Bias  due to differential attrition: what matters and what doesn’t

Background: Attrition in longitudinal randomized trials is common, and threatens validity of results, particularly in patient reported outcomes like quality of life. When attrition rates differ between treatment arms it is sometimes called differential attrition. The implications of differential attrition are often misunderstood: some believe that if attrition rates are similar between study arms estimated treatment effects will be unbiased. Or, if differential attrition occurs, results are biased. This study demonstrates that these beliefs are false.  

Methods: Quality of life data were simulated for a randomised 2 arm trial of 3 time points which were complete, missing completely at random, at random, and not at random (MNAR). Missingness was 30% in the equal case, and 40% and 20% for unequal missingness. Data were analysed with a t-test and simple approaches to missing data; and a mixed model contrast. 

Results: The simple approaches yielded biased estimates of the treatment effect, for nearly every type of attrition and missing data type. Mixed models yielded unbiased estimates for all scenarios except when data were MNAR.