Handling trial participants with missing outcome data for continuous outcomes in randomized controlled trials: systematic survey of the methods literature

Poster
2015 Vienna
Zhang Y1, Alyass A1, Vanniyasingam T1, Sadeghirad B1, Flórez ID2, Pichika SC1, Kennedy SA3, Abdulkarimova U4, Zhang Y1, Iljon T4, Morgano GP1, Colunga L5, Abu Bakar Aloweni F6, Lopes LC7, Yepes-Nuñez JJ1, Fei Y8, Wang L9, Kahale LA10, Meyre D11, Akl E12, Thabane L1, Guyatt G13
1Department of Clinical Epidemiology and Biostatistics, McMaster University, Canada
2Universidad de Antioquia (Medellin) AND McMaster University (Hamilton), Colombia, Canada
3Department of Medicine, McMaster University, Canada
4Department of Mathematics and Statistics, McMaster University, Canada, Canada
5Hospital Civil de Guadalajara, "Fray Antonio Alcalde", México
6Singapore General Hospital, Singapore
7Universidade de Sorocaba, São Paulo, Brazil
8Beijing University of Chinese Medicine, McMaster University, China, Canada
9Michael G DeGroote Institute for Pain Research and Care, McMaster University, Canada
10American University of Beirut, Lebanon
11Department of Clinical Epidemiology and Biostatistics, McMaster University, Canada; Department of Pathology and Molecular Medicine, McMaster University, Hamilton, Ontario, Canada, Canada
12Clinical Epidemiology Unit and Center for Systematic Reviews in Health Policy and Systems Research (SPARK), American University of Beirut, Lebanon
13Department of Medicine and Clinical Epidemiology and Biostatistics, McMaster University, Canada

Background:

Missing participant data (MPD) are a common source of bias in randomized controlled trials (RCTs). Sensitivity analyses can address whether results are robust under different assumptions regarding MPD. It is likely that some methods of conducting such sensitivity analysis are superior to others.

Objective:

To identify analytical approaches for dealing with MPD in continuous variables in RCTs for proposed and tested simulations of use.

Method:

We searched MEDLINE, the Cochrane Library, Web of Science, and Journal Storage for methodological articles up to January 2015. Reviewers conducted screening and data abstraction independently in duplicate. Data abstraction is ongoing. We will describe methods identified and their performance regarding bias, precision, coverage, accuracy, power, and type I error, and general ranking considering all the above aspects both in tabular and narrative formats.

Results:

From 15205 citations, we have identified 94 eligible papers reporting more than 60 methods. Simulated data were derived from real trial data from the following clinical areas: respirology, gastroenterology, psychology, dermatology, infectious diseases, cardiology, and rheumatology. The 25 studies we have abstracted to date, explored performance of methods under different missing mechanisms, proportions of missing data, effect size, sample size, and strength of associations between certain variables with missing data. Twenty of these specified general performance of the methods for ignorable and non-ignorable missing data, mixed-effect models (MEM; 4/20), multiple imputation (MI; 5/20), maximum likelihood (ML) method (4/20), or combination of MI and MEM (2/20) seem to be the most frequently mentioned best methods. Expectation maximization, weighted least squares estimator, and empirical bayes are also identified as possible optimal methods.

Conclusions:

The data abstracted so far suggest that using MEM or MI alone or as a combination, or ML may minimize the impact of ignorable and/or non-ignorable MPD in RCTs. The completion of data abstraction and synthesis will allow us to confirm or refute these preliminary conclusions.