Debiased machine learning for missing data

Context Research in economics, social and health sciences, and other areas of empirical study often involves statistical analysis of datasets. These datasets typically contain some missing values, for a variety of reasons. For example, in a cohort study which aims to follow up individuals at a series of visits over many years, some individuals may drop out from the study. Moreover, even when they attend a visit, some individuals may have some variables missing because, for example, they refused to answer a question in a questionnaire. Missing values complicate statistical analyses. They cause the results of an analysis to be less precise than they would have been had there been no missing values. Moreover, analyses of the incomplete data may lead to biased results when those with missing values differ systematically from those with complete data. The challenge to be addressed There exist a variety of statistical methods that can be used to accommodate missing values in a statistical analysis. One of the most popular is multiple imputation – for a given individual, plausible values for the missing variable (imputations) are generated based on the values of the observed data for that individual. These imputations are typically generated using simple statistical models, which may lead to biased results when these models make incorrect assumptions. In recent years, it has been shown that machine-learning techniques can sometimes outperform traditional statistical models in modelling complex relationships between variables. As such, there has been interest in trying to use machine-learning methods to impute or predict missing values. These approaches however lack theoretical justification, and results may be invalid due to an issue called regularisation bias. The latter occurs because machine-learning methods are optimised for prediction, rather than estimation of parameters of interest. Aims and objectives The overall aim of the project is to exploit recent fundamental developments in so-called debiased or double machine-learning to develop and implement machine-learning methods that flexibly accommodate missing data. We will do this first for the case where one variable is partially observed, before moving on to the more complex setting where multiple variables have missing values. We will compare the performance of the methods developed with that of existing multiple imputation methods through both simulation studies and application to exemplar datasets. To maximise the impact of our research, we will develop an open-source software package implementing our methods, a user guide with illustrative analyses, and run various dissemination events. Potential applications and benefits The methods we will develop will be able to be used by researchers faced with analysing incomplete datasets across a range of research fields, including social, health and business research. Unlike the existing imputation methods that are most commonly used, our methods will be less likely to generate biased results, because they flexibly model relationships between variables using machine-learning techniques. Our methods will mean researchers need not concern themselves with the tricky task of deciding how exactly to model the relationships between variables when imputing missing values. Moreover, unlike existing imputation approaches that use machine-learning methods, the approach we will develop will have a strong theoretical underpinning and superior statistical properties.