ordr integrates ordination analysis and biplot visualization into tidyverse workflows.
Wherever there is an SVD, there is a biplot.1
Ordination is a catch-all term for a variety of statistical techniques that introduce an artificial coordinate system for a data set in such a way that a few coordinates capture a large amount of the data structure 2. The branch of mathematical statistics called geometric data analysis (GDA) provides the theoretical basis for (most of) these techniques. Ordination overlaps with regression and with dimension reduction, which can be contrasted to clustering and classification in that they assign continuous rather than categorical values to data elements 3.
Most ordination techniques decompose a numeric rectangular data set into the product of two matrices, often using singular value decomposition (SVD). The coordinates of the shared dimensions of these matrices (over which they are multiplied) are the artificial coordinates. In some cases, such as principal components analysis, the decomposition is exact; in others, such as non-negative matrix factorization, it is approximate. Some techniques, such as correspondence analysis, transform the data before decomposition. Ordination techniques may be supervised, like linear discriminant analysis, or unsupervised, like multidimensional scaling.
Analysis pipelines that use these techniques may use the artificial coordinates directly, in place of natural coordinates, to arrange and compare data elements or to predict responses. This is possible because both the rows and the columns of the original table can be located, or positioned, along these shared coordinates. The number of artificial coordinates used in an application, such as regression or visualization, is called the rank of the ordination 4. A common application is the biplot, which positions the rows and columns of the original table in a scatterplot in 1, 2, or 3 artificial coordinates, usually those that explain the most variation in the data.
An extensive range of ordination techniques are implemented in R,
from classical multidimensional scaling (stats::cmdscale())
and principal components analysis (stats::prcomp() and
stats::princomp()) in the stats package
distributed with base R, across widely-used implementations of linear
discriminant analysis (MASS::lda()) and correspondence
analysis (ca::ca()) in general-use statistical packages, to
highly specialized packages that implement cutting-edge techniques or
adapt conventional techniques to challenging settings. These
implementations come with their own conventions, tailored to the
research communities that produced them, and it would be impractical
(and probably unhelpful) to try to consolidate them.
Instead, ordr provides a streamlined process by which the models output by these methods—in particular, the matrix factors into which the original data are approximately decomposed and the artificial coordinates they share—can be inspected, annotated, tabulated, summarized, and visualized. On this last point, most biplot implementations in R provide limited customizability. ordr adopts the grammar of graphics paradigm from ggplot2 to modularize and standardize biplot elements 5. Overall, the package is designed to follow the broader syntactic conventions of the tidyverse, so that users familiar with a this workflow can more easily and quickly integrate ordination models into practice.
ordr remains under development but is approaching a
CRAN release. For now, it can be installed from the (default)
main branch using remotes:
remotes::install_github("corybrunson/ordr")Morphologically, Iris versicolor is much closer to Iris virginica than to Iris setosa, though in every character by which it differs from Iris virginica it departs in the direction of Iris setosa.6
A very common illustration of ordination in R applies principal components analysis (PCA) to Anderson’s iris measurements. These data consist of lengths and widths of the petals and surrounding sepals from 50 each of three species of iris:
head(iris)
#> Sepal.Length Sepal.Width Petal.Length Petal.Width Species
#> 1 5.1 3.5 1.4 0.2 setosa
#> 2 4.9 3.0 1.4 0.2 setosa
#> 3 4.7 3.2 1.3 0.2 setosa
#> 4 4.6 3.1 1.5 0.2 setosa
#> 5 5.0 3.6 1.4 0.2 setosa
#> 6 5.4 3.9 1.7 0.4 setosa
summary(iris)
#> Sepal.Length Sepal.Width Petal.Length Petal.Width
#> Min. :4.300 Min. :2.000 Min. :1.000 Min. :0.100
#> 1st Qu.:5.100 1st Qu.:2.800 1st Qu.:1.600 1st Qu.:0.300
#> Median :5.800 Median :3.000 Median :4.350 Median :1.300
#> Mean :5.843 Mean :3.057 Mean :3.758 Mean :1.199
#> 3rd Qu.:6.400 3rd Qu.:3.300 3rd Qu.:5.100 3rd Qu.:1.800
#> Max. :7.900 Max. :4.400 Max. :6.900 Max. :2.500
#> Species
#> setosa :50
#> versicolor:50
#> virginica :50
#>
#>
#> ordr provides a convenience function to send a subset of columns to an ordination function, wrap the resulting model in the tibble-derived ‘tbl_ord’ class, and append both model diagnostics and other original data columns as annotations to the appropriate matrix factors:7
(iris_pca <- ordinate(iris, cols = 1:4, model = ~ prcomp(., scale. = TRUE)))
#> # A tbl_ord of class 'prcomp': (150 x 4) x (4 x 4)'
#> # 4 coordinates: PC1, PC2, ..., PC4
#> #
#> # Rows (principal): [ 150 x 4 | 1 ]
#> PC1 PC2 PC3 ... | Species
#> | <fct>
#> 1 -2.26 -0.478 0.127 | 1 setosa
#> 2 -2.07 0.672 0.234 ... | 2 setosa
#> 3 -2.36 0.341 -0.0441 | 3 setosa
#> 4 -2.29 0.595 -0.0910 | 4 setosa
#> 5 -2.38 -0.645 -0.0157 | 5 setosa
#> # … with 145 more rows
#> #
#> # Columns (standard): [ 4 x 4 | 3 ]
#> PC1 PC2 PC3 ... | name center scale
#> | <chr> <dbl> <dbl>
#> 1 0.521 -0.377 0.720 | 1 Sepal.Length 5.84 0.828
#> 2 -0.269 -0.923 -0.244 ... | 2 Sepal.Width 3.06 0.436
#> 3 0.580 -0.0245 -0.142 | 3 Petal.Length 3.76 1.77
#> 4 0.565 -0.0669 -0.634 | 4 Petal.Width 1.20 0.762Additional annotations can be added using several row- and column-specific dplyr-style verbs:
iris_meta <- data.frame(
Species = c("setosa", "versicolor", "virginica"),
Colony = c(1L, 1L, 2L),
Cytotype = c("diploid", "hexaploid", "tetraploid"),
Ploidy = c(2L, 6L, 4L)
)
(iris_pca <- left_join_rows(iris_pca, iris_meta, by = "Species"))
#> # A tbl_ord of class 'prcomp': (150 x 4) x (4 x 4)'
#> # 4 coordinates: PC1, PC2, ..., PC4
#> #
#> # Rows (principal): [ 150 x 4 | 4 ]
#> PC1 PC2 PC3 ... | Species Colony Cytotype Ploidy
#> | <chr> <int> <chr> <int>
#> 1 -2.26 -0.478 0.127 | 1 setosa 1 diploid 2
#> 2 -2.07 0.672 0.234 ... | 2 setosa 1 diploid 2
#> 3 -2.36 0.341 -0.0441 | 3 setosa 1 diploid 2
#> 4 -2.29 0.595 -0.0910 | 4 setosa 1 diploid 2
#> 5 -2.38 -0.645 -0.0157 | 5 setosa 1 diploid 2
#> # … with 145 more rows
#> #
#> # Columns (standard): [ 4 x 4 | 3 ]
#> PC1 PC2 PC3 ... | name center scale
#> | <chr> <dbl> <dbl>
#> 1 0.521 -0.377 0.720 | 1 Sepal.Length 5.84 0.828
#> 2 -0.269 -0.923 -0.244 ... | 2 Sepal.Width 3.06 0.436
#> 3 0.580 -0.0245 -0.142 | 3 Petal.Length 3.76 1.77
#> 4 0.565 -0.0669 -0.634 | 4 Petal.Width 1.20 0.762Following the broom
package, the tidy() method produces a tibble describing the
model components, in this case the principal coordinates, which is
suitable for scree plotting:
tidy(iris_pca) %T>% print() %>%
ggplot(aes(x = name, y = prop_var)) +
geom_col() +
labs(x = "", y = "Proportion of inertia") +
ggtitle("PCA of Anderson's iris measurements",
"Distribution of inertia")
#> # A tibble: 4 × 5
#> name sdev inertia prop_var quality
#> <fct> <dbl> <dbl> <dbl> <dbl>
#> 1 PC1 1.71 435. 0.730 0.730
#> 2 PC2 0.956 136. 0.229 0.958
#> 3 PC3 0.383 21.9 0.0367 0.995
#> 4 PC4 0.144 3.09 0.00518 1
Following ggplot2, the fortify() method
row-binds the factor tibbles with an additional .matrix
column. This is used by ggbiplot() to redirect row- and
column-specific plot layers to the appropriate subsets:8
ggbiplot(iris_pca, sec.axes = "cols", scale.factor = 2) +
geom_rows_point(aes(color = Species, shape = Species)) +
stat_rows_ellipse(aes(color = Species), alpha = .5, level = .99) +
geom_cols_vector() +
geom_cols_text_radiate(aes(label = name)) +
expand_limits(y = c(-3.5, NA)) +
ggtitle("PCA of Anderson's iris measurements",
"99% confidence ellipses; variables use top & right axes")
When variables are represented in standard coordinates, as typically in PCA, their rules can be rescaled to yield a predictive biplot:9
ggbiplot(iris_pca, axis.type = "predictive", axis.percents = FALSE) +
theme_biplot() +
geom_rows_point(aes(color = Species, shape = Species)) +
stat_rows_center(
aes(color = Species, shape = Species),
size = 5, alpha = .5, fun.data = mean_se
) +
geom_cols_axis(aes(label = name, center = center, scale = scale)) +
ggtitle("Predictive biplot of Anderson's iris measurements",
"Project a marker onto an axis to approximate its measurement")
aggregate(iris[, 1:4], by = iris[, "Species", drop = FALSE], FUN = mean)
#> Species Sepal.Length Sepal.Width Petal.Length Petal.Width
#> 1 setosa 5.006 3.428 1.462 0.246
#> 2 versicolor 5.936 2.770 4.260 1.326
#> 3 virginica 6.588 2.974 5.552 2.026The auxiliary package ordr.extra provides recovery methods for several additional ordination models—and has room for several more!
Any feedback on the package is very welcome! If you encounter
confusion or errors, do create an issue, with a minimal
reproducible example if feasible. If you have requests, suggestions,
or your own implementations for new features, feel free to create an
issue or submit a pull request. Methods for additional ordination
classes (see the methods-*.r scripts in the R
folder) are especially welcome, as are new plot layers. Please try to
follow the contributing
guidelines and respect the Code
of Conduct.
This package was originally inspired by the ggbiplot extension developed by Vincent Q. Vu, Richard J Telford, and Vilmantas Gegzna, among others. It probably first brought biplots into the tidyverse framework. The motivation to unify a variety of ordination methods came from several books and articles by Michael Greenacre, in particular Biplots in Practice. Several answers at CrossValidated, in particular by amoeba and ttnphns, provided theoretical insights and informed design choices. Thomas Lin Pedersen’s tidygraph prequel to ggraph finally induced the shift from the downstream generation of scatterplots to the upstream handling and manipulating of ordination models. Additional design elements and features have been informed by the monograph Biplots and the textbook Understanding Biplots by John C. Gower, David J. Hand, Sugnet Gardner–Lubbe, and Niel J. Le Roux, and by the volume Principal Components Analysis by I. T. Jolliffe.
Greenacre MJ (2010) Biplots in Practice. Fundacion BBVA, ISBN: 978-84-923846. https://www.fbbva.es/microsite/multivariate-statistics/biplots.html↩︎
The term ordination is most prevalent among ecologists; no catch-all term seems to be in common use outside ecology.↩︎
This is not a hard rule: PCA is often used to compress data before clustering, and LDA uses dimension reduction to perform classification tasks.↩︎
Regression and clustering models,
like classical linear
regression and k-means,
can also be understood as matrix decomposition approximations and even
visualized in biplots. Their shared coordinates, which are pre-defined
rather than artificial, are the predictor coefficients and the cluster
assignments, respectively. Methods for stats::lm() and
stats::kmeans(), for example, are implemented for the sake
of novelty and instruction, but are not widely used in practice.↩︎
Biplot elments must be chosen with care, and it is useful and appropriate that many model-specific biplot methods have limited flexibility. This package adopts the trade-off articulated in Wilkinson’s The Grammar of Graphics (p. 15): “This system is capable of producing some hideous graphics. There is nothing in its design to prevent its misuse. … This system cannot produce a meaningless graphic, however.”↩︎
Anderson E (1936) “The Species Problem in Iris”. Annals of the Missouri Botanical Garden 23(3), 457-469+471-483+485-501+503-509. https://doi.org/10.2307/2394164↩︎
The data must be in the form of a data frame that can be understood by the modeling function. Step-by-step methods also exist to build and annotate a ‘tbl_ord’ from a fitted ordination model.↩︎
The radiating text geom, like several other features, is adapted from the ggbiplot package.↩︎
This is an experimental feature only available for linear methods, namely eigendecomposition, singular value decomposition, and principal components analysis.↩︎