In mlindsk/molic: Multivariate Outlier Detection in Contingency Tables

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  message = FALSE,
  warnings = FALSE,
  fig.path = "man/figures/README-",
  out.width = "100%"
)

About molic

An R package to perform outlier detection in contingency tables (i.e. categorical data) using decomposable graphical models (DGMs); models for which the underlying association between all variables can be depicted by an undirected graph. molic are designed to work with undirected decomposable graphs returned from fit_graph in the ess package. Compute-intensive procedures are implemented using Rcpp/C++ for better run-time performance.

Installation

You can install the current stable release of the package by using the devtools package:

devtools::install_github("mlindsk/molic", build_vignettes = FALSE)

Articles

The Outlier Model: The "behind the scenes" of the outlier model.
Detecting Skin Diseases: An example of using the outlier model to detect skin diseases.
Outlier Detection in Genetic Data: An example of how to conduct an outlier analysis in genetic data.

Example of Usage

library(dplyr)
library(molic)
library(ess)   # For the fit_graph function
set.seed(7)    # For reproducibility

Psoriasis patients

d <- derma %>%
  filter(ES == "psoriasis") %>%
  select(-ES) %>%
  as_tibble()

Fitting the interaction graph

g <- fit_graph(d, trace = FALSE) # see package ess for details
plot(g, vertex.size = 15)

This plot shows how the variables are 'associated' in the psoriasis class; see ess for more information about fit_graph. The outlier model exploits this knowledge instead of assuming independence between all variables (which would clearly be a wrong assumption looking at the graph). The graph may look very different for other classes than psoriasis.

Example 1 - Testing which observations within the psoriasis class are outliers

We start by fitting an outlier model taking advantage of the fittet graph g which holds information about the psoriasis patients. The print method prints information about the distribution of the (deviance) test statistic.

m1 <- fit_outlier(d, g)
print(m1)

Notice that m1 is of class 'outlier'. This means, that the procedure has tested which observations within the data are outliers. This method is most often just referred to as outlier detection. The outliers, on a 5% significance level, can now be extracted as follows:

outs  <- outliers(m1)
douts <- d[which(outs), ]
douts

The following plot is the distribution of the test statistic corresponding to the information retrieved using the print method. One can think of a simple t-test, where the distribution of the test statistic is a t-distribution. In order to conclude on the hypothesis, one finds the critical value and verify if the test statistic is greater or less than this.

plot(m1)

Retrieving the observed test statistics for the individual observations:

x1   <- douts[1, ] %>% unlist() # an outlier
x2   <- d[1, ] %>% unlist()     # an inliner
dev1 <- deviance(m1, x1) # falls within the critical region in the plot (the red area)
dev2 <- deviance(m1, x2) # falls within the acceptable region in the plot
dev1
dev2

Retrieving the p-values:

pval(m1, dev1)
pval(m1, dev2)

Example 2 - Testing if a new observation is an outlier

An observation from class chronic dermatitis:

z <- derma %>%
  filter(ES == "chronic dermatitis") %>%
  select(-ES) %>%
  slice(1) %>%
  unlist()

Test if z is an outlier in class psoriasis:

m2 <- fit_outlier(d, g, z)
print(m2)
plot(m2)

Notice that m2 is of class 'novelty'. The term novelty detection is sometimes used in the litterature when the goal is to verify if a new unseen observation is an outlier in a homogeneous dataset. Retrieving the test statistic and p-value for z

dz <- deviance(m2, z)
pval(m2, dz)

How To Cite

If you want to cite the outlier method please use

@article{lindskououtlier,
  title={Outlier Detection in Contingency Tables Using Decomposable Graphical Models},
  author={Lindskou, Mads and Svante Eriksen, Poul and Tvedebrink, Torben},
  journal={Scandinavian Journal of Statistics},
  publisher={Wiley Online Library},
  doi={10.1111/sjos.12407},
  year={2019}
}

If you want to cite the molic package please use

@software{lindskoumolic,
  author       = {Mads Lindskou},
  title        = {{molic: An R package for multivariate outlier 
                   detection in contingency tables}},
  month        = oct,
  year         = 2019,
  publisher    = {Journal of Open Source Software},
  doi          = {10.21105/joss.01665},
  url          = {https://doi.org/10.21105/joss.01665}
}