Effects of Exposure-Confounder Misclassification and Criteria of Model Choice in Ecologic Studies.
- Author:
Sun Hee LEE
;
Chung Mo NAM
;
Hung Wok PARK
- Publication Type:Original Article
- MeSH:
Bias (Epidemiology);
Joints;
Odds Ratio;
Public Health
- From:Korean Journal of Epidemiology
1996;18(2):142-150
- CountryRepublic of Korea
- Language:English
-
Abstract:
Ecologic studies are widely used in all fields of public health on account of accesibility of data. However, two problems related to these studies have been brought up. The first is ecological fallacy occurred in the course of interpreting the ecologic level of exposure-disease associations into individual level. The second is exposure isclassification which leads to serious bias. Nevertheless there is few methodologic study dealing joint effects of the two problems in ecologic study. This study was conducted to suggest an ecologic model not having an ecologic fallacy due to model linkage failure and a methodology for correcting the misclassification bias due to exposure-confounder misclassification. Finally, we suggest a criteria for the ecologic model selection. Main results are as follows: 1. A linear ecologic regression model has a serious ecological fallacy due to model linkage failure and the misclassification bias due to the exposure-confounder misclassification. 2. An interaction ecologic regression model has no ecological fallacy due to model linkage failure, but it is affected seriously by the exposure misclassification. However misclassification bias could be removed mathematically if the information related to the misclassification was known. 3. A log-linear ecologic regression model has an ecological fallacy due to model linkage failure. It is seriously biased as the individual risk ratio are increased, but relatively less affected by the exposure misclassification than interaction ecologic regression model. 4. One of the two ecologic regression model-interaction ecologic regression model and log-linear ecologic regression model- would be selected according to the information of individual risk ratio and exposure misclassification. But using a linear ecologic regression model should be avoided in any circumstance. The above results are only valid in case that there is no other source of ecological fallacy except model linkage failure. Also exposure and confounder are independent each other, measured binary, and having nondifferential misclassification. Since the above assumptions are somewhat strong in considering the real situations of ecologic studies, it is necessary to extend the scope of this study.
