0
$\begingroup$

I am trying to reduce the dimension of a matrix of several hundred binary/indicator/boolean variables, and then use the reduced components in subsequent regression modeling.

For continuous variables, this would look typically involve using PCA on the covariance matrix, projecting the data onto some subset of the eigenvectors, and using these resulting components in a regression model, effectively reducing the number of parameters estimated while retaining most of the information in the data.

For a set of binary variables, correlation is not well-defined. Multiple correspondence analysis can be used instead. We take a Burt matrix, and apply MCA to it. I think I understand the math behind correspondence analysis, but what I cannot figure out is how to project the data down onto some subset of eigenvectors. That is, how are the components created in MCA for use in a regression analysis?

So, if I have $K$ binary variables, how do I get the closest $K-d, d\geq 1$ dimensional approximation to the data, for integers $d$ in the same way I would with PCA? Or, with MCA is this not the idea?

$\endgroup$
3
  • $\begingroup$ That sounds like a useful idea, see this paper which seems to be doing that! $\endgroup$ Commented Nov 12, 2021 at 22:13
  • $\begingroup$ Your question is not quite clear. Correspondence analysis is a dimensionality reduction method for nominal variables. It is the nominal variables that are converted internally to dummy ones and hence to Burt matrix, by the procedure. So are your features nominal? If all your features are just dichotomous, binary ones, then MCA will yield basically the same result as PCA done on these binary variables as if "scale" ones. $\endgroup$
    – ttnphns
    Commented Nov 13, 2021 at 8:34
  • $\begingroup$ About PCA on binary data stats.stackexchange.com/q/16331/3277 $\endgroup$
    – ttnphns
    Commented Nov 13, 2021 at 8:36

0

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.