Spatial implications of covariate adjustment on patterns of risk: Respiratory hospital admissions in Christchurch, New Zealand

Epidemiological studies that examine the relationship between environmental exposures and health often address other determinants of health that may influence the relationship being studied by adjusting for these factors as covariates. While disease surveillance methods routinely control for covariates such as deprivation, there has been limited investigative work on the spatial movement of risk at the intraurban scale due to the adjustment. It is important that the nature of any spatial relocation be well understood as a relocation to areas of increased risk may also introduce additional localised factors that influence the exposure–response relationship. This paper examines the spatial patterns of relative risk and clusters of hospitalisations based on an illustrative small-area example from Christchurch, New Zealand. A four-stage test of the spatial relocation effects of covariate adjustment was performed. First, relative risks for respiratory hospitalisations from 1999 to 2004 at the census area unit level were adjusted for age and sex. In three subsequent tests, admissions were adjusted for annual exposure to particulate matter less than 10μm in diameter (PM10), then for a deprivation index, and finally for both PM10 and deprivation. Spatial patterns of risk, disease clusters and cold and hot spots were generated using a spatial scan statistic and a Getis-Ord Gi* statistic. In all disease groups tested (except the control disease), adjustment for chronic PM10 exposure and deprivation modified the position of clusters substantially, as well as notably shifting patterns and hot/cold spots of relative risk. Adjusting for PM10 and/or for deprivation shifted clusters in a similar spatial fashion. In Christchurch, the resulting shift relocated the cluster from a purely residential area to a mixed residential/industrial area, possibly introducing new environmental exposures. Researchers should be aware of the potential spatial effects inherent in adjusting for covariates when considering study design and interpreting results.