Diametrically opposite to stratified sampling is cluster sampling. It consists of first selecting, at random, groups called clusters of individual items from the population and then choosing all or a sub-sample of the items within each cluster to make up the overall sample.
To get the best results in cluster sampling design, differences between clusters are made as small as possible.
In contrast, differences among individual items within each cluster are made as large as possible. Ideally, each cluster should be a miniature of the entire population, and thus a single cluster would be a satisfactory sample.
To illustrate how a cluster sampling works in practice, suppose that we need a random sample of n=200 households from a population of N=8,000 households of a city.
Since there does not exist any good list of the households, it would be a difficult job to sample the individual households. It would be at the same time too expensive to prepare such a list.
Instead, we can obtain a sample of blocks by dividing the entire area into several blocks and then selecting 200/8000=2.5% of the blocks. Suppose we make 80 blocks, each with 100 households.
Then 2.5% of 80 blocks implies 80×2.5%=2 sample blocks. These 2 blocks contain 200 households. These households located within the boundaries of the sample blocks comprise the sample.
In the above example, the blocks are identified as clusters, and hence they represent the sampling units. The households are the elementary units determined by the objective of the analysis.
Whether or not a group of elements would form a cluster depends on the circumstances. In the above example, a block was appropriately called a cluster, since it contained some households.
In another survey, the households might properly be called clusters if it represents a sampling unit, where the objectives were to study the characteristics of the individual members of the households.