Have an idea?

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

LC: Which segment size should I choose?

Hey,

according to Technical Paper (2019), the segment size with the lowest CAIC and/or the highest relative chi-square should be chosen.
In my LC the segment size with the lowest CAIC has at the same time the lowest relative chi-square. Additionally the segment size with the highest relative chi-square has the highest CAIC.
Does anyone have a solution for this problem?

Thanks a lot,
Simon
asked Aug 14, 2020 by breucko (160 points)

3 Answers

0 votes
The different measures of fit in LC do not always agree.  This is part of the art and science of Latent Class MNL modeling.   Often, there is some managerial/strategic reason to do the segmentation, so you should look at the different segmentation solutions and decide which one fulfills the managerial/interpretability purpose best while also showing acceptable fit.
answered Aug 14, 2020 by Bryan Orme Platinum Sawtooth Software, Inc. (181,965 points)
0 votes
I'll second Bryan's comments - I like the statistics to guide my choices about the number of segments, but not to determine them.  

On another note, in practice I use CAIC and BIC to make choices about the number of segments, and that's what I usually see in the academic literature as well.  I don't much see relative chi-square being used, so I would give that much less weight in my deliberations.
answered Aug 15, 2020 by Keith Chrzan Platinum Sawtooth Software, Inc. (99,300 points)
0 votes
Just try to share,
I'm doing LCA too righ now, the CAIC and BIC decreased until 5 segments more. I got confident solution after read the technical paper and this article: https://www.sciencedirect.com/science/article/abs/pii/S096969891830448X

So, there are some methods to determine segments in LCA:
1. Goodness of fit (CAIC and BIC value)
2. Inflection point (where the value decreased so much)
3. Tiny group size
4. The content

Because my CAIC and BIC was decrease continously, so I look for inflection point and remove segments that have small group size (<10%). Than look inside the segment members characteristic to view more differences and conclude segments.
answered Aug 15, 2020 by yustian (280 points)
edited Aug 15, 2020 by yustian
...