Revealing Facts and Avoiding Biases: A Review of Several Common Problems in Statistical Analyses of Epidemiological Data

• Main Authors: Lihan Yan, Youming Sun, Michael R Boivin, Paul O Kwon, Yuanzhang Li
• Format: Article
• Language: English
• Published: Frontiers Media S.A. 2016-10-01
• Series: Frontiers in Public Health
• Subjects: Epidemiology; Principal Component Analysis; regression; review; heteroscedasticity; Odds Ratio
• Online Access: http://journal.frontiersin.org/Journal/10.3389/fpubh.2016.00207/full

This paper reviews common challenges encountered in statistical analyses of epidemiological data for epidemiologists. We focus on the application of linear regression, multivariate logistic regression, and log-linear modeling to epidemiological data. Specific topics include: a) deletion of outliers, b) heteroscedasticity in linear regression, c) limitations of principal component analysis in dimension reduction, d) hazard ratio vs. odds ratio in a rate comparison analysis, e) log-linear models with multiple response data, and f) ordinal logistic vs. multinomial logistic models. As a general rule, a thorough examination of a model’s assumptions against both current data and prior research should precede its use in estimating effects.