poking a bit at apply_data and ModelCols idea...

src/StatsModels.jl

@@ -63,5 +63,6 @@ include("temporal_terms.jl")
 include("formula.jl")
 include("modelframe.jl")
 include("statsmodel.jl")
+include("data.jl")
 
 end # module StatsModels

src/data.jl

@@ -0,0 +1,29 @@
+# Notes:
+#
+# Add an apply_data step which puts a link to a table on each term, wrapping it
+# in a ModelCols struct.  Propogate missing mask up through the terms to allow
+# e.g. lead/lag to say that they're going to generate missings
+
+mutable struct ModelCols{T<:TermOrTerms,D}
+    term::T
+    data::D
+end
+
+function apply_data end
+
+apply_data(t::AbstractTerm, data) = ModelCols(t, data)
+apply_data(t::InteractionTerm, data) =
+    ModelCols(InteractionTerm(apply_data.(t.terms, Ref(data))), data)
+apply_data(t::FormulaTerm, data) = ModelCols(FormulaTerm(apply_data(t.lhs, data),
+                                                         apply_data(t.rhs, data)),
+                                             data)
+apply_data(t::MatrixTerm, data) = apply_data.(t.terms, Ref(data))
+
+
+width(mc::ModelCols) = width(mc.term)
+
+import Base.getindex
+
+getindex(mc::ModelCols{<:CategoricalTerm}, i, j) =
+    mc.term.contrasts[mc.data[mc.term.sym, i], j]
+

test/data.jl

@@ -0,0 +1,17 @@
+using StatsModels, StatsBase
+
+using StatsModels: apply_data, ModelCols
+
+f = @formula(y ~ 1 + a*b)
+
+d = (y = rand(20),
+     a = rand(20),
+     b = sample('a':'d', 5))
+
+ff = apply_schema(f, schema(d))
+
+fff = apply_data(ff, d)
+
+b = fff.term.rhs[end-1]
+
+b[:, :]