Representing categorical predictor variables

[torkar]

Hi,

perhaps @Max_Mantei’s post will make things more concrete (it did help me):

Blog post: Identifying non-identifiability

Cool, thanks! Maybe a minor addition on the first case that you present: In this example x and x^2 are actually the same thing (perfect multi-collinearity) and one could rewrite (ignoring subscripts) y\sim\text{Normal}(x+x^2,\sigma) to y\sim\text{Normal}(2x,\sigma), right? If you run set.seed(20180512) x = rep(c(0,1), times = 10) N = length(x) y = rnorm(N, x + x ^ 2, 1) # -1 for "no-intercept" lm(y ~ x + I(x^2) -1) R will just drop the second term: Call: lm(formula = y ~ x + I(x^2) - 1) …