OpenClustered: an R package with a benchmark suite of clustered datasets for methodological evaluation and comparison.

backgroundClustered data arise when observations are correlated within a group or sampling unit and frequently arise in epidemiology, social sciences, education, linguistics, econometrics, and medicine. Given growing interest in clustered data, we developed a data repository offering clustered datasets that can be used for methodologic comparison with open-source, publicly available data. Traditionally, data simulation studies are employed for methodology evaluation and comparison, which can be fraught with issues such as overly simplistic design and potential for bias. Excellent data repositories are available for standard (non-clustered) datasets, such as OpenML and the Penn Machine Learning Benchmark repository, but there is a paucity of resources available that have clustered data.

resultsIn this pilot study, we developed an R package called OpenClustered, which includes 19 clustered datasets with binary outcomes arising from various domains and varying in terms of their size and composition. We present tutorials for using OpenClustered, including examples for filtering and summarizing the datasets. We demonstrate the use of OpenClustered with a small benchmarking study comparing Frequentist and Bayesian implementations of generalized linear mixed models. All code and data are contained on the OpenClustered GitHub page.

conclusionThe OpenClustered R package is the start of a useful data resource for conducting benchmarking studies with open-source clustered data. It facilitates empirical methodologic guidance that is less prone to bias compared to data simulation studies, thereby improving rigor across diverse research fields. In the future, we plan to add more datasets, particularly those with continuous outcomes, as well as functionality for users to submit their clustered datasets to be included in the repository.