Loading...
Please wait, while we are loading the content...
Similar Documents
Commentary on Yu et al.: Opportunities and Challenges for Matching Methods in Large Datasets
| Content Provider | Semantic Scholar |
|---|---|
| Author | Stuart, Elizabeth A. Ackerman, Benjamin |
| Copyright Year | 2019 |
| Abstract | In their paper titled “Matching Methods for Observational Studies Derived from Large Administrative Databases,” authors Ruoqi Yu, Jeffrey Silber and Paul Rosenbaum [13] discuss matching methods in the age of “big data.” Matching methods, such as Mahalanobis distance matching, exact matching, and propensity score matching, are well established design strategies for reducing bias due to observed characteristics in non-experimental studies. Fundamentally, matching methods aim to create matched groups of units (often individuals) who are similar to one another on a (sometimes large) set of covariates. Exact matching aims to do that matching on each variable simultaneously (e.g., find for each treated individual a control individual with the same values of age, gender, baseline health status, etc.). Other matching methods summarize the covariates into a low-dimensional summary (e.g., the probability of receiving treatment, known as the propensity score [9]) and match units on that summary. The simplest form of matching is 1:1 matching, where each treated subject is matched to 1 comparison subject similar on the variables used in the matching. Many variations on 1:1 matching exist, including variable ratio matching, full matching, and methods that allow controls to be used as a match more than once (“matching with replacement”). Recent advances in matching include “fine balance,” which aims to exactly match the marginal distributions of the covariates rather than requiring each individual match exactly [8]. See Stuart (2010) [11] for the background and details of some of those approaches. (Note that while Yu et al. [13] use the term “control” to refer to the reference group, we prefer the term “comparison” group in the non-experimental context, to distinguish it from the control group in a randomized controlled trial. For current purposes readers can consider the “control” and “comparison” terms equivalent.) Weighting, in particular inverse probability of treatment weighting [1], is another common strategy for handling observed covariates in non-experimental studies. Some researchers view it as preferable to matching because it does not “throw away data” as matching can appear to do. (For example, a study with 1000 treated and 2000 comparison may end up using only 1000 of the comparison subjects in a 1:1 matched design). However, there are a number of benefits to considering matching. First, it is often an attractive approach for non-experimental study design because of its strong design aspects: it is straightforward, for example, to show the similarity of the matched groups and for even non-technical |
| File Format | PDF HTM / HTML |
| Alternate Webpage(s) | https://www.e-publications.org/ims/submission/STS/user/submissionFile/41389?confirm=cf15773d |
| Language | English |
| Access Restriction | Open |
| Content Type | Text |
| Resource Type | Article |
