Introduction au bootstrap

Échantillonnage en R

Richie Cotton

Data Evangelist at DataCamp

Avec ou sans

Échantillonnage sans remise

Cartes à jouer sur une table de casino.

Échantillonnage avec remise (« rééchantillonnage »)

Quatre dés qui roulent.

Échantillonnage en R

Échantillonnage aléatoire simple sans remise

Population

Grains de café disposés en rangées et colonnes.

Échantillon

Grains de café en rangées et colonnes, dont la plupart sont grisés.

Échantillonnage en R

Échantillonnage aléatoire simple avec remise

Population

Grains de café disposés en rangées et colonnes.

Échantillon

Un échantillon aléatoire de grains de café, dont certains en double.

Échantillonnage en R

Pourquoi échantillonner avec remise ?

  • Considérez les données coffee_ratings comme un échantillon d'une plus grande population de cafés.
  • Considérez chaque café de l'échantillon comme représentant plusieurs cafés absents de l'échantillon mais présents dans la population.
  • L'échantillonnage avec remise sert de substitut pour inclure différents membres de ces groupes dans l'échantillon.
Échantillonnage en R

Préparer les données sur le café

coffee_focus <- coffee_ratings %>%
  select(variety, country_of_origin, flavor) %>%
  rowid_to_column()
glimpse(coffee_focus)
Rows: 1,338
Columns: 4
$ rowid             <int> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, ...
$ variety           <chr> NA, "Other", "Bourbon", NA, "Other", NA, "Other", N...
$ country_of_origin <chr> "Ethiopia", "Ethiopia", "Guatemala", "Ethiopia", "E...
$ flavor            <dbl> 8.83, 8.67, 8.50, 8.58, 8.50, 8.42, 8.50, 8.33, 8.6...
Échantillonnage en R

Rééchantillonner avec slice_sample()

coffee_resamp <- coffee_focus %>%
  slice_sample(prop = 1, replace = TRUE)
# A tibble: 1,338 x 4
   rowid variety country_of_origin flavor
   <int> <chr>   <chr>              <dbl>
 1  1253 Bourbon Guatemala           6.92
 2   186 Caturra Colombia            7.58
 3  1185 Bourbon Guatemala           7.42
 4  1273 NA      Philippines         6.5 
 5  1042 Caturra Honduras            7.33
 6   195 Caturra Guatemala           7.75
 7  1219 Typica  Mexico              7   
 8   952 Caturra Honduras            7.5 
 9    41 Caturra Thailand            8.33
10   460 Caturra Honduras            7.67
# ... with 1,328 more rows
Échantillonnage en R

Cafés répétés

coffee_resamp %>% 
  count(rowid, sort = TRUE)
# A tibble: 844 x 2
   rowid     n
   <int> <int>
 1   704     5
 2   913     5
 3  1070     5
 4    16     4
 5   180     4
 6   230     4
 7   234     4
 8   342     4
 9   354     4
10   423     4
# ... with 834 more rows
Échantillonnage en R

Cafés manquants

coffee_resamp %>% 
  summarize(
    coffees_included = n_distinct(rowid),
    coffees_not_included = n() - coffees_included
  )
# A tibble: 1 x 2
  coffees_included coffees_not_included
             <int>                <int>
1              844                  494
Échantillonnage en R

Bootstrap

L'inverse de l'échantillonnage à partir d'une population.

Échantillonnage : passer d'une population à un plus petit échantillon.

Bootstrap : construire une population théorique à partir de votre échantillon.

Cas d'usage du bootstrap

  • Comprendre la variabilité d'échantillonnage à partir d'un seul échantillon.

Une botte de cowboy.

Échantillonnage en R

Processus de bootstrap

  1. Créer un rééchantillon de même taille que l'échantillon initial.
  2. Calculer la statistique d'intérêt pour cet échantillon bootstrap.
  3. Répéter les étapes 1 et 2 plusieurs fois.

Les statistiques obtenues sont des statistiques bootstrap et, lorsqu'on les examine pour leur variabilité, forment une distribution bootstrap.

Échantillonnage en R

Bootstrap de la saveur moyenne du café

# Step 3. Repeat many times
mean_flavors_1000 <- replicate(
  n = 1000,
  expr = {
    coffee_focus %>%
      # Step 1. Resample
      slice_sample(prop = 1, replace = TRUE) %>%
      # Step 2. Calculate statistic
      summarize(mean_flavor = mean(flavor, na.rm = TRUE)) %>% 
      pull(mean_flavor)
  })
Échantillonnage en R

Histogramme de la distribution bootstrap

bootstrap_distn <- tibble(
  resample_mean = mean_flavors_1000
)
ggplot(bootstrap_distn, aes(resample_mean)) +
  geom_histogram(binwidth = 0.0025)

Histogramme de la distribution bootstrap de la moyenne d'échantillon.

Échantillonnage en R

Passons à la pratique !

Échantillonnage en R

Preparing Video For Download...