Handling missing data in observational clinical studies concerning cardiovascular risk: an evaluation of alternative approaches

In observational clinical studies, subjects’ health status is empirically assessed according to research protocols that prescribe aspects to investigate and methods for investigation. Commonly to many fields of research, such studies are frequently affected by incompleteness of information, a problem that, if not duly handled, may seriously invalidate conclusions drawn from investigations. Regarding cardiovascular risk assessment, usual coronary risk factors (e.g. high blood pressure) and proxies of neurovegetative domain (e.g. heart rate variability) are individually evaluated through direct measurements taken in laboratory. Apart from subjects refusing to undergo tests, a major cause of missingness can be ascribed to the fact that overall sets of collected data typically derive from aggregation of a multitude of sub-studies, undertaken at different times and under slightly different protocols that might not involve the same variables. Data on certain variables can thus be missing if such variables were not included in all protocols. Referring to a specific case study, this issue is addressed by first introducing diagnostic tools for assessing the patterns of missingness compared to the complete part of data, and then detecting the most adequate imputation methods by comparing the performance of alternative (both parametric and data-driven) approaches through a MC simulation study.