Statistics How To

Bipartite Matching: Definition, Examples

Share on

Randomized Clinical Trial >

Matching places participants in observational studies into comparable, homogeneous groups or strata at the beginning of a study. It is one way to avoid selection bias (Cochran and Chambers, 1965). Matching designs can be bipartite matching, or non-bipartite matching, which are terms borrowed from graph theory.

bipartite matching

Bipartite Matching

Bipartate matching—also called conventional two-group matching—creates pairs from two distinct groups. On the left in the above image, connections are made between the treatment and controls. Bipartate matching is equivalent to sampling without replacement.

Graphically, you would match each node in the treatment group to a single node from the control group. A node is a point or circle (red in the left image, blue on the right). It is the fundamental unit from which graphs are made. The match is represented by an edge in the graph. An edge is a line that connects two nodes. A weight is assigned to each edge; Basically, this weight is the difference for some aspect of the pairs, like difference in age, height, or BMI. Well matched pairs ideally have very little difference between them. Therefore, smaller weights are preferable. Although the concept is simple (i.e. create matched pairs with the smallest differences), the calculations are not—especially if you have multiple covariates. Algorithms like the Greedy Matching Algorithm have been developed to create ideal weights between nodes.

One drawback to this type of matching is that it can only be used for fairly simple designs.

Non-Bipartite Matching

A non-bipartate design—or multi-group matched design—produces pairs from multiple groups. It is equivalent to sampling with replacement. Bipartite designs are more common, but non-bipartite designs are available for the rare case when you want to reuse a member; For example, if you use the same control as a match for two or more treatment group participants.

In the above graph, each node in the right hand box is in a separate group. Non-bipartite designs are available for when you want to reuse a member; For example, if you use the same control as a match for two or more treatment group. Augmenting path algorithms like the Blossom V algorithm (available here) are available for creating non-bipartite matches. They are technically complex, which may be a reason why biparite matches are often preferred.

Cochran WG, Chambers SP. The Planning of Observational Studies of Human Populations. Journal of Royal Statistical Society, Ser A. 1965;128:234–266.
Lu et. al. Optimal Nonbipartite…and Its Statistical Applications. Am Stat. 2011; 65(1): 21–30.

Stephanie Glen. "Bipartite Matching: Definition, Examples" From Elementary Statistics for the rest of us!

Need help with a homework or test question? With Chegg Study, you can get step-by-step solutions to your questions from an expert in the field. Your first 30 minutes with a Chegg tutor is free!

Comments? Need to post a correction? Please post a comment on our Facebook page.