Merge pull request #15 from s-ccs/feature-unpairedt

feature unpairedt

Author: behinger

src/utils.jl

@@ -73,7 +73,24 @@ function sign_permute!(rng, x::AbstractArray)
     return x
 end
 
-function batch_unpaired_ttest_unequal_var(x::AbstractArray, group)
+"""
+    studentt_unpaired(x::AbstractArray, group)
+
+Implements a unpaired two groups t-test with unequal variances. 10x as fast as HypothesisTests because we don't allocate that much
+Use like this:
+
+```julia
+studentt_unpaired([1,2,1,1,4,5,4],[false,false,false,false,true,true,true])
+```
+
+To use with ClusterDepth, you have to "bake-in" the group membership by defining your own method:
+```julia
+grp = ["bla","bla","bla","bla","blub","blub","blub"] .== "blub"
+my_statfun = x->studentt_unpaired(x,grp)
+```
+
+"""
+function studentt_unpaired(x::AbstractArray, group)
     x_reshaped = reshape(x, :, size(x, ndims(x)))
     x₁ = x_reshaped[:, group]
     x₂ = x_reshaped[:, .!group]

test/utils.jl

@@ -24,6 +24,9 @@ end
 @testset "studentt" begin
     x = randn(StableRNG(1), 10000, 50)
     t = ClusterDepth.studentt(x)
+
+    t_true = [HypothesisTests.OneSampleTTest(r).t for r in eachrow(x)]
+    @test all(t .≈ t_true)
     @test length(t) == 10000
     @test maximum(abs.(t)) < 10 # we'd need to be super lucky ;)
     @test mean(abs.(t) .> 2) < 0.06
@@ -40,6 +43,7 @@ end
     t = rand(10000)
     ClusterDepth.studentt!(t, x)
     @test t ≈ ClusterDepth.studentt(x)
+
     @test length(t) == 10000
     @test maximum(abs.(t)) < 10 # we'd need to be super lucky ;)
     @test mean(abs.(t) .> 2) < 0.06
@@ -54,3 +58,27 @@ end
     data = randn(StableRNG(1), 3, 4, 5)
     @test size(ClusterDepth.studentt(data)) == (3, 4)
 end
+
+
+@testset begin
+    "ttest unpaired"
+    _x = randn(StableRNG(1), 10000, 50, 2)
+    x1 = _x[:, :, 1]
+    x2 = _x[:, :, 2]
+    x = hcat(x1, x2)
+    group = repeat([false, true], inner=size(x1, 2))
+    t = ClusterDepth.studentt_unpaired(x, group .== 1)
+
+    t_true = [HypothesisTests.UnequalVarianceTTest(r[group], r[.!group]).t for r in eachrow(x)]
+    @test all(t .≈ t_true)
+    @test length(t) == 10000
+
+
+    _x = randn(StableRNG(1), 10000, 50, 30, 2)
+    x1 = _x[:, :, :, 1]
+    x2 = _x[:, :, :, 2]
+    x = cat(x1, x2, dims=3)
+    group = repeat([false, true], inner=size(x1, 3))
+    t = ClusterDepth.studentt_unpaired(x, group .== 1)
+    @test size(t) == (1000, 50)
+end