# Feature standardization for polynomial regression with categorical data

Assuming I have a set of p=n_features, here set to 3 independent variables, each with n=n_samples, without any missing values, defining my design matrix $$X$$ as follows:

$$X = \begin{bmatrix} x_{11} & \dots & x_{1p} \\ \vdots & \ddots & \vdots \\ x_{n1} & \dots & x_{np} \end{bmatrix}$$
For my dataset with p=3 features:

$$X=\left[\vec{x_1},\ \vec{x_2},\ \vec{x_3}\right]$$

The variables are of the following kinds:

• $$y$$, the dependent variable: continuous numeric variable
• $$x_1$$ and $$x_2$$: continuous numeric variables, with different value ranges requiring standardization/scaling due to l1/l2 regularization
• $$x_3$$: categorical numeric variable with the 3 levels $$\left[0,1,2\right]$$, requiring dummy coding/one hot encoding into $$k-1=2$$ binary dummy variables

I want to feed this dataset into a polynomial regression of second degree with interaction terms (also regularization is applied), meaning my linear model to fit is of the following form:

$$y=c + c_1x_1 + c_2x_2 + c_3x_3 + c_4x_1x_2 + c_5x_1x_3 + c_6x_2x_3 + c_7x_1^2 + c_8x_2^2 + c_9x_3^2 + \vec{\epsilon}$$

with the intercept $$c$$, the coefficients $$c_1\dots c_9$$ and the error $$\vec{\epsilon}$$.
A polynomial transformation of the design matrix yields the transformed design matrix $$X^*$$:
$$X^*=\left[\vec{x_1^*},\ \vec{x_2^*},\ \vec{x_3^*},\ \vec{x_4^*},\ \vec{x_5^*},\ \vec{x_6^*},\ \vec{x_7^*},\ \vec{x_8^*},\ \vec{x_9^*}\right]$$
with $$\vec{x_1^*}=\vec{x_1},\quad \dots,\quad \vec{x_4^*}=\vec{x_1}\vec{x_2},\quad \vec{x_5^*}=\vec{x_1}\vec{x_3},\quad \dots \vec{x_9^*}=\vec{x_3^2}$$

## Problem description

We now have interaction terms between continuous and categorical variables, namely $$c_5x_1x_3$$ and $$c_6x_2x_3$$.
Dummy coding of the categoric variable has not yet been performed! (More polynomial terms if done before transformation.)
Standardization of the cont. indep. variables still needs to be done!
Having a model only consisting of continuous variables, I'd standardize after poly. transformation in most cases. In this case, with mixed types of indep. variables, I'd standardize the continuous variables and dummy code the categorical variables before polynomial transformation.

## Questions

1. Should I standardize and dummy code after polynomial transformation?
2. If yes, how to deal with the interaction terms of categorical and continuous variables?
3. If yes, how serious are the disadvantages introduced with standardizing/dummy coding before poly. transf.?
4. In general: How to avoid alternating signs (making "random" negative values) by subtracting the mean and multiplying for interaction terms (f.i. $$x_1x_2$$ where both $$x_1$$ and $$x_2$$ were positive before standardization, but afterwards $$x_1$$ is negative)? Just scale by the standard deviation $$\sigma$$ (and possibly min-max-scale afterwards)?
• A few questions to clarify what you're trying to accomplish. Why do you need to standardize the predictor variables, if this is an ordinary linear regression? What are you trying to accomplish with the $x_3^2$ term, with $x_3$ the categorical predictor? Why are you using a polynomial model instead of, say, restricted cubic splines for modeling the non-linear aspects of the relationships?
– EdM
May 1, 2020 at 14:30
• @EdM thanks for your contribution! Forgot to mention that I am mainly using l1/l2 regularized algorithms and also often implement PCA or alike prior to fitting, thus standardization is important. The $x_3^2$ term was just there to fully show the effect of poly. transf.. Also most poly. transf. algorithms don't automatically exclude cat. vars. from transformation, thus it will be present in the design matrix, of not excluded explicitly. May 1, 2020 at 15:13
• The last question is a bit more difficult to answer: As far as I know my tool of choice sklearn/scikit-learn does not support restricted cubic splines. This is also the reason why I don't have any experience with cubic spline fitting. Or is there any restricted cubic spline method in sklearn, which does not require splitting up the paths separately, with f.i. decision trees? May 1, 2020 at 15:18
• I don't use sklearn, so I can't comment on that. In R the rms package provides restricted cubic splines easily. Think carefully about whether and how to standardize the categorical predictor; see this answer for an introduction to the problems, which are even greater with more than 2 levels, and its links for further study. Also think about what you hope to gain from including all the high-order interaction terms. Sometimes applying knowledge of the subject matter can simplify the model.
– EdM
May 1, 2020 at 15:44
• Thanks for the links! Yeah, as stated in the bullet point list in my question, I only want to standardize the continuous independent variables resp. predictors. The categorical predictors are not to be standardized (well, perhaps I could scale by unit variance without subtracting the mean to get a $\sigma=1$ also for the categories, but this is currently not planned and I guess also not beneficial from a statistics point of view). Yes, the last point is imho really important and I will consider it when fitting my "real" model. I left it out of the question to keep it simple. May 1, 2020 at 15:56

## 1 Answer

When a LASSO model includes a categorical predictor with more than 2 levels, you usually want to ensure that all levels of the predictor are selected together as with the group LASSO. When a LASSO model includes interaction terms, it's important to maintain the hierarchy of the interactions. That is, if LASSO selects an interaction term it should also select the terms of the individual predictors contributing to the interaction. That's discussed briefly here and with more rigor by Bien, Taylor and Tibshirani in "A lasso for hierarchical interactions", Ann. Stat. 41; 1111–1141, 2013.

For your questions 1 and 3, Bien, Taylor and Tibshirani seem to deal directly with your question:

It is common with the lasso to standardize the predictors so that they are on the same scale. In this paper, we standardize X [matrix of individual predictors] so that its columns have mean 0 and standard deviation 1; we then form Z [matrix of interaction terms] from these standardized predictors and, finally, center the resulting columns of Z.

As the quadratic terms in your model are essentially self-interactions it would seem that you would be advised to proceed similarly. That is, standardize the continuous predictors $$x_1$$ and $$x_2$$ (subtract mean, divide by standard deviation), form the polynomial and interaction terms from the standardized predictors, then only center the polynomial and interaction terms. (As I understand it the centering of the interactions isn't necessary but does simplify interpretations of coefficients.) The corresponding R hierNet package by Bien and Tibshirani provides those choices as defaults: center features, standardize main effects, and don't standardize interactions. The hierNet() function does allow for other choices, if you want to play with other possibilities.

With respect to question 2, as noted in a comment it's not clear whether or how best to standardize a categorical predictor, particularly with more than 2 levels. Provided you handle it with group LASSO and respect the hierarchy of interactions, however, there isn't any problem in "deal[ing] with the interaction terms of categorical and continuous variables." If you choose treatment coding of the categorical predictor then the coefficients of the continuous predictors and their interactions with each other represent those values when the categorical predictor is at its reference level. The corresponding interaction terms with the other levels of the predictor are the differences of the coefficients from those values for the reference level. I see nothing to be gained by incorporating powers of the dummy variables representing the categorical predictor.

With respect to question 4, the "alternating signs" in interaction values after centering are features, not bugs. See this page for example. Leave them alone.

• Great answer, many thanks! Unluckily I cannot use group LASSO, since it is not included in sklearn and I am restricted to using sklearn. Concerning forming the polynomial: It might sound like a stupid question, but until now I've been including the categorical predictors in the process of forming the polynomial. In your second paragraph it sounds like forming the polynomial without the cat.-predictors and then merge the polynomial with the untransformed cat. predictors. Is that correct? May 1, 2020 at 18:21
• Concerning paragraph 3 with respect to question 2: Then I'll go with treatment coding of the cat. predictors. That means I can treat the interaction terms of cat. predictors and cont. predictors just like normal continuous predictor interaction terms, since the coefficients will make up the differences? Yes, power of the dummy variables are useless and are just included in my description for completeness of the formulation. Concerning question 4: That was exactly what I was hoping to head (resp. read)! Thanks! May 1, 2020 at 18:26
• @Scotty1- there might be an add-on package lightning that implements group LASSO, but again I have no sklearn experience. I think that treatment coding simplifies things here for coding and interpretation. Yes, start with your polynomials and interactions for the continuous predictors. Those coefficient values will be for the reference level of the categorical. Then, for each of those terms based on the continuous predictors, add an interaction with one of the 2 remaining categorical levels. I'd recommend the same for L2 penalty, too.
– EdM
May 1, 2020 at 19:03
• @Scotty1- also check out pyglmnet for a Python implementation of group lasso.
– EdM
May 1, 2020 at 19:08