Code and tutorials for our propensity-based causal inference methods designed to estimate treatment effects for massively multi-arm studies using real-world data.
The core methodology is a novel neural network approach which uses
a custom loss function to explicitly balance covariates in addition
to specifying propensity for IPTW ATE estimations and is described
in the manuscript titled:
"Minimizing Bias in Massive Multi-Arm Observational Studies with BCAUS:
Balancing Covariates Automatically Using Supervision".
The code can be found in bcaus.py and a demonstration of its usage
can be found in BCAUS_demo.ipynb
We've extended the core methodology by including a network meta-analysis
(NMA) after BCAUS that combines pair-wise observations of
control-treatment experiments to leverage both direct and indirect
comparisons. Our NMA implementation is validated using R's netmeta
package (see NMA_demo.ipynb). This enhanced approached is what was
used in the manuscript entitled
"Causal Deep Learning on Real-world Data Reveals the Comparative
Effectiveness of Anti-hyperglycemic Treatments"
(henceforth referred to as paper)
(assuming a conda installation is available)
git clone https://github.com/gstef80/bcaus_nma.gitconda env create -f environment.yml(will create a conda environment named 'bcaus_nma')jupyter notebook(will start a Jupyter notebook server)
# 10 SME-defined clinical cohorts
for cohort in clinical_cohorts:
# treatments with cohort sizes above threshold
treatments = cohort_treatments(cohort, threshold=35)
ate_estimates = []
for ct in treatments:
for tx in treatments:
# perform pair-wise BCAUS experiments
ate = BCAUS(cohort, ct, tx)
ate_estimates.append(ate)
# perform NMA on ATEs
ranks = NMA(ate_estimates)- trace_samples=100000
- burn_ratio=0.5
- num_chains=4
The choice of these values was based on NICE DSU TECHNICAL SUPPORT DOCUMENT 2 (http://nicedsu.org.uk/wp-content/uploads/2017/05/TSD2-General-meta-analysis-corrected-2Sep2016v2.pdf).