Obtaining causal estimates of therapeutic effects in observational studies : the usefulness and validity of physician’s preference as an instrumental variable

This thesis investigates the validity and usefulness of physician's preference-based instrumental variable analysis in clinical epidemiological studies. Chapter 2 describes a survey amongst general practitioners, showing substantial variation in prescribing preference and showing prescribing patterns which suggest the stochastic monotonicity assumption may be plausible for physician's preference as an instrumental variable. Chapter 3 describes an application of physician's preference-based instrumental variable analysis in a moderate-sized study, showing uninformatively wide confidence intervals which limit the usefulness of instrumental variable analysis in this setting. Chapter 4 focuses on the bias-variance trade-off of instrumental variable analysis in comparison to conventional analyses, using simulations and theoretical derivations. Chapter 5 compares instrumental variable and conventional estimates of the effect of third versus second generation oral contraceptives on occurrence of venous thromboembolism. The similarity of these estimates under different sets of assumptions suggests major confounding is unlikely. Chapter 6 contains a suggestion for an additional step for reporting of instrumental variable analyses. The focus shifts to Mendelian randomisation in the second part of the thesis. Chapter 7 reviews the methodological approaches used in Mendelian randomisation studies and the quality of reporting. Chapter 8 shows that collider-stratification bias may exist in Mendelian randomisation studies in elderly populations.