Error in Terms when no intercept is requested

[andreasnoack]

This does not work

df = DataFrame(A = 1:4, B = ["M", "F", "F", "M"]);
ModelMatrix(ModelFrame(A~B-1, df))
ERROR: no method convert(Type{Expr}, Array{Any,1})
 in unshift! at array.jl:511
 in Terms at /Users/andreasnoackjensen/.julia/v0.3/DataFrames/src/formula/formula.jl:145
 in ModelFrame at /Users/andreasnoackjensen/.julia/v0.3/DataFrames/src/formula/formula.jl:174

but this does

julia> ModelMatrix(ModelFrame(A~B+(-1), df))
ModelMatrix{Float64}(4x1 Array{Float64,2}:
 1.0
 0.0
 0.0
 1.0,[1])

[ararslan]

This has been fixed now, hasn't it?

[nalimilan]

Still doesn't work AFAICT (even after converting B to a CategoricalArray).

[kleinschmidt]

It's the -1 that's the issue I think. +0 should work (and have the same effect).

[nalimilan]

But do we still want to support -1?

[ararslan]

It might make more sense to have a single, consistent means of requesting no intercept, and IMO +0 is nicer than -1.

[kleinschmidt]

I agree. And, in fact, we might go so far as to make no intercept the default, so you'd need to explicitly have +1 to add an intercept.

[ararslan]

On the one hand, I like that because it requires being explicit (which is almost always a good thing) and will require people to think harder about the exact model they're fitting, but on the other hand I think it could be surprising for some people since implicit intercept is the default in R, SAS, and likely others as well. So... I'm undecided on that, but leaning toward +1 require +1.

[dmbates]

I think that forcing the user to use an explicit +1 is a good idea. The fact that earlier languages made the intercept the default doesn't mean that the choice is appropriate now. Including 1 + in the formula is just a matter of typing a few characters. @StefanKarpinski has mentioned before that having 1 + x == x and 0 + x != x in the formula language looks like a wart.

[nalimilan]

We've had that discuss in many different places already, but I'll restate my preferred solution: require people to either say 1 + or 0 +, else raise an error with an explanation. That's the only way to get rid of the implicit intercept without trapping people into fitting weird models.

[ararslan]

+1 to Milan's proposal. And if we end up not doing that, +1 to Doug's.

[johnmyleswhite]

Explicitness is good by me, so I like Milan's solution.