2
$\begingroup$

I'm doing binary logistic regression. I want to predict the chance of being in an advanced class. There no multicollinearity among my variables. I have three predictors:

  1. If you passed the test or not (a binary variable >60 or <=60),
  2. The number of correct answers you had on the test (a continuous variable between 0 and 100),
  3. The number of times you did a practice test before taking the test officially (a continuous variable).

There is no multicollinearity, but I'm wondering, can use variables 1 and 2 since 1 is created from 2? (That is, if you had more than 60 good answers on variable 2, your code on variable 1 would be >60.) I'm wondering if I'm not using the same concept twice.

$\endgroup$
3
  • $\begingroup$ There is certain level of multicollinearity given the construction of binary variable. $\endgroup$ – Subhash C. Davar Apr 14 '16 at 14:10
  • $\begingroup$ Even do when I test for it there none apparent for the test, like more a theoric multicollinearity for you? What about the other answer given? $\endgroup$ – lili Apr 14 '16 at 14:13
  • $\begingroup$ If you are working on binary logistic regression, do not mix variables - continous with binary. $\endgroup$ – Subhash C. Davar Apr 14 '16 at 14:28
4
$\begingroup$

Yes, this is perfectly acceptable. In effect, you are testing to see if there is a "jump" or level shift associated with crossing the 61-answer threshold for passing the test. When tests have a minimum score to pass (which pretty much all tests do), they implicitly assume that there is a meaningful difference between the knowledge / ability of a person who scores a 61 and someone who only scores a 60. However, it is reasonable to imagine that someone whose score is 61 knows about the same amount as someone who scored 60. Moreover, we might believe that the increase in the amount of knowledge represented by 62 vs. 61 is the same as the increase in knowledge represented by 61 vs. 60 (and 60 vs. 59, etc.). This possibility is somewhat inconsistent with the test having a hard threshold. Thus, you might want to test if the threshold is meaningful. That is what you are doing when you include variable 1 in addition to variable 2.


For further study, what you have is the simplest possible form of a spline term. The next step up in complexity would be a linear spline, in which you would have a variable indicating how far past 60 someone's score is (i.e., either 0, or +1, ..., +40). A linear spline would let you test if increasing scores above the threshold represent a larger increment than the scores below. You can take these ideas further with various cubic splines and additional thresholds (called "knots"). I provided an introductory discussion of those ideas in my answer here: What are the advantages / disadvantages of using splines, smoothed splines, and Gaussian process emulators?

$\endgroup$
2
  • $\begingroup$ What if you have bonus point on your test that you calculate on you predictor 2 but not one 1. Like the >60 or <=60 are juste for you test score, and the number of good answar calculate you good answer at the test and at the bonus question. Would it be still be ok to use the two predictor since one doesn't cover all of the 2? $\endgroup$ – lili Apr 13 '16 at 0:33
  • $\begingroup$ That would still be fine, @lili, but it would slightly affect the interpretation of the results. $\endgroup$ – gung - Reinstate Monica Apr 13 '16 at 0:38

Your Answer

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

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