allow specifying type argument in GLMM predict on original data

NEWS.md

@@ -9,6 +9,7 @@ MixedModels v5.0.0 Release Notes
 - Internal code around optimization in profiling has been restructuring so that fitting done during calls to `profile` respect the `backend` and `optimizer` settings. [#853]
 - The `prfit!` convenience function has been removed. [#853]
 - The `dataset` and `datasets` functions have been removed. They are now housed in `MixedModelsDatasets`.[#854]
+- One argument `predict(::GeneralizedLinearMixedModel)`, i.e. prediction on the original data, now supports the `type` keyword argument. [#856]
 
 MixedModels v4.38.0 Release Notes
 ==============================

src/predict.jl

@@ -86,6 +86,11 @@ function StatsAPI.predict(
     return type == :linpred ? y : broadcast!(Base.Fix1(linkinv, Link(m)), y, y)
 end
 
+function StatsAPI.predict(m::GeneralizedLinearMixedModel; type=:response)
+    type in (:linpred, :response) || throw(ArgumentError("Invalid value for type: $(type)"))
+    return type == :response ? fitted(m) : m.resp.eta
+end
+
 # β is separated out here because m.β != m.LMM.β depending on how β is estimated for GLMM
 # also β should already be pivoted but NOT truncated in the rank deficient case
 function _predict(m::MixedModel{T}, newdata, β; new_re_levels) where {T}

test/fit.jl

@@ -77,11 +77,10 @@ end
 @testset "generalized" begin
     gm1 = fit(MixedModel, @formula(use ~ 1 + urban + livch + age + abs2(age) + (1 | dist)),
         MixedModels.dataset(:contra), Bernoulli(); progress=false)
-    @test deviance(gm1) ≈ 2372.7286 atol = 1.0e-3
-
     gm2 = glmm(@formula(use ~ 1 + urban + livch + age + abs2(age) + (1 | dist)),
         MixedModels.dataset(:contra), Bernoulli(); progress=false)
-    @test deviance(gm2) ≈ 2372.7286 atol = 1.0e-3
+
+    @test deviance(gm1) ≈ deviance(gm2)
 end
 
 @testset "Normal-IdentityLink" begin

test/predict.jl

@@ -197,6 +197,8 @@ end
             # we can skip a lot of testing if the broad strokes work because
             # internally this is punted off to the same machinery as LMM
             @test predict(gm0) ≈ fitted(gm0)
+            @test predict(gm0; type=:response) ≈ fitted(gm0)
+            @test predict(gm0; type=:linpred) ≈ GLM.linkfun.(Ref(Link(gm0)), fitted(gm0))
             # XXX these tolerances aren't great but are required for fast=false fits
             @test predict(gm0, contra; type=:linpred) ≈ gm0.resp.eta rtol = 0.1
             @test predict(gm0, contra; type=:response) ≈ gm0.resp.mu rtol = 0.01