Statistical Methods for bias controlling in the analysis of rich data.

点击翻译按钮获取中文摘要

Innovations in digital technology and use of electronic devices generate an increasing amount of rich data sets, such as large administrative databases, and intensive electronic diary data collected through real-time mobile data-capturing devices (handheld computers, smartphones, wearable devices etc.). Despite the richness of such new data sets, their utility remains limited due to various common data limitations that can introduce significant bias into statistical inference. The size and richness of these new kinds of data often outpace the computing resources needed for traditional approaches to controlling bias. In response to the strong demand from across different industries and sectors for developing appropriate analytic techniques for use with these new kinds of data, this research program aims to develop a set of novel and principled statistical methods for bias control that are also scalable for use in today's rich data environment. The proposed research will build on my work on developing new bias controlling methods for rich data that have been published in statistics and quantitative science journals, and have already had impact in various applied domains, such as life sciences, biomedical engineering, biostatistics, social and health sciences, economics and business management. The short-term objectives are to develop novel and tractable methods to: A) quantify the sensitivity of causal inference to nonignorable missingness, B) quantify the sensitivity to nonignorable censoring in the analysis of clustered survival data, C) perform distribution-free multiple imputation with variable selection to handle missing values in rich data applications, and D) overcome the issue of unmeasured key variables. The long-term goal of this proposed research program is to develop novel, general, robust and computationally feasible methodology to increase the quality, reliability, usability and accessibility of rich data. The methodological approach will include: i) analytical derivations for both simple and generalized models, ii) a study of performance through computer simulation experiments and iii) applications to real data sets. Training HQPs and disseminating new research results are two important aspects of the proposed research program. The research program will provide ample opportunities for interdisciplinary training in all aspects of statistical research and in developing and applying innovative statistical methods to unique data sets that span many industries and sectors, including government agencies, firms, nonprofit organizations and academic institutions. The proposed work will motivate and contribute new analytical methods for big data, and improve how researchers and practitioners in sciences and engineering in Canada and internationally can analyze and make use of rich data sets.