Continuing our work on developing epidemiology methods we recently published a paper in Annals of Epidemiology describing a new approach to matching, that we call “flex matching”, in nested case-control studies. We show that flex matching prevents over matching, which is common with traditional approaches to matching, is statistically efficient, and does not increase Type I error.

Traditional Individual matching in case-control studies can improve statistical efficiency over random selection of controls but in practice matched case-control studies are often over matched which produces biased effect estimates and lower statistical efficiency. (The classic paper by Sholom Wacholder on matching is here and an excellent discussion of matching and data analysis by Neil Pearce is here) The issue with traditional matching is that when the matching criteria are complex or restrictive, often controls that meet the matching criteria are not available and the case for which the investigator is seeking a matched control must be dropped from the analysis. When cases are dropped from the analyses bias often occurs and statistical efficiency drops rapidly with each unmatched case that has to be removed from the analytical data set.     

Flex matching uses sequential rounds of matching with loosening of matching criteria in each round. In the first round of matching, controls are matched to cases using the investigator’s preferred matching variables and criteria (e.g., +/- 1 years of age).  In subsequent rounds, controls are sought for cases for which matched controls were not found in the prior rounds using the same matching variables, but with relaxed criteria (e.g., +/- 2 years of age and then +/- 3 years).  After several rounds of increasingly relaxed matching, remaining unmatched cases are randomly assigned a control.  The data are then analyzed using conditional logistic regression with matching variables included as covariates.  We show that this approach reduces bias compared to traditional matching approaches that require the investigator to remove cases from the analytical data set and is more statistically efficient than selecting controls without matching.