ananke exploration

ananke.py

# pip install ananke-causal
from ananke import graphs
from ananke import identification
from ananke.estimation import CausalEffect

import numpy as np
import pandas as pd

# Simulate front-door situation with confounder Z
N = 100000
z = np.random.normal(size=N)
x = .8 * z + np.random.normal(size=N) > 0
x = x.astype(int)

m = .5 * x + np.random.normal(size=N)
y = .7 * z + 1.2 * m + np.random.normal(size=N)

df = pd.DataFrame({'X': x, 'Y': y, "M": m, "Z": z})

# Note that this is a front door graph where M is the mediator
# DAG example with a single confounder and a front-door path
vertices = ['X', 'Z', 'Y', 'M']
edges = [
    ('X', 'M'), ('M', 'Y'),  # Mediation path
    ('Z', 'X'), ('Z', 'Y'),  # Confounding path
]
dag = graphs.DAG(vertices, edges)
dag_graph = dag.draw(direction='LR')  # Need Graphviz installed
dag_graph.view(filename="front_door")  # Wait for the browser to open (20 seconds)
id_pya = identification.OneLineID(
    graph=dag,
    treatments=['X'],
    outcomes=['Y']
)
id_pya.id()  # Is it identified?
id_pya.functional()  # The Functional (have no idea what this is)

ACE_estimand = CausalEffect(graph=dag, treatment='X', outcome='Y')
ace = ACE_estimand.compute_effect(df, "eff-aipw")
print(f"truth = {1.2 * .5} vs est = {np.round(ace, 4)} for {N=}")

# ADMG front-door example with covariance instead of observed Z
vertices = ['X', 'Y', 'M']
di_edges = [('X', 'M'), ('M', 'Y')]  # Mediation path
bi_edges = [('X', 'Y')]  # Confounding path
admg = graphs.ADMG(vertices, di_edges=di_edges, bi_edges=bi_edges)
digraph = admg.draw(direction='LR')
digraph.view(filename="front_door_ADMG")  # Wait for the browser to open (20 seconds)
id_pya = identification.OneLineID(graph=admg, treatments=['X'],
                                  outcomes=['Y'])
id_pya.id()  # Is it identified?
id_pya.functional()

ACE_estimand2 = CausalEffect(graph=admg, treatment='X', outcome='Y')

ace = ACE_estimand2.compute_effect(df, "eff-aipw")  # This doesn't work
ace2 = ACE_estimand2.compute_effect(df, "p-ipw")
ace3 = ACE_estimand2.compute_effect(df, "d-ipw")
ace4 = ACE_estimand2.compute_effect(df, "apipw")
print(f"truth = {1.2 * .5} vs est = {np.round(ace3, 3)} for {N=}")