start values in tidyLPA

[Jennifer Schauer]

Dear listers,

 

I read a lot about Mplus using different sets of start values and number of iterations in LPA to examine local maxima. Is there something equivalent in tidyLPA (using mclust), or to put it in other words: How can I use sets of different start values in tidyLPA? I wonder if set.seed() does address this.

However, when specifying different values in set.seed, I get meaningful differences in profile shape. So another more practical question is: How should I deal with varying results depending on start values?

 

Any thoughts are appreciated. Thanks a lot.

Jennifer

 

 

 

 

 

______________________________________________

Jennifer Schauer, M.A.

Wissenschaftliche Mitarbeiterin

 

Professur für Berufspädagogik

Institut für Berufspädagogik und Berufliche Didaktiken

Fakultät Erziehungswissenschaften

Technische Universität Dresden

01062 Dresden

Mail: jennifer...@tu-dresden.de

Website: https://tu-dresden.de/ew/bp

 

 

 

[Caspar van Lissa]

Dear Jennifer,

tidyLPA runs Mclust or Mplus. As explained in the Mclust documentation, they use a preliminary clustering procedure instead of multiple random seeds. set.seed() indeed renders that preliminary clustering procedure deterministic.

The "problem" behind unstable results (probably the most frequently asked question here) is likely that your model is too complex, or sample too small, or your data do not meet the assumptions for LPA. Try a simpler model with fewer clusters, and make sure you are using continuous indicator variables, not Likert scales.

best,

Caspar

[Jennifer Schauer]

Many thanks, Caspar, for making this clear! I will skip some variables for exploration. Indeed, I have Likert scales as indicators (which I cannot skip for theoretical reasons). Do you know of a different package that is appropriate when there are both continuous and discrete indicators?