Title:Linear Rescaling to Accurately Interpret Logarithms

Abstract:The natural logarithm transformation is an important tool in statistical analysis. The standard approach to interpreting a linear change of $p$ in $\ln(X)$ is as a $(1+p)$ proportional change in $X$. This is an approximation that is only valid for small values of $p$. In this paper I suggest the use of base-$(1+p)$ logarithms (linearly rescaling $\ln(X)$ by $1/\ln(1+p)$), where $p$ is chosen ahead of time. Then, a one-unit change in $\log_{1+p}(X)$ is exactly equivalent to a $(1+p)$ proportional change in $X$. This method avoids an approximation being applied outside its useful range, offers an easier way of reporting exact interpretations, improves approximation quality when approximations are used, produces a logarithm more similar to other regression variables by making the change of interest a one-log-unit change, and reduces error from the use of $\log(1+X)$ in cases where $X$ can be 0.