Segments

Segments is a term for sample units in area probability sampling (a specifie kind of cluster sampling). Most often, segments are sample units in the second stage of area probability sampling and are more formally referred to as secondary sampling units (SSUs).

As second stage sample units, segments are neighborhoods or blocks (either census-defined or practically defined by field workers) within the selected primary sampling units, which are often counties or whole metropolitan areas). Occasionally, segments can refer to the first-stage sample units or to larger areas than neighborhoods or blocks, such as entire census tracts or even counties, but this entry describes them in their more common usage as second-stage sample units.

Individual units (often housing units, or sometimes clusters of housing units) within the selected segments are selected for inclusion in the sample. Traditionally, field workers are sent out to the selected segments to list every housing unit. New lists are becoming available for segments in urban areas built from postal address lists. These lists are not yet available in rural areas, but as rural addresses in the United States get [Page 803]converted to city-style addresses for 911 reasons, the postal address lists available commercially will continue to increase their coverage.

Segments are usually defined only within primary sampling units that have been selected in the first stage of sample selection. Segments are designed to be as contiguous as possible because this reduces interviewer travel between selected units, but if selected using census data, consecutive-numbered blocks may not be strictly contiguous. There are two key decisions to be made in defining segments. The first is how large to make the segments, and the second is how many segments to select within each primary sampling unit.

Deciding how large to make the segments involves a trade-off of survey cost versus variance. Under traditional listing, larger segments will cost more to list every housing unit. Larger segments will also necessitate more travel between individually selected sample units, which increases survey costs. However, smaller segments result in a more homogenous sample, as measured by larger intraclass correlations, often represented by the Greek letter rho (p). Larger rho values reduce the effective sample size, which results in more variance. The rho differs for each variable, depending on how similar people who live near each other are on any particular characteristic. As a general rule, socioeconomic characteristics (e.g. income) tend to have higher rho values than do behavioral variables. This is because people who live near each other tend to have similar financial situations, but even people in similar financial situations tend to have different opinions and behaviors.

Morris H. Hansen, William N. Hurwitz, and William G. Madow suggested a logarithmic relationship between the average cluster size and rho:

where a and m are different parameters for different variables, and N represents the average segment size. The assumption in the formula is that m < 0 so that as the average cluster size increases, the value of rho will decrease. This is explained by the fact that smaller areas tend to be more similar than larger areas (e.g. one neighborhood vs. an entire city).

...