Have an idea?

Visit Sawtooth Software Feedback to share your ideas on how we can improve our products.

How to determine the number of groups?


Whilst conducting a LCA, I wonder now, how many groups, I should go for. Firstly, I know, I should run the LCA with different kind of starting seeds and compare the results. When I compare them, I look on 'results of best solution'. I look on AIC, CAIC, BIC and Percent Certainty. And my AIC, CAIC, BIC drop with a single group more and at the same time my Percent Certainty goes up. As too many groups are from an economical view not good to analyse, I need to think of a solution with no more than, let's say 5 groups. Now, my questions are:
- What pararmeter is most important? AIC, CAIC, BIC, Percent Certainty? Plus any advice on papers for the problem? Or a hint, how to find a solution?

- In terms of growth, is there a rule of thumb, which says a growth of less than 5% whatever for AIC is not much and thus, I can go for less groups?

- What number of percent certainty is desired? Which percentage is okay for segmentation?

- My average respondent probability is around 95% - can I consider that as a good model fit?

Thanks a lot!
asked Oct 17, 2017 by Ben

1 Answer

0 votes

BIC and CAIC seem to be the preferred measures.  Most users plot them and read them sort of like a scree plot in factor analysis, looking for where they start to level off.  Even more folks decide in advance that they can't do much with more than 4-5 segments an that 2 is too few, so they look at 3, 4 and 5 segment solutions.  

I am not aware of a rule of thumb about percentage differences in the fit measures.  I don't see percent certainty used much as a diagnostic for LC solutions.  

It's not uncommon to find most respondents with VERY HIGH respondent probabilities - that seems to be par for the course with LC models.
answered Oct 18, 2017 by Keith Chrzan Platinum Sawtooth Software, Inc. (62,500 points)