Abstract:
|
In built-environment studies, accurate information about environment features is required to obtain unbiased exposure assessment and correct estimation of their health effect. Thanks to the availability of diverse commercial business listings, data about the location of community resources (e.g., food outlets) can be collected from multiple sources. However, ascertainment error arises in merging multi-source information since these databases often provide conflicting claims about the number and location of these businesses. In this work, we propose a method that accounts for erroneous count exposures from different databases when used as predictors in regression models. Given knowledge of source reliability, we can more effectively combine multiple sources and derive complete information about latent true exposures using a multinomial model with partially known margins. In our motivating example, we illustrate discovery of the true association between children’s BMI and the concentration of food outlets near their schools when both National Establishment Time Series database and the Reference USA database are available.
|