I don’t think these are actually opposite camps, just two related but different issues. Sure, functions can be nonlinear, but we shouldn’t create arbitrary nonlinearities due to the selection of x values.
What I was thinking of was measurements at baseline and 3,6,9,12,18,24 mo into treatment (as in ISwR::alkfos). Fitting a linear trend with the factor codes will de facto assume a kink at 12 months. The jump from time 0 to 3mo may well be deviate from a linear trend (AFAIR, it does), but that is another issue. - Peter > On 25 Oct 2025, at 17.15, Therneau, Terry M., Ph.D. <[email protected]> wrote: > > Peter wrote: > A related issue that has bugged me "forever" is that we treat ordered factors > as if their levels are equidistant even when that is patently untrue. The use > of polynomial contrasts for ordered factors reflects this - it would really > be more sensible to use e.g. successive differences or (ick!) Helmert > contrasts for the ordered case and reserve poly() for factors with actual > numerical levels. To me, this effectively makes ordered factors conceptually > useless. > I actually fall into the opposite camp. We teach students that they *must* > code factors as having non-equal increments but that continuous variables > are okay s is (the old interval vs. ordinal dichotomy). In my medical > work, a lot of the categorical variables actually are close to evenly spaced, > a disease grade for instance, since that is what the original authors of the > scale were trying to do. It is the continuous variables that violate equal > spacing most violently. A cardiac ejection fraction drop from 70 to 60 is > "meh", a drop from 30 to 20 is "I hope your affairs are in order". We > don't check this nearly often enough. > Also, I use as.integer(factor) quite when creating a an analysis data set > from input data. It's just another tool for creating new variables. > Terry T. ______________________________________________ [email protected] mailing list https://stat.ethz.ch/mailman/listinfo/r-devel
