Simple imputation methods were inadequate for missing not at random (MNAR) quality of life data

Abstract

Objective: QoL data were routinely collected in a randomised controlled trial (RCT), which employed a reminder system, retrieving about 50% of data originally missing. The objective was to use this unique feature to evaluate possible missingness mechanisms and to assess the accuracy of simple imputation methods.

Methods: Those patients responding after reminder were regarded as providing missing responses. A hypothesis test and a logistic regression approach were used to evaluate the missingness mechanism. Simple imputation procedures were carried out on these missing scores and the results compared to the actual observed scores.

Results: The hypothesis test and logistic regression approaches suggested the reminder data were missing not at random (MNAR). Reminder-response data showed that simple imputation procedures utilising information collected close to the point of imputation (last value carried forward, next value carried backward and last-and-next), were the best methods in this setting. However, although these methods were the best of the simple imputation procedures considered, they were not sufficiently accurate to be confident of obtaining unbiased results under imputation.

Conclusion: The use of the reminder data enabled the conclusion of possible MNAR data. Evaluating this mechanism was important in determining if imputation was useful. Simple imputation was shown to be inadequate if MNAR are likely and alternative strategies should be considered.

Figure 1

Bias results of EQ5D imputation at the 24 month follow up.

Figure 2

PV results of EQ5D imputation at the 24 month follow up.