Crumble

Description

Flexible and general causal mediation analysis using Riesz representers and deep learning.

Usage

crumble(
    data::DataFrame;
    trt::Vector{String},
    outcome::String,
    mediators::Vector{String},
    moc::Union{Vector{String}, Nothing}=nothing,
    covar::Vector{String},
    obs::Union{String, Nothing}=nothing,
    id::Union{String, Nothing}=nothing,
    d0::Union{Function, Nothing}=nothing,
    d1::Union{Function, Nothing}=nothing,
    effect::String="RT",
    weights::Vector{Float64}=ones(nrow(data)),
    learners::Vector{String}=["glm"],
    nn_module=sequential_module(),
    control::CrumbleControl=crumble_control()
)

Arguments

Argument	Type	Description
data	DataFrame	Data in wide format
trt	Vector{String}	Treatment variable names
outcome	String	Outcome variable name
mediators	Vector{String}	Mediator variable names
moc	Union{Vector{String}, Nothing}	Mediator-outcome confounder names (required for RT, RI)
covar	Vector{String}	Baseline covariate names
obs	Union{String, Nothing}	Outcome observation indicator (0/1)
id	Union{String, Nothing}	Cluster ID
d0	Union{Function, Nothing}	Shift function for control
d1	Union{Function, Nothing}	Shift function for treatment
effect	String	Effect type: “N”, “O”, “RI”, or “RT”
weights	Vector{Float64}	Survey weights
learners	Vector{String}	MLJ learner algorithms
nn_module	Function	Neural network architecture
control	CrumbleControl	Control parameters

Value

Returns a CrumbleResult with:

• estimates — Dictionary of estimates with SE, CI, p-values
• outcome_reg — Outcome regression predictions
• alpha_n — Natural density ratio estimates
• alpha_r — Randomized density ratio estimates
• fits — Fitted values
• effect — Effect type

Examples

using Crumble
using DataFrames

data = DataFrame(
    A = rand([0, 1], 100),
    Y = rand(100),
    M = rand([0, 1], 100),
    W1 = randn(100),
)

result = crumble(
    data = data,
    trt = ["A"],
    outcome = "Y",
    mediators = ["M"],
    covar = ["W1"],
    effect = "N"
)

print(result)