Abstracts: ISEE 21st Annual Conference, Dublin, Ireland, August 25-29, 2009: Oral Presentations
Background and Objective:
Valid model-based estimation typically requires several assumptions including no residual confounding, correct model specification and no measurement error. Here, we present a method for assessing whether these assumptions hold. It depends on availability of an evaluation factor that meets the criterion: it is independent of the outcome conditional on measured exposures in the absence of confounding, misspecification and correlated measurement errors.
Using directed acyclic graphs, we show one can identify violations of model assumptions using an evaluation factor F that satisfies the stated criterion. Validity to partially correct estimates follows from probability arguments in a simple situation when the only violation is residual confounding. We apply the method to a time-series study of emergency department visits for asthma and ozone levels over the preceding 2 days in Atlanta.
When F was the ozone level two days following ED visits, its association with ED visits was weak and non-significant in the base model. However, when several variables judged a priori to be confounders were excluded, F was strongly associated with ED visits-correctly indicating the presence of a violation of an assumption, here residual confounding. In contrast, the AIC incorrectly suggested these variables might be omitted.
This method provides a potentially useful way to identify important sources of bias-residual confounding, model mis-specification or correlated measurement error. Its sensitivity and performance will differ by context. The example illustrates that use of some statistical approaches, such as AIC, can be misleading. Pollution levels or emissions after the outcome has occurred may be suitable evaluation factors in some situations.