Latent Class analysis

Question

Hi,

I am running a large study in two countries (A and B), I am interested in comparing the differences between the two countries. When I run a 3-group solution using latent class analysis, I get slightly different segments for both countries A and B (two of the three segments are the same, while the remaining segment differs slightly between country A and B in terms of the attributes that are preferred).

Would it be incorrect to run latent class analysis on both countries as a whole (A+B) to generate a 3-group solution, then divide the membership of individuals by their country of origin before calculating HB utilities by country and by segment so that the segments are comparable? Or would this be incorrect because in using both countries to generate the universal segments, one country may in fact influence certain segments more than the other?

With much thanks,
Victoria

Answer 1

If your goal is to compare the part-worth utilities (preference scores) between countries, then (the Bayesians would argue) HB analysis with a covariate set to country membership is a very strong approach.  You compare the history of "alphas" for the used draws associated with the covariate (saved in an alpha file separate from the utility estimates) and count for what percent of the respondents the country covariate for a particular parameter is larger or smaller than for the other country.  We describe this process in the paper found at: http://www.sawtoothsoftware.com/download/techpap/HBCovariates.pdf

Frequentist folks would simply take the final utilities (point estimates from HB) for the two groups of respondents, and after normalizing using zero-centered Diffs (you can export these to this normalization using SMRT's Analysis + Run Manager + Export), using t-tests or F-tests to compare the two groups on the utilities.

Now, if your goal is to test whether the two countries have the same segmentation structure, then that would seem to leverage Latent Class analysis.  Your suggested approach seems intuitively nice, but I don't know of a formal statistical test of whether the segmentation structures are different within a country vs. the pooled countries.  Maybe somebody with more expertise in Latent Class can comment on that.

Comments

Thank you very much Bryan, I think you answered my question. I was more interested in differences between countries, but since the majority of studies publish on combined LC/HB results, I wasn't sure if skipping the LC component was acceptable for my study. Thanks!

---

I've got a different view . . That's not to say its  right of course.  I would just whack the calibration of the choice model into a LC routine . . Then if you profile the resultant segments and you find that country turns out to discriminate  significantly, or it may not.  Or am I missing the point of the question ?  btw entering these comments via an android keypad is a real pain . . Is there a mobile-friendly  version of the forum ?