Interpretable Subclass-Wise Measures for Interaction Detection as an Alternative to Interaction Tests

• Activity Number: 442 - Disease Prediction, Statistical Methods for Genetic Epidemiology and Mis
• Type: Contributed
• Date/Time: Thursday, August 12, 2021 : 4:00 PM to 5:50 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Eiji Nakatani* and Yoko Sato and Seiichiro Yamamoto
• Companies: Shizuoka Graduate University of Public Health and Shizuoka Graduate University of Public Health and Shizuoka Graduate University of Public Health
• Keywords: interaction test; risk difference; risk ratio; Bootstrap

Abstract

In general, we determine that there is an interaction when there is a difference between the coarse effect of the first variable affecting the outcome (drug use, presence of disease, etc.) and the effect of any subclass divided by the second variable such as a specific background variable. When the first variable is drug use, the criterion for determining whether there is an interaction is usually to determine that there is a drug effect based on the null hypothesis (zero difference) for each subclass. However, if the coarse effect for first variable is a large difference, the criterion for determining whether there is a difference in subclass is better to examine how much difference there is around the coarse effect. With the above in mind, we propose a statistical method to compare the coarse effect with the effect of each subclass. It is the difference between the risk difference as the effect in the whole population and the risk difference in the subclasses, as well as the ratio of the risk ratios performed in the same way (distinguished difference of risk difference [dDRD] and distinguished ratio of risk ratio [dRRR]). An example application of the proposed method is shown.