tidy_mahalanobis: Calculate Mahalanobis Distance

In JoeyStanley/joeyr: Functions for Vowel Data

Calculate Mahalanobis Distance

Description

This is a tidyverse-tidyverse-compatible version of the stats::mahalanobis function. It just makes it easier to include it as part of a dplyr::mutate.

Usage

tidy_mahalanobis(...)

Arguments

...	Names of columns that should be included in the Mahalanobis distance. For vowel data, this is typically your F1 and F2 columns.

Details

Typically you'll want to group your data (using dplyr::group_by) by speaker and vowel class so that you get the distance from vowel centroids.

I won't tell you what to do with those distances, but if you might consider looking at tokens where the square root of the Mahalanobis distance is greater than around 2. However, to be clear, the exact cutoff will vary depending on the size and variability of your data. You can see how you might isolate these points visually in the example code below.

One small modification that this function does that stats::mahalanobis does not do is that if there are fewer than 5 measurements in a group, tidy_mahalanobis returns them all as having a distance of zero. I found that this prevents some fatal errors from crashing the script when running this function on smaller datasets.

Note that this function requires the MASS package to be installed to work, but you don't need to load it.

Value

A vector that contains the Mahalanobis distances for each observation.

Examples

suppressPackageStartupMessages(library(tidyverse))
df <- joeysvowels::midpoints

# Calculate the distances
m_dists <- df %>%
  group_by(vowel) %>%
  mutate(mahal_dist = tidy_mahalanobis(F1, F2))

# Take a peek at the resulting dataset
m_dists %>%
  select(vowel, F1, F2, mahal_dist) %>%
  head()

# Plot potential outliers
ggplot(m_dists, aes(F2, F1, color = sqrt(mahal_dist) > 2)) +
   geom_point() +
   scale_x_reverse() +
   scale_y_reverse()

# You can include whatever numeric variables you want, like duration
df %>%
  group_by(vowel) %>%
  mutate(dur = end - start) %>%
  mutate(mahal_dist = tidy_mahalanobis(F1, F2, dur)) %>%
  ggplot(aes(F2, F1, color = sqrt(mahal_dist) > 2.5)) +
  geom_point() +
  scale_x_reverse() +
  scale_y_reverse()

# Data cannot contain NAs. Remove them before running.
df[1,]$F1 <- NA
df %>%
  group_by(vowel) %>%
  mutate(mahal_dist = tidy_mahalanobis(F1, F2))
df %>%
  group_by(vowel) %>%
  filter(!is.na(F1)) %>%
  mutate(mahal_dist = tidy_mahalanobis(F1, F2)) %>%
  select(vowel_id, vowel, mahal_dist, F1, F2) %>%
  head()

JoeyStanley/joeyr documentation built on April 7, 2023, 8:37 p.m.