1
$\begingroup$

I have a large dataset in which only Y and one of the independent variables are continuous. There are 12 binary independent variables and 2 other categorical independent variables (each with 8 categories).

I want to use the ACE algorithm to find the transformations. Should I consider all the independent variables for ACE? Do I need to transform binary/categorical variables?

Improve this question
$\endgroup$
3
  • 1
    $\begingroup$ Please explain what you mean by "the transformations." It is rare for a numerical transformation of any binary variable to be meaningful. $\endgroup$
    – whuber ♦
    Commented Mar 9, 2018 at 16:24
  • $\begingroup$ thanks for your answer. the correlation between DV and the IVs is very low, however, based on the domain it should not. the adjusted R-squared for linear model is just 0.05 and I want to try using ACE to see if transformation can help. $\endgroup$
    – Maryam
    Commented Mar 9, 2018 at 16:34
  • $\begingroup$ It won't help. For a list of just some of the things you might address to improve the model, see stats.stackexchange.com/questions/332430. $\endgroup$
    – whuber ♦
    Commented Mar 9, 2018 at 16:40

2 Answers 2

Reset to default
2
$\begingroup$

Any transformation of say 0 and 1 to any two distinct values would change the coefficient estimates but could not improve model fit.

Think of this in geometric terms: imagine a plot of your outcome against a binary predictor coded 0 and 1. Changing 0 and 1 to any other pair of values is just re-labelling points on the predictor axis. The most you could do is reverse them so that low and high values change places. That would change the sign of a coefficient but the fit could be no better. So, if 0 1 were mapped to 42 and 666 or 666 and 42, there could be no gain, and a real loss in that interpreting the coefficient estimates is harder. The example is silly to make the point emphatic.

Having other predictors too doesn't undermine this principle.

Categorical variables could be re-coded but the same applies. The usual way to handle categorical variables entails treating them through a set of indicator variables, so that any arbitrary coding does not bite.

The only exception I can think of would be if ordered (ordinal) categories were treated numerically, then it's likely that a transformation would change fit, but you would need a good reason for it too.

Improve this answer
$\endgroup$
0
$\begingroup$

Partially answered in comments:

Please explain what you mean by "the transformations." It is rare for a numerical transformation of any binary variable to be meaningful.

– whuber

( The correlation between DV and the IVs is very low, however, based on the domain it should not. The adjusted R-squared for linear model is just 0.05 and I want to try using ACE to see if transformation can help. – Maryam )

It won't help. For a list of just some of the things you might address to improve the model, see Variables supposed to be kept show low significance.

– whuber

Improve this answer
$\endgroup$

Your Answer

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

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