Types Without Taxa

A Matched-Null Test of Whether the Clusters a Pipeline Reports Are Real, With an Application to Personality Types

Do the "personality types" that clustering algorithms report reflect real categories of people, or are they artifacts of the analysis?

A preregistered reanalysis, in R, of three large public personality datasets (~1.0M respondents). When you run a clustering algorithm on personality data it almost always returns some number of "types," whether or not real types exist. This study asks how to tell the difference, and answers it with a matched null: a synthetic population built to share the real data's trait distributions and its full correlation structure, but with no latent types by construction. If a typeless null produces as many "types" as the real data, the types were never there.

The point is methodological. Change one routine setting, the covariance assumption, and the number of types found slides from one to ten. Held against the matched null and a set of preregistered criteria, none of the three instruments shows separated categories, and the test gives the same verdict whichever clustering engine you plug in. Personality is just where I tried the test out. The test is the thing I'd want others to pick up and use: it ships as an R package, matchednull, now on CRAN. The manuscript is under review at Behavior Research Methods.

Author: Miura Meng

Read it

• Read the paper (web)
• The paper as a formatted PDF
• R package: https://CRAN.R-project.org/package=matchednull
• Preregistration (OSF): https://osf.io/2ekcg

What's here

Path	Contents
R/	analysis scripts 00–08, run in order; 15–17 are the supplementary simulations (engine check, dimension sweep, t-copula calibration)
prereg/	the preregistered decision rule, specification grid, knowledge-of-data, and the sealed hold-out indices
figures/	the figures
paper/	the manuscript: Quarto source, plus rendered PDF and HTML
data/	SOURCES.md, where to obtain the public datasets (the data are not redistributed here)

Method, in one paragraph

Each dataset is modeled as a Gaussian mixture; "types" are the components selected by BIC across mclust's fourteen covariance parameterizations. To ask whether an apparent clustering exceeds what the data's own margins and covariance produce, each dataset is compared to a Gaussian-copula matched null that fixes every univariate marginal and the full covariance and randomizes only the higher-order dependence. A taxometric analysis (CCFI) provides an independent categorical-versus- continuous test. A decision rule, fixed before the held-out data were opened, declares the structure continuous only if all four preregistered criteria hold. See prereg/decision_rule.md.

Reproduce

Requires R (≥ 4.4) with mclust and RTaxometrics. The datasets are public but large and are not redistributed here; obtain them via data/SOURCES.md, score them to the trait/facet level, then run, in order:

Rscript R/00_seal_holdout.R                       # 50/50 split, seed 20260620
Rscript R/01_make_spec_grid.R                     # the specification grid
Rscript R/02_real_grid.R                          # selected k across 14 covariance models, matched n
Rscript R/03_copula_null.R <slice> <n> <reps>     # matched-null distribution, per slice
Rscript R/04_ccfi.R && Rscript R/05_ccfi_extract.R  # taxometric CCFI per trait
Rscript R/06_criteria.R                           # confident-assignment + first-split shares
Rscript R/07_figures.R                            # Figures 1 and 2
Rscript R/08_positive_control.R                   # Figure 3 (positive-control power check; self-contained synthetic data)

The scripts are research code: paths and slice subsamples reflect the author's setup and may need adjusting. Seed of record: 20260620. The exact held-out respondents are listed in prereg/holdout_confirm_*.txt.

Data sources

All public. Each dataset belongs to its providers; cite them as the data source. See data/SOURCES.md.

• IPIP-NEO-120 (Johnson): 5 domains, 30 facets
• IPIP Big-Five 50-item (Open-Source Psychometrics): 5 domains
• IPIP-HEXACO: 6 domains, 24 facets

Preregistration

Registered on OSF before the held-out data were analyzed (https://osf.io/2ekcg). The decision rule and specification grid in prereg/ are the registered versions; the confirmatory verdict appended to prereg/decision_rule.md was computed after the rule was locked.

Citation

If you use this material, please cite the registration:

Miura Meng. (2026). Types Without Taxa: A Covariance-Matched-Null Multiverse Test of Categorical versus Continuous Personality Structure. OSF. https://osf.io/2ekcg

(Manuscript in preparation.)

License

Code is released under the MIT License (see LICENSE). The data are not covered by that license and remain subject to the terms of their original providers.