bart: Fit BART models to select confounders and estimate treatment...

In bartcs: Bayesian Additive Regression Trees for Confounder Selection

Fit BART models to select confounders and estimate treatment effect

Description

Fit Bayesian Regression Additive Trees (BART) models to select relevant confounders among a large set of potential confounders and to estimate average treatment effect E[Y(1) - Y(0)].

Usage

separate_bart(
  Y, trt, X,
  trt_treated     = 1,
  trt_control     = 0,
  num_tree        = 50,
  num_chain       = 4,
  num_burn_in     = 100,
  num_thin        = 1,
  num_post_sample = 100,
  step_prob       = c(0.28, 0.28, 0.44),
  alpha           = 0.95,
  beta            = 2,
  nu              = 3,
  q               = 0.95,
  dir_alpha       = 5,
  parallel        = FALSE,
  verbose         = TRUE
)

single_bart(
  Y, trt, X,
  trt_treated     = 1,
  trt_control     = 0,
  num_tree        = 50,
  num_chain       = 4,
  num_burn_in     = 100,
  num_thin        = 1,
  num_post_sample = 100,
  step_prob       = c(0.28, 0.28, 0.44),
  alpha           = 0.95,
  beta            = 2,
  nu              = 3,
  q               = 0.95,
  dir_alpha       = 5,
  parallel        = FALSE,
  verbose         = TRUE
)

Arguments

Y	A vector of outcome values.
trt	A vector of treatment values. Binary treatment works for both model and continuous treatment works for single_bart(). For binary treatment, use 1 to indicate the treated group and 0 for the control group.
X	A matrix of potential confounders.
trt_treated	Value of trt for the treated group. The default value is set to 1.
trt_control	Value of trt for the control group. The default value is set to 0.
num_tree	Number of trees in BART model. The default value is set to 100.
num_chain	Number of MCMC chains. Need to set num_chain > 1 for the Gelman-Rubin diagnostic. The default value is set to 4.
num_burn_in	Number of MCMC samples to be discarded per chain as initial burn-in periods. The default value is set to 100.
num_thin	Number of thinning per chain. One in every num_thin samples are selected. The default value is set to 1.
num_post_sample	Final number of posterior samples per chain. Number of MCMC iterations per chain is burn_in + num_thin * num_post_sample. The default value is set to 100.
step_prob	A vector of tree alteration probabilities (GROW, PRUNE, CHANGE). Each alteration is proposed to change the tree structure. The default setting is ⁠(0.28, 0.28, 0.44)⁠.
alpha, beta	Hyperparameters for tree regularization prior. A terminal node of depth d will split with probability of alpha * (1 + d)^(-beta). The default setting is ⁠(alpha, beta) = (0.95, 2)⁠ from Chipman et al. (2010).
nu, q	Values to calibrate hyperparameter of sigma prior. The default setting is ⁠(nu, q) = (3, 0.95)⁠ from Chipman et al. (2010).
dir_alpha	Hyperparameter of Dirichlet prior for selection probabilities. The default value is 5.
parallel	If TRUE, model fitting will be parallelized with respect to N = nrow(X). Parallelization is recommended for very high n only. The default setting is FALSE.
verbose	If TRUE, message will be printed during training. If FALSE, message will be suppressed.

Details

separate_bart() and single_bart() fit an exposure model and outcome model(s) for estimating treatment effect with adjustment of confounders in the presence of a large set of potential confounders (Kim et al. 2023).

The exposure model E[A|X] and the outcome model(s) E[Y|A,X] are linked together with a common Dirichlet prior that accrues posterior selection probabilities to the corresponding confounders (X) on the basis of association with both the exposure (A) and the outcome (Y).

There is a distinction between fitting separate outcome models for the treated and control groups and fitting a single outcome model for both groups.

• separate_bart() specifies two "separate" outcome models for two binary treatment levels. Thus, it fits three models: one exposure model and two separate outcome models for A = 0, 1.

• single_bart() specifies one "single" outcome model. Thus, it fits two models: one exposure model and one outcome model for the entire sample.

All inferences are made with outcome model(s).

Value

A bartcs object. A list object contains the following components.

mcmc_list

A mcmc.list object from coda package. mcmc_list contains the following items.

• ATE Posterior sample of average treatment effect E[Y(1) - Y(0)].

• Y1 Posterior sample of potential outcome E[Y(1)].

• Y0 Posterior sample of potential outcome E[Y(0)].

• dir_alpha Posterior sample of dir_alpha.

• sigma2_out Posterior sample of sigma2 in the outcome model.

var_prob	Aggregated posterior inclusion probability of each variable.
var_count	Number of selection of each variable in each MCMC iteration. Its dimension is num_post_sample * ncol(X).
chains	A list of results from each MCMC chain.
model	separate or single.
label	Column names of X.
params	Parameters used in the model.

References

Chipman, H. A., George, E. I., & McCulloch, R. E. (2010). BART: Bayesian additive regression trees. The Annals of Applied Statistics, 4(1), 266-298. \Sexpr[results=rd]{tools:::Rd_expr_doi("10.1214/09-AOAS285")}

Kim, C., Tec, M., & Zigler, C. M. (2023). Bayesian Nonparametric Adjustment of Confounding, Biometrics \Sexpr[results=rd]{tools:::Rd_expr_doi("10.1111/biom.13833")}

Examples

data(ihdp, package = "bartcs")
single_bart(
  Y               = ihdp$y_factual,
  trt             = ihdp$treatment,
  X               = ihdp[, 6:30],
  num_tree        = 10,
  num_chain       = 2,
  num_post_sample = 20,
  num_burn_in     = 10,
  verbose         = FALSE
)
separate_bart(
  Y               = ihdp$y_factual,
  trt             = ihdp$treatment,
  X               = ihdp[, 6:30],
  num_tree        = 10,
  num_chain       = 2,
  num_post_sample = 20,
  num_burn_in     = 10,
  verbose         = FALSE
)

bartcs documentation built on June 22, 2024, 6:48 p.m.