Comments on Statistical Issues in January 2015

In this section, we address the problem of multicollinearity in multiple regression analysisthat appeared in the article titled, " Correlation between frailty and cognitive function in non-demented community dwelling older Koreans, " published in November 2014 by Kim et al. This is one of the most frequent comments made about articles usingmultiple regression analysis. Multicollinearity indicates that independent (explanatory) variables are not mutually independent, but have some linearly correlated relationship. When we performthe multiple regression analysis,it is usually assumed that independent variables are deemed as given (fixed) values and only the dependent (response) variable is regarded as a random variable. This is why we adopt the term multicollinearity, instead of saying correlation,which denotes an association occurring among random variables. Depending on the situation, the term multicollinearitymay representa relationship of perfect linear combinations between independent variables. However,we confine its meaning tothe case of near-linear dependence. It is foredoomed thatsome degree of associationwill exist among the independent variables in a multiple regression analysis. The very reason we perform a multiple regression analysis is to controlfor the interdependency among independent variables. If there is no associationamong the independent variables, we do not have to conduct the multiple regression analysis. The results from the multiple regression analysis will be the sameas those from simple regression analyses using independent variables one by one. We do not have to worry about the multicollinearity problem when the degree of association between independent variables is not too high. However, in the case of an extremely high degree of association, some regression coefficients or their standard errors cannot be correctly calculated (estimated); that is, the phenomenaare such that no regression coefficients can be obtained, or extremely large standard errors in the analysis results might occur. In these cases, we say, " We could not obtain proper estimates from the multiple regression model due to the multicollinearity (near-linear dependency). " Hence, the reviewer's comment, " Did you check the multicollinearity? " precisely implies that, " You Commentary