在 GLM 适合 R 之前组织因子变量

Question

我在 R 中使用 GLM 做了很多工作，在相当大的数据集上（通常在模型拟合中包括几十个变量）。 为了能够在拟合模型后生成某种类型的图形输出，我发现在拟合之前“准备”任何我打算作为因子拟合的变量（名称以f_开头的变量）很有用模型拟合。 我的意思是：

(i) 因为我在拟合 GLM 之前对每个因子重新排序，使得参考级别等于权重最大的级别，所以我想在relevel()命令之前保留级别排序；

(ii) 为了以后在图表中突出参考水平，我喜欢将其记录在一个单独的变量中。

我已经使用内置的mtcars数据集将这种方法放在一起示例。

目前我有这个代码：

library(dplyr)

data(mtcars)

# tidy up and make easier to read

df <- mtcars # built in data set

# let's make it a bit easier to follow

df <- df %>%
  select(mpg,
         f_cylinders = cyl,
         c_displacement = disp,
         c_hp = hp,
         c_final_drive_ratio = drat,
         c_weight = wt,
         c_qtr_mile_time = qsec,
         f_v_or_straight = vs,
         f_transmission = am,
         f_gears = gear,
         f_num_carbs = carb)

df$f_v_or_straight <- ifelse(df$f_v_or_straight == 0, "V", "Straight")
df$f_transmission <- ifelse(df$f_transmission == 0, "Automatic", "Manual")

df$glm_weight <- 1

# organise factors - levels, reference level, weights

my_list <- list()

df$f_cylinders <- as.factor(df$f_cylinders)
my_list$f_cylinders_levels <- levels(df$f_cylinders) 
my_list$f_cylinders_weights <- df %>% group_by(f_cylinders) %>% summarise(glm_weight = sum(glm_weight)) %>% ungroup() %>% pull(glm_weight)
my_list$f_cylinders_ref <- "8"
df$f_cylinders <- df$f_cylinders %>% relevel(ref = my_list$f_cylinders_ref)

df$f_v_or_straight <- as.factor(df$f_v_or_straight)
my_list$f_v_or_straight_levels <- levels(df$f_v_or_straight) 
my_list$f_v_or_straight_weights <- df %>% group_by(f_v_or_straight) %>% summarise(glm_weight = sum(glm_weight)) %>% ungroup() %>% pull(glm_weight)
my_list$f_v_or_straight_ref <- "V"
df$f_v_or_straight <- df$f_v_or_straight %>% relevel(ref = my_list$f_v_or_straight_ref)

df$f_transmission <- as.factor(df$f_transmission)
my_list$f_transmission_levels <- levels(df$f_transmission) 
my_list$f_transmission_weights <- df %>% group_by(f_transmission) %>% summarise(glm_weight = sum(glm_weight)) %>% ungroup() %>% pull(glm_weight)
my_list$f_transmission_ref <- "Automatic"
df$f_transmission <- df$f_transmission %>% relevel(ref = my_list$f_transmission_ref)

df$f_gears <- as.factor(df$f_gears)
my_list$f_gears_levels <- levels(df$f_gears) 
my_list$f_gears_weights <- df %>% group_by(f_gears) %>% summarise(glm_weight = sum(glm_weight)) %>% ungroup() %>% pull(glm_weight)
my_list$f_gears_ref <- "3"
df$f_gears <- df$f_gears %>% relevel(ref = my_list$f_gears_ref)

df$f_num_carbs <- as.factor(df$f_num_carbs)
my_list$f_num_carbs_levels <- levels(df$f_num_carbs) 
my_list$f_num_carbs_weights <- df %>% group_by(f_num_carbs) %>% summarise(glm_weight = sum(glm_weight)) %>% ungroup() %>% pull(glm_weight)
my_list$f_num_carbs_ref <- "4"
df$f_num_carbs <- df$f_num_carbs %>% relevel(ref = my_list$f_num_carbs_ref)

这段代码工作正常，但是……在现实世界中，我正在处理数十个因子变量，而不仅仅是上面的 5 个。 所以如果我有 50 个因子变量，我会重复做同样的事情 50 次。 我想将此准备工作捆绑到一个函数调用中，基本上是说：

对于名称以f_开头的每个字段（即看起来像f_xxx ）：

1. 把它从chr / int /whatever 变成一个因子f_xxx ；

2. 计算权重f_xxx_weights

3. 计算出参考水平f_xxx_ref （如果领先，不确定该怎么做）；

4. 将当前因子水平存储在f_xxx_levels ；

5. 重新排列因子水平，使f_xxx_ref成为列表中的第一个。

我在这里问了很多……但任何能推动我前进的东西都会非常感激。

谢谢你。

Answer 1

考虑在用户定义的方法中概括您的relevel过程，然后使用purrr::map_df映射调用您的过程的f_列（与 tidy 保持一致）：

数据

library(dplyr)
library(purrr)

df <- mtcars %>%
  select(mpg,
         f_cylinders = cyl,
         c_displacement = disp,
         c_hp = hp,
         c_final_drive_ratio = drat,
         c_weight = wt,
         c_qtr_mile_time = qsec,
         f_v_or_straight = vs,
         f_transmission = am,
         f_gears = gear,
         f_num_carbs = carb) %>%
  mutate(f_v_or_straight = ifelse(f_v_or_straight == 0, 
                                  "V", 
                                  "Straight"),
         f_transmission = ifelse(f_transmission == 0, 
                                 "Automatic", 
                                 "Manual"),
         glm_weight = 1)

rlevel进程（使用table频率）

proc_rlevel <- function(col) {
    agg <- df %>% group_by_at(col) %>% 
      summarise(glm_weight = sum(glm_weight)) %>%
      arrange(desc(glm_weight))

    f_ref <- df[[col]] %>% 
      as.character() %>%
      as.factor() %>%
      relevel(ref = paste(agg[[col]][1]))

    return(f_ref)
}

# REPLACING ORIGINAL f_cols WITH TWO WAY PIPES
df[grep("f_", names(df))] %<>% 
  names() %>%
  setNames(identity(.)) %>%
  map_df(proc_rlevel)

查看更改

# ORIGINAL LEVELS
df %>% 
  select(starts_with("f_")) %>%
  map_df(as.factor) %>%
  map(levels)
# $f_cylinders
# [1] "4" "6" "8"
# 
# $f_v_or_straight
# [1] "Straight" "V"       
# 
# $f_transmission
# [1] "Automatic" "Manual"   
# 
# $f_gears
# [1] "3" "4" "5"
# 
# $f_num_carbs
# [1] "1" "2" "3" "4" "6" "8"


# ADJUSTED LEVELS
df %>% 
  select(starts_with("f_")) %>%
  map_df(as.factor) %>%
  map(levels)
# $f_cylinders
# [1] "8" "4" "6"
# 
# $f_v_or_straight
# [1] "V"        "Straight"
# 
# $f_transmission
# [1] "Automatic" "Manual"   
# 
# $f_gears
# [1] "3" "4" "5"
# 
# $f_num_carbs
# [1] "2" "1" "3" "4" "6" "8"