Unit

Sampling, variance, analysis, reporting, and dissemination units cover most of the unit types of interest in survey methodology. Sampling units can be primary, secondary, tertiary, and beyond—also known as first-stage, second-stage, and so on. These refer to a hierarchy of smaller and smaller geographic or organizational structures that have been exploited in the sample design to improve efficiency. In face-to-face surveys of the general population, the stages are [Page 925]typically county (or count group), block (or cluster of neighboring blocks), housing unit (or other quarters), and person. In student surveys, the stages are typically school district, school, classroom, and student. In random-digit dialing surveys, the stages are typically just household and person, although sometimes a 100-bank is also a stage.

Variance units are often the same as primary sampling units (PSUs) but can be different. Sometimes PSUs are collapsed into super PSUs to reduce the number of replicate weights that need to be created for variance estimation via resampling methods such as the bootstrap or jackknife. Other times, PSUs are split into pseudo PSUs to improve the stability of variance estimates even if it means a downward bias in the variance estimates. Both operations can also be done to make it harder for data snoopers to discover anything about known sample members from public use files.

Analysis units are often people but can be households or families. However, these other constructs usually are well-defined only at a particular point in time. Efforts to define stable interpersonal groupings over time have been largely unsuccessful, in the United States at least. Even the concept of "parent" is hard to keep stable over time, as varying constellations of adults can provide parenting services to children during their minor years. Where family structures and parents have been of interest over time, the most successful approach has been to follow the persons of interest and report the characteristics of their families or parents as attributes of themselves.

Reporting units are generally the same as ultimate stage sample units, but in surveys of businesses and other larger social constructs such as governmental bodies, a "large" sample unit may need to be split into multiple reporting units. For example, in a survey of energy usage in commercial buildings, the individual tenants might be the reporting units for a multi-tenant building with separately metered utilities.

Dissemination units are the smallest geographic or organizational structures for which formal estimates are published. Typical examples in the United States include the four census regions, the states, and counties, depending on the size of the survey. Sometimes, there is interest in the primary sampling units as analysis or dissemination units. Metropolitan areas constitute one example where attempts have been made to use them both as sampling units and as dissemination units. These attempts are not, however, very satisfying. Sample sizes are generally too small within individual PSUs to satisfy user expectations for the precision of formally published estimates. Also, it is hard to explain why, if such estimates are important or useful, they are only available for a few dozen of them.

...