Information about patients and the care they receive is routinely collected in public hospitals and general practices. Clinicians and government agencies use this information to complement evidence from clinical trials to evaluate the benefits of different treatment options. Routine data are not collected primarily for research purposes, and hence the investigator has no control over the way patients are allocated to different treatment groups. In these settings, one approach that can help ensure that different groups of patients are comparable is to allocate patients according to a risk factor threshold. For example, anti-hypertensive drugs may be only given to patients whose blood pressure exceeds a certain level. Patients whose blood pressure is just below or above that level are expected to have similar characteristics, but may receive a different treatment. As a result, any differences in health outcome (e.g. mortality) between the treatment groups can be directly attributed to differences in treatment (anti-hypertensive drug). This study approach, known as 'regression discontinuity (RD) design', has received little attention in clinical research. This study aims to develop appropriate guidance for the use of RD designs to study the effects of medical treatments on patients' health. By doing this, we seek to clarify in which circumstances RD designs can be credibly used in clinical research.
Routinely-collected data are increasingly used to complement trial-based evidence for evaluating treatment effects of health interventions. The regression discontinuity (RD) design has been identified as a key tool for drawing causal inferences in such settings. While the number of applications of RD designs for studying the effects of health interventions is increasing, appropriate diagnostic tests for investigating the validity of RD designs in clinical settings have received little attention. In addition, statistical methods currently used in practice for the estimation of causal effects and corresponding uncertainty measures are sub-standard. This study will devise methodological guidelines for the use regression discontinuity designs to estimate treatment effects from routinely-collected data. Firstly, drawing on insights from the econometric literature, we will propose appropriate diagnostics tests to help assess whether a specific RD design is applicable. Secondly, we will investigate the required sample sizes for power analyses as a function of bandwidth estimates. Thirdly, we will critically assess alternative estimation approaches, including bandwidth estimation, for drawing causal inferences in RD designs. We will illustrate the methods in a study of the effects of statins on a wide range of outcomes, such as LDL cholesterol level, mortality and health care costs, using CPRD linked data. We will consider statin initiation as a function of cardiovascular disease risk score according to NICE clinical guidelines ('treatment rules') for the UK.
Mortality (all-cause and cardiovascular-related, identified through ONS data); LDL cholesterol (identified from CPRD); Hospital admissions/length-of-stay (HES inpatient data); Cardiovascular events (identified from CPRD/HES diagnoses).
Luke Keele - Chief Investigator - Georgetown University
Manuel Gomes - Corresponding Applicant - University College London ( UCL )
Michail Katsoulis - Collaborator - Farr Institute of Health Informatics Research
HES Admitted Patient Care;ONS Death Registration Data;Patient Level Index of Multiple Deprivation