當一個因子被分層時，在 RStudio 中使用 ggforest 繪制 Cox PH model？

Question

當需要對 model 變量中的一個變量進行分層時，我正在嘗試找到一種方法來從 Cox-PH model 中創建風險比的森林 plot。 對於非分層 model， ggforest() function 非常好。 運行一些示例代碼

library(survival)
library(survminer)

model <- coxph(Surv(time, status) ~ sex + rx + adhere,
               data = colon )
ggforest(model)

colon <- within(colon, {
  sex <- factor(sex, labels = c("female", "male"))
  differ <- factor(differ, labels = c("well", "moderate", "poor"))
  extent <- factor(extent, labels = c("submuc.", "muscle", "serosa", "contig."))
})
bigmodel <-
  coxph(Surv(time, status) ~ sex + rx + adhere + differ + extent + node4,
        data = colon )
ggforest(bigmodel)

生成此圖

但是，如果我必須糾正分層的非比例性

stratamodel <- coxph(Surv(time, status) ~ sex + strata(rx) + adhere + differ + extent + node4,
                     data = colon )
ggforest(stratamodel)

出現以下錯誤消息：

“ [.data.frame (data, , var) 中的錯誤：選擇了未定義的列

另外：警告信息：

in.get_data(model, data = data)：未提供data參數。 數據將從 model 擬合中提取。”

關於如何從地層 model 獲取 ggforest 需要的信息以便它可以生成 plot 的任何建議？ 謝謝你的幫助！

Answer 1

我認為所需的森林圖是簡單地跳過模型公式中分層 RX 變量的圖。 如果是這樣，我們可以簡單地在其中插入一個 if 子句，以忽略與數據中的列名不完全對應的公式部分（例如“strata(rx)”不是列名）。

方法一

如果您對 R 足夠滿意以修改函數，請運行trace(ggforest, edit = TRUE)並將彈出窗口中的allTerms <- lapply(...) （大約第 10-25 行）替換為以下版本：

allTerms <- lapply(seq_along(terms), function(i) {
  var <- names(terms)[i]
  if(var %in% colnames(data)) {
    if (terms[i] %in% c("factor", "character")) {
      adf <- as.data.frame(table(data[, var]))
      cbind(var = var, adf, pos = 1:nrow(adf))
    }
    else if (terms[i] == "numeric") {
      data.frame(var = var, Var1 = "", Freq = nrow(data), 
                 pos = 1)
    }
    else {
      vars = grep(paste0("^", var, "*."), coef$term, 
                  value = TRUE)
      data.frame(var = vars, Var1 = "", Freq = nrow(data), 
                 pos = seq_along(vars))
    }
  } else {
    message(var, "is not found in data columns, and will be skipped.")
  }
  
})

ggforest(stratamodel) # this should work after the modification

修改不會持續到后續的 R 會話。 如果您想在當前會話中撤銷修改，只需運行untrace(ggforest) 。

方法二

如果您希望擁有該函數的永久修改版本以供將來使用/不想亂用庫的函數，請保存以下函數：

ggforest2 <- function (model, data = NULL, main = "Hazard ratio", 
                       cpositions = c(0.02, 0.22, 0.4), fontsize = 0.7, 
                       refLabel = "reference", noDigits = 2) {
  conf.high <- conf.low <- estimate <- NULL
  stopifnot(class(model) == "coxph")
  data <- survminer:::.get_data(model, data = data)
  terms <- attr(model$terms, "dataClasses")[-1]
  coef <- as.data.frame(broom::tidy(model))
  gmodel <- broom::glance(model)
  allTerms <- lapply(seq_along(terms), function(i) {
    var <- names(terms)[i]
    if(var %in% colnames(data)) {
      if (terms[i] %in% c("factor", "character")) {
        adf <- as.data.frame(table(data[, var]))
        cbind(var = var, adf, pos = 1:nrow(adf))
      }
      else if (terms[i] == "numeric") {
        data.frame(var = var, Var1 = "", Freq = nrow(data), 
                   pos = 1)
      }
      else {
        vars = grep(paste0("^", var, "*."), coef$term, 
                    value = TRUE)
        data.frame(var = vars, Var1 = "", Freq = nrow(data), 
                   pos = seq_along(vars))
      }
    } else {
      message(var, "is not found in data columns, and will be skipped.")
    }    
  })
  allTermsDF <- do.call(rbind, allTerms)
  colnames(allTermsDF) <- c("var", "level", "N", 
                            "pos")
  inds <- apply(allTermsDF[, 1:2], 1, paste0, collapse = "")
  rownames(coef) <- gsub(coef$term, pattern = "`", replacement = "")
  toShow <- cbind(allTermsDF, coef[inds, ])[, c("var", "level", "N", "p.value", 
                                                "estimate", "conf.low", 
                                                "conf.high", "pos")]
  toShowExp <- toShow[, 5:7]
  toShowExp[is.na(toShowExp)] <- 0
  toShowExp <- format(exp(toShowExp), digits = noDigits)
  toShowExpClean <- data.frame(toShow, pvalue = signif(toShow[, 4], noDigits + 1), 
                               toShowExp)
  toShowExpClean$stars <- paste0(round(toShowExpClean$p.value, noDigits + 1), " ", 
                                 ifelse(toShowExpClean$p.value < 0.05, "*", ""), 
                                 ifelse(toShowExpClean$p.value < 0.01, "*", ""), 
                                 ifelse(toShowExpClean$p.value < 0.001, "*", ""))
  toShowExpClean$ci <- paste0("(", toShowExpClean[, "conf.low.1"], 
                              " - ", toShowExpClean[, "conf.high.1"], ")")
  toShowExpClean$estimate.1[is.na(toShowExpClean$estimate)] = refLabel
  toShowExpClean$stars[which(toShowExpClean$p.value < 0.001)] = "<0.001 ***"
  toShowExpClean$stars[is.na(toShowExpClean$estimate)] = ""
  toShowExpClean$ci[is.na(toShowExpClean$estimate)] = ""
  toShowExpClean$estimate[is.na(toShowExpClean$estimate)] = 0
  toShowExpClean$var = as.character(toShowExpClean$var)
  toShowExpClean$var[duplicated(toShowExpClean$var)] = ""
  toShowExpClean$N <- paste0("(N=", toShowExpClean$N, ")")
  toShowExpClean <- toShowExpClean[nrow(toShowExpClean):1, ]
  rangeb <- range(toShowExpClean$conf.low, toShowExpClean$conf.high, 
                  na.rm = TRUE)
  breaks <- axisTicks(rangeb/2, log = TRUE, nint = 7)
  rangeplot <- rangeb
  rangeplot[1] <- rangeplot[1] - diff(rangeb)
  rangeplot[2] <- rangeplot[2] + 0.15 * diff(rangeb)
  width <- diff(rangeplot)
  y_variable <- rangeplot[1] + cpositions[1] * width
  y_nlevel <- rangeplot[1] + cpositions[2] * width
  y_cistring <- rangeplot[1] + cpositions[3] * width
  y_stars <- rangeb[2]
  x_annotate <- seq_len(nrow(toShowExpClean))
  annot_size_mm <- fontsize * 
    as.numeric(grid::convertX(unit(theme_get()$text$size, "pt"), "mm"))
  p <- ggplot(toShowExpClean, 
              aes(seq_along(var), exp(estimate))) + 
    geom_rect(aes(xmin = seq_along(var) - 0.5, 
                  xmax = seq_along(var) + 0.5, 
                  ymin = exp(rangeplot[1]), 
                  ymax = exp(rangeplot[2]), 
                  fill = ordered(seq_along(var)%%2 + 1))) +
    scale_fill_manual(values = c("#FFFFFF33",  "#00000033"), guide = "none") + 
    geom_point(pch = 15, size = 4) + 
    geom_errorbar(aes(ymin = exp(conf.low), ymax = exp(conf.high)), 
                  width = 0.15) + 
    geom_hline(yintercept = 1, linetype = 3) + 
    coord_flip(ylim = exp(rangeplot)) + 
    ggtitle(main) + 
    scale_y_log10(name = "", labels = sprintf("%g", breaks), 
                  expand = c(0.02, 0.02), breaks = breaks) + 
    theme_light() +
    theme(panel.grid.minor.y = element_blank(), 
          panel.grid.minor.x = element_blank(), 
          panel.grid.major.y = element_blank(), 
          legend.position = "none", 
          panel.border = element_blank(), 
          axis.title.y = element_blank(), 
          axis.text.y = element_blank(), 
          axis.ticks.y = element_blank(), 
          plot.title = element_text(hjust = 0.5)) + 
    xlab("") + 
    annotate(geom = "text", x = x_annotate, y = exp(y_variable), 
             label = toShowExpClean$var, fontface = "bold", 
             hjust = 0, size = annot_size_mm) + 
    annotate(geom = "text", x = x_annotate, y = exp(y_nlevel), hjust = 0, 
             label = toShowExpClean$level, 
             vjust = -0.1, size = annot_size_mm) + 
    annotate(geom = "text", x = x_annotate, y = exp(y_nlevel), 
             label = toShowExpClean$N, fontface = "italic", hjust = 0, 
             vjust = ifelse(toShowExpClean$level == "", 0.5, 1.1),
             size = annot_size_mm) + 
    annotate(geom = "text", x = x_annotate, y = exp(y_cistring), 
             label = toShowExpClean$estimate.1, size = annot_size_mm, 
             vjust = ifelse(toShowExpClean$estimate.1 == "reference", 0.5, -0.1)) + 
    annotate(geom = "text", x = x_annotate, y = exp(y_cistring), 
             label = toShowExpClean$ci, size = annot_size_mm, 
             vjust = 1.1, fontface = "italic") + 
    annotate(geom = "text", x = x_annotate, y = exp(y_stars), 
             label = toShowExpClean$stars, size = annot_size_mm, 
             hjust = -0.2, fontface = "italic") +
    annotate(geom = "text", x = 0.5, y = exp(y_variable), 
             label = paste0("# Events: ", gmodel$nevent, 
                            "; Global p-value (Log-Rank): ", 
                            format.pval(gmodel$p.value.log, eps = ".001"), 
                            " \nAIC: ", round(gmodel$AIC, 2), 
                            "; Concordance Index: ", round(gmodel$concordance, 2)), 
             size = annot_size_mm, hjust = 0, vjust = 1.2, fontface = "italic")
  gt <- ggplot_gtable(ggplot_build(p))
  gt$layout$clip[gt$layout$name == "panel"] <- "off"
  ggpubr::as_ggplot(gt)
}

它是當前ggforest函數的快照，並進行了與上述相同的修改。 如果包的創建者將來對包進行修改，這可能會損壞或過時。 但就目前而言， ggforest2(stratamodel)將產生與方法 1 相同的結果。

Answer 2

正如這個答案所述， ggforest查看attr(model$terms, "dataClasses")[-1]以獲取 model 條款。

所以讓我們給它想要的東西？

library(dplyr)

ggforest(stratamodel, data = colon %>% mutate(`strata(rx)` = rx))

通過與summary(stratamodel) output 進行交叉檢查，這似乎是正確的。

不幸的是，我們留下了不必要的分層行。 也許有辦法解決這個問題。

編輯：有！

我們嘗試不同的策略，我們更改 model 的屬性，以便ggforest function 不會搜索這些術語。

attr(stratamodel$terms, "dataClasses") <- attr(stratamodel$terms, "dataClasses")[-3]
ggforest(stratamodel, data = colon)

我們有我們需要的東西。