Comparing Old and New Partial Derivative Estimates from Nonlinear Nonparametric Regressions

Abstract

Partial derivatives have a special place in economics since the marginal revolution of the 1850s. We present results from multivariate partial derivative estimates using nonlinear non-parametric regressions in a finite difference method, accessible via the R-package NNS. Numerical partial derivatives are notoriously unstable, but NNS always correctly estimates their sign and comes closest to the correct magnitude compared to the coefficients in multiple linear regressions, and compared to the gradients from the popular np package for non-parametric kernel regressions.