Hi,
I'm discussing with a customer if one could use Latent Class analysis to inform survey analysis.
Imagine you have questions about "Trust" in whatever, the answers are "A lot", "Some", "not too much", "not at all", "refused". Then you have a lot of other questions about personal habbits, Confidence in public organizations, local community, overall country specific information, Health, Buying patterns and so forth from other studies not necessarily related with each other, demographics. This is e.g done in the 2016 PEW RESEARCH CENTER’S AMERICAN TRENDS PANEL. Others said, there are some questions all get, some just a subgroup and others another subgroup and so forth.
Would it make sense to first cluster by questions all have to answer before looking into the details of the other questions? I think about using a Latent Class Clustering as a prestep and then doing e.g. MCA, Categorical, Modeling and others but using the clusters.
Jian has a great example for a survey analysis using Factor Analysis to get predictors from a survey which could be linked to car properties. I rate this kind of similar.
However what I'm unsure about is: Is this statistical correct, or do I hit any assumptions I should take care of? E.g. doing a latent class and then a Categorical analysis, could it be that I added some bias or unwanted weighting through the Latent Class Clusters?
I know it is a more general question, but if this would be feasible one could combine surveys from different sources as long as you can link the participants to those general clusters, and make conclusions from one topic to a total different topic for a participant without asking this person on the different topic, becuse this person belongs to a cluster (like more confident than others) who can be linked to the second survey's conclusions (is there a problem with local schools, ...).
(Honestly this is kind of scary, as you might not have to answer any question but could be rated based on other people's answers, but that's probably already true for those googes and amazones in the world).
Long write short question: Is there any statistical reason, not to follow such a methodology?
Thanks,
Martin
Attached some data from the above mentioned study. There are three Forms each with several questions. More information and a table can be found here: below in my second response is a link to the users2 drive
Hi Jian,
thanks for your thoughts. So using LCA as a X for a categorical analyis is fine. For the use of modeling rather not. This helps already!
I've attached the cleanded and modified data table and tried to make the grouping variables as descriptable as possible, so it's easier to follow.
I've put the files you couldn't see on the User's drive: \\jmpsmd\Users2\Demel\Community
Martin,
LCA places subjects into groups based on their responses to a set of categorical questions. For example, using your data, you can cluster participants based on their answers to Knowledge questions in the survey, or cluster them based on Confidence questions. In other words, a cluster is tied to how their responses to a set of questions; you may see two subjects that are in same Confidence cluster appear in two different Knowledge clusters.
Of course you could plot Confidence questions (or any other variables) with clustering results derived from knowledge questions. That’s just like using them as supplemental variables.
Unlike PCA, FA and fPCA, I do not see LCA used as a data prep for modeling.
Did I missunderstand your questions?
BTW, I cannot find the forms you put up in your One Drive Folder.
