Bayesian causal inference models for healthcare treatment using real world patient data
Computer implemented methods, systems, and computer readable medium are provided for performing causal inference analyses to determine the more effective treatment among alternative treatments in the healthcare setting using real world observational data. Both binary treatment and adaptive treatment strategies are considered in the analysis. The methods comprise generating a Bayesian marginal structural model and performing a single step of Bayesian regression that incorporates matching, weighting, and estimation processes and in which the matching process is performed using a Guassian process (“GP”) prior covariance function.
1. A method for treating juvenile idiopathic arthritis (JIA) in a subject in need thereof, the method comprising
implementing a Bayesian nonparametric causal inference method for determining the more effective treatment among two alternative JIA treatments administered to a plurality of JIA patients by a non-transitory machine readable storage medium storing at least one program that when executed by at least one processor, causes the at least one processor to implement the Bayesian nonparametric causal inference method, the method comprising generating a marginal structural model and performing a single step of Bayesian regression on patient data of the plurality of patients, wherein the method incorporates matching, weighting, and estimation processes into the single regression step and wherein the matching process is performed using a Gaussian process (“GP”) prior covariance function; and
administering to the subject the JIA treatment determined by the Bayesian nonparametric causal inference method to be more effective, wherein the JIA treatment is either a combination of biologic disease modifying antirheumatic drugs (bDMARDs) and conventional synthetic disease modifying antirheumatic drugs (csDMARDs) or csDMARDs alone, without administration of bDMARDs, wherein the csDMARDs comprise methotrexate and the bDMARDs comprise adalimumab or etanercept.
2. The method of claim 1 , wherein the matching process is performed in the absence of a known matching structure.
3. The method of claim 2 , wherein the GP prior covariance function is a squared exponential (“SE”) function.
4. The method of claim 3 , wherein potential outcomes are modeled nonparametrically and missing potential outcomes are estimated by a weighted sum of observed data.
5. The method of claim 4 , wherein the method comprises generating a model by initializing a treatment effect parameter and a GP covariance matrix in which, for each individual patient (i-th) in the patient data, the GP prior allocates different weights to information obtained from other individual (j-th) patients, based on confounding variables specified in GP prior.
6. The method of claim 4 , further comprising fitting the model using Bayesian Markov chain Monte Carlo (MCMC) sampling of posterior distributions.
7. The method of claim 4 , further comprising performing the matching process by determining for each patient in the plurality a similarity to every other patient in the plurality and assigning a matching weight to each patient.
8. The method of claim 7 , wherein the matching process comprises calculating weighted average covariates after weighting and before weighting and determining a balance for all covariates across their distributions by treatment group.
9. The method of claim 7 , wherein the balance is determined by the mean absolute difference and median absolute deviance.
10. The method of claim 7 , wherein if the covariates are unbalanced, repeating the step of fitting the model using Bayesian MCMC, thereby generating a weighted sum of observed data.
11. The method of claim 4 , wherein the weighted sum of the observed data is used to estimate missing potential outcomes for each patient utilizing data from other ‘matched’ patients who are sufficiently similar.
12. The method of claim 1 , wherein the method further comprises deriving a propensity score using the covariate balancing propensity score (CBPS) method to ensure balance on identified clinically important covariates.