Bayesian Regression Tree Models for Causal Inference: Regularization, Confounding, and Heterogeneous Effects (with Discussion)

Hahn, P. Richard; Murray, John S.; Carvalho, Carlos M.

doi:10.1214/19-ba1195

Cited by 179 publications

(203 citation statements)

References 103 publications

Supporting

Mentioning

200

Contrasting

Unclassified

Order By: Relevance

“…Both the model for potential outcomes and PS (i.e., π(X i ) = E(Z i |X i )) are fitted by BART with the default setting. Direct/ps-BART incorporates the estimate of PS in the specification of the outcome model, implicitly inducing a covariate-dependent prior to the regression function, which even outperforms the BART model ( [25,40]. .…”

Section: Additional Methodsmentioning

confidence: 99%

See 1 more Smart Citation

Estimating the effect of adjuvant chemo-therapy for colon-cancer using registry data: a method comparison and validation

Shen¹,

Visser²,

Wilt³

et al. 2020

Preprint

View full text Add to dashboard Cite

Background: Although randomized controlled trials (RCT) are the gold standard to estimate treatment effects, they are often criticized in terms of generalizability. Observational data might alleviate this problem by being readily available in large quantities. However, observational data are potentially confounded. In this methodological study we use a large-scale RCT as the gold standard to examine the performance of various statistical methods to control for potential confounding in observational data. Methods: In this paper we compare three types of methods that allow researchers to correct for such potential confounding: direct methods, inverse probability weighting (IPW) methods and doubly robust (DR) methods. We uniquely compare estimates obtained from the population-wide Netherlands Cancer Registry (NCR) on colon cancer (n=52621) with estimates obtained from a large-scale RCT. As the RCT differs from the observational data both in its sampling mechanism and in its treatment assignment mechanism, we first resample the NCR data to reflect the distribution in RCT data. Next, we correct for potential confounding using three alternative types of methods and consequentially evaluate their estimates to those obtained in the RCT. Results: We find that while all estimators qualitatively approximate to findings in the RCT, methods that can flexibly model the response (i.e., direct estimation and DR estimation) performed consistently superior to the inverse propensity score method. Subgroup analysis indicates that relatively simple models allow us to properly estimate the treatment effect. However, these simple models do not properly identify heterogeneous treatment effects in stage2 colon cancer. Careful sensitivity analysis using more flexible models demonstrates both the uncertainty and the potential heterogeneous treatment effect in stage2 cancer and provides robust estimation of treatment effect in stage3 cancer. Conclusions: Our results suggest that both the direct method and the DR method, when executed with care, can be used to reliably estimate treatment effects based on registry data. This methodological validation opens the door to more extensive use of registry data for the estimation of (individual) treatment effects. Additionally, our identification of potentially meaningful subgroups of stage2 colon cancer patients who, based on our analysis seem to benefit from chemotherapy, should be further explored.

show abstract

Section: Additional Methodsmentioning

confidence: 99%

“…Recently, more flexible semi-parametric and non-parametric models have demonstrated good performance [14]. In particular, BART seems well suited to uncover the true treatment effect when the treatment assignment is confounded [14,[25][26][27].…”

Section: Introductionmentioning

confidence: 99%

Estimating the effect of adjuvant chemo-therapy for colon-cancer using registry data: a method comparison and validation

Shen¹,

Visser²,

Wilt³

et al. 2020

Preprint

View full text Add to dashboard Cite

show abstract

“…The latter authors introduced XGBoost, an especially fast and scalable version that has gone on to dominate in a number of public data science competitions (Gorman 2017). Bayesian versions of boosting have also been developed (Chipman et al 2010) with extensions to causal inference (Hahn et al 2017;Hill 2011), survival analysis (Sparapani et al 2016), and high-dimensional modeling (Linero 2018;Linero and Yang 2018).…”

Section: Boostingmentioning

confidence: 99%

The Rhetoric and Reality of Anthropomorphism in Artificial Intelligence

Watson

2019

Minds & Machines

View full text Add to dashboard Cite

Artificial intelligence (AI) has historically been conceptualized in anthropomorphic terms. Some algorithms deploy biomimetic designs in a deliberate attempt to effect a sort of digital isomorphism of the human brain. Others leverage more general learning strategies that happen to coincide with popular theories of cognitive science and social epistemology. In this paper, I challenge the anthropomorphic credentials of the neural network algorithm, whose similarities to human cognition I argue are vastly overstated and narrowly construed. I submit that three alternative supervised learning methods-namely lasso penalties, bagging, and boosting-offer subtler, more interesting analogies to human reasoning as both an individual and a social phenomenon. Despite the temptation to fall back on anthropomorphic tropes when discussing AI, however, I conclude that such rhetoric is at best misleading and at worst downright dangerous. The impulse to humanize algorithms is an obstacle to properly conceptualizing the ethical challenges posed by emerging technologies.

show abstract

“…For example, Hill () proposed using Bayesian additive regression trees (BART) to flexibly model the response surface. To mitigate the regression‐induced confounding issue (Hahn et al, ) resulting in biased estimates of treatment effects in regularized models such as BART used in Hill (), Hahn et al () proposed a Bayesian random forest model which represents the regression as a sum of two components: one is a function of covariates and estimated propensity score (PS) as prognostic effect and the other is a function on treatment effects that depends on covariates and estimated PS. The model also allows the estimates of heterogeneous treatment effects.…”

Section: Introductionmentioning

confidence: 99%

A Bayesian Nonparametric Approach to Causal Inference on Quantiles

Daniels

Winterstein

2018

Biometrics

View full text Add to dashboard Cite

We propose a Bayesian nonparametric approach (BNP) for causal inference on quantiles in the presence of many confounders. In particular, we define relevant causal quantities and specify BNP models to avoid bias from restrictive parametric assumptions. We first use Bayesian additive regression trees (BART) to model the propensity score and then construct the distribution of potential outcomes given the propensity score using a Dirichlet process mixture (DPM) of normals model. We thoroughly evaluate the operating characteristics of our approach and compare it to Bayesian and frequentist competitors. We use our approach to answer an important clinical question involving acute kidney injury using electronic health records.

show abstract

Bayesian Regression Tree Models for Causal Inference: Regularization, Confounding, and Heterogeneous Effects (with Discussion)

Cited by 179 publications

References 103 publications

Estimating the effect of adjuvant chemo-therapy for colon-cancer using registry data: a method comparison and validation

Estimating the effect of adjuvant chemo-therapy for colon-cancer using registry data: a method comparison and validation

The Rhetoric and Reality of Anthropomorphism in Artificial Intelligence

A Bayesian Nonparametric Approach to Causal Inference on Quantiles

Contact Info

Product

Resources

About