Stratified Sample

Often, survey samples are designed in a way that takes into account knowledge about the population structure. This may involve an attempt to ensure that the sample has the same structure as the population in important respects or an attempt purposively to deviate from the population structure in order to increase the representation of certain subgroups. Such sample designs are referred to as stratified sampling, and the outcome of implementing the design is a stratified sample.

There are two types of stratified sampling: proportionate and disproportionate. Proportionate stratified sampling involves controlling the sample proportions in each stratum to equal the population proportions. If the strata are correlated with the survey measures, this will have the effect of increasing the precision of survey estimates (see design effects). Proportionate stratification can be achieved by either creating explicit strata and sampling independently from each, or sorting the sampling frame units into a meaningful order and then sampling systematically from the ordered list (see systematic sampling). These methods are known, respectively, as “explicit stratified sampling” and “implicit stratified sampling.” Proportionate stratification cannot have an adverse effect on the precision of estimates. The worst scenario is that strata turn out to have zero correlation with a particular survey measure, in which case stratification has no effect on precision and the sample is equivalent to a simple random sample (see simple random sampling). Selection probabilities are not affected by proportionate stratification, so no bias can be introduced by a poor choice of strata.

[Page 1086]

Disproportionate stratification involves applying different sampling fractions (see SAMPLING FRACTION) in different strata. The objective is often to increase the sample size of one or more important subgroups for which separate estimates are required. In this situation, disproportionate stratification typically reduces the precision of estimates relating to the total study population (see DESIGN EFFECTS), but increases the precision of estimates for the over sampled stratum or strata. Disproportionate stratification can, however, also be used to increase the precision of total population estimates in situations where inherent variability differs greatly between the strata and can be estimated in advance. In such situations, precision is maximized by use of Neyman optimum allocation:

The optimum allocation rule states that the sampling fraction in stratum h,nh/Nh, should be set proportional to the stratum standard deviation, sh, and inversely proportional to the square root of the unit cost of data collection in the stratum, Ch.

In quantitative surveys, the choice of proportionate or disproportionate sampling is an important and explicit part of the design. In QUALITATIVE studies, there is no attempt to quantify the relationship between sample and population, so the distinction between proportionate and disproportionate sampling is not relevant. But a sample can still be said to be stratified if it has been designed deliberately to reflect certain aspects of population structure.