Statistical models for the control phase of clinical monitoring

The rise in the prevalence of chronic conditions means that these are now the leading causes of death and disability worldwide, accounting for almost 60% of all deaths and 43% of the global burden of disease. Management of chronic conditions requires both effective treatment and ongoing monitoring. Although costs related to monitoring are substantial, there is relatively little evidence on its effectiveness. Monitoring is inherently different to diagnosis in its use of regularly repeated tests, and increasing frequency can result in poorer rather than better statistical properties because of multiple testing in the presence of high variability. We present here a general framework for modelling the control phase of a monitoring programme, and for the estimation of quantities of potential clinical interest such as the ratio of false to true positive tests. We show how four recent clinical studies of monitoring cardiovascular disease, hypertension, diabetes and HIV infection can be thought as special cases of this framework; as well as using this framework to clarify the choice of estimation and calculation methods available. Noticeably, in each of the presented examples over-frequent monitoring appears to be a greater problem than under-frequent monitoring. We also present recalculations of results under alternative conditions, illustrating conceptual decisions about modelling the true or observed value of a clinical measure.