Replies to This Discussion

Permalink Reply by Bhaswati Sengupta on September 7, 2010 at 11:49pm

Tom, I am not sure about "Latent Class" analysis.

• ▶ Reply

Permalink Reply by Ralph Winters on September 7, 2010 at 8:18pm

I wouldn't recommend recoding categorical variables into numerics. I would stick with decision trees, correspondence analysis, or latent class analysis. You cannot do latent class analysis in SAS using EG, but there is a PROC LCA which will do the trick.

-Ralph Winters

• ▶ Reply

Permalink Reply by Bhaswati Sengupta on September 7, 2010 at 11:51pm

Could you please elaborate the reason for not converting categorical variables to numeric ?
Does Proc LCA work in SAS EG ?
Also if you have write ups on Latent Class analysis it would be really great!

• ▶ Reply

Permalink Reply by Ralph Winters on September 8, 2010 at 11:37am

Main reason is that nominal categorical variables do not have order. for others, you are assigning them arbitrarily. The dummy variable technique is fine for regression where the effects are additive, but am not sure how I would interpret them in a cluster analysis with multi levels. Maybe adding with 1 binary variable would be OK.

Haven't tried Proc LCA in SAS EG, but it might work in the code node.

-Ralph Winters

• ▶ Reply

Permalink Reply by Tom Wolfer on September 8, 2010 at 8:28am

Hi Ralph. What are the benefits of Latent Class analysis?

• ▶ Reply

Permalink Reply by Ralph Winters on September 8, 2010 at 11:55am

Tom. The two main benefits are ability to mix different scaled variables within the model framework, and no distribution assumptions. LCA does not assign to only 1 group, it computes the posterior probabilities of an observation belonging to all of the groups. Like cluster or factor analysis, the theory seems to be that the variation is explained via "hidden" groups (latent classes) rather than thru the variables themselves.

I can't really recommend a good book on LCA, since it is a relatively new field, and I'm still looking for one myself. Suggest you follow the documentation for whatever package you use.

-Ralph Winters

• ▶ Reply

Permalink Reply by Tom Wolfer on September 8, 2010 at 12:42pm

Ah, so, in other words, save for the two assumptions that you mentioned above, LCA is essentially a form of Factor analysis. Except, Factor Analysis would be a data reduction technique used on a bunch of same-scale (eg. 1-10) attitudinal variables; Latent Class Analysis (LCA) would be a data reduction technique that might include a $ spent variable, an attitudnal (1-10 scale variable), and a a index variable all in one 'Latent Variable'. But, the probability table that is outputted is pretty much the same as Factor Analysis one, is it not?

In essence, a 'Factor' in Factor Analysis is akin to a 'Latent Variable' in Latent Class Analysis?

• ▶ Reply

Permalink Reply by Ralph Winters on September 9, 2010 at 7:21am

Yes, you are correct. A factor is similar to a Latent Class variable. Good synopsis.

-Ralph Winters

• ▶ Reply

Permalink Reply by K.Kalyanaraman on September 21, 2010 at 6:20pm

This problem is very simple now. There is a procedure called "Two Step Cluster Analysis", where we can use both categorical and continuous variables. All leading software have this procedure included in their list. Please look into the assumptions. This technique is useful for a wide class of problems.

• ▶ Reply

Permalink Reply by Simon Huang on September 23, 2010 at 4:37am

Yes. the most important part of cluster analysis is the measure of "statistical distance" between two data points, which has numerous forms for either numerical or categorical variables. Try to google some keywords like "distance measure of categorical variables", I am sure you will find something useful.

• ▶ Reply

Permalink Reply by Diogo Taurinho Prado on February 12, 2011 at 11:14am

I've never heard the name 'Latent Class Analysis', but from this discussion, it seems to me that it is a Structural Equation Model. Am I right?

Anyway, there are a lot of distance measures for categorical data, just like Simon said. You can surely use it, just make sure to take a look at some of them before using as they vary depending on your final objectives and might need (most likely WILL need) some data recoding.

• ▶ Reply

Permalink Reply by Ravirajan.K on February 17, 2011 at 11:23pm

Hi ,

for treating categorical varible in segmentation. Do first canonical discriminant analysis. After gettin " Can " result through ncan in SAS. Do final cluster for segmentation

• ▶ Reply

• ‹ Previous
• 1
• 2
• 3
• Next ›
• Page