Title
Can Standard Statistical Methods Remove Selection Bias in Observational Studies?
Speaker
Thérèse Stukel, Institute for Clinical Evaluative Sciences and University of Toronto
Abstract
Comparisons of outcomes of treated and untreated patients in observational studies may be biased due to differences in patient prognosis, often as a result of unobserved biases in patient selection for treatment. We analyzed a cohort of 122,124 heart attack patients, linking medical chart data with health administrative data. Our objective was to compare four analytic methods for removing the effects of selection bias in observational studies: multivariable model risk adjustment, propensity score risk adjustment, propensity-based matching, and an econometric technique that removes confounding due to unmeasured variables, instrumental variable analysis. Invasive (surgical) versus conservative (pharmacological) cardiac management was associated with a 50% relative decrease in mortality using standard multivariate Cox survival modeling, propensity score risk adjustment, and propensity-based matching. Instrumental variable analysis showed a 16% relative decrease in mortality, using regional cardiac catheterization rate as the instrument. Randomized clinical trials have shown the survival benefits of routine invasive care to be between 8 percent and 21 percent. The sensitivity to analytic method is due to unmeasured patient variables as well as survival bias, since short term mortality rates are high and patients may die before receiving treatment. Standard risk-adjustment methods have similar limitations regarding removal of unmeasured treatment selection biases. Caution is advised in analyzing the effects of invasive or surgical treatments using standard statistical methods with observational data, as they are particularly susceptible to such biases.
Meet the speaker in Room 212 Cockins Hall at 4:30 p.m. Refreshments will be served.
