Skip to contents

Create a reproducible calibration/validation split

Source: R/nomo_split.R
nomo_split.Rd

nomo_split() creates an explicit random split for workflows in which EFA and CFA should be evaluated on different observations when the available sample permits it. The function does not claim that sample splitting is always preferable: dividing a modest sample reduces precision in both subsets, and an external validation sample is generally stronger evidence when one is available.

Usage

nomo_split(
  data,
  validation_prop = 0.5,
  seed = 1234L,
  guidance = nomo_defaults()
)

Arguments

data

A non-empty data frame.

validation_prop

Proportion of rows assigned to the validation sample. Must be strictly between 0 and 1.

seed

Integer seed used to make the split reproducible.

guidance

Guidance settings from nomo_defaults().

Value

A nomo_split object. The fields to read are:

  • calibration and validation: the two samples.

  • assignment: which sample each row went to.

  • n_total, n_calibration, n_validation, and validation_prop_realized.

  • seed and decision_log.

Other fields record the call, the settings used, and intermediate engine results. They may change between releases and are not part of the stable interface (see ?nomologR).

Details

The caller's random-number-generator state is restored after the split so that using nomo_split() does not silently alter later stochastic analyses.

References

Fokkema, M., & Greiff, S. (2017). How performing PCA and CFA on the same data equals trouble. European Journal of Psychological Assessment, 33(6), 399-402. doi:10.1027/1015-5759/a000460

MacCallum, R. C., Roznowski, M., & Necowitz, L. B. (1992). Model modifications in covariance structure analysis: The problem of capitalization on chance. Psychological Bulletin, 111(3), 490-504. doi:10.1037/0033-2909.111.3.490

Examples

split <- nomo_split(nomo_demo_continuous, validation_prop = 0.40, seed = 2026)
split
#> <nomo_split> Calibration and validation split
#> Rows: 500 total | 300 calibration | 200 validation
#> Validation proportion: 0.400 requested | 0.400 realized | Seed: 2026
#> Use splitting only when the gain in independence justifies the loss of
#> precision.
nrow(split$calibration)
#> [1] 300
nrow(split$validation)
#> [1] 200