This packages provides a simple python implementation of Invariant Causal Prediction (ICP) [1].
The source code for the actual algorithm resides in ./src/icp/ICP.py.
See also the original implementation in the R package InvariantCausalPrediction.
pip install ICPyimport icpy as icpy
import numpy as np
np.random.seed(seed=1)
n = 100
noise = 0.1
E = np.repeat([0, 1, 2], np.ceil(n / 3.0))[0:n] # "Environment"
A = np.random.normal(scale=noise, size=[n]) + np.equal(E, 1) # Node A
B = A + np.random.normal(scale=noise, size=[n]) / 3 + np.equal(E, 2) # Node B
C = B + np.random.normal(scale=noise, size=[n]) # Node C
# /--->---\
# E -> A -> B -> C
icpy.invariant_causal_prediction(X=np.column_stack((A, B)), y=C, z=E) # test if A or B are parents of COutput
ICP(S_hat=array([1], dtype=int64), # Column 1 = Node B was (correctly) identified as parent of C
p_values=array([ 1.51508232e-01, 4.59577055e-37]), # error levels at which A and B would/are indentied as parent of C
p_value=0.16416488336322549) # p-value for testing against violation of the model assumptions (e.g. a direct effect of E on C)v0.0.003 (2020-05-15)
- fix failing import (thanks to @lgmoneda, #1)
- fix issues when environments are not subsequent whole numbers starting at 0 (thanks to @lgmoneda, #1)
[1] J. Peters, P. Bühlmann, N. Meinshausen, Causal inference by using invariant prediction: identification and confidence intervals, J. R. Stat. Soc. Ser. B Stat. Methodol. 78 (2016) 947-1012. doi:10.1111/rssb.12167.
