Discriminant analysis

[Kajori Banerjee]

Hi, I am trying to perform discriminant analysis (using "candisc" command) using a dependent variable with two categories. On performing DA I am getting one structured canonical function as there are two categories in the dependent variable. How shall I plot a graph to show which variables are best discriminate the two categories of dependent variables. Kindly help me out!

[Dick Campbell]

I'm not sure you want to do it this way. There is a close relationship between logistic regression and discriminant analysis. Most analysts these days use the former. See Press, S. J and Wilson, S, "Choosing between Logistic Regression and Discriminant Analysis" Journal of the American Statistical Association Volume 73, Issue 364, 1978. There are standard graphs that show the change in estimated probabilities for unit changes in X following logistic regression. See the most recent edition of Long and Freese:Regression Models for Categorical Dependent Variables Using Stata

[Marcos Almeida]

Hello Kajori,

I gather you may want to employ commands related to "scoreplot", "loadingplot" and "screeplot".

Best,

Marcos

[Kajori Banerjee]

Thank you for your responses! I intend to locate the variables that best discriminate between the two categories of the dependent variables. Logistic regression will help us understand the ratio between odds of the occurrence of one even to another. I have employed that technique in my work too but discriminant analysis is being used to understand which variables have the maximum discriminating power. Am I wrong in using DA?

Scoreplot or loading plot is giving me a scattered diagram and since I have only one discriminant score function the plots are not giving a meaningful graph. Can you please suggest another method that will help me graphically represent the variables that best discriminate between the two categories of dependent variables?

[Dick Campbell]

I take it that you want to have something that shows the "importance" of your various variables in discriminating between the two groups. At the risk of telling you what you already know, discriminant analysis is a special case of canonical correlation, and if you are going to do it, you should use Stata's candisc command. If you were doing this in SAS or SPSS you would be able to get standardized coefficients just as you can in OLS. Some people would interpret those as measures of importance, although their usefulness breaks down when the predictor variable is dichotomous, e.g. an indicator variable for sex. In any case, Stata does not appear to have this feature. .

I can't tell you if what you are doing is wrong or right but it is often hard to assess the relative importance of variables. Both methods allow you to assess the probability that a given case is in one group or the other as a function of the variables in the model. Although logistic regression reports odds ratios, Stata will calculate estimated probabilities for any model you chose. You can see how those predicted probabilities change as you vary the predictors and probably work out a graph that shows this.