|
Abstract:
|
Much research with government data can be framed as the study of trends. In trend analysis, researchers combine individual observations into groups: often time and space. Group level variation is then used to make precise inferences. Aggregation bias occurs when researchers incorrectly extend those group level inferences to individual observations. Statisticians fall into two camps on this practice. Hardliners maintain that any such extension is speculative. Others, including us, argue that these extensions are valid if the variation between groups greatly exceeds the variation within groups. However, overestimation of the flexibility offered by this position has likely encouraged its widespread misapplication. We address aggregation bias found in recent work on mortality trends. We examine data from the CDC Wonder Database and the National Longitudinal Mortality Study. Hierarchical models provide a natural framework to account for group level variation, and we discuss criteria for quantifying aggregation bias due to changes in group composition. We finally connect our work to the current dialogue on the widely reported trajectory of mortality rates among middle-aged Americans
|
