2009 Atlantic Causal Modeling Conference

Research Interests

Matched Designs for Observational Studies

In an observational study of a treatment effect, a researcher might perform pair matching to produce samples of treated and control subjects that are similar on observed covariates. What if the researcher would like to go beyond pair matching and match more than one control to a treated subject? How many controls should be matched to a treated subject? I have developed a new design for an observational study in which treated subjects are matched to a number of controls, using entire matching.

Using the propensity score in a new way, I determine the optimal number of controls to be matched to each treated subject. Entire matching accomplishes two important objectives: (1) it produces treated and control samples that have similar distributions of observed covariates; and (2) subject to attaining the first objective, it maximizes the size of the control sample. A design that achieves these two objectives would produce an estimate of a treatment effect that (1) is unbiased under strong ignorability and, subject to that, (2) attains the minimum standard error. While pair matching can attain the first objective, it will by definition never attain the second. On the other hand, matching with a fixed number of controls may produce treated and control samples that are not similar on observed covariates, meaning that the first objective is not attained.

Entire matching, which uses an optimal number of controls, outperforms both traditional methods of matching by meeting both objectives. In my work I demonstrate this theoretically, by simulation, and through an example of an observational study of treatments for skin cancer.

Here is an R workspace with the functions for entire matching already loaded entire_match.RData with a user's guide. The software comes with no promises or guarantees, but please email me with any problems or questions.

Manuscripts & Publications

Gimotty, P. A., Yoon, F., Hammond, R., Rosenbaum, P. R., and Guerry, D. "Therapeutic Effect of Sentinel Lymph Node Biopsy in Melanoma Remains an Open Question." August 10, 2009. Journal of Clinical Oncology, 27.

Yoon, F., Gimotty, P. A., Guerry, D., and Rosenbaum, P. R. "Using Multiple Control Groups as Evidence About Unobserved Biases in an Observational Study of Treatments for Melanoma." Oct 2008. [pdf]

Iqbal, M., Kim, S., Yoon, F. "An Investigation Into Differential Diagnosis of Pulp and Periapical Pain: A PennEndo Database Study." Journal of Endodontics (33:5). May 2007. [pdf]