Better missing handling

src/contrasts.jl

@@ -229,8 +229,13 @@ function termnames(C::AbstractContrasts, levels::AbstractVector, baseind::Intege
     levels[not_base]
 end
 
-Base.getindex(contrasts::ContrastsMatrix, rowinds, colinds) =
-    getindex(contrasts.matrix, getindex.(Ref(contrasts.invindex), rowinds), colinds)
+function Base.getindex(contrasts::ContrastsMatrix{C,T}, rowinds, colinds) where {C,T}
+    # allow rows to be missing
+    rows = get.(Ref(contrasts.invindex), rowinds, missing)
+    # create a row of nothing but missings for missing values
+    mrow = reduce(vcat, [missing for c in getindex(contrasts.matrix, 1, colinds)])
+    vcat([r === missing ? mrow : getindex(contrasts.matrix, r, colinds) for r in rows])
+end
 
 # Making a contrast type T only requires that there be a method for
 # contrasts_matrix(T,  baseind, n) and optionally termnames(T, levels, baseind)

src/modelframe.jl

@@ -78,11 +78,11 @@ function ModelFrame(f::FormulaTerm, data::ColumnTable;
         throw(ArgumentError(msg))
     end
 
-    data, _ = missing_omit(data, f)
-
     sch = schema(f, data, contrasts)
     f = apply_schema(f, sch, M)
-
+
+    data, _ = missing_omit(data, f)
+
     ModelFrame(f, sch, data, model)
 end
 

src/schema.jl

@@ -198,6 +198,8 @@ concrete_term(t::Term, x, hint::AbstractTerm) = hint
 concrete_term(t, d, hint) = t
 
 concrete_term(t::Term, xs::AbstractVector{<:Number}, ::Nothing) = concrete_term(t, xs, ContinuousTerm)
+# and for missing values
+concrete_term(t::Term, xs::AbstractVector{Union{Missing,T}} where T<:Number, ::Nothing) = concrete_term(t, xs, ContinuousTerm)
 function concrete_term(t::Term, xs::AbstractVector, ::Type{ContinuousTerm})
     μ, σ2 = StatsBase.mean_and_var(xs)
     min, max = extrema(xs)

src/terms.jl

@@ -506,6 +506,13 @@ lazy_modelcols(x, d) = modelcols(x, d)
 
 
 
+# this is weird, but using import Base: copy leads to exporting type piracy
+# for non missing values, the compiler should hopefully optimize down the extra
+# layer of indirection
+function copy end
+copy(x::Any) = Base.copy(x)
+copy(m::Missing) = m
+
 modelcols(t::ContinuousTerm, d::NamedTuple) = copy.(d[t.sym])
 
 modelcols(t::CategoricalTerm, d::NamedTuple) = t.contrasts[d[t.sym], :]

test/terms.jl

@@ -31,6 +31,17 @@ StatsModels.apply_schema(mt::MultiTerm, sch::StatsModels.Schema, Mod::Type) =
         @test t0.min == 1.0
         @test t0.max == 3.0
 
+        vals0m = [3, missing, 1]
+        t0m = concrete_term(t, vals0m)
+        @test string(t0m) == "aaa"
+        @test mimestring(t0m) == "aaa(continuous)"
+        # compute all these values to make sure the behavior of terms matches
+        # the behavior of other relevant packages
+        @test isequal(t0m.mean, mean(vals0m))
+        @test isequal(t0m.var, var(vals0m))
+        @test isequal(t0m.min, min(vals0m...))
+        @test isequal(t0m.max, max(vals0m...))
+
         t1 = concrete_term(t, [:a, :b, :c])
         @test t1.contrasts isa StatsModels.ContrastsMatrix{DummyCoding}
         @test string(t1) == "aaa"