The conclusions from a statistical analysis depend heavily on the validity of the fitted model. Thus a careful examination of the adequacy of the model should always be a part of any statistical inference procedure. There is at present no commercial package that provides a unified way to check lack-of-fit for a broad class of regression models. This proposal is intended to fill that need. Its goal is to develop commercial software which will test for model mis-specification of generalized linear models with or without correlated error terms, non- linear regression models, the Cox proportional hazards model, and a variety of semi-parametric and parametric survival models. The approach is based on comparing cumulative sums of residuals for the observed data to corresponding cumulative sums obtained by simulation under the assumed model. It incorporates both graphical and numerical model checking procedures. Although the software relies on fairly advanced mathematics the user is shielded from these complexities by a friendly and flexible user interface running on the familiar Windows-95 platform. This user interface makes good use of dynamic graphics and encourages interactive model selection, fitting and checking. The software will be available as a stand-alone Windows implementation, as an external SAS procedure, and as an external Splus function.
Very little software is available in commercial packages for checking the adequacy of the model assumptions for generalized linear models, nonlinear models and survival models. Yet these models are used extensively for biomedical, economic, and managerial applications. A software package implementing both model fitting and model testing procedures in a flexible interactive Windows environment should have great commercial appeal.
