Discussion of “A selective review of statistical methods using calibration information from similar studies” and some remarks on data integration

Abstract

Qin, Liu and Li (henceforth QLL) review methods for combining information using empirical likelihood and related approaches; many of these ideas originated in the earlier work of Jing Qin. I thank the authors for their review, and for the opportunity to contribute to its discussion. I have little to say about technical aspects, which are well established but will comment briefly on broader aspects of data integration, and implications for methods like those in the article. I will focus on settings where there is a response variable Y and covariates X , Z and assume the target of inference is either the distribution f ( y | x , z ) of Y given X , Z or the ‘marginal’ distribution f m ( y | x ) of Y given X . In health research Y might represent (time to) the occurrence of some specific event, and X , Z covariates, exposures or interventions. The distribution f ( y | x , z ) is important for individual-level decisions; in settings where X represents interventions f m ( y | x ) is relevant in randomized trials and comparative effectiveness research. The authors consider two main topics in data integration: (i) the use of external auxiliary data to augment the analysis of a specific ‘internal’ study, and (ii) the combination of data from separate studies with a view to for common parameters or They focus on where,