1
$\begingroup$

Using a linear regression model, I can calculate standardized beta coefficients. I can then use this to find the strength of each of the individual independent variables to the dependent variable.

For instance, I can create a model such as Sales = f(TV Spend, Digital Spend, Radio Spend)

Now, the resulting standardized beta values from this model will help me their relative importance in driving sales.

However, linear regression has some drawbacks and it does not take into account non-linearity, independent variables being correlated, etc.

Is there another statistical method or technique that can help quantify variable importance but does not have the drawbacks of linear regression.

Improve this question
$\endgroup$
5
  • $\begingroup$ You could look at the mutual information between two variables. This only makes sense in a particular framework, however. $\endgroup$
    – Olivier
    Commented Feb 7, 2018 at 0:44
  • 1
    $\begingroup$ It's also not quite clear what you mean by "variable importance". Usually, this concept is only defined relatively to a particular model. I don't think it's possible to make sense of it in complete generality. $\endgroup$
    – Olivier
    Commented Feb 7, 2018 at 0:45
  • 1
    $\begingroup$ There are really two issues here - models with non-linear relations, and the "importance" of variables. Generalized additive models will address the non-linearity problem. But the idea of "importance" is very difficult to solve (and even linear regression does not really do it cleanly for you, if the explanatory variables are correlated). $\endgroup$
    – Peter Ellis
    Commented Feb 7, 2018 at 0:46
  • 2
    $\begingroup$ Please explain what "variable importance" is intended to mean. $\endgroup$
    – whuber ♦
    Commented Feb 7, 2018 at 1:00
  • $\begingroup$ @whuber changed the question so that we focus only on calculating the relative strength of each independent variable in impacting the dependent variable. Essentially, I want to calculate an equivalent of standard betas in linear regression using another method. Hope this helps $\endgroup$
    – Sharath G
    Commented Feb 7, 2018 at 2:54

0

Reset to default

Your Answer

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

Browse other questions tagged or ask your own question.