pyplsc 0.0.25

Python implementation of partial least squares correlation (PLSC)

Python implementation of partial least squares correlation (PLSC)

Background

PLSC is a multivariate statistical technique used in neuroscience (McIntosh et al., 1994; McIntosh & Lobaugh, 2004; Krishnan et al., 2011), among other fields. It uses compact singular value decomposition (SVD) to analyze relationships between a multivariate data array and a design matrix. When the object of study is brain-behaviour correlations or functional connectivity, this method is referred to as "behaviour PLSC" or "seed PLSC". In pyplsc, these are implemented by the PLSC model class.

Multivariate categorical differences across experimental conditions can also be analyzed by applying SVD to matrices of condition-wise averages. This approach is called "mean-centred PLSC" or "barycentric discriminant analysis" (BDA; Abdi et al., 2018) and is implemented in pyplsc by the BDA model class.

Finally, associations between continuous variables within participants (e.g., trial-by-trial ratings versus brain data) can be analyzed using within-participants PLSC (Roberts et al., 2016) as implemented in the WPLSC model class.

Installation

pyplsc can be installed from PyPI with:

pip install pyplsc

pyplsc is tested with Python 3.10 and above but may also work with earlier versions.

Usage

pyplsc replicates the statistical functionality of the PLS Matlab package, much like the pyls library. A major difference is that pyplsc uses a scikit-learn-style model-fitting syntax and accepts tabular (pandas.DataFrame) input:

from pyplsc import PLSC, BDA

mod = PLSC(random_state=123)
mod.fit(data=data_array, covariates=cov_table)

Permutation testing and bootstrap resampling are then run as separate steps (possibly in parallel using the n_jobs parameter):

perm_dist = mod.permute(n_perm=1000, n_jobs=3)
boot_dist = mod.bootstrap(n_boot=1000)

In contrast to other PLS implementations, pyplsc does not require data to be pre-sorted by (between-participant) group and (within-participant) condition:

mod = BDA()
mod.fit(data=data_array,
		design=design_matrix_dataframe,
		between='group',
		within='cond',
		participant='subj')

See the documentation for more details and examples.