Work out how many to take from each group so that a sample matches the proportions in the population, and round the results sensibly.
Last stop the fix was a list and a generator: number all 900 members, let chance pick 60. Every member gets the same chance, and that is fair. It can still land oddly. One draw might come back with four volunteers when twelve is their share, because chance cannot read the membership list. Stratified sampling settles the shares first. Split the population into groups that do not overlap and leave nobody out, called strata, and one of them is a stratum. Then decide how many places each group gets before a single person is picked.
A group's share of the population is the group's size divided by the total. Multiply that share by the sample size and you have the group's places. A group of 240 out of 900 has a share of 240 / 900, so in a sample of 60 it gets 240 / 900 x 60 = 16 places. The same sum works the other way round: 60 out of 900 is 1 in 15, and 240 / 15 = 16. It runs backwards: 32 park places from 480 members is 1 in 15, so the sample is 900 / 15 = 60.
Wren Lane has 240 plot holders, 180 volunteers and 480 park members, 900 in all. For a sample of 60 the three sums give 240 / 900 x 60 = 16, 180 / 900 x 60 = 12 and 480 / 900 x 60 = 32. Check the total: 16 + 12 + 32 = 60. A second check: the groups sit in the ratio 4:3:8, and 16:12:32 is that ratio times four. A group with more places than its share is over-represented, and none is here.
Real numbers rarely divide so kindly. Take the same three groups for a sample of 50. 240 / 900 x 50 comes to about 13.3, 180 / 900 x 50 is exactly 10, and 480 / 900 x 50 is about 26.7. Nobody can ask part of a person, so round each to the nearest whole number: 13, 10 and 27, which totals 50. Round both decimals up instead and you get 14, 10 and 27, which is 51. Rounding pushed the 13.3 furthest, so pull it back to 13 and the total is 50 again. Round, add up, adjust.
Two things stratifying does not do. It fixes the shares of the groups you chose to split by and nothing else, so a difference you never thought of can still be missing from the sample. And somebody still has to pick the 16 and the 12 inside each group. If that picking happens at the Saturday gate, the bias from last stop walks straight back in. Strata earn their keep when the groups tend to answer differently, so splitting by something unrelated to the question just gives you groups that are copies of each other.
A group's share is its size divided by the population, and that share times the sample size gives the group's places. Add the places up to check they make the sample size, and compare the ratio with the population's. When the division leaves a decimal, round each group and then fix the total. Stratifying sets the shares; picking at random inside each group is what keeps them honest.
The gate is gone and the office list is out. Wren Lane's 900 members are 240 plot holders, 180 volunteers and 480 park members, and the crew wants 60 of them. The quickest suggestion in the room is 20 each: three groups, 60 places, done.
Before that goes on the board, check it against the site.
This opens up
Jobs that lean on this skill. Follow one to see everything it is built on.
24 questions wait behind this lesson, each with its answer explained. Every answer feeds the sky: stars light as they are learned, and dim when it is time to come back.