0
$\begingroup$

I am confusing regarding the differences between poly and contr.poly in regression.

Both should generate the orthonormal polynomial transformation of a vector. But poly is influenced by the length of the vector while contr.poly only by the number of unique elements in the vector, giving therefore different results.

This is shown passing a vector of 1:10 to the functions:

poly(1:10, 1) %>% unique() %>% sort()
 [1] -0.49543369 -0.38533732 -0.27524094 -0.16514456 -0.05504819  0.05504819  0.16514456  0.27524094  0.38533732  0.49543369

attr(C(as.ordered(1:10), ,1), 'contrasts') %>% as.vector()
 [1] -0.49543369 -0.38533732 -0.27524094 -0.16514456 -0.05504819  0.05504819  0.16514456  0.27524094  0.38533732  0.49543369

Same result. But if I join the vector twice with c(1:10, 1:10) I get different results:

poly(c(1:10, 1:10), 1) %>% unique() %>% sort()
 [1] -0.35032452 -0.27247463 -0.19462474 -0.11677484 -0.03892495  0.03892495  0.11677484  0.19462474  0.27247463  0.35032452

attr(C(as.ordered(c(1:10,1:10)), ,1), 'contrasts') %>% as.vector()
 [1] -0.49543369 -0.38533732 -0.27524094 -0.16514456 -0.05504819  0.05504819  0.16514456  0.27524094  0.38533732  0.49543369

I guess it's because being the vector a factor in the second example the elements are considered only once.

Therefore, which is the difference in interpretation of the two in a regression setting? Which one should be used and in general how one should interpret the coefficients after an orthonormal transformation, what does a change of 1 mean?

Thanks

Improve this question
$\endgroup$
3
  • $\begingroup$ It is instructive to read the help pages and to test the software while you do so. Compare, for instance, the values of poly(1:d, degree=d-1) and contr.poly(d). Start with d<-3 and try it for several more (larger integral) values of d. $\endgroup$
    – whuber ♦
    Jul 3, 2017 at 20:03
  • $\begingroup$ Uhm, I tried it and as long there are only unique values both functions give the same results. In the manual pages doesn't say much. So I don't know which to use and how to interpretate the relative coefficients in a regression setting. $\endgroup$
    – Bakaburg
    Jul 4, 2017 at 7:17
  • $\begingroup$ You will find useful information about poly at stats.stackexchange.com/questions/31858. As you have learned through your experimentation, contr.poly implements a special case of orthogonal polynomials. $\endgroup$
    – whuber ♦
    Jul 4, 2017 at 13:45

0

Reset to default

Your Answer

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