A type of random sample which is designed by first classifying the overall population into sub-groups or strata according to some principle, and then taking a simple random sample within each stratum. Its purpose is to ensure that the sample is, in fact, representative of the various groups which make up the population, a property which a simple random sample need by no means possess. For example, in order to design a sample to estimate the expenditure by families on house rents, we might divide the total population into broad income groups, decide on the sample size appropriate for each group, and then choose the families in each group according to the same procedure as would be used for a simple random sample. If this stratification were not done, the sample chosen might, by chance, contain an over-representative figure for one income group and an under-representative figure for another, and thus the estimate of average expenditure on rent per household would be biased upward or downward respectively. In designing a stratified sample we need some information about the population to be available before the sample is carried out. This information is then used to increase the representativeness of the sample. The stratified sample is still a random sample, since each item in the population has a known chance of being selected. The basis of the stratification, or the ‘stratification factor’, is chosen so as to be relevant to the problem in hand. As well as increasing the representativeness, and therefore accuracy, of the sample, stratification may also be used because it is of interest to see how the results differ between the strata, e.g. how rent expenditures vary between income groups.
Reference: The Penguin Dictionary of Economics, 3rd edt.