belwhat / dragonnet

Introduction

This repository contains software and data for "Adapting Neural Networks for the Estimation of Treatment Effects".

The paper describes approaches to estimating causal effects from observational data using neural networks. The high-level idea is to modify standard neural net design and training in order to induce a bias towards accurate estimates.

Requirements and setup

You will need to install tensorflow 1.13, sklearn, numpy 1.15, keras 2.2.4 and, pandas 0.24.1

Data

1. IHDP This dataset is based on a randomized experiment investigating the effect of home visits by specialists on future cognitive scores. It is generated via the npci package https://github.com/vdorie/npci (setting A) For convenience, we have also uploaded a portion of the simulated data in the dat folder. This can be used for testing the code.

2. ACIC ACIC is a collection of semi-synthetic datasets derived from the linked birth and infant death data (LBIDD)

• Here is the full dataset description https://www.researchgate.net/publication/11523952_Infant_Mortality_Statistics_from_the_1999_Period_Linked_BirthInfant_Death_Data_Set
• Here is the GitHub repo associated with the competition https://github.com/IBM-HRL-MLHLS/IBM-Causal-Inference-Benchmarking-Framework/blob/master/data/LBIDD/scaling_params.csv
• For access to the ACIC 2018 competition data: Please see here https://www.synapse.org/#!Synapse:syn11294478/wiki/486304

Reproducing neural net training for IHDP experiments

The default setting would let you run Dragonnet, TARNET, and NEDnet under targeted regularization and default mode

You'll run the from src code as ./experiment/run_ihdp.sh Before doing this, you'll need to edit run_ihdp.sh and change the following: data_base_dir= where you stored the data output_base_dir=wherer you want the result to be

If you only want to run one of the frameworks, delete the rest of the options in run_ihdp.sh

Reproducing neural net training for the ACIC experiment

Same as above except you run the from src code as ./experiment/run_acic.sh

Computing the ATE

All of the estimators functions are in semi_parametric_estimation.ate

To reproduce the table in the paper: i) get the neural net predictions; ii) update the output file location in ihdp_ate.py iii) run ihdp_ate.py. The make_table function should generate the mean absolute error for each framework.

Note: the default code use all the data for prediction and estimation. If you want to get the in-sample or out-sample error: i) change the train_test_split criteria in ihdp_main.py; ii) rerun the neural net training; iii) run ihdp_ate.py with apporiate in-sample data and out-sample data.