LC Identification of the relevant variables

Question

Hello,
I want to do a latent class analysis to see which variables cause the more homogeneous groups. Accordingly, I want to achieve a high assignment to the groups based on a combination of variables.

I have now determined the number of groups. For the next step I saw that on the last sheet of the LC results the test persons with the assignment to the groups can be found. So I would import this assignment into SPSS and then do a demographic analysis. For example, group 1 is predominantly male, over 30 etc. Would a new LC analysis be possible afterwards to improve the values if I found out that gender and age were the main reasons for the groups?

Answer 1

Our Latent Class option doesn't allow you to add in variables beyond the conjoint or MaxDiff choice data, sorry.  If you had access to our CCEA software, you could potentially use individual scores from an HB run plus other variables to come up with an ensemble solution that incorporated lots of different variables, or potentially do something similar with another clustering software, or a package for R, etc.  I believe the Latent Gold software does have a more feature-rich Latent Class implementation that would allow you to add in additional variables beyond the choice data as well.

Comments

Thanks for your answer. But if I found a variable (e.g. salary) from my survey, which mainly causes the homogeneous groups. Isn't there a possibility to examine it again, for example to get a better allocation to the groups?

https://sawtoothsoftware.com/resources/technical-papers/latent-class-technical-paper
In the technical paper I found on page 14 the hint that I can split my sample and run the LC analysis again. Could I use it to determine if my identified variables allow a better assignment of the groups? If so, how can I split the sample for a new analysis?

---

There's no way to formally include your salary variable in the Latent Class analysis, no.

In Lighthouse Studio you can create a filter under the Analysis menu, or you can formally turn a filter into a new variable in your data by then going to the Segments and Weights area.  Once a filter exists, you can activate it for a utility run by clicking on the gear icon next to the dropdown box where you choose to run Latent Class.

---

Many thanks for the information!

I would examine a subject's group membership according to demographic characteristics and then, based on this, determine the variables that are likely to have the greatest influence on group membership. In the next step I would filter my sample for the relevant variables and view the LC results. If the LC results in two groups and at least 60% assignment to one group, I would consider this as proof of successful segmentation. Would this be a legitimate way?

Is there a source or at least an indication, from how many subjects on an LC analysis is reasonable?

---

Clustering is a mixture of art and science.  Generally speaking, you want a clustering solution that makes logical sense (has strategic value for example, people in cluster 1 tend to be older than 2) and that has no groups have a small number of respondents.  Most people would probably say no groups of less than 5% or 10% of the total sample.

I don't think there is really a certain size you are going for.  A 2-group solution might produce an 80/20 split, a 60/40 split or a 50/50 split.  The size is not a good indicator of whether or not its useful.

I'm not sure if there is a good rule of thumb for minimum sample sizes for Latent Class, sorry.

---

Thank you very much for your support!