Practicing imputing with different models

Dealing With Missing Data in R

Nicholas Tierney

Statistician

Lesson overview

Imputation using the simputation package
Use linear model to impute values with impute_lm
Assess new imputations
Build many imputation models
Compare imputations across different models and variables

How imputing using a linear model works

df

# A tibble: 5 x 3
      y    x1    x2
  <dbl> <dbl> <dbl>
1  2.67  2.43  3.27
2  3.87  3.55  1.45
3 NA     2.90  1.49
4  5.21  2.72  1.84
5 NA     4.29  1.15

df %>% 
   bind_shadow(only_miss = TRUE) %>%
   add_label_shadow() %>%
   impute_lm(y ~ x1 + x2)

# A tibble: 5 x 7
     y   x1    x2   y_NA any_missing
 <dbl> <dbl> <dbl> <fct> <chr>      
1 2.67  2.43  3.27 !NA   Not Missing
2 3.87  3.55  1.45 !NA   Not Missing
3 5.54  2.90  1.49 NA    Missing    
4 5.21  2.72  1.84 !NA   Not Missing
5 2.56  4.29  1.15 NA    Missing

Using impute_lm

aq_imp_lm <- airquality %>% bind_shadow() %>% add_label_shadow() %>%
  impute_lm(Solar.R ~ Wind + Temp + Month) %>%
  impute_lm(Ozone ~ Wind + Temp + Month)

aq_imp_lm

# A tibble: 153 x 13
    Ozone Solar.R  Wind  Temp Month   Day Ozone_NA Solar.R_NA
 *  <dbl>   <dbl> <dbl> <int> <int> <int> <fct>    <fct>     
 1  41       190    7.4    67     5     1 !NA      !NA       
 2  36       118    8      72     5     2 !NA      !NA       
 3  12       149   12.6    74     5     3 !NA      !NA       
 4  18       313   11.5    62     5     4 !NA      !NA       
 5  -9.04    138.  14.3    56     5     5 NA       NA        
 6  28       178.  14.9    66     5     6 !NA      NA           
# ... with 147 more rows, and 5 more variables: Wind_NA <fct>,
#   Temp_NA <fct>, Month_NA <fct>, Day_NA <fct>,
#   any_missing <chr>

Tracking missing values

aq_imp_lm <-
airquality %>%
  bind_shadow() %>%
  add_label_missings() %>%
  impute_lm(Solar.R ~ Wind + Temp + 
            Month) %>%
  impute_lm(Ozone ~ Wind + Temp + 
            Month)

ggplot(aq_imp_lm,
       aes(x = Solar.R,
           y = Ozone,
           color = any_missing)) +
  geom_point()

Evaluating imputations: evaluating and comparing imputations

aq_imp_small <- airquality %>%
  bind_shadow() %>%
  impute_lm(Ozone ~ Wind + Temp) %>%
  impute_lm(Solar.R ~ Wind + Temp) %>%
  add_label_shadow()

aq_imp_large <- airquality %>%
  bind_shadow() %>%
  impute_lm(Ozone ~ Wind + Temp + Month + Day) %>%
  impute_lm(Solar.R ~ Wind + Temp + Month + Day)  %>%
  add_label_shadow()

Evaluating imputations: binding and visualizing many models

bound_models <- bind_rows(small = aq_imp_small,
                            large = aq_imp_large,
                            .id = "imp_model")
bound_models

     imp_model     Ozone  Solar.R Wind Temp Month Day
  1:     small  41.00000 190.0000  7.4   67     5   1
  2:     small  36.00000 118.0000  8.0   72     5   2
  3:     small  12.00000 149.0000 12.6   74     5   3
 ...                                                
304:     large  14.00000 191.0000 14.3   75     9  28
305:     large  18.00000 131.0000  8.0   76     9  29
306:     large  20.00000 223.0000 11.5   68     9  30

Evaluating imputations: exploring many imputations

ggplot(bound_models,
       aes(x = Ozone,
           y = Solar.R,
           color = any_missing)) +
  geom_point() + 
  facet_wrap(~ imp_model)

bound_models_gather <- bound_models %>%
  select(Ozone, Solar.R, any_missing, imp_model) %>%
  gather(key = "variable", value = "value", -any_missing, -imp_model)
bound_models_gather

     any_missing imp_model variable     value
  1: Not Missing     small    Ozone  41.00000
  2: Not Missing     small    Ozone  36.00000
  3: Not Missing     small    Ozone  12.00000
  4: Not Missing     small    Ozone  18.00000
  5:     Missing     small    Ozone -11.67673
 ...                                         
608: Not Missing     large  Solar.R 193.00000
609:     Missing     large  Solar.R 145.00000
610: Not Missing     large  Solar.R 191.00000
611: Not Missing     large  Solar.R 131.00000
612: Not Missing     large  Solar.R 223.00000

Explore imputations in multiple variables and models

ggplot(bound_models_gather,
       aes(x = imp_model,
           y = value)) +
  geom_boxplot() + 
  facet_wrap(~ key)

Explore imputations in multiple variables and models

bound_models_gather %>%
  filter(any_missing == "Missing") %>%
  ggplot(aes(x = imp_model,
             y = value)) +
  geom_boxplot() + 
  facet_wrap(~ key)

Let's practice!

Dealing With Missing Data in R