Take random sample by group

Question

I have a data frame made by almost 50,000 rows spread in 15 different IDs (every ID has thousands of observations). Data frame looks like:

        ID  Year    Temp    ph
1       P1  1996    11.3    6.80
2       P1  1996    9.7     6.90
3       P1  1997    9.8     7.10
...
2000    P2  1997    10.5    6.90
2001    P2  1997    9.9     7.00
2002    P2  1997    10.0    6.93

I want to take 500 random rows for every ID (so 500 for P1, 500 for P2,....) and create a new df. I try:

new_df<-df[df$ID %in% sample(unique(dfID),500),]

But it takes randomly one ID, while I need 500 random rows for every ID.

Answer 1

Try this:

library(plyr)
ddply(df,.(ID),function(x) x[sample(nrow(x),500),])

Answer 2

This is available as the slice_sample function in dplyr:

library(dplyr)
new_df <- df %>% group_by(ID) %>% slice_sample(n=500)

In older versions of R, the function was called sample_n, which has been deprecated.

Answer 3

Here is one approach in base R.

First, the prerequisite sample data to work with:

set.seed(1)
mydf <- data.frame(ID = rep(1:3, each = 5), matrix(rnorm(45), ncol = 3))
mydf
#    ID         X1          X2          X3
# 1   1 -0.6264538 -0.04493361  1.35867955
# 2   1  0.1836433 -0.01619026 -0.10278773
# 3   1 -0.8356286  0.94383621  0.38767161
# 4   1  1.5952808  0.82122120 -0.05380504
# 5   1  0.3295078  0.59390132 -1.37705956
# 6   2 -0.8204684  0.91897737 -0.41499456
# 7   2  0.4874291  0.78213630 -0.39428995
# 8   2  0.7383247  0.07456498 -0.05931340
# 9   2  0.5757814 -1.98935170  1.10002537
# 10  2 -0.3053884  0.61982575  0.76317575
# 11  3  1.5117812 -0.05612874 -0.16452360
# 12  3  0.3898432 -0.15579551 -0.25336168
# 13  3 -0.6212406 -1.47075238  0.69696338
# 14  3 -2.2146999 -0.47815006  0.55666320
# 15  3  1.1249309  0.41794156 -0.68875569

Second, the sampling:

do.call(rbind, 
        lapply(split(mydf, mydf$ID), 
               function(x) x[sample(nrow(x), 3), ]))
#      ID         X1          X2         X3
# 1.2   1  0.1836433 -0.01619026 -0.1027877
# 1.1   1 -0.6264538 -0.04493361  1.3586796
# 1.5   1  0.3295078  0.59390132 -1.3770596
# 2.10  2 -0.3053884  0.61982575  0.7631757
# 2.9   2  0.5757814 -1.98935170  1.1000254
# 2.8   2  0.7383247  0.07456498 -0.0593134
# 3.13  3 -0.6212406 -1.47075238  0.6969634
# 3.12  3  0.3898432 -0.15579551 -0.2533617
# 3.15  3  1.1249309  0.41794156 -0.6887557

---

There is also strata from the sampling package, which is convenient when you want to sample different sizes from each group:

# install.packages("sampling")
library(sampling)
set.seed(1)
x <- strata(mydf, "ID", size = c(2, 3, 2), method = "srswor")
getdata(mydf, x)
#            X1          X2         X3 ID ID_unit Prob Stratum
# 2   0.1836433 -0.01619026 -0.1027877  1       2  0.4       1
# 5   0.3295078  0.59390132 -1.3770596  1       5  0.4       1
# 6  -0.8204684  0.91897737 -0.4149946  2       6  0.6       2
# 8   0.7383247  0.07456498 -0.0593134  2       8  0.6       2
# 9   0.5757814 -1.98935170  1.1000254  2       9  0.6       2
# 14 -2.2146999 -0.47815006  0.5566632  3      14  0.4       3
# 15  1.1249309  0.41794156 -0.6887557  3      15  0.4       3

Answer 4

In case you have big datasets, a data.table solution could go like this:

library(data.table)

# Generate 26 mil rows random data
set.seed(2019)
dt <- data.table(c1 = sample(length(LETTERS)*10^6), 
                 c2 = sample(LETTERS, replace = TRUE))

# For each letter, sample 500 rows
dt_sample <- dt[, .SD[sample(x = .N, size = 500)], by = c2]

# We indeed sampled 500 rows for each letter
dt_sample[, .N, by = c2][order(c2)]
#>     c2   N
#>  1:  A 500
#>  2:  D 500
#>  3:  G 500
#>  4:  I 500
#>  5:  M 500
#>  6:  N 500
#>  7:  O 500
#>  8:  P 500
#>  9:  Q 500
#> 10:  R 500
#> 11:  S 500
#> 12:  T 500
#> 13:  U 500
#> 14:  V 500
#> 15:  W 500
#> 16:  Y 500
#> 17:  Z 500

^{Created on 2019-04-23 by the reprex package (v0.2.1)}

In case your data is unbalanced in the sense that some groups happen to be smaller (as number of rows) than your desired sample size, then you need to set a defensive trick like sample size should be min(500, .N) - see sample random rows within each group in a data.table. So like:

dt[, .SD[sample(x = .N, size = min(500, .N))], by = c2]

Answer 5

An approach if on of the IDs is < 500. Here I used the mtcars set:

n <- 8
df <- mtcars
df$ID <- df$cyl

FUN <- function(x, n) {
    if (length(x) <= n) return(x)
    x[x %in% sample(x, n)]
}

df[unlist(lapply(split(1:nrow(df), df$ID), FUN, n = 8)), ]

Answer 6

Here's an elegant solution based on data.table. You can randomly draw IDs from a panel data set (balanced or unbalanced) in three simple steps:

Step 1: Store unique IDs from your original data set in a vector (my data set is called "main" and the identifier is called "id"):

ids <- unique(main$id)

Step 2: Randomly draw IDs from the vector from step 1. In the example below, I randomly draw 50 IDs from the vector "ids" and store them in the new vector "draw":

draw <- ids %>% sample(50)

Step 3: Subset rows in your original data set based on matches with the IDs drawn in step 2.

rsample <- main[main$id %in% draw, ]

Answer 7

mydata1 is your original data(not tested)

mydata2<- split(mydata1,mydata1$ID)
names(mydata2)<-paste0("mydata2",1:length(levels(ID))) 
mysample<-Map(function(x) x[sample((1:nrow(x)),size=500,replace=FALSE),], mydata2)

library(plyr)# for rbinding the mysample
ldply(mysample)

Answer 8

Although this is not very elegant solution, but it may work.

library(data.table)
df <- data.table(df)
f <- list()
for(i in unique(df1$ID)){
 f[[i]] <- df1[id == i][sample(.N,(500))]
  }
 dfnew <- rbindlist(f)

Answer 9

library(data.table) #1
df <- data.table(df) #2
df[,group_num := sample(2,.N,replace = TRUE,prob = c(500,.N-500)/.N),by = "ID"] #3
df_sample = df[group_num == 1,] #4

or you can change line #3 and #4 to:

df[,random_num := sample(.N,.N),by="ID"]
df_sample  = df[random_num <=500,]