wrapper returns more now

README.md

@@ -5,3 +5,33 @@
 [![Coverage Status](https://coveralls.io/repos/magerton/OrderedResponse.jl/badge.svg?branch=master&service=github)](https://coveralls.io/github/magerton/OrderedResponse.jl?branch=master)
 
 [![codecov.io](http://codecov.io/github/magerton/OrderedResponse.jl/coverage.svg?branch=master)](http://codecov.io/github/magerton/OrderedResponse.jl?branch=master)
+
+Estimates ordered logit & probit models. Returns a tuple with the negative log likelihood, gradient, hessian, and the `Optim.MultivariateOptimizationResults` object.
+
+# Example
+```julia
+using OrderedResponse
+using DataFrames
+using Optim
+
+# ------- make some data ----------
+
+srand(1234)
+
+n,k = 1000,2
+X = randn(n, k)
+ϵ = randn(n)
+
+β = [0.2, -1.0]
+γ = [-0.4, 0.5]
+
+ystar = X*β + ϵ
+y = map((yi) -> searchsortedfirst(γ,yi), ystar)
+
+df = convert(DataFrame, X)
+df[:y] = y
+
+# ------- run model ----------
+@show orlm(@formula(y ~ 0 + x1 + x2), df, :logit)
+@show orlm(@formula(y ~ 0 + x1 + x2), df, :probit)
+```

src/OrderedResponse.jl

@@ -1,3 +1,5 @@
+__precompile__()
+
 module OrderedResponse
 
 using StatsFuns

src/likelihood.jl

@@ -46,16 +46,16 @@ function LLi!(grad::Vector{T}, hess::Matrix{T}, tmpgrad::Vector, l::Integer, η:
   η2 = l == L+1 ?  Inf : γ[l]   - η
 
   # LL takes a short-cut to improve accuracy if l == 1 or L+1
-  # if l == 1
-  #     F = cdf(η2, model)
-  #     LL = logcdf(η2, model)
-  # elseif l == L+1
-  #     F = ccdf(η1, model)
-  #     LL = logccdf(η1, model)
-  # else
+  if l == 1
+      F = cdf(η2, model)
+      LL = logcdf(η2, model)
+  elseif l == L+1
+      F = ccdf(η1, model)
+      LL = logccdf(η1, model)
+  else
       F = cdf(η2, model) - cdf(η1, model)
       LL = log(F)
-  # end
+  end
 
   if length(grad) > 0
       dlogF1 = pdf(η1, model) / F

src/outer-wrapper.jl

@@ -5,7 +5,12 @@ response_vec(y::DataVector{<:Integer}) = Vector(y)
 response_vec(mf::ModelFrame) = response_vec(mf.df[:,mf.terms.eterms[1]])
 
 """
-    Estimate ordered response of variety `model ∈ (:logit, :probit)` with specification `fm` and `data`. Optim options passed with `kwargs`
+    orlm(fm::Formula, data::DataFrame, model::Symbol; method=Optim.Newton(), kwargs...)
+
+Estimate ordered response model of variety `model ∈ (:logit, :probit)` with specification
+`fm` and `data`. Optim options passed with `kwargs`. Set the optimizer with `method`.
+Returns a tuple with negative log likelihood, gradient, hessian, and optimization results
+`(LL::Real, grad::Vector, hess::Matrix, opt::Optim.MultivariateOptimizationResults)`
 """
 function orlm(fm::Formula, data::DataFrame, model::Symbol; method=Optim.Newton(), kwargs...)
 
@@ -38,6 +43,14 @@ function orlm(fm::Formula, data::DataFrame, model::Symbol; method=Optim.Newton()
         (hess::Matrix, θ::Vector) -> LL!(zeros(T,KLm1),         hess, tmpgrad , η, y, X, θ[1:k], θ[k+1:end], Val{model}, -1.0)
     )
 
-    return Optim.optimize(td, θ0, method, Optim.Options(;kwargs...))
+    opt =  Optim.optimize(td, θ0, method, Optim.Options(;kwargs...))
+
+    # final eval
+    hess = zeros(T, KLm1, KLm1)
+    grad = similar(tmpgrad)
+    θfinal = opt.minimizer
+    LL = LL!(grad, hess, tmpgrad, η, y, X, θfinal[1:k], θfinal[k+1:end], Val{model}, -1.0)
+
+    return (LL, grad, hess, opt)
 
 end