Measurement of outcomes related to the quality of physical and mental health states in population-based cancer studies has increased in recent years as more and more researchers realize the importance of such endpoints. These endpoints are measured alongside conventional clinical outcomes and for the most part rely on patient self- report. A key problem has been missing data as patients may die or may be too sick to complete the study. This loss of information represents, besides the loss of efficiency, a potentially large threat to validity of the study results. There is strong evidence that such data are not missing at random, and cannot be ignored without introducing bias. Two extreme views on how to deal with incomplete data are (1) to delete cases with incomplete information altogether and (2) to construct complicated joint models for the measurement and missingness processes. These extreme views are surrounded with problems, largely due to the untestable nature of the assumptions one has to make regarding the missingness mechanism. A more versatile methodology that embeds the treatment of incomplete data within a sensitivity analysis is then required. Developing such a methodology necessitates extensive knowledge of biology and epidemiology of cancer. The K01 mechanism will help integrate mentoring and formal basic training in cancer research with prior training in Biostatistics and Population Health by building on a solid foundation in the development of new statistical methodologies for handling missing data. The research plan integrates training and mentoring to study, for example, how baseline and time dependent characteristics impact cancer patients'functional and mental states across time with missing data adjustment. Our approach is to develop a family of non ignorable models with sensitivity parameters that can be interpretable by subject matter experts. A global sensitivity analysis for the proposed non-ignorable models will be developed and implemented in the context of estimation and hypothesis testing using the classical frequentist approach and the Bayesian posterior predictive check principle. And finally, central theoretical questions about the proposed methods will be investigated using both analytic techniques and Monte Carlo simulations. A key goal of this K01 grant mechanism is to improve our ability to help, through collaborations, design complex clinical trials and observational studies in cancer research, analyze the generated data while adjusting for dropouts and missing data, and interpret the findings to public health professionals and the public.
