Regression Discontinuity Designs

• Published in: Nursing economic Vol. 38; no. 2; pp. 98 - 102
• Main Authors: Perraillon, Marcelo Coca, Welton, John M, Hamer, Mika K, Myerson, Rebecca M
• Format: Journal Article
• Language: English
• Published: Pitman Anthony J. Jannetti, Inc 01.03.2020
• Subjects: Age groups; Americans with Disabilities Act 1990-US; Bias; Birth weight; Cancer; Clinical outcomes; Health care policy; Health economics; Insurance coverage; Medical research; Medicare; Mortality; Research methodology; Statistical analysis; Students; United States > US
• ISSN: 0746-1739

Regression discontinuity design is a statistical method used to estimate causal effects in observational studies. The method requires observing a variable and a cutoff point that either ultimately determines treatment assignment or is a strong predictor of treatment. Regression discontinuity requires large datasets since, often, not all observations are used in the analysis. Despite its strong internal validity, regression discontinuity remains underutilized in clinical research. The fundamental challenge in estimating causal effects from observational data is to correctly account for factors that determine the receipt of treatment. In classic randomization, units (e.g., patients) are assigned to treatment solely by chance. Random assignment ensures units in the treatment and control groups are comparable on all observed and unobserved characteristics. For many research questions, it may be infeasible or unethical to randomize units to treatment. In these cases, observational data can be used, but it is often not possible to know or observe all the factors that account for units receiving treatment. If it were possible to identify and observe all these factors, regression adjustment could be used to estimate unbiased treatment effects. In previous articles (Perraillon, Lindrooth, & Welton, 2019; Perraillon, Welton, & Jenkins, 2019), we described some research designs developed to estimate causal effects from observational data. In this article, the fundamental aspects of regression discontinuity design (RDD), a method developed to assess causal effects when the variable used to assign a treatment or policy change - along with a cutoff or threshold - is explored. RDD has become a standard empirical tool in health economics and health services research (Lee & Lemieux, 2010), but it remains underutilized in other fields (Moscoe et al., 2015; Venkataramani et al., 2016). As with other methods for causal inference with observational data, some assumptions can be verified with data, while others require indepth subject knowledge to assess their validity. Unlike other observational methods, RDD requires large datasets since only a portion of the data is used.