Abarin, Taraneh (Memorial University of Newfoundland)

Purpose

In statistical association studies, outcomes are determined by combinations of error-prone and accurately measured variables that interact with each other. Association models with interaction terms are widely applied in many areas, such as genetics, engineering, economics, education and epidemiology. For example, an interaction model is considered to assess whether or not the magnitude of an error-prone variable (such as self-reported average number of cigarettes smoked a day) on an outcome (e.g. risk of heart attack) is modified by accurately measured variables (e.g. age, gender, presence or absence of family history of a heart attack). As individuals in the study may have imprecise recall, the variable “average number of cigarettes smoked a day” is considered imprecise or measured with error. In order to improve the accuracy in the assessment of the effects of these variables on the outcome, one needs to take into account these errors. As techniques dealing with interaction terms when error-prone variables are involved, are quite challenging, in practice, we often see the use of additive models that ignore the interaction effects. However, erroneously omitting interactions in those models significantly reduces efficiency of these studies.