A note on using ratio variables in regression analysis

- We revisits the problem of choosing ratio variables in regression analysis.
- Linear regressions with the ratio variable, its reciprocal or logarithm are rejected.
- We suggest to use nonlinear regressions on ratio variables.
- Empirical evidence prefers a semiparametric partially linear model and the logarithm of the ratio variable.

This paper revisits the problem of choosing ratio variables in regression analysis in Musumeci and Peterson (2011). In the application we examined, linear regressions with the ratio variable, its reciprocal or logarithm have been rejected. To avoid model misspecifications, we suggest to use nonlinear regressions on ratio variables. Our empirical evidence shows that a semiparametric partially linear model could be a robust solution. In particular, the logarithm of the ratio variable performs slightly better than the ratio variable and its reciprocal.