4
$\begingroup$

I'm looking at the Lending Club data from Kaggle and I'm just building a pretty simple model to predict defaults.

The data has a large amount of both continuous and categorical variables (I have converted them to binary variables) in it, and now I want to look at doing some dimensionality reduction on it.

Whether it will add any value I'm unsure, but the sheer number of continuous variables makes me feel like an approach like this could be useful.

Background: I'm familiar with the PCA approach but not any other approaches that deal with categorical data.

So in terms of sticking with my strengths, could I use the PCA approach to reduce the dimensionality of the continuous data, combine the dimensionality reduced data with the dummy data and then run my model over the top?

I know we might be getting into some scaling issues (as the PCA will require scaling but the dummy variables won't), but are there any theoretical issues you can see cropping up?

$\endgroup$
  • $\begingroup$ You could do that, there is no problem with it. Whether it's going to end up being useful, nobody can say in advance. $\endgroup$ – amoeba May 12 '16 at 19:03
  • 1
    $\begingroup$ LASSO and ridge regression can also deal with categorical data as predictors when you need to reduce dimensionality. The dummy variables are just coded numeric 0/1 and can be standardized like the continuous variables if you want to adjust for scaling that way. Ridge regression is essentially principal-components regression, except that the components are weighted in a graded way rather than yes/no for inclusion in the final model. But at @amoeba put it, "Whether it's going to end up being useful, nobody can say in advance." $\endgroup$ – EdM May 12 '16 at 21:20
  • $\begingroup$ @EdM that's a really good point. I might do all of the above and and compare the CV score at the end. $\endgroup$ – plumbus_bouquet May 12 '16 at 21:23
  • $\begingroup$ @EdM (+1) Consider posting this as an answer. $\endgroup$ – amoeba May 12 '16 at 21:27
4
$\begingroup$

If your interest is in data reduction, PCA, LASSO and ridge regression can all handle categorical predictors in principle. The default is typically to code the dummies as 0/1 numeric variables and standardize them like continuous numeric variables for scaling. Principal-components regression and ridge regression are fundamentally similar, with an all-or-none choice of components in the former and a graded combination produced by the latter.

Your idea to do PCA on the continuous variables and then combine with the dummy data, however, is also reasonable, as I said in a recent answer. As @amoeba noted in a comment on the present page, however, in either case it's true that: "Whether it's going to end up being useful, nobody can say in advance."

For example, for scaling of categorical variables coded 0/1, Frank Harrell notes in Regression Modeling Strategies, second edition, page 209 that "high prevalence cells [get] more shrinkage than low prevalence ones because the high prevalence cells will dominate the penalty function." That might or might not pose a problem for you.

And as he says on the following page:

For a categorical predictor having c levels, users of ridge regression often do not recognize that the amount of shrinkage and the predicted values from the fitted model depend on how the design matrix is coded. For example, one will get different predictions depending on which cell is chosen as the reference cell when constructing dummy variables.

So if you've pre-coded a multi-level categorical variable already as multiple binaries, you have to consider that issue. Penalizing the squared difference of all k regression coefficients for a k-level categorical variable from the mean of the coefficients (including a coefficient of 0 for the reference level) can help with that, as he points out.

An alternative you might consider is to use penalized maximum likelihood, where instead of penalizing the regression coefficients of pre-standardized predictor variables, you maximize $$ \log L - \frac{1}{2}\lambda\sum_{i=1}^{p}\left(s_i\beta_i\right)^2$$ where each $s_i$ is chosen to make $s_i\beta_i$ unitless and $\lambda$ is the penalty. That allows work in the original scale of variables, and if there are some variables you don't want to penalize you can just set their scale factors to 0 in the penalized likelihood. Harrell's rms package in R provides for this.

A few final notes. I've tried to provide a generally useful answer for future reference here, but I don't have experience with datasets of the scale you are considering and I can't say how efficient these approaches may be. Second, if you are going to use cross-validation or bootstrapping to compare among approaches, as always be sure to do this validation on all the steps of the model-building process.

$\endgroup$
  • $\begingroup$ +1 Excellent answer. I forgot about the recent question that you also answered to which you link to. Looking at it now, I realize that these two questions are arguably duplicates; should we perhaps consider voting to close one of them as such? $\endgroup$ – amoeba May 12 '16 at 23:11
  • $\begingroup$ @amoeba : I had forgotten about my answer from a week ago until I started working on this one. That happens at my age. My sense is that although the answers end up pretty similar, the questions come from different stated perspectives, the first for dealing with multicollinearity, this one for dimension reduction. I'll defer to your judgment on this, however, as you have a good deal more experience with such issues on this site than I do. $\endgroup$ – EdM May 12 '16 at 23:19

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.