TukeySens

This package realizes the sensitivity analysis method based on Tukey’s factorization, which allows flexible models for observed data and clean separation of the identified and unidentified parts of the sensitivity model. For theoretical details, please see the full paper. Specifically, TukeySens provides graphical tools that investigators can use to refine their sensitivity analyses and visualize the results.

How to install the package

To install the development version on GitHub make sure you have the package devtools installed.

# install.packages("devtools") 
devtools::install_github("JiajingZ/TukeySens")

Basic Usage

# load package
library(TukeySens)

Calibration

# Observed Pre-treatment Variables 
x = NHANES %>% select(-one_of("trt_dbp", "ave_dbp"))
# Treatment 
trt = NHANES %>% select(trt_dbp)
# Outcomes 
y = NHANES %>% select(ave_dbp)
# Sensitivity Parameter Sequence 
gamma = seq(0.01, 0.1, by = 0.001)
# plot 
caliplot(x, trt, y, gamma)

Based on the above plot, we may want to limit the magnitude of sensitivity parameter (\gamma_t) with respect to body mass index (bmi), one of the most important predictors in terms of partial variance explained. Hence, we set (|\gamma_t| \ \leq \ 0.05) for the following analysis.

Model fitting

Fit the BART outcome model. largest_gamma is the absolute magnitude of the largest sensitivity parameter for Y(0) and Y(1). joint is a logical which determines whether we infer Y(0) and Y(1) models independently or together.

# Observed data in treatment group
NHANES_trt <- NHANES %>% dplyr::filter(trt_dbp == 1)
x_trt <- NHANES_trt %>% select(-one_of("trt_dbp", "ave_dbp"))
y_trt <- NHANES_trt %>% select(ave_dbp)

# Observed data in control group 
NHANES_ctrl <- NHANES %>% dplyr::filter(trt_dbp == 0)
x_ctrl <- NHANES_ctrl %>% select(-one_of("trt_dbp", "ave_dbp"))
y_ctrl <- NHANES_ctrl %>% select(ave_dbp)

tukey_out <- fit_outcome(x_trt, y_trt, x_ctrl, y_ctrl, largest_gamma = 0.05, joint=FALSE)

Visualizing the results

We can plot a heatmap of the resulting model fit using the standard plot function with type="ate". Alternatively, we can visualize ribbon plots of the quantile treatment effects with type="ate". Note that in the current implementation QTE visualization can be slow for large datasets, since we are computing quantiles of n-component normal mixtures.

ATE by Heatmap

# ATE Heatmap 

plot(tukey_out, type="ate")

QTE by Ribbon Plot

# Ribbon Plot of QTE
gamma_select = rbind(c(0, 0), c(-0.03, 0.05), c(0.05, -0.02), c(-0.01, 0.01))

plot(tukey_out, type="qte", gamma_select = gamma_select)