Missing data: a review of current methods and applications in epidemiological research

Abstract

Purpose of review 

Researchers inevitably confront missing data. In cross-sectional studies, nonresponse to specific items causes item-level missing data. Longitudinal studies pose a greater likelihood of item nonresponse and introduce unit nonresponse when data for an individual are missing because that person was not available for assessment. The need to adequately deal with missing data remains, regardless of whether missing data result from item nonresponse, participant attrition, or sporadic availability of respondents. The wealth of missing data techniques available to researchers often produces uncertainty regarding which to use. Our purpose is to discuss the applicability of general methods for dealing with missing data and to review current advances associated with specific missing data techniques.

Recent findings 

Traditional missing data methods such as complete case analysis often produce bias and inaccurate conclusions. Similar problems extend to single imputation techniques commonly thought of as improvements over complete case methods. Research demonstrates that procedures such as multiple imputation, which incorporate uncertainty into estimates for missing data, often provide significant improvements over traditional methods.

Summary 

Recent work suggests that multiple imputation and specific modeling techniques offer general methods for dealing with missing data that perform well across many types of missing data situations. In addition, advances in desktop computers and the development of user-friendly software make these techniques accessible to researchers in all fields. Future research will undoubtedly result in further refinements and extensions of these techniques, making them applicable to difficult but common situations in which missing data arise.