How to interpret coefficients of nominal independent variables in Weka?

Question

I'm struggling a bit with interpreting the output of a linear regression in Weka. This is my model:

  0.1063 * checking_status=0<=X<200,>=200,no checking +
  0.1329 * checking_status=>=200,no checking +
  0.0593 * checking_status=no checking +
  0.2201 * credit_history=existing paid,delayed previously,critical/other existing credit +
  0.0963 * credit_history=critical/other existing credit +
  0.3752 * purpose=other,new car,repairs,business,domestic appliance,furniture/equipment,radio/tv,used car,retraining +
 -0.2895 * purpose=new car,repairs,business,domestic appliance,furniture/equipment,radio/tv,used car,retraining +
  0.0988 * purpose=business,domestic appliance,furniture/equipment,radio/tv,used car,retraining +
  0.0953 * purpose=used car,retraining +
 -0      * credit_amount +
  0.103  * saving_status=no known savings,500<=X<1000,>=1000 +
  0.0839 * employment=4<=X<7 +
  0.0619 * personal_status=male mar/wid,male single +
  0.0022 * age +
  0.0344

Could someone explain how should I interpret different coefficients for different values of the same attribute? For example, checking_status has been divided into three separate "brackets", each with its own different coefficient. Does it mean that if "checking_status=>=200,no checking" has a coefficient of 0.1329, then it is somewhat "more important" than "checking_status=no checking" that has a coefficient of 0.0593?

Thanks

Answer 1

It's different effect coding than you may be expecting. Let's suppose we have three possible values: High, Medium, and Low. We can dummy code this as

        Code as:    
        high    med
Instance:           
High        1   0
Medium      0   1
Low         0   0

That's fine, but it loses the ordinal aspect that High > Medium > Low.

As a default, this Weka module uses coding that combines attribute values and tries to preserve the ordinal aspect, e.g.

        Code as:    
        high    highORmed
Instance:           
High        1   1
Medium      0   1
Low         0   0

So, if you are looking at the effect of medium (versus low) you'd use the highORmed coefficient. If you are looking at the effect of high, you'd combine both the high and the highORmed coefficients.