stRoke/R/ci_plot.R

utils::globalVariables(c("vname", "lo", "or", "ord", "up"))
#' Confidence interval plot with point estimate
#' 
#' Horizontal forest plot of point estimate with confidence intervals. 
#' Includes dichotomous or olr, depending on number of levels in "x".
#' Title and axis labels can be added to the ggplot afterwards.
#'
#' @param ds main input, either data set or logistic model
#' @param x text string of main exposure variable
#' @param y text string of outcome variable
#' @param vars variables for multivariate analysis.
#' @param dec Decimals in labels
#' @param lbls Labels for variable names
#' @param title Plot title. Can be specified later.
#' @param method Character vector. The method for the regression. 
#' Can be c("auto", "model").
#'
#' @return ggplot element
#' @export
#'
#' @import ggplot2
#' @importFrom MASS polr
#' @importFrom stats as.formula coef confint lm quantile reorder binomial glm
#'
#' @examples
#' # Auto plot
#' data(talos)
#' talos[,"mrs_1"]<-factor(talos[,"mrs_1"],ordered=TRUE)
#' ci_plot(ds = talos, x = "rtreat", y = "mrs_1", 
#' vars = c("hypertension","diabetes"))
#' ## Model plot
#' # iris$ord<-factor(sample(1:3,size=nrow(iris),replace=TRUE),ordered=TRUE)
#' # lm <- MASS::polr(ord~., data=iris, Hess=TRUE, method="logistic")
#' # ci_plot(ds = lm, method="model")
ci_plot <-
  function(ds,
           x = NULL,
           y = NULL,
           vars = NULL,
           dec = 3,
           lbls = NULL,
           title = NULL, 
           method = "auto") {
    
    if (!method %in% c("auto", "model"))
      stop("Method has to either 'auto' or 'model'")
    
    if (method == "auto") {
      if (!is.factor(ds[, y]))
        stop("Outcome has to be factor")
      
      # Formula
      ci_form <- as.formula(paste0(y, "~", x, "+."))
    
      # Ordinal logistic regression for non-dichotomous factors
      if (length(levels(ds[, y])) > 2) {
        m <- MASS::polr(
          formula = ci_form,
          data = ds[, unique(c(x, y, vars))],
          Hess = TRUE,
          method = "logistic"
        )
        if (is.null(title))
          title <- "Ordinal logistic regression"
      }
      
      # Binary logistic regression for dichotomous factors
      if (length(levels(ds[, y])) == 2) {
        m <- glm(formula = ci_form,
                 data = ds[unique(c(x, y, vars))],
                 family = binomial())
        if (is.null(title))
          title <- "Binary logistic regression"
      }
    } else if (method == "model") {
      if (is.data.frame(ds)) {
        stop("Method is 'model', but input is data.frame")
      }
      m <- ds
    }
    
    odds <- data.frame(cbind(exp(coef(m)), exp(confint(m))))
    
    names(odds) <- c("or", "lo", "up")
    rodds <- round(odds, digits = dec)
    
    if (is.null(lbls)) {
      odds$vname <- paste0(row.names(odds),
                           " \n",
                           paste0(rodds$or, "
                              [", rodds$lo, ":", rodds$up, "]"))
    } else {
      odds$vname <- paste0(lbls,
                           " \n",
                           paste0(rodds$or,
                                  " [", rodds$lo, ":", rodds$up, "]"))
      
    }
    
    odds$ord <- rev(seq_len(nrow(odds)))
    
    ggplot2::ggplot(data = odds,
                    mapping = ggplot2::aes(y = or, x = reorder(vname, ord))) +
      ggplot2::geom_point() +
      ggplot2::geom_errorbar(mapping = ggplot2::aes(ymin = lo, ymax = up),
                             width = 0.2) +
      ggplot2::scale_y_log10() +
      ggplot2::geom_hline(yintercept = 1, linetype = 2) +
      ggplot2::labs(title = title) +
      ggplot2::coord_flip()
  }