I have a multiple regression that contains multiple variables, many of them ordinal with 3 or 4 possible choices. (e.g: Are you currently experiencing stress? 1: No, not at all. 2: Somewhat. 3: Yes, quite a lot.)

My question is a follows - should I start these variables at 1, 0 or doesn't it matter? The advantage of starting at 0 is that the intercept of the regression has more meaning (e.g, the base value at the absence of all other factors) - but starting at 1 is easier for people at questionnaires.

  • 6
    $\begingroup$ If it matters in the regression, you are treating the variable as interval-scaled and not just ordinal. $\endgroup$ – Frank Harrell Apr 15 '18 at 12:02
  • 5
    $\begingroup$ This question is founded on a misconception: numbers like $0, 1,$ etc are merely codes used to indicate the sequence of values an ordinal variable. You may freely replace them by any other sequence of numbers that preserve the order, such as $-10^6, 0.1, \pi$ instead of $1,2,3.$ Recognizing this fact will create opportunities for deeper and more useful analysis, because it shows that you are free to select these numerical codes in any order-preserving manner that helps with understanding or prediction. $\endgroup$ – whuber Apr 15 '18 at 16:01
  • 1
    $\begingroup$ Start at 1 in questionnaires and transform to start at 0 for the regression. $\endgroup$ – Alecos Papadopoulos Apr 15 '18 at 18:58
  • $\begingroup$ @whuber - thanks for your response! If I understand correctly I will still have to increment the effect with the position of the code in the sequence right? For example, if the sequence is [ -10, 4, 340 ], and the regression show that the increase of each step is 0.5. A value of '4' would correspond to '2x --> 2 * 0.5'? $\endgroup$ – Jasper Apr 16 '18 at 15:44

Your Answer

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

Browse other questions tagged or ask your own question.