Some best practices for anticlustering

Martin Papenberg

Source: vignettes/Best_practices.Rmd

This vignette documents some “best practices” for anticlustering using the R package anticlust. In many cases, the suggestions pertain to overriding the default values of arguments of anticlustering(), which seems to be a difficult decision for users. However, I advise you: Do not stick with the defaults; check out the results of different anticlustering specifications; repeat the process; play around; read the documentation (especially ?anticlustering); change arguments arbitrarily; compare the output. Nothing can break.1

This document uses somewhat imperative language; nuance and explanations are given in the package documentation, the other vignettes, and the papers by Papenberg and Klau (2021; https://doi.org/10.1037/met0000301) and Papenberg (2024; https://doi.org/10.1111/bmsp.12315). Note that deciding which anticlustering objective to use usually requires substantial content considerations and cannot be reduced to “which one is better”. However, some hints are given below.

  • If speed is not an issue (it usually is not), use method = "local-maximum" instead of the default method = "exchange". It is unambiguously better.
  • If speed is not an issue (it usually is not), use several repetitions.
  • Use standardize = TRUE instead of the default standardize = FALSE.2
  • Do not use the default objective = "diversity" when the group sizes are not equal (preferably, use objective = "kplus").
  • If you only care about similarity in mean values, use objective = "variance".
  • You should (probably) not only care about similarity in mean values: prefer objective = "kplus" over objective = "variance" (or check out the function kplus_anticlustering()).
  • If you (only) care about similarity in means and standard deviations, use objective = "kplus" instead of the default objective = "diversity".
  • With k-plus anticlustering, always use standardize = TRUE.
  • If you want to apply anticlustering on a large data set, read the vignette “Speeding up Anticlustering”.
  • If you have a very large3 data set, do not use the default objective = "diversity" because computing the distance matrix may overwhelm your computer. Instead use objective = "variance" or objective = "kplus".
  • For the optimal anticlustering algorithms (i.e., method = "ilp"): Install the SYMPHONY solver instead of the GLPK solver.
  • The objective = "dispersion" is the only objective where you should first try out the optimal method instead of a heuristic, i.e., use method = "ilp" or the function optimal_dispersion().4

References

Papenberg, M., & Klau, G. W. (2021). Using anticlustering to partition data sets into equivalent parts. Psychological Methods, 26(2), 161–174. https://doi.org/10.1037/met0000301.

Papenberg, M. (2024). K-plus Anticlustering: An Improved k-means Criterion for Maximizing Between-Group Similarity. British Journal of Mathematical and Statistical Psychology, 77 (1), 80–102. https://doi.org/10.1111/bmsp.12315

  1. Well, actually your R session can break if you use an optimal method (method = "ilp") with a data set that is too large.↩︎

  2. You might ask why standardize = TRUE is not the default. Actually, there are two reasons. First, the argument was not always available in anticlust and changing the default behaviour of a function when releasing a new version is oftentimes undesirable. Second, it seems like a big decision to me to just change users’ data by default (which is done when standardizing the data). In doubt, just compare the results of using standardize = TRUE and standardize = FALSE and decide for yourself which you like best. Standardization may not be the best choice in all settings.↩︎

  3. What is considered very large may depend on the specifications of your computer. I used to have problems computing a distance matrix for \(N > 20000\).↩︎

  4. The maximum dispersion problem can be solved optimally for rather large data sets, especially for \(K = 2\). For \(K = 3\) and \(K = 4\), several 100 elements can usually be processed, especially when installing the SYMPHONY solver.↩︎