Annex 8. Complex sampling design characteristics of household surveys

Complex sampling design characteristics of household surveys include stratification, clustering, multistage sampling and weighting.

Stratification consists of dividing the population into subgroups (strata), and then selecting the sample for each subpopulation separately. Stratification can be based on geographic location (e.g. regions or districts) or certain population characteristics (e.g. age or socioeconomic status). Estimates can then be produced for each stratum. This helps to increase the accuracy of estimates if the variable of interest is heterogeneous among subpopulations. For example, if a population is stratified into four regions, a sample could be taken from within each region. With proportionate stratification, the sample size for each region depends on the population size of each region. With disproportionate stratification, the sample size in each region is fixed to guarantee a certain level of precision (so large strata do not need proportionately larger sample sizes). The sample from each region can be considered as an independent sample, and so representative estimates can be produced for each region.

Clustering is the process of dividing the population into smaller groups (clusters), and then taking a random sample of clusters to draw a sample. The goal of clustering is to make data collection more cost-effective by geographically concentrating the sample, while maintaining representativeness of the overall population. For example, if a population is partitioned into clusters, a sample of these clusters can be selected, and a random sample of households within these clusters can be taken. A disadvantage of clustering is that no data are collected for the non-sampled clusters, necessitating reliance on estimation and modelling techniques to generate indicator estimates for those areas.

In stratified sampling, strata are constructed such that populations are homogeneous within them but heterogeneous between them – that is, the population is similar within the strata but the populations between the strata are different. On the other hand, clusters are constructed such that populations are heterogeneous within them but homogeneous between them – that is, the populations within the cluster are different enough that they can be representative of the wider population, but the populations between clusters are similar enough that the sampled clusters can be representative of the non-sampled clusters.

Unless stratification and clusters are designed specifically to be representative of certain population subgroups (e.g. if urban and rural areas are stratified and sampled separately to produce estimates that are representative in urban and rural areas, or if certain small population groups such as migrants are purposefully oversampled to produce estimates that are representative of migrant populations), sample sizes are usually based on achieving representativeness of the overall population and the population within the strata. Therefore, they may not be representative for specific population subgroups, if certain groups are not sampled at all or if sample sizes are too small to produce reliable estimates. Multistage sampling involves selecting a sample through several stages, usually involving both stratification and clustering. At each stage, the selected strata or clusters are divided into smaller strata or clusters. For example, a population could be stratified into four regions, and then each region partitioned into clusters, from which a random selection of clusters is used to identify the sample population. Representative estimates can then be produced for each region. Two-stage cluster sampling takes this a step further by randomly selecting clusters and then randomly selecting a sample from within those clusters. A primary sampling unit, such as postal codes, is the first unit to be sampled. A secondary sampling unit, such as households within the selected postal codes, is the second unit to be sampled. Individuals are then selected within households as the final sample units.

Sampling weights are used to produce estimates that reflect the situation of the whole population. Weights account for unequal selection probabilities that arise from stratification or other sampling design features, including oversampling (which may be critical for inequality analysis of smaller population subgroups). The weight for a specific observation reflects how many people that record represents. For example, a weight of 2 means the observation represents two people and will be counted twice in the analysis. Survey weights can be fractions, but they are always positive.