Seam Effect

The seam effect, also called the seam bias, a phenomenon specific to longitudinal panel surveys, refers to the tendency for estimates of change, as measured across the “seam” between two successive survey administrations (or “waves”), to far exceed change estimates that are measured within a single survey wave—often by a factor of 10 or more. Seam effects have been found in virtually every panel survey examined, regardless of the characteristics under study, the data collection methods, or the length of the recall period. Seam bias almost always signals the presence of serious measurement error, which can severely compromise the statistical utility of estimates of change. A considerable amount of research over the past two decades has documented the existence of seam effects in longitudinal surveys and also has shed light on their essential nature—too little change is observed within the reference period of a single interview wave, and too much is observed at the seam.

Figure 1 presents a typical seam bias profile. It shows month-to-month transitions in reported receipt of Food Stamps and Social Security retirement benefits from the first three interview waves of the 1984 panel of the U.S. Census Bureau's Survey of Income and Program Participation (SIPP). SIPP waves occur at 4-month intervals and collect data about the preceding 4-month period; thus Months 4 and 5, and [Page 802]Months 8 and 9, comprise the “seams” between Waves 1 and 2 and Waves 2 and 3, respectively, which are reflected by the large spikes in Figure 1.

Figure 1 Survey of income and program participation month-to-month transition rates for receipt of social security benefits and food stamps

Source: Adapted from Burkhead and Coder, 1985, pp. 355–356.

Many factors have been cited as potential contributors to seam effect phenomena, including the following:

• Data processing actions—for example, strategies for assigning missing values and errors in linking cases across interview waves can create spurious transitions at the seam.
• Interviewer, coder, or respondent inconsistencies—any kind of interviewer error or inconsistency across successive survey waves is a possible cause of seam bias, as are coder inconsistencies in classifying open-ended questions and respondent inconsistencies in applying labels to phenomena of interest.
• Self or proxy response status—spurious change at the seam may result from the fact that respondents can change across successive waves of panel surveys; questionnaire design—unlike most response errors, seam effects characterize phenomena (i.e. month-to-month changes) that generally are not measured directly from respondents' reports but rather are derived in the analysis stage from those data.
• Memory issues—memories for more recent portions of the response period of one wave are likely to be of different quality and to result from different recall strategies (e.g. direct recall vs. estimation), as compared to memories for the most remote portion of the response period of the subsequent wave.
• Satisficing—in response to a difficult or burdensome recall task, respondents may adopt short-cut strategies such as constant wave responding, in which the same answer is reported for all months of an interview wave's reporting period.

Most evidence, however, discounts the relative importance of the initial, more “extrinsic” factors in the preceding list and suggests instead that questionnaire design, respondent memory issues, and recall strategies play the predominant roles in producing seam effects.

...