useful:

Asparouhov, T. and Muthén, B. (2012). Using Mplus TECH11 and TECH14 to test the number of latent classes. Mplus Web Notes: No. 14. May 22, 2012.
 Michelle Colder Carras posted on Wednesday, October 09, 2013 - 1:14 pm
Hi Drs. Muthen,

Could you please suggest some computer specs for a new computer for the office that would be running MPlus? I understand that models can be run with 8 processors but am not sure what else would be optimal. Any suggestions would be appreciated.

Thanks,

Michelle
 Bengt O. Muthen posted on Wednesday, October 09, 2013 - 1:35 pm
We don't make specific recommendations partly because things change quickly, but we did discuss hardware that we happen to use in the early part of our training in Utrecht last year:

http://mplus.fss.uu.nl/files/2012/09/V7Part1.pdf
 Michelle Colder Carras posted on Thursday, October 10, 2013 - 8:26 am
That's perfect! Thanks--

Michelle
 nidhi gupta posted on Thursday, May 17, 2018 - 7:08 am
Dear Prof Muthen
i am running a latent profile analysis where i want to free the variance across classes.
this is my model
CLASSES = c (4);
ANALYSIS: TYPE = MIXTURE;
starts=0;
PROCESSORS = 4(STARTS);
optseed=605161;
LRTSTARTS = 0 0 1000 200;
OUTPUT: TECH14;
model: %OVERALL%
%c#1%
V1-V8*;

[V1*];
[V2*];
[V3*];
[V4*];
[V5*];
[V6*];
[V7*];
[V8*];

%c#2%

V1-V8*;

[V1*];
[V2*];
[V3*];
[V4*];
[V5*];
[V6*];
[V7*];
[V8*];


%c#3%
V1-V8*;

[V1*];
[V2*];
[V3*];
[V4*];
[V5*];
[V6*];
[V7*];
[V8*];


%c#4%
V1-V8*;

[V1*];
[V2*];
[V3*];
[V4*];
[V5*];
[V6*];
[V7*];
[V8*];

However i get warning
"OF THE 5 BOOTSTRAP DRAWS, 4 DRAWS HAD BOTH A SMALLER LRT VALUE THAN THE OBSERVED LRT VALUE AND NOT A REPLICATED BEST LOGLIKELIHOOD VALUE FOR THE 4-CLASS MODEL. THIS MEANS THAT THE P-VALUE MAY NOT BE TRUSTWORTHY DUE TO LOCAL MAXIMA. INCREASE THE NUMBER OF RANDOM STARTS USING THE LRTSTARTS OPTION."

I tried to follow the instructions given in the webnote 14 but it is not helping.
Do you have any suggestion on how to solve this issue?
 Tihomir Asparouhov posted on Thursday, May 17, 2018 - 10:05 am
This is probably the result of classes being too similar / or classes being too small or both. You can try arranging the classes so the first class is the smallest. This example is too simple to have such a rate of failure - there must be an obvious reason for this to happen. Maybe look for variance=0 problems. You can switch to using BIC or tech11.
 nidhi gupta posted on Tuesday, May 22, 2018 - 4:43 am
Hi Tihomir
Thanks for your reply.
I am a beginner in mplus. Thus, could you please tell, how exactly i should look into variance=0 problem. How to write codes for this?
Regards
Nidhi
 Tihomir Asparouhov posted on Wednesday, May 23, 2018 - 1:13 pm
There is no code to write. Inspect the variance reported for each class and make sure it is not near zero.
Back to top
Add Your Message Here
Post:
Username: Posting Information:
This is a private posting area. Only registered users and moderators may post messages here.
Password:
Options: Enable HTML code in message
Automatically activate URLs in message
Action: