Power calculators for studies with designed missingness

点击翻译按钮获取中文摘要

Missing data (missingness) is a common issue in any data driven research. Missingness occurs in data sets when cells to be measured are missing due to reasons either related or unrelated to the study. In any data driven research, missing data poses a real problem since it hinders the ability to conduct research, and mostly importantly can severely bias the results when missing data are not handled appropriately. With sample size, the focus is always on collecting (complete) “large” sample sizes for the purpose of detecting signals that are of research interest. However, the detection of a signal (if such a signal exist in the population) reaches a saturation point at some sample size, beyond which any more data collection is unnecessary and is an inefficient use of resources. Additionally, these larger (than needed) sample sizes come at the cost of (for example) subject participation burden, data collection burden, data storage, etc, which can result in further (unplanned) missing data. So, a study design that lessens the burden of conducting the study with sufficient/adequate sample size will result in savings and efficient use of resources. This proposal aims at determining the optimum sample size for study designs where data cells are missing by design (i.e. prior to data collection, random cells in data are designed to be missing). The benefit of such a design is two folds: (1) surveys with designed missingness lessen the time needed to complete it, and data collection, (2) the cells not measured by design result in monetary savings, which can then be used in other aspects of the research or extend the study scope.******The first challenge with such designs is the loss of information due to missing cells and consequently affecting statistical inference. The second challenge is determination of sample size such that statistical inference remains reliable given that the data has missingness. With this proposal, I will address these aforementioned challenges by using methods of Sequential Analysis (SA) in the context of Multiple Imputation (MI). MI will helps in “recovering” the information lost due to designed missingness, and SA will help determine to the adequate number of imputations (and later sample size) needed for the recovery. For the aspects of optimization discussed in this proposal, there is no existing literature (in MI or SA) that has proposed such a novel and natural application of methods from SA in MI. For example for number of imputations, existing literature in MI involves results based on simulation studies. My proposed ideas will provide an imputation size (and sample size) calculator that (unlike traditional cutoffs) guarantee good statistical properties (via theorems from SA). The findings from this research will benefit Canadian and international researchers from any discipline by providing study design calculators that save time and money without compromising statistical inference.**