A widely used approach on handling missing outcome data in longitudinal studies is the so-called ‘last observation carried forward’ (LOCF) approach. This approach can be implemented for participants who left the study before the final measurement but provided one or more intermediate measurements.
The LOCF imputation scheme assumes that for patients dropping prematurely out of the study the last observed measurement is representative and can be used instead of an actual observation at the end of the study. However, the outcome of patients who drop out may change over time and LOCF can lead to incorrect estimates for the treatment effects. Although LOCF has been widely critiqued and can be restrictive as it assumes that the outcome remains constant after its last measurement and till the end of the study (introducing systematic error), it is an approach extensively used; in a recent meta-analysis of 212 trials comparing antipsychotics, the vast majority reported results according to the LOCF.
Choosing between different approaches to deal with missing outcome data can be critical in certain areas of medicine, such as psychiatry, where (due to the nature of the conditions and/or treatments), high drop-out rates are often observed in most of the studies leading to significant amounts of missing data, the incorrect handling of which will have a great impact on the study-specific and meta-analytic treatment effects. Most trial reports in mental health present complete cases analysis alongside the LOCF imputed effect sizes and meta-analysts often choose to extract and synthesize the latter to gain precision. However, LOCF does not account for error or uncertainty in the imputation process, leading to biased or over-precise results.
Motivated by the latter, we are in need of a meta-analytic model that accounts for possible error in the effect sizes estimated in studies with LOCF imputed patients. As individual-patient data are not genuinely available for most meta-analyses we restrict ourselves to the general case of aggregated data meta-analysis. Assuming a dichotomous outcome, we regard LOCF as a process characterised by sensitivity and specificity with respect to the true, unobserved outcome at studied endpoint. Our model is based on decomposing the probability of a successful outcome after incorporating the sensitivity and specificity of the imputation process. We fit our models within a Bayesian framework and we explore the use of different formulations for the prior distributions of sensitivity and specificity.
The paper describing our work has been published by Statistics in Medicine, Dimitrakopoulou V, Eftimiou O, Leucht S and Salanti G. "Accounting for imperfect 'Last Observation Carried Forward' outcome imputation in a meta-analysis model" , doi: 10.1002/sim.6364
O. Efthimiou presented this work during the «36th Conference of the International Society for Clinical Biostatistics (ISCB 2015)». Link to the poster here.