Variations in outcomes TidyLPA

[a.e.smal...@rug.nl]

Dear members,

I am currently exploring the TidyLPA package in R studio for the first time to get aquiainted with the approach. I really like the possibilities it provides. However, I do have a question: I noticed that when I ran the same script several times, this resulted in different outcomes. For instance, the entropy value varied from about 0.56 to 0.42 (both not good - I know - it was just an exploration). Can you explain where such differences come from? How should this affect my interpretations? I hope you can find time to clarify this so I can understand the package better.

Kind regards, Annemieke Smale

[Rosenberg, Joshua]

Hi Annemieke, 

Are you using the default package (if you aren't passing a package argument, this is the case; I'm curious whether you are adding package = 'MplusAutomation' or package = 'mplus' (either work).

In short, this matters because the two potential packages ('mclust' or 'MplusAutomation') have slight differences in how the estimation works. For mclust, first, a hierarchical cluster analysis is carried out, and then the results of the clustering are used to provide starting values for the maximum likelihood estimation. For mplus, random starting values are used, but a number of starts are run, and then the fitted model associated with the lowest log-likelihood is considered to be the model used for/presented in the output. 

So, knowing about which package you are using will at least help me/us to zero in on what may be going on.

Best,

Josh

-- 

 

Joshua M. Rosenberg

Assistant Professor, STEM Education

 

The University of Tennessee, Knoxville

Dept. of Theory and Practice in Teacher Education

420 Claxton, 1122 Volunteer Boulevard, Knoxville, TN 37996

 

jmros...@utk.edu

(865) 974-5973

---

From: tid...@googlegroups.com <tid...@googlegroups.com> on behalf of a.e.smal...@rug.nl <a.e.smal...@rug.nl>
Sent: Monday, June 17, 2019 10:38 AM
To: tidyLPA
Subject: Variations in outcomes TidyLPA

 

Dear members,

I am currently exploring the TidyLPA package in R studio for the first time to get aquiainted with the approach. I really like the possibilities it provides. However, I do have a question: I noticed that when I ran the same script several times, this resulted in different outcomes. For instance, the entropy value varied from about 0.56 to 0.42 (both not good - I know - it was just an exploration). Can you explain where such differences come from? How should this affect my interpretations? I hope you can find time to clarify this so I can understand the package better.

Kind regards, Annemieke Smale

--
You received this message because you are subscribed to the Google Groups "tidyLPA" group.
To unsubscribe from this group and stop receiving emails from it, send an email to tidylpa+u...@googlegroups.com.
To post to this group, send email to tid...@googlegroups.com.
Visit this group at https://groups.google.com/group/tidylpa.
To view this discussion on the web visit https://groups.google.com/d/msgid/tidylpa/e8d66324-e9fb-4ec2-9fd6-81052d2b2c2b%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Smale-Jacobse, A.E.]

We are using the mclust package. Unfortunately, I don not have Mplus available here.

Op ma 17 jun. 2019 om 20:07 schreef Rosenberg, Joshua <jmros...@utk.edu>:

[Jingguang Li]

Dear all,

I also found this issue in my LPA data-analyses with 1815 participants, 2 indicators, and 4 profiles. 

--------

fit3 <- dateset %>%

  select(indicator1,  indicator2) %>%

  single_imputation() %>%

  scale() %>%

  estimate_profiles(4, models = 1) # change factors and models 

get_fit(fit3)

data.lpa <- get_data(fit3)

-------- 

By running the above code repeatly, the results change slightly. For example：

# the first run: profile 1 (N = 1554); profile 2 (N = 30); profile 3 (N = 28); profile 4 (N = 203)

# the second run : profile 1 (N = 1446); profile 2 (N = 76); profile 3 (N = 33); profile 4 (N = 260)

# the third run: profile 1 (N = 1461); profile 2 (N = 71); profile 3 (N = 251); profile 4 (N = 32)

......

Thank you for your help. 

Best,

Jingguang

[Puzzle Of Danish]

Hi all,

I have the same issue — the number of individuals in each profile changes dramatically if I run the same model multiple times. Is there a reason for this? I'm using the default package for estimation.

Hope you can help me figure this out :)

Best,

Fabio Trecca

Aarhus University

Denmark

[MC]

Has this been resolved? 

[Caspar van Lissa]

I see that my response went directly to the OP, not to this discussion thread. There is likely a problem with the model, see below:

Caspar van Lissa <c.j.va...@gmail.com>

Thu, Sep 24, 3:17 PM

to Puzzle

Not sure if this has been answered here or elsewhere, but the problem here is most likely instability of the model.

Potential solutions are:

* estimating a simpler model with fewer parameters per group 

* estimating a smaller number of classes

* increasing sample size

* checking your multivariate distributions to make sure you're not trying to fit a mixture model to a normally distributed variable

* the opposite: checking for major violations of assumptions, for example, using categorical indicators (likert scales) 

[May Conley]

Hi Caspar,

Thanks so much for getting back to me!

Hmm…I’m using a model with 3 variables with equal variances and covariances fixed to 0 (default model, model 0) to estimate 1:6 profiles using mclust. My overall n=~7000 and it seems like samples that large can be an issue for tidyLPA/mclust? Is that correct? All of my indicators are continuous and they are all positive skewed. 

Any insight would be greatly appreciated.

You received this message because you are subscribed to a topic in the Google Groups "tidyLPA" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/tidylpa/TFHUub5HZGw/unsubscribe.
To unsubscribe from this group and all its topics, send an email to tidylpa+u...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/tidylpa/b6a14fea-6c21-4f02-a6c4-918b06f0cacen%40googlegroups.com.