如何用GTSUMMARY的tbl_ard_summary实现多分组汇总表？

问题背景 #

已通过tbl_strata+tbl_summary实现按trt分层、每层内按grade分组的age/ttdeath连续变量汇总表，代码如下：

# Summary by TRT & GRADE
trial %>%
  mutate(grade = paste("Grade", grade)) %>%
  tbl_strata(
    strata = trt,
    .tbl_fun = ~.x %>%
      tbl_summary(
        by = grade,
        include = c(age, ttdeath),
        type = c(age, ttdeath) ~ "continuous2",
        statistic = c(age, ttdeath) ~ c("{N_obs}", "{mean} ({sd})",
                                        "{median} ({p25}, {p75})", "{min}, {max}")) %>%
    add_n(),
    .header = "**{strata}**, N = {n}"
  )

尝试用CARDS包生成ARD，再结合tbl_ard_summary复现相同结果时失败。已生成ARD的代码：

# ARD table for AGE, TTDEATH by TRT & GRADE
ard.cont <- ard_continuous(
  trial,
  by = c(trt, grade),
  variables = c(age, ttdeath)) %>%
  bind_ard(
    ard_attributes(
      trial,
      variables = c(trt, grade, age, ttdeath)))

失败的调用代码：

# Summary from ARD
trial %>% tbl_strata(
  strata = trt,
  ~.x %>%
    tbl_ard_summary(
      cards = ard.cont,
      by = grade,
      include = c(age, ttdeath),
      type = c(age, ttdeath) ~ "continuous2",
      statistic = c(age, ttdeath) ~ c("{N}", "{mean} ({sd})", "{median} ({p25}, {p75})", "{min}, {max}")
    ))

解决方案 #

可以通过ARD实现需求，问题核心在于ARD数据与分层子集不匹配：tbl_strata会按trt拆分数据，但传入的ard.cont包含所有trt的完整统计，tbl_ard_summary无法自动过滤当前分层的trt，同时需保证grade的格式统一、统计项映射一致。以下是两种可行方案：

方案1：分层内生成对应ARD #

在tbl_strata的每层中，针对当前trt子集生成专属ARD，避免数据不匹配：

# 用ARD复现分层汇总表
trial %>%
  mutate(grade = paste("Grade", grade)) %>%
  tbl_strata(
    strata = trt,
    .tbl_fun = ~.x %>%
      # 为当前分层生成对应ARD
      ard_continuous(
        by = grade,
        variables = c(age, ttdeath)
      ) %>%
      bind_ard(ard_attributes(.x, variables = c(grade, age, ttdeath))) %>%
      tbl_ard_summary(
        by = grade,
        include = c(age, ttdeath),
        type = c(age, ttdeath) ~ "continuous2",
        statistic = c(age, ttdeath) ~ c("{N_obs}", "{mean} ({sd})",
                                        "{median} ({p25}, {p75})", "{min}, {max}")
      ) %>%
      add_n(),
    .header = "**{strata}**, N = {n}"
  )

方案2：提前生成完整ARD，分层内过滤 #

先生成包含所有trt和grade的完整ARD，在分层内过滤出当前trt的统计数据：

# 提前生成完整ARD（统一grade格式）
ard.cont <- trial %>%
  mutate(grade = paste("Grade", grade)) %>%
  ard_continuous(
    by = c(trt, grade),
    variables = c(age, ttdeath)
  ) %>%
  bind_ard(
    ard_attributes(
      trial %>% mutate(grade = paste("Grade", grade)),
      variables = c(trt, grade, age, ttdeath)
    )
  )

# 分层调用时过滤当前trt的ARD数据
trial %>%
  mutate(grade = paste("Grade", grade)) %>%
  tbl_strata(
    strata = trt,
    .tbl_fun = function(x) {
      current_trt <- unique(x$trt)
      ard.cont %>%
        filter(across(trt, ~.x == current_trt)) %>%
        tbl_ard_summary(
          by = grade,
          include = c(age, ttdeath),
          type = c(age, ttdeath) ~ "continuous2",
          statistic = c(age, ttdeath) ~ c("{N_obs}", "{mean} ({sd})",
                                          "{median} ({p25}, {p75})", "{min}, {max}")
        ) %>%
        add_n()
    },
    .header = "**{strata}**, N = {n}"
  )

关键注意点 #

1. grade格式统一：在生成ARD和处理原始数据时，都要执行mutate(grade = paste("Grade", grade))，保证分组标签一致；
2. 统计项映射：ARD中连续变量的样本量统计项为N_obs，需将statistic中的{N}改为{N_obs}，避免匹配失败；
3. ARD过滤逻辑：若提前生成完整ARD，必须在分层内过滤当前trt的子集，否则tbl_ard_summary会加载所有trt的统计数据，导致结果混乱。

内容的提问来源于stack exchange，提问作者Ranjan Karmakar