Leitenstorfer, Florian; Tutz, Gerhard
Generalized Monotonic Regression Based on B-Splines with an Application to Air Pollution Data.
Collaborative Research Center 386, Discussion Paper 444
In many studies where it is known that one or more of the certain covariates have monotonic effect on the response variable, common fitting methods for generalized additive models (GAM) may be affected by a sparse design and often generate implausible results. A fitting procedure is proposed that incorporates the monotonicity assumptions on one or more smooth components within a GAM framework. The flexible likelihood based boosting algorithm uses the monotonicity restriction for B-spline coefficients and provides componentwise selection of smooth components. Stopping criteria and approximate pointwise confidence bands are derived. The method is applied to data from a study conducted in the metropolitan area of Sao Paulo, Brazil, where the influence of several air pollutants like SO_2 on respiratory mortality of children is investigated.