使用 R 中数据帧列表中的 tidymodels 进行引导

Question

我正在运行使用 tidymodels 创建一个模型，其中按组拆分数据并对每个单独的数据帧进行运行回归。这很好用。但是，现在我还需要 bootstrap 我的结果。我不确定如何将它构建到我现有的代码中。

我的原始代码如下所示：

library(dplyr)

year <- rep(2014:2018, length.out=10000)
group <- sample(c(0,1,2,3,4,5,6), replace=TRUE, size=10000)
value <- sample(10000, replace=T)
female <- sample(c(0,1), replace=TRUE, size=10000)
smoker <- sample(c(0,1), replace=TRUE, size=10000)
dta <- data.frame(year=year, group=group, value=value, female=female, smoker=smoker)

# cut the dataset into list
table_list <- dta %>%
  group_by(year, group) %>%
  group_split()

# fit model per subgroup
model_list <- lapply(table_list, function(x) glm(smoker ~ female, data=x,
                                                 family=binomial(link="probit")))
# predict
pred_list <- lapply(model_list, function(x) predict.glm(x, type = "response"))

我想 bootstrap 替换以获得 bootstrapped 预测值。我的直觉是，我应该在创建 table_list 时通过创建随机样本来进一步拆分数据集。但我该怎么做呢？

感谢您的帮助。

Answer 1

这相当复杂，有分组和 bootstrapping，所以我可能会接近它 like this，使用 map() 两层深：

library(tidyverse)
library(tidymodels)
#> Registered S3 method overwritten by 'tune':
#>   method                   from   
#>   required_pkgs.model_spec parsnip

year <- rep(2014:2018, length.out=10000)
group <- sample(c(0,1,2,3,4,5,6), replace=TRUE, size=10000)
value <- sample(10000, replace=T)
female <- sample(c(0,1), replace=TRUE, size=10000)
smoker <- sample(c(0,1), replace=TRUE, size=10000)
dta <- tibble(year=year, group=group, value=value, female=female, smoker=smoker)


glm_boot_mods <- 
  dta %>%
  nest(data = c(-year, -group)) %>%
  mutate(boots = map(
    data,  
    ~ bootstraps(., times = 20) %>%
      mutate(model = map(.$splits, ~ glm(smoker ~ female, data = analysis(.x),
                                         family = binomial(link = "probit"))),
             preds = map2(model, .$splits, ~predict(.x, newdata = assessment(.y))))
    ))


glm_boot_mods
#> # A tibble: 35 × 4
#>     year group data               boots                
#>    <int> <dbl> <list>             <list>               
#>  1  2014     1 <tibble [288 × 3]> <bootstraps [20 × 4]>
#>  2  2015     4 <tibble [273 × 3]> <bootstraps [20 × 4]>
#>  3  2016     3 <tibble [301 × 3]> <bootstraps [20 × 4]>
#>  4  2017     2 <tibble [282 × 3]> <bootstraps [20 × 4]>
#>  5  2018     0 <tibble [276 × 3]> <bootstraps [20 × 4]>
#>  6  2014     3 <tibble [279 × 3]> <bootstraps [20 × 4]>
#>  7  2016     2 <tibble [314 × 3]> <bootstraps [20 × 4]>
#>  8  2018     1 <tibble [296 × 3]> <bootstraps [20 × 4]>
#>  9  2014     0 <tibble [304 × 3]> <bootstraps [20 × 4]>
#> 10  2015     6 <tibble [288 × 3]> <bootstraps [20 × 4]>
#> # … with 25 more rows

第一个 map() 为每个分组创建 bootstrap 重采样，然后我们更深一层 每个重采样 拟合模型并预测对于该重新抽样的保留意见。您可以在这里看到第一组的内部结构：

glm_boot_mods %>%
  head(1) %>% 
  pull(boots)
#> [[1]]
#> # Bootstrap sampling 
#> # A tibble: 20 × 4
#>    splits            id          model  preds      
#>    <list>            <chr>       <list> <list>     
#>  1 <split [288/111]> Bootstrap01 <glm>  <dbl [111]>
#>  2 <split [288/93]>  Bootstrap02 <glm>  <dbl [93]> 
#>  3 <split [288/103]> Bootstrap03 <glm>  <dbl [103]>
#>  4 <split [288/106]> Bootstrap04 <glm>  <dbl [106]>
#>  5 <split [288/109]> Bootstrap05 <glm>  <dbl [109]>
#>  6 <split [288/109]> Bootstrap06 <glm>  <dbl [109]>
#>  7 <split [288/92]>  Bootstrap07 <glm>  <dbl [92]> 
#>  8 <split [288/111]> Bootstrap08 <glm>  <dbl [111]>
#>  9 <split [288/99]>  Bootstrap09 <glm>  <dbl [99]> 
#> 10 <split [288/111]> Bootstrap10 <glm>  <dbl [111]>
#> 11 <split [288/102]> Bootstrap11 <glm>  <dbl [102]>
#> 12 <split [288/104]> Bootstrap12 <glm>  <dbl [104]>
#> 13 <split [288/115]> Bootstrap13 <glm>  <dbl [115]>
#> 14 <split [288/111]> Bootstrap14 <glm>  <dbl [111]>
#> 15 <split [288/108]> Bootstrap15 <glm>  <dbl [108]>
#> 16 <split [288/110]> Bootstrap16 <glm>  <dbl [110]>
#> 17 <split [288/110]> Bootstrap17 <glm>  <dbl [110]>
#> 18 <split [288/111]> Bootstrap18 <glm>  <dbl [111]>
#> 19 <split [288/103]> Bootstrap19 <glm>  <dbl [103]>
#> 20 <split [288/109]> Bootstrap20 <glm>  <dbl [109]>

^{由 reprex package (v2.0.1)}

创建于 2021-11-02

请注意，对于每个重新采样的保留观察结果都有预测。根据您要执行的操作，您可以在接下来需要处理的 glm_boot_mods 列上使用 unnest()。

使用 R 中数据帧列表中的 tidymodels 进行引导

Bootstrapping using tidymodels from a list of dataframes in R

r

tidymodels