Adaptive sampling in behavioral surveys.

Abstract

Studies of populations such as drug users encounter difficulties because the members of the populations are rare, hidden, or hard to reach. Conventionally designed large-scale surveys detect relatively few members of the populations so that estimates of population characteristics have high uncertainty. Ethnographic studies, on the other hand, reach suitable numbers of individuals only through the use of link-tracing, chain referral, or snowball sampling procedures that often leave the investigators unable to make inferences from their sample to the hidden population as a whole. In adaptive sampling, the procedure for selecting people or other units to be in the sample depends on variables of interest observed during the survey, so the design adapts to the population as encountered. For example, when self-reported drug use is found among members of the sample, sampling effort may be increased in nearby areas. Types of adaptive sampling designs include ordinary sequential sampling, adaptive allocation in stratified sampling, adaptive cluster sampling, and optimal model-based designs. Graph sampling refers to situations with nodes (for example, people) connected by edges (such as social links or geographic proximity). An initial sample of nodes or edges is selected and edges are subsequently followed to bring other nodes into the sample. Graph sampling designs include network sampling, snowball sampling, link-tracing, chain referral, and adaptive cluster sampling. A graph sampling design is adaptive if the decision to include linked nodes depends on variables of interest observed on nodes already in the sample. Adjustment methods for nonsampling errors such as imperfect detection of drug users in the sample apply to adaptive as well as conventional designs.