Order-Preserving Nonparametric Regression, with Applications to Conditional Distribution and Quantile Function Estimation

In some regression problems we observe a "response" Y _ti to level t of a "treatment" applied to an individual with level X_i of a given characteristic, where it has been established that response is monotone increasing in the level of the treatment. A related problem arises when estimating conditional distributions, where the raw data are typically independent and identically distributed pairs (X _i, Z_i), and Y_ti denotes the proportion of Z_i's that do not exceed t. We expect the regression means g _t(x) = E(Y_tiX_i = x) to enjoy the same order relation as the responses, that is, g_t ≤ g_s whenever s ≤ t. This requirement is necessary to obtain bona fide conditional distribution functions, for example. If we estimate g_t by passing a linear smoother through each dataset Χ_t = {(X_i, Y _ti) : 1 ≤ i ≤ n}, then the order-preserving property is guaranteed if and only if the smoother has nonnegative weights. However, in such cases the estimators generally have high levels of boundary bias. On the other hand, the order-preserving property usually fails for linear estimators with low boundary bias, such as local linear estimators, or kernel estimators employing boundary kernels. This failure is generally most serious at boundaries of the distribution of the explanatory variables, and ironically it is often in just those places that estimation is of greatest interest, because responses there imply constraints on the larger population. In this article we suggest nonlinear, order-invariant estimators for nonparametric regression, and discuss their properties. The resulting estimators are applied to the estimation of conditional distribution functions at endpoints and also changepoints. The availability of bona fide distribution function estimators at endpoints also enables the computation of changepoint diagnostics that are based on differences in a suitable norm between two estimated conditional distribution functions, obtained from data that fall into one-sided bins.