NB: this is now deprecated in favor of the similarly named ldamatch, written in pure R and introducing additional features.

Introduction

lda-match2.py and lda-match3.py are Python 2 and Python 3 versions (respectively) of scripts which allow the user to generate a subsample of data (represented by a comma-separated values file) in which two groups do not differ (according to a two-tailed unequal variance t-test) on an arbitrary number of real-valued measures.

The approach here is similar to the "greedy" algorithm used by van Santen et al. (2010; Autism) but in general results in larger subgroups, as it uses linear discriminant analysis (LDA) to identify outliers. To a first approximation, when the t-test assumptions (normality or large samples), the assumptions of LDA will also hold, and therefore this will give something close to an "optimal" subsample according to a criterion which favors large subsamples of approximately the same size.

Installation instructions

This code requires Python (either 2 or 3) and two additional packages, numpy and rpy2. You will need all the following:

• Python (try python --version)

• R (try R --version)

• C compiler (try cc --version)

• pip, the Python package manager (try pip)

• numpy and rpy2:

  sudo pip install numpy rpy2

Usage

Users must specify the labels for the two groups (-a, -b), the column name containing the groups (-g), the column name(s) of the feature(s) to match on (-m), and the location of the input file. Users may specify that observations are only to be removed from the first group (-d), the two-tailed alpha level (-p, by default .2), and the location for the output file (-o; by default, results are printed to STDOUT).

For more information, refer to the worked examples below.

Worked examples

• Match ALN and ALI children in DX.csv on chronological age (CA), ADOS severity score (ADOS), and SCQ total score (SCQ) at two-tailed alpha >= 0.2, and write the resulting set to a file called TD-SLI.csv:

  python lda-match2.py -a TD -b SLI -g DX -m CA -m ADOS -m SCQ -p 0.2 -o TD-SLI.csv DX.csv

• Match TD and ALN/ALI children in DX.csv on chronological age (CA) and non-verbal IQ (NVIQ) at two-tailed alpha >= .5 and write the resulting set to a file called TD-ASD-p5.csv:

  python lda-match2.py -a TD -b ALN -b ALI -g DX -m CA -m NVIQ -p 0.5 -o TD-ASD-p5.csv DX.csv

• Alternative conventions for generating the previous set:

  python lda-match2.py -aTD -bALN,ALI -gDX -mCA,NVIQ -p.05 DX.csv > TD-ASD-p5.csv

Some results on the ERPA data

• TD (36) vs. SLI (20): CA! [+1]
• ALN (25) vs. TD (28): CA, NVIQ, VIQ! [+1]
• ALN (23) vs. ALI (24): CA, ADOS!, SCQ [+2]
• TD (42) vs. ASD (ALN = 24, ALI = 19): CA, NVIQ! [0]
• LN (ALN = 26, TD = 39) vs. LI (ALI = 26, SLI = 20): CA! [+2]
• ALI (26) vs. SLI (20): CA, NVIQ, VIQ (no matching necessary) [+3]
• ASD (ALN = 25, ALI = 26) vs. nASD (TD = 44, SLI = 20): CA, NVIQ (no matching necessary) [+6]

Key:

• !: last feature matched
• [+N]: change in overall subsample size compared to the "greedy" method operating at the same alpha level

License

BSD-like (see the source)

Author

Kyle Gorman ([email protected]), with thanks to Steven Bedrick