The Performance of Mixed Model with Repeated Measures (MMRM) and Last Observation Carried Forward (LOCF) in an Observational Study – Results from Interim Data

Thomas Kumke, Senior Biotatistcian, UCB BIOSCIENCES GmbH

The statistician working on study X was asked the question “which method are you going to use for the analysis of missing data on HDL cholesterol, the primary endpoint?” At that point the statistician did not know how to answer the question. Luckily for her sMissing data imputation in clinical trials typically follows some basic principles. Among those, the imputed value should be based on a sound statistical model and it should be as close as possible to the true value. On the other hand, imputed values should be conservative and not be in favour of the active treatment. This has led to the frequent use of the Last Observation Carried Forward (LOCF) approach for continuous data in clinical trials. However, the use of LOCF assumes that the missing value is independent on both the observed and unobserved outcome of the variable. This is an unrealistic assumption in clinical trials. Missing data imputation with a Mixed Model with Repeated Measures (MMRM) has the more realistic assumption that missingness depends on the observed outcome and is independent of the unobserved outcome of the variable. The dependency of missing data on the observed outcome is very often the case in clinical trials and, especially, in observational studies.

Reasons for missingness in observational studies may be manifold, very frequently the patients may drop out due to lack of efficacy and adverse events, or data are missing due to incomplete data documentation. The aims of this analysis are (i) to compare the results of LOCF and MMRM imputation and (ii) to analyze the performance of missing data imputation in dependence of the reason for missingness.

Data from an interim analysis of a non-interventional study were used and missing data for two continuous variables were imputed using LOCF and using a MMRM with selected Baseline variables as predictors. Summaries of the variables show that the MMRM imputed data are distinctly closer to the observed data compared to LOCF imputed data. Differences between LOCF and MMRM imputed data are largest for patients that dropped out due to lack of efficacy, followed by drop-out due to adverse events. The differences between LOCF and MMRM imputed data for patients that have not completed the trial cut-off date yet are relatively close to zero indicating that both methods perform very similar when a reason for missingness is absent.

The results show that MMRM imputed data do reflect the observed data distinctly better than LOCF imputed data. The MMRM approach imputes missing data more realistically especially in situations when patients drop out of the study due to lack of efficacy.