--- title: "Demonstration of the COMMA R Package" author: 'Created by Kimberly A. Hochstedler Webb. Contact: kah343@cornell.edu' date: "2024-04-23" output: rmarkdown::html_vignette vignette: > %\VignetteEngine{knitr::knitr} %\VignetteIndexEntry{Demonstration of the COMMA R Package} %\usepackage[UTF-8]{inputenc} --- ```{r setup, include=FALSE} knitr::opts_chunk$set(echo = TRUE, warning = FALSE, error = FALSE, message = FALSE, fig.align = "center") library(ggplot2) library(kableExtra) #devtools::install_github("kimberlywebb/COMMA") ``` \raggedright In this vignette, we provide a demonstration of the R Package *COMMA* (correcting misclassified mediation analysis). This package provides methods for estimating a mediation analysis when the binary mediator is potentially misclassified. Technical details about estimation are not included in this demonstration. For additional information on the methods used in this R Package, please consult ``Effect estimation in the presence of a misclassified binary mediator'' by Kimberly A. Hochstedler Webb and Martin T. Wells. # Model and Conceptual Framework Let $X$ denote a predictor of interest. $C$ denotes a matrix of covariates. $Y$ is the outcome variable of interest. The relationship between $X$ and $Y$ may be mediated by $M$. $M = m$ denotes an observation's true mediator value, with $m \in \{1, 2\}$. We do not observe $M$ directly. Instead, we observe $M^*$, a potentially misclassified (noisy) version of $M$. Given $M$, a subject's observed mediator value $M^*$ depends on a set of predictors $Z$. The true mediator, observed mediator, and outcome mechanisms are provided below. $$\text{True mediator mechanism: } \text{logit}\{ P(M = 1 | X, \boldsymbol{C} ; \boldsymbol{\beta}) \} = \beta_{0} + \beta_{X} X + \boldsymbol{\beta_{C} C}$$ $$\text{Observed mediator mechanisms: } \text{logit}\{ P(M^* = 1 | M = 1, \boldsymbol{Z} ; \boldsymbol{\gamma}) \} = \gamma_{110} + \boldsymbol{\gamma_{11Z} Z}, \\ \text{logit}\{ P(M^* = 1 | M = 2, \boldsymbol{Z} ; \boldsymbol{\gamma}) \} = \gamma_{120} + \boldsymbol{\gamma_{12Z} Z}$$ $$\text{Outcome mechanism: } E(Y | X, \boldsymbol{C}, M) = \theta_0 + \theta_X X + \boldsymbol{\theta_C C} + \theta_M M + \theta_{XM} XM$$ If we have a Bernoulli outcome that we model with a logit link, the *outcome mechanism* can be $\text{logit}\{P(Y = 1 | X, \boldsymbol{C}, M) \} = \theta_0 + \theta_X X + \boldsymbol{\theta_C C} + \theta_m M + \theta_{XM} XM$. # Simulate data - Normal outcome We begin this demonstration by generating data using the `COMMA_data()` function. The binary mediator simulated by this scheme is subject to misclassification. The predictor related to the true outcome mechanism is "x" and the predictor related to the observation mechanism is "z". ```{r} library(COMMA) library(dplyr) set.seed(20240422) sample_size <- 10000 n_cat <- 2 # Number of categories in the binary mediator # Data generation settings x_mu <- 0 x_sigma <- 1 z_shape <- 1 z_scale <- 1 c_shape <- 1 # True parameter values (gamma terms set the misclassification rate) true_beta <- matrix(c(1, -2, .5), ncol = 1) true_gamma <- matrix(c(1.8, 1, -1.5, -1), nrow = 2, byrow = FALSE) true_theta <- matrix(c(1, 1.5, -2.5, -.2), ncol = 1) # Generate data. my_data <- COMMA_data(sample_size, x_mu, x_sigma, z_shape, c_shape, interaction_indicator = FALSE, outcome_distribution = "Normal", true_beta, true_gamma, true_theta) # Save list elements as vectors. Mstar = my_data[["obs_mediator"]] Mstar_01 <- ifelse(Mstar == 1, 1, 0) outcome = my_data[["outcome"]] x_matrix = my_data[["x"]] z_matrix = my_data[["z"]] c_matrix = my_data[["c"]] ``` ## Effect estimation We propose estimation methods using the Expectation-Maximization algorithm (EM) and an ordinary least squares (OLS) correction procedure. The proposed predictive value weighting (PVW) approach detailed in Webb and Wells (2024) is currently only available for Bernoulli outcomes in *COMMA*. Each method checks and corrects instances of label switching, as described in Webb and Wells (2024). In the code below, we provide functions for implementing these methods. ### EM algorithm ```{r} # Supply starting values for all parameters. beta_start <- coef(glm(Mstar_01 ~ x_matrix + c_matrix, family = "binomial"(link = "logit"))) gamma_start <- matrix(rep(1,4), ncol = 2, nrow = 2, byrow = FALSE) theta_start <- coef(lm(outcome ~ x_matrix + Mstar_01 + c_matrix)) # Estimate parameters using the EM-Algorithm. EM_results <- COMMA_EM(Mstar, outcome, outcome_distribution = "Normal", interaction_indicator = FALSE, x_matrix, z_matrix, c_matrix, beta_start, gamma_start, theta_start, sigma_start = 1) EM_results$True_Value <- c(true_beta, c(true_gamma), true_theta, 1) EM_results$Estimates <- round(EM_results$Estimates, 3) EM_results ``` ### OLS correction ```{r} # Estimate parameters using the OLS correction. OLS_results <- COMMA_OLS(Mstar, outcome, x_matrix, z_matrix, c_matrix, beta_start, gamma_start, theta_start) OLS_results$True_Value <- c(true_beta, c(true_gamma), true_theta[c(1,3,2,4)]) OLS_results$Estimates <- round(OLS_results$Estimates, 3) OLS_results ``` ### Compare results ```{r} NormalY_results <- data.frame(Parameter = rep(c("beta_x", "theta_x", "theta_m"), 3), Method = c(rep("EM", 3), rep("OLS", 3), rep("Naive", 3)), Estimate = c(EM_results$Estimates[c(2, 9, 10)], OLS_results$Estimates[c(2, 9, 10)], beta_start[2], theta_start[c(2,3)]), True_Value = rep(c(true_beta[2], true_theta[2], true_theta[3]), 3)) ggplot(data = NormalY_results) + geom_point(aes(x = Method, y = Estimate, color = Method), size = 3) + geom_hline(aes(yintercept = True_Value), linetype = "dashed") + facet_wrap(~Parameter) + theme_minimal() + scale_color_manual(values = c("#DA86A5", "#785E95", "#409DBE")) + ggtitle("Parameter estimates from EM, OLS, and Naive approaches", subtitle = "Normal outcome model with a misclassified mediator") ``` # Simulate data - Bernoulli outcome Next, we generate data with a Bernoulli outcome using the `COMMA_data()` function. Once again, the binary mediator simulated by this scheme is subject to misclassification. The predictor related to the true outcome mechanism is "x" and the predictor related to the observation mechanism is "z". ```{r} library(COMMA) library(dplyr) set.seed(20240423) sample_size <- 10000 n_cat <- 2 # Number of categories in the binary mediator # Data generation settings x_mu <- 0 x_sigma <- 1 z_shape <- 1 z_scale <- 1 c_shape <- 1 # True parameter values (gamma terms set the misclassification rate) true_beta <- matrix(c(1, -2, .5), ncol = 1) true_gamma <- matrix(c(1.8, 1, -1.5, -1), nrow = 2, byrow = FALSE) true_theta <- matrix(c(1, 1.5, -2.5, -.2, .5), ncol = 1) # Generate data. my_data <- COMMA_data(sample_size, x_mu, x_sigma, z_shape, c_shape, interaction_indicator = TRUE, outcome_distribution = "Bernoulli", true_beta, true_gamma, true_theta) # Save list elements as vectors. Mstar = my_data[["obs_mediator"]] Mstar_01 <- ifelse(Mstar == 1, 1, 0) outcome = my_data[["outcome"]] x_matrix = my_data[["x"]] z_matrix = my_data[["z"]] c_matrix = my_data[["c"]] ``` ## Effect estimation We propose estimation methods using the Expectation-Maximization algorithm (EM) and a predictive value weighting (PVW) approach. The ordinary least squares (OLS) correction procedure is only appropriate for Normal outcome models. Each method checks and corrects instances of label switching, as described in Webb and Wells (2024). In the code below, we provide functions for implementing these methods. ### EM algorithm ```{r} # Supply starting values for all parameters. beta_start <- coef(glm(Mstar_01 ~ x_matrix + c_matrix, family = "binomial"(link = "logit"))) gamma_start <- matrix(rep(1,4), ncol = 2, nrow = 2, byrow = FALSE) xm_interaction <- x_matrix * c_matrix theta_start <- coef(glm(outcome ~ x_matrix + Mstar_01 + c_matrix + xm_interaction, family = "binomial"(link = "logit"))) # Estimate parameters using the EM-Algorithm. EM_results <- COMMA_EM(Mstar, outcome, outcome_distribution = "Bernoulli", interaction_indicator = TRUE, x_matrix, z_matrix, c_matrix, beta_start, gamma_start, theta_start) EM_results$True_Value <- c(true_beta, c(true_gamma), true_theta) EM_results$Estimates <- round(EM_results$Estimates, 3) EM_results ``` ### PVW approach ```{r} PVW_results <- COMMA_PVW(Mstar, outcome, outcome_distribution = "Bernoulli", interaction_indicator = TRUE, x_matrix, z_matrix, c_matrix, beta_start, gamma_start, theta_start) PVW_results$True_Value <- c(true_beta, c(true_gamma), true_theta) PVW_results$Estimates <- round(PVW_results$Estimates, 3) PVW_results ``` ### Compare results ```{r} BernoulliY_results <- data.frame(Parameter = rep(c("beta_x", "theta_x", "theta_m", "theta_xm"), 3), Method = c(rep("EM", 4), rep("PVW", 4), rep("Naive", 4)), Estimate = c(EM_results$Estimates[c(2, 9, 10, 12)], PVW_results$Estimates[c(2, 9, 10, 12)], beta_start[2], theta_start[c(2,3,5)]), True_Value = rep(c(true_beta[2], true_theta[2], true_theta[3], true_theta[5]), 3)) ggplot(data = BernoulliY_results) + geom_point(aes(x = Method, y = Estimate, color = Method), size = 3) + geom_hline(aes(yintercept = True_Value), linetype = "dashed") + facet_wrap(~Parameter) + theme_minimal() + scale_color_manual(values = c("#DA86A5", "#785E95", "#ECA698")) + ggtitle("Parameter estimates from EM, PVW, and Naive approaches", subtitle = "Bernoulli outcome model with a misclassified mediator") ```