Title: Beyond the tilde -- what models can we specify easily?
Abstract:
A generalized linear model can be specified by a link function,
variance function, and design matrix. The first two are chosen from a
fixed list or supplied as an object, and for the last we have a
standard notation. Given a good general-purpose optimiser this can be
extended to allow a linear predictor for each parameter in a
likelihood. Hierarchical models can be specified by a series of model
formulas and likelihoods, a la BUGS.
The problem becomes more difficult when the model is not specified by
a likelihood, and less consideration has been given in these cases:
examples include marginal models for correlated data, and models for
survey-sampled data. I will discuss more and less general interfaces
for specifying regressions and when they might be useful