Repeated Cross-Sectional Design

Many important cross-sectional surveys are repeated at regular or irregular intervals so that estimates of changes can be made at the aggregate or population level. Examples include monthly labor force surveys, retail trade surveys, television and radio ratings surveys, and political opinion polls. These surveys are designed to give good estimates for the current population and the changes or movements that have occurred since the last survey or previous surveys. Typically surveys are conducted on a monthly, quarterly, or annual basis, although other intervals are possible, such as daily or weekly in the case of TV ratings and opinion polls. Surveys may also be conducted at longer intervals, such as 3 years, or repeated on an irregular basis, but in all cases there will be interest in estimating and analyzing changes at the population level and also various subgroups of the [Page 715]population, which are often defined geographically or in terms of sociodemographic variables.

Repeated cross-sectional surveys differ from longitudinal surveys, which are designed specifically to permit analysis of change at the individual or micro level and usually involve following an initial sample over several waves even if respondents move location. The need to follow respondents contributes to the cost and complexity of a longitudinal survey. In a longitudinal survey, there may be no interest in ensuring good cross-sectional estimates for each wave of the survey, and it may be difficult to do so. Longitudinal surveys are subject to attrition bias and conditioning effects but are valuable when the main aim of the survey is to understand changes at the individual level. In a repeated cross-sectional design, there is a strong emphasis placed on maintaining good sample representation to produce unbiased estimates for each time period. This can be done without following respondents over time.

In a repeated survey, an independent sample may be selected on each occasion, and so there will be essentially no overlap in the samples between time periods. There is then no possibility of conditioning effects or respondent fatigue, although there are the costs involved in making the initial contact with respondents and obtaining their cooperation. Valid estimates of changes at the population level can be calculated from independent samples. If yt is the estimate for the population for time t and yt − s the estimate for the population for time t − s, then the change or movement between the two time periods can be estimated by yt - yt − s With independent samples, there will be differences between estimates for different periods because they are based on different samples. The sampling variance of the estimate of change will be the sum of the sampling variance on each of the estimates, so . If the sampling variances of each of the estimates are approximately equal, then the sampling variance of the estimate of change will be twice that of the cross-sectional estimates, and hence the standard error will be about 40% higher. Reliable estimates of changes can be obtained provided the sample sizes at each period are large enough and an efficient sample design is used, which produces unbiased estimates for each period. There is no need for the same sample size or design to be used at each occasion although it is usually efficient to do so.

...