Zero Imputation Methods for Log-transformation of Independent Variables
The Korean Data Analysis Society, Seo Young Park · The Korean Data Analysis Society · 2023
In studies in medical science or public health, it is common to apply log-transformation to independent variables that have severely right-skewed distribution. Among such variables, lab values have skewed distribution but have zeros due to limit of detection, which makes it impossible to apply log-transformation directly. There have been various methods proposed to deal with such situation, but most of these methods are rarely used in practice because they usually require additional step in calculation. In this study, five methods that removes zeros (shift by 1, shift by half of the smallest nonzero, shift by square root of the smallest nonzero, replace zeros with half of the smallest nonzero, replace zeros with square root of the smallest nonzero are presented. To compare performances of these methods we performed a simulation study based on randomly generated data with independent variable with log-normal distribution and response variable which has linear relationship with the independent variable, Shift by 1 method has the worst performance, and overall shift by half of the smallest nonzero method and replace zeros with half of the smallest nonzero method showed most stable performances.