Methods/Statistics
Methods/Statistics Emma Kileel* Emma Kileel Alana Brennan Matthew Fox Jennifer Weuve Jacob Bor
Background: The regression discontinuity design, a quasi-experimental technique used to estimate the causal effect of an intervention when random assignment is not possible, is gaining traction in epidemiologic research. This method leverages measurement error, typically regarded as a source of misclassification and bias in epidemiology, to emulate a random treatment assignment. It is unclear how the degree of measurement error impacts the ability of the regression discontinuity method to obtain valid and precise estimates of effect.
Methods: We simulated a dataset of 50,000 observations each with a value representing an HbA1c test result, the assignment variable, measured without error. At the diabetes diagnosis value of 6.5%, we simulated a discontinuity in the probability of outcome, Y, of 20%, representing the “true” estimate of effect. We simulated nondifferential measurement error by generating error terms randomly sampled from a normal distribution with specified means and standard deviations. Error terms were added to simulated HbA1c values to represent measured HbA1c values. True versus measured HbA1c values were compared to calculate sensitivity (Se) and specificity (Sp) of the assignment variable. Local linear regression was used to estimate the intent-to-treat (ITT) effect (risk difference).
Results: The ITT estimate did not vary substantially across degrees of measurement error (Table). When HbA1c was measured with very little error (Error SD=0.05, Se=99.5%, Sp=99.6%), the ITT estimate of effect was 20.7% (95% CI: 17.9, 23.5%). Even in scenarios of extreme measurement error (Sp<50%) the ITT estimate was 20.9% (95% CI: 18.1, 23.8%). When the Se was less than 50% the ITT estimate was 22.6% (95% CI: 19.5, 25.7%).
Conclusion: Using simulation methods, we demonstrated that even in the presence of large degrees of measurement error and corresponding misclassification, there was no bias in the ITT effect estimate when using a regression discontinuity design.
