R: Random sampling an even number of observations from a range of categories

Question

I previously took a random sample of postcodes from my dataframe and then realised that I wasn't sampling across all higher level statistical units. I have around 1 million postcodes and 7000 middle output statistical units. I want the sample to have roughly the same number of postcodes from each statistical unit.

How do I randomly sample 35 postcodes from each higher level statistical unit?

I used the following code previously to randomly sample 250,000 postcodes:

total.sample <- total[sample(1:nrow(total), 250000,
                           replace=FALSE),] 

How do I specify a random sample quota of postcodes based on another column variable (eg such as the higher level statistical unit (see msoa.rank in the dataframe structure below))?

Database stucture:

'data.frame':   1096289 obs. of  25 variables:
$ pcd                : Factor w/ 986055 levels "AL100AB","AL100AD",..: 282268 282258 
$ mbps2              : int  0 1 0 0 0 1 0 0 0 0 ...
$ averagesp          : num  16 7.8 7.8 9.5 9.4 3.2 11.1 19.4 10.5 11.8 ...
$ mediansp           : num  18.2 8 7.8 8.1 8.5 3.2 8.1 18.7 9.7 8.9 ...
$ nga                : int  0 0 0 0 0 0 0 0 0 0 ...
$ x                  : int  533432 532192 533416 533223 532866 531394 532899 532744 
$ total.dps          : int  11 91 10 7 9 10 3 5 21 12 ...
$ connections.density: num  7.909 0.747 3.1 7.714 1.889 ...
$ urban              : int  1 1 1 1 1 1 1 1 1 1 ...
$ gross.pay          : num  36607 36607 36607 36607 36607 ...
$ p.tert             : num  98.8 98.8 98.8 98.8 98.8 ...
$ p.kibs             : num  70.3 70.3 70.3 70.3 70.3 ...
$ density            : num  25.5 25.5 25.5 25.5 25.5 25.5 25.5 25.5 25.5 25.5 ...
$ p_m_s              : num  93.5 93.5 93.5 93.5 93.5 ...
$ p_m_l              : num  6.52 6.52 6.52 6.52 6.52 ...
$ p.edu              : num  62.6 62.6 62.6 62.6 62.6 ...
$ p.claim            : num  1.58 1.58 1.58 1.58 1.58 ...
$ p.non.white        : num  21.4 21.4 21.4 21.4 21.4 21.4 21.4 21.4 21.4 21.4 ...
$ msoa.rank          : int  2 2 2 2 2 2 2 2 2 2 ...
$ oslaua.rank        : int  321 321 321 321 321 321 321 321 321 321 ...
$ nuts2.rank         : int  22 22 22 22 22 22 22 22 22 22 ...
$ gor.rank           : int  8 8 8 8 8 8 8 8 8 8 ...
$ cons               : int  1 1 1 1 1 1 1 1 1 1 ...

pcd = postcode

msoa.rank = the ordinal variable of each middle output statistical unit

Answer 1

This is easy will hopefully soon be fixed in the dplyr (thanks, @Henrik!)

library(dplyr)
yourSample <- yourData %>%
    group_by(msoa.rank) %>%
    sample_n(size = 35) # currently buggy

Until then, here's a workaround, (thanks to @beginneR)

yourSample <- yourData %>%
    group_by(msoa.rank) %>%
    do(sample_n(., size = 35))

or stick with plain old plyr for the grouping.

library(plyr)
yourSample <- ddply(yourData, "msoa.rank", .fun = function(x) {sample_n(x, size = 35)})

Answer 2

Does every msoa.rank have at least 35 postcodes? This will be fast with data.table

#Create a data.table object
require(data.table)
total <- data.table(total)

#Sample by each msoa.rank group (take a sample that is size min(35,total size of msoa grp)
total.sample <- total[ , .SD[sample(1:.N,min(35,.N))], by=msoa.rank]

So here is how to example would work using the classic iris dataset.

iris < data.table(iris)
set.seed(2014)
iris.sample <- iris[ , .SD[sample(1:.N,min(10,.N))], by=Species]
summary(iris.sample$Sepal.Length)

Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
4.400   5.000   5.850   5.797   6.525   7.200 

Here is another sample and summary to see the difference

iris.sample2 <- iris[ , .SD[sample(1:.N,min(10,.N))], by=Species]
summary(iris.sample2$Sepal.Length)

Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 
4.400   5.100   5.850   5.743   6.275   7.300