Differential Attrition

Panel studies are subject to attrition, which is unit nonresponse after the initial wave of data collection. Attrition affects the results of analyses based on panel data by reducing the sample size and thereby diminishing the efficiency of the estimates. In addition, and more important, attrition also may be selective; differential or selective attrition occurs when the characteristics of the panel members who drop out of the panel because of attrition differ systematically from the characteristics of panel members who are retained in the panel. Differential attrition may introduce bias in survey estimates. However, the amount of bias depends both on the amount of attrition and on the selectivity of attrition, or in other words, on the association between the variables from which the estimate is constructed and the attrition propensity of the panel units. If an estimate is not associated at all with the attrition propensity, then the data are not biased. However, if an estimate is associated with the propensity to participate in the panel, the data are biased.

The propensity to participate in a panel survey (or alternatively, the propensity to be contacted, and given contact, the propensity to agree to participate in the panel survey) is influenced by many different factors, from characteristics of the survey design, survey-taking climate, and neighborhood characteristics to sociodemographic characteristics of the sample persons, the sample persons' knowledge of the survey topic, and their prior wave experiences. For example, the “at-home” patterns of a household and its members, and thus also their propensity to be contacted, are a function of sociodemographic attributes (e.g. number of persons in household) and lifestyle (e.g. working hours, social activities). If one person lives alone in a housing unit, contact is completely dependent on when he or she is at home. Likewise, the lifestyles of younger people may involve more out-of-home activities than those of other groups, and this also means that they will be harder to contact. Consequently, for example, when studying the extent of and changes in social contacts as teenagers grow into adulthood and later when they start their own families, the results are likely to be biased because the survey disproportionally loses (due to attrition) young individuals with more out-of-house activities. A similar logic underlies how error related to refusals is generated. For example, some studies of panel attrition provide evidence that a pleasant survey experience enhances the chance that people will participate in subsequent surveys, whereas those without such an experience are less likely to participate. Participating in a survey is a negative experience when one lacks the cognitive ability to perform the respondent task. We can assume that respondents with low socioeco-nomic status, including lower educational attainment, might have more difficulties in performing the respondent task; consequently, the interview is an unpleasant or bad experience, and these respondents will be less motivated to participate again in the panel survey. Since socioeconomic status is an important explanatory variable in many panel data analyses, it may be expected that at least some of the conclusions of these studies will be based on biased estimates due to the resulting differential attrition.

...