Contrast coding systems for categorical variables

[awellis]

Are there any plans to add more contrast coding systems for categorical variables? At the moment, contr_treatment seems to be the only one.

[johnmyleswhite]

I don't know if anyone is planning to work on that soon, but it seems clearly worth having eventually. @dmbates might have more to say. I'm still planning on rewriting our underlying representation of categorical variables, but that should be orthogonal.

[awellis]

Another thing:
in R, omitting the intercept will create summary variables for all levels:

df <- as.data.frame(c("A", "A", "B", "C", "C", "B"))
names(df) <- "var"
df <- model.matrix(~ 0 + df$var)

will return

  df$varA df$varB df$varC
1       1       0       0
2       1       0       0
3       0       1       0
4       0       0       1
5       0       0       1
6       0       1       0

In Julia, if I omit the intercept

df = DataFrame(var = ["A", "A", "B", "C", "C", "B"], y = randn(6))
pool!(df, [:var])
mm = ModelMatrix(ModelFrame( y ~ var + 0, df))
ModelMatrix{Float64}(6x2 Array{Float64,2}:
 0.0  0.0
 0.0  0.0
 1.0  0.0
 0.0  1.0
 0.0  1.0
 1.0  0.0,[1,1])

No dummy variable is created for the reference level ("A"). Is this intended?

[johnmyleswhite]

That seems like a bug to me.

[awellis]

should I open a separate issue?

[kleinschmidt]

The contrast generation code was pretty primitive last time I looked, the most obvious thing being that it doesn't check for whether an intercept is present or not. Things get tricky when you have more than one categorical variable (in R, only the first categorical variable gets the full complement of predictors), or interactions between categorical predictors and other predictors (you need to include more levels when an interaction is included but the lower-order terms are not).

I think someone had opened an issue or pull request about other contrast types a long time ago but obviously it didn't make it in.

[nalimilan]

Fixed by #870.