A1676
Title: Pooled likelihood inference for secondary phenotype models in multiple case-control studies
Authors: Zifu Wei - Sichuan University (China) [presenting]
Abstract: Secondary outcomes from case-control studies are scientifically useful but statistically awkward because retrospective ascertainment hides the source-population case fraction. In a single study, intercept-like secondary regression parameters are generically confounded with that hidden prevalence and are therefore not identified without external prevalence information or an equivalent normalization. Pooling several case-control studies under a common source-population covariate law changes the problem: cross-study distortion contrasts identify the hidden prevalence vector through a linear system, and the conditioning of that system yields a natural weak-identification index. On that foundation, a pooled profile-likelihood framework is developed that profiles over the common covariate distribution and treats the prevalence vector as part of the inferential target. The framework is analyzed in three model cells: a baseline linear-mean model with known error density, an extension with unknown common error density, and an extension with unknown mean surface and known error density. Several extensions to this framework are also discussed, including dimension reduction and extreme regions.
