MODELLING LONGITUDINAL COUNT DATA WITH NON-IGNORABLE DROPOUTS. AN APPLICATION TO CD4+ COUNT IN HIV-INFECTED PATIENTS

Abstract

Background: Incomplete data are not unusual in longitudinal studies as some subjects may not be available for measurements at all times. Obviously, they pose a challenge in the analysis as there is a loss of information in the data. Several principled approaches are now available for analysing such incomplete data. However, in many such approaches, inferences are valid when missing data mechanism is ignorable. Thus, in situations where missing data mechanism is non ignorable, such inferences may not be valid, hence analysing such data requires more complex models which incorporate the missing data mechanism. In this study, we analyse CD4+ count data subject to non-ignorable dropouts. Methods: The study adopts the use of likelihood-based methods, motivated by time-to-dropout approach where a Joint model is used to model the counts and the events simultaneously. The methods are applied to a longitudinal data from the database of Thyolo district hospital and its partners. In the dataset, CD4+ counts are measured from HIV-infected patients at intervals of 6 months from the baseline. We compare the results of a model with only complete cases and that of a joint model. Results: In measuring the effect of the underlying longitudinal outcome (CD4+ count) to the risk of a patient dropping out of the study, we observed a strongly significant effect (p-value < 0.001). That is, a unit increase in CD4+ count level reduces the risk of patients’ dropout by a factor of 0.24 (e−1.4267), which indi cates the high importance of CD4+ count level in quantifying the risk to dropout. A comparison between the model with complete data and the joint model di vulged some interesting features. We observed that some variables which were either marginally or not significant in the former model either became highly or marginally significant in the latter model. This signified a greater competence of vi the joint model in analysing longitudinal data subject to informative dropouts. Conclusion: In longitudinal data analysis, inferences from models based on only observed data may not be valid when dropouts from studies are non-ignorable. Therefore, models that incorporate the missing data mechanism tend to provide valuable information.