Our goal is to develop a set of new statistical modeling tools to allow biomedical researchers to easily develop better, and predictively, more accurate models. The basis for these tools is the recent advancement, Multivariate Adaptive Regression Splines (MARS). Our package will apply MARS to cross-section regression. and extend it to longitudinal regression, logistic regression, and matched-sample case-control models. Because MARS requires few of the assumptions implicit in classical statistical modeling. more flexible models can be developed. MARS-based empirical models can be expected to be predictively more accurate, and reliably include variables and interactions that might otherwise be missed. In particular. MARS technology allows subtle effects to be easily detected, such as how risk of exposure can vary over time or how exposure risk might interact with demographic characteristics. The graphically displayed are easy to comprehend and are recorded in reusable programming statements compatible with SAS, and other major statistics packages. We have two major sub-goals. First. to make the existing MARS procedure fully accessible to the bio-medical research community by providing a familiar Windows-style GUI interface. Second, to make key enhancements to MARS to render the procedure directly applicable to the most common model types encountered in biomedical research.
The considerable, potential commercial market for a comprehensively modified MARS procedure includes analysts looking for advice. suggestions. or diagnostics that could improve a parametric model, researchers willing to add new exploratory data analysis tools to their methods chests, and investigators interested in nonparametric methods. As a standalone package capable of directly reading from and writing to virtually all major statistical package and database formats, it will be usable by' researchers in any UNIX or PC environment.
