--- title: "Introduction to tern.rbmi" date: "2022-05-16" output: rmarkdown::html_document: theme: "spacelab" highlight: "kate" toc: true toc_float: true vignette: > %\VignetteIndexEntry{Introduction to tern.rbmi} %\VignetteEngine{knitr::rmarkdown} %\VignetteEncoding{UTF-8} editor_options: markdown: wrap: 72 --- ```{r setup, include = FALSE} suggested_dependent_pkgs <- c("rstan") knitr::opts_chunk$set( collapse = TRUE, comment = "#>", eval = all(vapply( suggested_dependent_pkgs, requireNamespace, logical(1), quietly = TRUE )) ) ``` ## Introduction to tern.rbmi --------- This vignette shows the general purpose and syntax of the `tern.rbmi` R package. The `tern.rbmi` provides an interface for Reference Based Multiple Imputation (`rbmi`) within the tern framework. For details of the `rbmi` package, please see [Reference Based Multiple Imputation (rbmi)](https://github.com/insightsengineering/rbmi). The basic usage of `rbmi` core functions is described in the `quickstart` vignette: ```{r eval=FALSE} vignette(topic = "quickstart", package = "rbmi") ``` --------- ## Example of using `tern.rbmi` The `rbmi` package consists of 4 core functions (plus several helper functions) which are typically called in sequence: * `draws()` - fits the imputation models and stores their parameters * `impute()` - creates multiple imputed datasets * `analyse()` - analyses each of the multiple imputed datasets * `pool()` - combines the analysis results across imputed datasets into a single statistic ### The Data We use a publicly available example dataset from an antidepressant clinical trial of an active drug versus placebo. The relevant endpoint is the Hamilton 17-item depression rating scale (`HAMD17`) which was assessed at baseline and at weeks 1, 2, 4, and 6. Study drug discontinuation occurred in 24% of subjects from the active drug and 26% of subjects from placebo. All data after study drug discontinuation are missing and there is a single additional intermittent missing observation. ```{r} library(tern.rbmi) library(dplyr) ``` ```{r} data <- antidepressant_data levels(data$THERAPY) <- c("PLACEBO", "DRUG") # This is important! The order defines the computation order later missing_var <- "CHANGE" vars <- list( id = "PATIENT", visit = "VISIT", expand_vars = c("BASVAL", "THERAPY"), group = "THERAPY" ) covariates <- list( draws = c("BASVAL*VISIT", "THERAPY*VISIT"), analyse = c("BASVAL") ) data <- data %>% dplyr::select(PATIENT, THERAPY, VISIT, BASVAL, THERAPY, CHANGE) %>% dplyr::mutate(dplyr::across(.cols = vars$id, ~ as.factor(.x))) %>% dplyr::arrange(dplyr::across(.cols = c(vars$id, vars$visit))) ``` ```{r} # Use expand_locf to add rows corresponding to visits with missing outcomes to the dataset data_full <- do.call( expand_locf, args = list( data = data, vars = c(vars$expand_vars, vars$group), group = vars$id, order = c(vars$id, vars$visit) ) %>% append(lapply(data[c(vars$id, vars$visit)], levels)) ) data_full <- data_full %>% dplyr::group_by(dplyr::across(vars$id)) %>% dplyr::mutate(!!vars$group := Filter(Negate(is.na), .data[[vars$group]])[1]) # there are duplicates - use first value data_full <- data_full %>% dplyr::group_by(dplyr::across(c(vars$id, vars$group, vars$visit))) %>% dplyr::slice(1) %>% dplyr::ungroup() # need to have a single ID column data_full <- data_full %>% tidyr::unite("TMP_ID", dplyr::all_of(vars$id), sep = "_#_", remove = FALSE) %>% dplyr::mutate(TMP_ID = as.factor(TMP_ID)) ``` #### Creating intercurrent event data Set the imputation strategy to MAR for each patient with at least one missing observation. ```{r} data_ice <- data_full %>% dplyr::arrange(dplyr::across(.cols = c("TMP_ID", vars$visit))) %>% dplyr::filter(is.na(.data[[missing_var]])) %>% dplyr::group_by(TMP_ID) %>% dplyr::slice(1) %>% dplyr::ungroup() %>% dplyr::select(all_of(c("TMP_ID", vars$visit))) %>% dplyr::mutate(strategy = "MAR") ``` ### Draws The `rbmi::draws()` function fits the imputation models and stores the corresponding parameter estimates or Bayesian posterior parameter draws. The three main inputs to the `rbmi::draws()` function are: * data - The primary longitudinal data.frame containing the outcome variable and all covariates. * data_ice - A data.frame which specifies the first visit affected by an intercurrent event (ICE) and the imputation strategy for handling missing outcome data after the ICE. At most one ICE which is to be imputed by a non-MAR strategy is allowed per subject. * method - The statistical method used to fit the imputation models and to create imputed datasets. #### Define key variables Define the names of key variables in our dataset and the covariates included in the imputation model using `rbmi::set_vars()`. Note that the covariates argument can also include interaction terms. ```{r} debug_mode <- FALSE draws_vars <- rbmi::set_vars( outcome = missing_var, visit = vars$visit, group = vars$group, covariates = covariates$draws ) draws_vars$subjid <- "TMP_ID" ``` Define which imputation method to use, then create samples for the imputation parameters by running the `draws()` function. ```{r} draws_method <- method_bayes() draws_obj <- rbmi::draws( data = data_full, data_ice = data_ice, vars = draws_vars, method = draws_method ) ``` ### Impute The next step is to use the parameters from the imputation model to generate the imputed datasets. This is done via the `rbmi::impute()` function. The function only has two key inputs: the imputation model output from `rbmi::draws()` and the reference groups relevant to reference-based imputation methods. Its usage is thus: ```{r} impute_references <- c("DRUG" = "PLACEBO", "PLACEBO" = "PLACEBO") impute_obj <- rbmi::impute( draws_obj, references = impute_references ) ``` ### Analyze The next step is to run the analysis model on each imputed dataset. This is done by defining an analysis function and then calling `rbmi::analyse()` to apply this function to each imputed dataset. ```{r} # Define analysis model analyse_fun <- ancova ref_levels <- levels(impute_obj$data$group[[1]]) names(ref_levels) <- c("ref", "alt") analyse_obj <- rbmi::analyse( imputations = impute_obj, fun = analyse_fun, vars = rbmi::set_vars( subjid = "TMP_ID", outcome = missing_var, visit = vars$visit, group = vars$group, covariates = covariates$analyse ) ) ``` ### Pool The `rbmi::pool()` function can be used to summarize the analysis results across multiple imputed datasets to provide an overall statistic with a standard error, confidence intervals and a p-value for the hypothesis test of the null hypothesis that the effect is equal to 0. ```{r} pool_obj <- rbmi::pool( results = analyse_obj, conf.level = 0.95, alternative = c("two.sided", "less", "greater"), type = c("percentile", "normal") ) ``` ### Create output Finally create output with `rtables` and `tern` packages ```{r} library(broom) df <- tidy(pool_obj) df ``` Final product, reshape `rbmi` final results to nicely formatted `rtable` object. ```{r} basic_table() %>% split_cols_by("group", ref_group = levels(df$group)[1]) %>% split_rows_by("visit", split_label = "Visit", label_pos = "topleft") %>% summarize_rbmi() %>% build_table(df) ```