William Heavlin, Affiliation: Google Inc On the shoulders of Gauss and Bessel: links, chunks, spheres, and conditional models For generalized linear models (GLMs), we define the conditional likelihood, which partitions (or chunks) data into subsets. To efficiently compute such a likelihood, which has combinatorial complexity, we introduce the spherical approximation. The resulting model estimates are highly linearized and computationally attractive. Further, these estimates refine our understanding of standard (unconditional) GLMs.