GEE solvers: case studies in DSC design and implementation

GEEs (generalized estimating equations) provide a framework for flexibly modeling clustered data. Generalized linear model (GLM) components are used to specify marginal mean and variance functions, and “working” covariance models specify multivariate structure. The indefiniteness of the estimation and inference framework is a basis for criticism from theoretical quarters (Crowder, Bka 1995), but also a basis for interesting interface design challenges and opportunities. I will comprehensively describe the redesign of S4/R-targeted GEE solvers. Basic issues include a) choice of language, b) representation of complex clustered data structures to accommodate, e.g., responses and predictors obtained on discordant timing sequences; c) inheritance from and interoperation with existing tools for multivariate modeling; d) choice of class/method decomposition to support recognition of statistical data types; d) weak implementation methods to ease retargeting to DSC platforms as they mature. Peripheral issues include a) exploitation of XMLbased literate programming methods; b) automatic generation of javadoc-like hypertext doc for S4/R classes and methods.