Exploratory factor analysis using categorical data

[Guest]

Hi,

I am trying to carry out an exploratory factor analysis on a household data set where information was collected via family surveys. I have 10 categorical variables that revolve around decision making in the household (who makes decisions regarding purchases, children, medical treatment etc.). The variables are coded such that:

1= respondent (the wife)
2= Husband and wife jointly
3= Husband

I am using a polychoric correlation matrix followed by the factormat command, however, my results are only showing (using kaiser and scree plot) that there is only 1 factor to extract, which doesn't make much sense for my research question. Am I going about the analysis the wrong way by using a polychoric matrix? I have attempted the factor analysis using principle factor analysis and I am getting similar results.

thank you in advance.

[William Lisowski]

Welcome to Statalist.

The output of help polychoric tells us that the polychoric correlation is defined for ordinal categorical data. I understand the sense in which you data is coded in an ordinal fashion - with unilateral decision making by each spouse at the extremes, and shared decision making in the center - but perhaps that is an oversimplification of the reality. It could be the case that households are divided into those that predominantly make decisions jointly and those that do so unilaterally, and the latter then divided by topic into which spouse takes responsibility for which decisions.

With that said, I cannot confidently recommend an approach to modeling this. It would be interesting to see if it is possible to divide the households by decision-making style in a "natural" way - is the distribution of number of types of decisions made jointly U-shaped, with peaks at the ends and very few households making 5 types of decisions jointly and 5 unilaterally? Perhaps another approach would be to think about two variables for each of the ten decisions - is the respondent involved, is the husband involved?

[Guest]

Thank you William, I see where you're coming from and it might explain the results I am getting. I will look more into it.

[Red Owl]

Guest,

Latent Class Analysis (LCA) is an alternative to Exploratory Factor Analysis when the indicator variables are either binary or categorical. You can see an example of LCA at https://methodology.psu.edu/ra/lca/example

Stata 15 added a native command for LCA using -gsem-. See Example 50g at

help sem examples

If you are using a version of Stata older than 15 on a Windows-based system, you can use the LCA plugin from The Methodology Center at Penn State (https://methodology.psu.edu/downloads).

You can also use the user-written -gllamm- ado program (from SSC) on any operating system. A worked example using -gllamm- is at http://www.gllamm.org/examples.html.

Hope that helps.

Red Owl
Stata/IC 15.1, Windows 10 (64-bit)

[William Lisowski]

I refrained from recommending sem in the (realized) hope that someone else could make the recommendation more knowledgeably than I could. My thanks to Red Owl for picking up the story and advancing it.

[Guest]

Thank you very much Red Owl, from the links you've provided it appears that LCA would be the most appropriate method. You've both been very helpful.