Attrition

Attrition is a term used to describe the process by which a SAMPLE reduces in size over the course of a survey data collection process due to NONRESPONSE and/or due to units ceasing to be eligible. The term is also used to refer to the numerical outcome of the process.

In PANEL surveys, the number of sample units having responded at every wave decreases over waves due to the cumulative effects of nonresponse and changes in eligibility status. Such attrition may introduce unquantifiable bias into survey estimates due to the nonrandom nature of the dropout. For this reason, considerable efforts are often made to minimize the extent of sample attrition. The extent of attrition is affected also by factors such as the number of waves of data collection, the intervals between waves, the burden of the data collection exercise, and so on. Weighting and other adjustment methods may be used at the analysis stage to limit the detrimental impacts of attrition.

Although perhaps most pronounced in the context of panel surveys, attrition can affect any survey with multiple stages in the data collection process. Many CROSS-SECTIONAL surveys have multiple data collection instruments—for example, a personal interview followed by a SELF-COMPLETION (“drop-off”) QUESTIONNAIRE or an interview followed by medical examinations. In such cases, there will be attrition between the stages: Not all persons interviewed will also complete and return the self-completion questionnaire, for example. Even with a single data collection instrument, there may be multiple stages in the survey process. For example, in a survey of schoolchildren, [Page 43]it may be necessary first to gain the cooperation of a sample of schools to provide lists of pupils from which a sample can be selected. It may then be necessary to gain parental consent before approaching the sample children. Thus, there are at least three potential stages of sample dropout: nonresponse by schools, parents, and pupils. This, too, can be thought of as a process of sample attrition.

In attempting to tackle attrition, either by reducing it at the source or by adjusting at the analysis stage, different component subprocesses should be recognized and considered explicitly. There is an important conceptual difference between attrition due to nonresponse by eligible units and attrition due to previously eligible units having become ineligible. The latter need not have a detrimental effect on analysis if identified appropriately. Each of these two main components of attrition splits into distinct subcomponents. Nonresponse could be due to a failure to contact the sample unit (“noncontact”), inability of the sample unit to respond (for reasons of language, illness, etc.), or unwillingness to respond (“refusal”). Change of eligibility status could be due to death, geographical relocation, change in status, and so forth. The different components have different implications for reduction and adjustment. Many panel surveys devote considerable resources to “tracking” sample members over time to minimize the noncontact rate. The experience of taking part in a wave of the survey (time taken, interest, sensitivity, cognitive demands, etc.) is believed to be an important determinant of the refusal rate at subsequent waves; consequently, efforts are made to maximize the perceived benefits to respondents of cooperation relative to perceived drawbacks.

...