`bayesmsm` for longitudinal data without right-censoring"

In bayesmsm: Fitting Bayesian Marginal Structural Models for Longitudinal Observational Data

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  fig.align = 'center',
  fig.width = 9,
  fig.height = 6,
  warning = F,
  message = F
)

Introduction

• The bayesmsm package enables easy implementation of the Bayesian marginal structural models (BMSMs) for longitudinal data. The methodology of BMSMs can be divided into 2 estimation steps:

• Step 1. Bayesian treatment effect weight estimation

• Step 2. Bayesian non-parametric bootstrap to maximize the utility function with respect to the causal effect

• For Step 1, we estimate treatment weights $w_{ij}$ using posterior samples of the $\alpha$ and $\beta$ via fitting a series of logistic regressions in a Bayesian framework. The package is able to handle longitudinal data without and with right-censoring. For Step 2, $P_n(v_{ij})$ is estimated via non-parametric Bayesian bootstrap with $Dir(1,...,1)$ sampling weights.

• The main functions in this package include:

• bayesweight: Calculates Bayesian weights for subject-specific treatment effects.

• bayesweight_cen: Calculates Bayesian weights for subject-specific treatment effects with right-censored data.
• bayesmsm: Estimates marginal structural models using the calculated Bayesian weights.
• plot_ATE: Plots the estimated Average Treatment Effect (ATE).
• plot_APO: Plots the estimated Average Potential Outcome (APO).
• plot_est_box: Plots the distribution of estimated treatment effects.
• summary_bayesmsm: Summarizes the model results from bayesmsm.

• simData: Generates synthetic longitudinal data with optional right-censoring

• Installation

• To install the bayesmsm package, you can use the devtools package to install it directly from GitHub.

if (!requireNamespace("bayesmsm", quietly = TRUE)) {
  stop("The package 'bayesmsm' is required to run this vignette. Please install it manually using:
       devtools::install_github('Kuan-Liu-Lab/bayesmsm')")
} else {
  library(bayesmsm)
}

Simulated longitudinal observational data without right-censoring

In this vignette, we demonstrate how to simulate a clean longitudinal dataset without censoring using simData(). We generate data for 200 individuals observed over 2 visits. At each visit, two normally distributed covariates (L1_j, L2_j) and a binary treatment assignment (A_j) are created. No censoring is applied, so all covariates, treatments, and the final outcome Y are fully observed. The outcome Y is continuous, generated from a linear model on the covariate and treatment history.

| Variable | Description | | --------------- | --------------------------------------- | | L1_1, L2_1 | Baseline covariates (continuous) | | L1_2, L2_2 | Time-varying covariates at visit 2 | | A1, A2 | Binary treatments at visits 1 and 2 | | Y | Continuous end-of-study outcome |

# 1) Define coefficient lists for 2 visits
amodel <- list(
  # Visit 1: logit P(A1=1) = -0.3 + 0.4*L1_1 - 0.2*L2_1
  c("(Intercept)" = -0.3, "L1_1" = 0.4, "L2_1" = -0.2),
  # Visit 2: logit P(A2=1) = -0.1 + 0.3*L1_2 - 0.1*L2_2 + 0.5*A_prev
  c("(Intercept)" = -0.1, "L1_2" = 0.3, "L2_2" = -0.1, "A_prev" = 0.5)
)

# 2) Define binary outcome model: logistic on treatments and last covariates
ymodel <- c(
  "(Intercept)" = -0.8,
  "A1"          = 0.2,
  "A2"          = 0.4,
  "L1_2"        = 0.3,
  "L2_2"        = -0.3
)

# 3) Load package and simulate data without censoring
simdat <- simData(
  n                = 100,
  n_visits         = 2,
  covariate_counts = c(2, 2),
  amodel           = amodel,
  ymodel           = ymodel,
  y_type           = "binary",
  right_censor     = FALSE,
  seed             = 123
)

# 4) Inspect first rows
head(simdat)

• Suppose the causal parameter of interest is the average treatment effect between always treated and never treated,

$$ ATE = E(Y \mid A_1 = 1, A_2 = 1) - E(Y \mid A_1 = 0, A_2 = 0) $$

Bayesian treatment effect weight estimation using bayesweight

• The following code calls the function bayesweight to run JAGS and calculate the weights.
• Non-parallel computing requires that n.chains = 1. Parallel MCMC requires at least 2 chains because computing is running on 1 core per chain, and we recommend using at most 2 chains less than the number of available cores on your computer.

• Running this function automatically saves a JAGS model file in the working directory, which the user can check to review the model specifications.

• Parameters Description:

• trtmodel.list: A list of formulas corresponding to each time point with the time-specific treatment variable on the left hand side and pre-treatment covariates to be balanced on the right hand side. Interactions and functions of covariates are allowed.
• data: The dataset containing all the variables specified in trtmodel.list.
• n.chains: Number of MCMC chains to run. For non-parallel execution, this should be set to 1. For parallel execution, it requires at least 2 chains.
• n.iter: Total number of iterations for each chain (including burn-in).
• n.burnin: Number of iterations to discard at the beginning of the simulation (burn-in).
• n.thin: Thinning rate for the MCMC sampler.
• seed: Seed to ensure reproducibility.
• parallel: Logical flag indicating whether to run the MCMC chains in parallel. Default is TRUE.

weights <- bayesweight(
  trtmodel.list = list(
    A1 ~ L1_1 + L2_1,
    A2 ~ L1_2 + L2_2 + A1),
  data = simdat,
  n.chains = 1,
  n.iter = 200,
  n.burnin = 100,
  n.thin = 1,
  seed = 890123,
  parallel = FALSE)

summary(weights)

• It returns a list containing:
• weights: The calculated weights for subject-specific treatment effects, computed by taking the average of the weights across all MCMC iterations.

• model_string: A character of the JAGS model based on input of the argument trtmodel.list.

• We can print and view the JAGS model stored in model_string from the above function output:

cat(weights$model_string)

Bayesian non-parametric bootstrap to maximize the utility function with respect to the causal effect using bayesmsm

The function bayesmsm estimates causal effect of time-varying treatments. It uses subject-specific treatment assignment weights weights calculated using bayesweight, and performs Bayesian non-parametric bootstrap to estimate the causal parameters.

• Parameters Description:
• ymodel: A formula representing the outcome model, which can include interactions and functions of covariates.
• nvisit: Specifies the number of visits or time points considered in the model.
• reference: The baseline or reference intervention across all visits, typically represented by a vector of zeros indicating no treatment (default is a vector of all zeros).
• comparator: The comparison intervention across all visits, typically represented by a vector of ones indicating full treatment (default is a vector of all ones).
• treatment_effect_type: A character string specifying the type of treatment effect to estimate. Options are "sq" for sequential treatment effects, which estimates effects for specific treatment sequences across visits, and "cum" for cumulative treatment effects, which assumes a single cumulative treatment variable representing the total exposure. The default is "sq".
• family: Specifies the outcome distribution family; use "gaussian" for continuous outcomes or "binomial" for binary outcomes (default is "gaussian").
• data: The dataset containing all variables required for the model.
• wmean: A vector of treatment assignment weights. Default is a vector of ones, implying equal weighting.
• nboot: The number of bootstrap iterations to perform for estimating the uncertainty around the causal estimates.
• optim_method: The optimization method used to find the best parameters in the model (default is 'BFGS').
• seed: A seed value to ensure reproducibility of results.
• parallel: A logical flag indicating whether to perform computations in parallel (default is TRUE).
• ncore: The number of cores to use for parallel computation (default is 4).

model <- bayesmsm(ymodel = Y ~ A1 + A2,
  nvisit = 2,
  reference = c(rep(0,2)),
  comparator = c(rep(1,2)),
  family = "binomial",
  data = simdat,
  wmean = weights$weights,
  nboot = 50,
  optim_method = "BFGS",
  parallel = TRUE,
  seed = 890123,
  ncore = 2)
str(model)

• It returns a model object which contains:
• mean, sd, quantile: the mean, standard deviation and 95% credible interval of the estimated causal effect (ATE). From the above results, the mean of ATE is approximately -3.161, which indicates that the expected outcome for always treated patients is, on average, 3.161 units less than that for never treated patients.
• bootdata: a data frame containing the bootstrap samples for the reference effect, comparator effect, and average treatment effect (ATE).

• reference, comparator: the reference level and comparator level the user chooses to compare. Here the reference level is never treated (0,0), and the comparator level is always treated (1,1).

• Summary function to generate result table from bayesmsm

• The summary_bayesmsm function automatically generates a summary table of the model output from the function bayesmsm.

summary_bayesmsm(model)

Visualization functions: plot_ATE, plot_APO, plot_est_box

The bayesmsm package also provides several other functions for visualizing the above results: plot_ATE, plot_APO, and plot_est_box. These functions help the user better interpret the estimated causal effects.

• Plotting the Average Treatment Effect (ATE)
• The plot_ATE function generates a plot of the estimated ATE with its 95% credible interval.

plot_ATE(model)

• Plotting the Average Potential Outcome (APO)
• Similarly, the plot_APO function plots the estimated APO for both the reference and comparator level effects.

plot_APO(model, "effect_reference")
plot_APO(model, "effect_comparator")

• Plotting the Distribution of Estimated Treatment Effects
• The plot_est_box function generates an error bar plot of the estimated treatment effects (APO and ATE) from the bootstrap samples.

plot_est_box(model)

Reference

• Saarela, O., Stephens, D. A., Moodie, E. E. M., & Klein, M. B. (2015). On Bayesian estimation of marginal structural models. Biometrics, 71(2), 279–288. https://doi.org/10.1111/biom.12269
• Robins, J. M., Hernán, M. A., & Brumback, B. (2000). Marginal structural models and causal inference in epidemiology. Epidemiology, 11(5), 550–560. https://doi.org/10.1097/00001648-200009000-00011
• Liu, K., Saarela, O., Feldman, B. M., & Pullenayegum, E. (2020). Estimation of causal effects with repeatedly measured outcomes in a Bayesian framework. Statistical Methods in Medical Research, 29(9), 2507–2519. https://doi.org/10.1177/0962280219900362

bayesmsm documentation built on June 17, 2025, 9:08 a.m.