Stratified Sampling in R with dplyr

stratify.r

# Uses a subset of the Iris data set with different proportions of the Species factor
set.seed(42)
iris_subset <- iris[c(1:50, 51:80, 101:120), ]

stratified_sample <- iris_subset %>%
  group_by(Species) %>%
  mutate(num_rows=n()) %>%
  sample_frac(0.4, weight=num_rows) %>%
  ungroup

# These results should be equal
table(iris_subset$Species) / nrow(iris_subset)
table(stratified_sample$Species) / nrow(stratified_sample)

# Success!
#     setosa versicolor  virginica 
#       0.5        0.3        0.2 

Thank you for posting this. What if you wanted to stratify sampling based on two conditions? For example, if you wanted an equal proportions of both species and condition, using these data:

    iris_subset$condition <- rep(seq(1,5,by=1), 20)
    # next line does not run, but I'm wondering how it could. 
    stratified_sample <- iris_subset %>%
      group_by(c(Species,condition)) %>%
      mutate(num_rows=n()) %>%
      sample_frac(0.4, weight=num_rows) %>%
      ungroup

Thanks for the post. A question about a previous stage: How can I get/calculate de optimum sample size?

This is nice, introduced sample_frac to me thanks!

However, I might not have entirely understood now having read the docs ... why do we need to have the num_rows and weight by them? The docs say subset_frac honours any grouping so I would have thought this also achieves a stratified sample:

stratified_sample <- iris_subset %>%
  group_by(Species) %>%
  sample_frac(0.4) %>%
  ungroup

Indeed, I think sample_frac will by definition see the same weight in each group so that the weight has no effect after grouping?

Big thanks! Don't know how much it helped me on my projects! :D

Thank you for posting this. What if you wanted to stratify sampling based on two conditions? For example, if you wanted an equal proportions of both species and condition, using these data:

    iris_subset$condition <- rep(seq(1,5,by=1), 20)
    # next line does not run, but I'm wondering how it could. 
    stratified_sample <- iris_subset %>%
      group_by(c(Species,condition)) %>%
      mutate(num_rows=n()) %>%
      sample_frac(0.4, weight=num_rows) %>%
      ungroup

maybe you could filter it after grouping? like that:

  stratified_sample <- iris_subset %>%
   group_by(c(Species)) %>%
   filter(condition == TRUE) %>%
   mutate(num_rows=n()) %>% ...

This is nice, introduced sample_frac to me thanks!

However, I might not have entirely understood now having read the docs ... why do we need to have the num_rows and weight by them? The docs say subset_frac honours any grouping so I would have thought this also achieves a stratified sample:

stratified_sample <- iris_subset %>%
  group_by(Species) %>%
  sample_frac(0.4) %>%
  ungroup

Indeed, I think sample_frac will by definition see the same weight in each group so that the weight has no effect after grouping?

I have the same question as above-- any responses on this?