Within-Unit Selection

In survey research, a sample of households usually must be converted into a sample of individuals. This often is accomplished by choosing respondents within households using methods intended to yield a sample similar to the population of interest. Ideally, this is done with a probability technique because all unit members will have a known, nonzero chance of selection, thus allowing generalization to a population. Probability methods, however, tend to be time consuming and relatively intrusive because they ask about household composition, potentially alienating prospective respondents and therefore increasing nonresponse.

Most researchers have limited resources, so often they need quicker, easier, and less expensive quasi-probability or nonprobability methods that they believe will yield samples adequately resembling the population being studied. Although surveyors wish to minimize nonresponse and reduce noncoverage, they have to balance these choices to fit their goals and resources. They have to decide if the benefits of probability methods outweigh their possible contributions to total survey costs or if departures from probability selection will contribute too much to total survey error. This does not mean that they tolerate substandard practices but that they consider which trade-offs will be the most acceptable within their budget and time restrictions. One paramount goal is to gain the respondent's cooperation in as short a time as possible, and a second is to obtain a reasonable balance of demographic (usually age and gender) distributions. Most households are homogeneous in other demographic characteristics, such as education and race. Many well-respected polling and other survey research organizations have to make these kinds of choices.

Selection of respondents within households is particularly important in telephone surveys because most refusals occur at the inception of contact. Although it might seem advantageous to take the first person who answers the phone or to let interviewers choose respondents, those persons who are often the most available or most willing to be interviewed also tend to be disproportionately female or older, which may bias results. In addition, patterns of telephone answering are not random; they can vary by gender and by region. Researchers need to control respondent selection in systematic ways, therefore, even if methods are quasi-probability or nonprobability. The well-known Kish selection procedure is almost a pure probability technique. Interviewers list all adult males in the household and their relationships to others in order of decreasing age, then make a similar list of adult females. Interviewers then randomly select one person by consulting a set of tables. This technique is criticized as being time consuming for large households and potentially threatening to informants, especially women who may be concerned about their safety.

The "last-birthday," or "most recent birthday" method is a popular quasi-probability selection scheme, considered to be less intrusive and time consuming than the Kish method. Interviewers ask to speak to the adult in the household who had the last birthday. A variation is the less frequently used "next-birthday" method. In theory (but not necessarily in practice), the birthday methods are probability methods because they assume the first stage of a two-stage selection process is birth (expected to be a random event), and a second stage is selection into the sample.

...