Fixed bounds issues with inference

src/estimators.jl

@@ -145,7 +145,7 @@ mutable struct GComputation <: CausalEstimator
     """Weights learned during training"""
     β::Array{Float64}
     """The effect of exposure or treatment"""
-    causal_effect::Vector{Float64}
+    causal_effect::Float64
 
 """
 GComputation(X, Y, T, task, quantity_of_interest, regularized, activation, temporal, 
@@ -233,7 +233,7 @@ mutable struct DoublyRobust <: CausalEstimator
     """Predicted outcomes for the treatment group"""
     μ₁::Array{Float64}
     """The effect of exposure or treatment"""
-    causal_effect::Vector{Float64}
+    causal_effect::Float64
 
     """
 DoublyRobust(X, Xₚ, Y, T, task, quantity_of_interest, regularized, activation, 
@@ -362,7 +362,7 @@ function estimatecausaleffect!(g::GComputation)
     end
 
     g.β = fit!(g.learner)
-    g.causal_effect = [sum(predict(g.learner, Xₜ) - predict(g.learner, Xᵤ))/size(Xₜ, 1)]
+    g.causal_effect = sum(predict(g.learner, Xₜ) - predict(g.learner, Xᵤ))/size(Xₜ, 1)
 
     return g.causal_effect
 end
@@ -446,7 +446,7 @@ function estimatecausaleffect!(DRE::DoublyRobust)
         DRE.μ₁ = reduce(vcat, treatment_predictions)
     end
 
-    DRE.causal_effect = [mean(fold_level_effects)]
+    DRE.causal_effect = mean(fold_level_effects)
     return DRE.causal_effect
 end
 

src/inference.jl

@@ -4,7 +4,7 @@ module Inference
 using CausalELM: mean
 using ..Metalearners: Metalearner
 using ..Estimators: CausalEstimator, InterruptedTimeSeries, GComputation, DoublyRobust, 
-    estimatecausaleffect!, mean
+    estimatecausaleffect!
 
 import CausalELM: summarize
 
@@ -226,7 +226,7 @@ function generatenulldistribution(e::Union{CausalEstimator, Metalearner}, n::Int
     for iter in 1:n 
         m.T = float(rand(0:1, nobs))
         estimatecausaleffect!(m)
-        results[iter] = ifelse(e isa Metalearner, mean(m.causal_effect), m.causal_effect[1])
+        results[iter] = e isa Metalearner ? mean(m.causal_effect) : m.causal_effect
     end
     return results
 end
@@ -311,9 +311,9 @@ julia> quantitiesofinterest(g_computer, 1000)
 (0.114, 6.953133617011371)
 ```
 """
-function quantitiesofinterest(model::Union{CausalEstimator, Metalearner}, n::Integer=1000)
-    local null_dist = generatenulldistribution(model, n)
-    local avg_effect = mean(model.causal_effect)
+function quantitiesofinterest(m::Union{CausalEstimator, Metalearner}, n::Integer=1000)
+    local null_dist = generatenulldistribution(m, n)
+    local avg_effect = m isa Metalearner ? mean(m.causal_effect) : m.causal_effect
 
     extremes = length(null_dist[abs(avg_effect) .< abs.(null_dist)])
     pvalue = extremes/n

src/model_validation.jl

@@ -1,7 +1,7 @@
 module ModelValidation
 
-using ..Estimators: InterruptedTimeSeries, estimatecausaleffect!
-using CausalELM: mean
+using ..Estimators: InterruptedTimeSeries, estimatecausaleffect!, GComputation
+using CausalELM: mean, var
 using LinearAlgebra: norm
 
 """
@@ -112,7 +112,7 @@ end
 See how an omitted predictor/variable could change the results of an interrupted time series 
 analysis.
 
-This method reestimates interrupted time series models with normal random variables with 
+This method reestimates interrupted time series models with normal random variables and 
 uniform noise. If the included covariates are good predictors of the counterfactual outcome, 
 adding a random variable as a covariate should not have a large effect on the predicted 
 counterfactual outcomes and therefore the estimated average effect.
@@ -260,6 +260,12 @@ function pval(x::Array{Float64}, y::Array{Float64}, β::Float64; n::Int=1000,
     return p
 end
 
+function counterfactualconsistency(g::GComputation)
+    treatment_covariates, treatment_outcomes = g.X[g.T == 1, :], g.Y[g.T == 1]
+    ŷ = treatment_covariates\treatment_outcomes
+    observed_residual_variance = var(ŷ)
+end
+
 """
     ned(a, b)
 
@@ -283,6 +289,8 @@ function ned(a::Vector{T}, b::Vector{T}) where T <: Number
             a = reduce(vcat, (a, zeros(abs(length(a)-length(b)))))
         end
     end
+
+    # Changing NaN to zero fixes divde by zero errors
     @fastmath norm(replace(sort(a)./norm(a), NaN=>0) .- replace((sort(b)./norm(b)), NaN=>0))
 end
 

test/test_estimators.jl

@@ -81,10 +81,10 @@ end
 
 @testset "G-Computation Estimation" begin
     @test isa(g_computer.β, Array)
-    @test isa(g_computer.causal_effect, Vector{Float64})
+    @test isa(g_computer.causal_effect, Float64)
 
     # Check that the estimats for ATE and ATT are different
-    @test g_computer.causal_effect[1] !== gcomputer_att.causal_effect[1]
+    @test g_computer.causal_effect !== gcomputer_att.causal_effect
 end
 
 @testset "Doubly Robust Estimation Structure" begin
@@ -127,23 +127,23 @@ end
     @test dr.ps isa Array{Float64}
     @test dr.μ₀ isa Array{Float64}
     @test dr.μ₁ isa Array{Float64}
-    @test dr.causal_effect isa Vector{Float64}
+    @test dr.causal_effect isa Float64
 
     # No regularization
     @test dr_noreg.ps isa Array{Float64}
     @test dr_noreg.μ₀ isa Array{Float64}
     @test dr_noreg.μ₁ isa Array{Float64}
-    @test dr_noreg.causal_effect isa Vector{Float64}
+    @test dr_noreg.causal_effect isa Float64
 
     # Using the ATT
     @test dr_att.ps isa Array{Float64}
     @test dr_att.μ₀ isa Array{Float64}
-    @test dr_att.causal_effect isa Vector{Float64}
+    @test dr_att.causal_effect isa Float64
 
     # Using the ATT with no regularization
     @test dr_att_noreg.ps isa Array{Float64}
     @test dr_att_noreg.μ₀ isa Array{Float64}
-    @test dr_att_noreg.causal_effect isa Vector{Float64}
+    @test dr_att_noreg.causal_effect isa Float64
 end
 
 @testset "Quanities of Interest Errors" begin