Principal Component Analysis (PCA) and Factor Analysis (FA)

Source: R/factor_analysis.R, R/principal_components.R, R/utils_pca_efa.R

The functions principal_components() and factor_analysis() can be used to perform a principal component analysis (PCA) or a factor analysis (FA). They return the loadings as a data frame, and various methods and functions are available to access / display other information (see the Details section).

factor_analysis(
  x,
  n = "auto",
  rotation = "none",
  sort = FALSE,
  threshold = NULL,
  standardize = TRUE,
  cor = NULL,
  ...
)

principal_components(
  x,
  n = "auto",
  rotation = "none",
  sort = FALSE,
  threshold = NULL,
  standardize = TRUE,
  ...
)

closest_component(pca_results)

rotated_data(pca_results)

# S3 method for parameters_efa
predict(object, newdata = NULL, names = NULL, keep_na = TRUE, ...)

# S3 method for parameters_efa
print(x, digits = 2, sort = FALSE, threshold = NULL, labels = NULL, ...)

# S3 method for parameters_efa
sort(x, ...)

Arguments

x

A data frame or a statistical model.
n

Number of components to extract. If n="all", then n is set as the number of variables minus 1 (ncol(x)-1). If n="auto" (default) or n=NULL, the number of components is selected through n_factors() resp. n_components(). In reduce_parameters(), can also be "max", in which case it will select all the components that are maximally pseudo-loaded (i.e., correlated) by at least one variable.
rotation

If not "none", the PCA / FA will be computed using the psych package. Possible options include "varimax", "quartimax", "promax", "oblimin", "simplimax", or "cluster" (and more). See psych::fa() for details.
sort

Sort the loadings.
threshold

A value between 0 and 1 indicates which (absolute) values from the loadings should be removed. An integer higher than 1 indicates the n strongest loadings to retain. Can also be "max", in which case it will only display the maximum loading per variable (the most simple structure).
standardize

A logical value indicating whether the variables should be standardized (centered and scaled) to have unit variance before the analysis (in general, such scaling is advisable).
cor

An optional correlation matrix that can be used (note that the data must still be passed as the first argument). If NULL, will compute it by running cor() on the passed data.
...

Arguments passed to or from other methods.
pca_results

The output of the principal_components() function.
object

An object of class parameters_pca or parameters_efa
newdata

An optional data frame in which to look for variables with which to predict. If omitted, the fitted values are used.
names

Optional character vector to name columns of the returned data frame.
keep_na

Logical, if TRUE, predictions also return observations with missing values from the original data, hence the number of rows of predicted data and original data is equal.
digits, labels

Arguments for print().

Value

A data frame of loadings.

Details

Methods and Utilities

  • n_components() and n_factors() automatically estimates the optimal number of dimensions to retain.

  • check_factorstructure() checks the suitability of the data for factor analysis using the sphericity() and the sphericity() KMO measure.

  • performance::check_itemscale() computes various measures of internal consistencies applied to the (sub)scales (i.e., components) extracted from the PCA.

  • Running summary returns information related to each component/factor, such as the explained variance and the Eivenvalues.

  • Running get_scores() computes scores for each subscale.

  • Running closest_component() will return a numeric vector with the assigned component index for each column from the original data frame.

  • Running rotated_data() will return the rotated data, including missing values, so it matches the original data frame.

  • Running plot() visually displays the loadings (that requires the see package to work).

Complexity

Complexity represents the number of latent components needed to account for the observed variables. Whereas a perfect simple structure solution has a complexity of 1 in that each item would only load on one factor, a solution with evenly distributed items has a complexity greater than 1 (Hofman, 1978; Pettersson and Turkheimer, 2010) .

Uniqueness

Uniqueness represents the variance that is 'unique' to the variable and not shared with other variables. It is equal to 1 – communality (variance that is shared with other variables). A uniqueness of 0.20 suggests that 20% or that variable's variance is not shared with other variables in the overall factor model. The greater 'uniqueness' the lower the relevance of the variable in the factor model.

MSA

MSA represents the Kaiser-Meyer-Olkin Measure of Sampling Adequacy (Kaiser and Rice, 1974) for each item. It indicates whether there is enough data for each factor give reliable results for the PCA. The value should be > 0.6, and desirable values are > 0.8 (Tabachnick and Fidell, 2013).

PCA or FA?

There is a simplified rule of thumb that may help do decide whether to run a factor analysis or a principal component analysis:

  • Run factor analysis if you assume or wish to test a theoretical model of latent factors causing observed variables.

  • Run principal component analysis If you want to simply reduce your correlated observed variables to a smaller set of important independent composite variables.

(Source: CrossValidated)

Computing Item Scores

Use get_scores() to compute scores for the "subscales" represented by the extracted principal components. get_scores() takes the results from principal_components() and extracts the variables for each component found by the PCA. Then, for each of these "subscales", raw means are calculated (which equals adding up the single items and dividing by the number of items). This results in a sum score for each component from the PCA, which is on the same scale as the original, single items that were used to compute the PCA. One can also use predict() to back-predict scores for each component, to which one can provide newdata or a vector of names for the components.

Explained Variance and Eingenvalues

Use summary() to get the Eigenvalues and the explained variance for each extracted component. The eigenvectors and eigenvalues represent the "core" of a PCA: The eigenvectors (the principal components) determine the directions of the new feature space, and the eigenvalues determine their magnitude. In other words, the eigenvalues explain the variance of the data along the new feature axes.

References

  • Kaiser, H.F. and Rice. J. (1974). Little jiffy, mark iv. Educational and Psychological Measurement, 34(1):111–117

  • Hofmann, R. (1978). Complexity and simplicity as objective indices descriptive of factor solutions. Multivariate Behavioral Research, 13:2, 247-250, doi: 10.1207/s15327906mbr1302_9

  • Pettersson, E., & Turkheimer, E. (2010). Item selection, evaluation, and simple structure in personality data. Journal of research in personality, 44(4), 407-420, doi: 10.1016/j.jrp.2010.03.002

  • Tabachnick, B. G., and Fidell, L. S. (2013). Using multivariate statistics (6th ed.). Boston: Pearson Education.

Examples


library(parameters)

# \donttest{
# Principal Component Analysis (PCA) -------------------
if (require("psych")) {
  principal_components(mtcars[, 1:7], n = "all", threshold = 0.2)
  principal_components(mtcars[, 1:7],
    n = 2, rotation = "oblimin",
    threshold = "max", sort = TRUE
  )
  principal_components(mtcars[, 1:7], n = 2, threshold = 2, sort = TRUE)

  pca <- principal_components(mtcars[, 1:5], n = 2, rotation = "varimax")
  pca # Print loadings
  summary(pca) # Print information about the factors
  predict(pca, names = c("Component1", "Component2")) # Back-predict scores

  # which variables from the original data belong to which extracted component?
  closest_component(pca)
  # rotated_data(pca)  # TODO: doesn't work
}
#>  mpg  cyl disp   hp drat 
#>    1    1    1    1    2 

# Automated number of components
principal_components(mtcars[, 1:4], n = "auto")
#> # Loadings from Principal Component Analysis (no rotation)
#> 
#> Variable |  PC1  | Complexity
#> -----------------------------
#> mpg      | -0.93 |    1.00   
#> cyl      | 0.96  |    1.00   
#> disp     | 0.95  |    1.00   
#> hp       | 0.91  |    1.00   
#> 
#> The unique principal component accounted for 87.55% of the total variance of the original data.
#> 
# }



# Factor Analysis (FA) ------------------------
if (require("psych")) {
  factor_analysis(mtcars[, 1:7], n = "all", threshold = 0.2)
  factor_analysis(mtcars[, 1:7], n = 2, rotation = "oblimin", threshold = "max", sort = TRUE)
  factor_analysis(mtcars[, 1:7], n = 2, threshold = 2, sort = TRUE)

  efa <- factor_analysis(mtcars[, 1:5], n = 2)
  summary(efa)
  predict(efa)
# \donttest{
  # Automated number of components
  factor_analysis(mtcars[, 1:4], n = "auto")
# }
}
#> Warning: Could not retrieve information about missing data. Returning only complete
#>   cases.
#> # Loadings from Factor Analysis (no rotation)
#> 
#> Variable |  MR1  | Complexity | Uniqueness
#> ------------------------------------------
#> mpg      | -0.90 |    1.00    |    0.19   
#> cyl      | 0.96  |    1.00    |    0.08   
#> disp     | 0.93  |    1.00    |    0.13   
#> hp       | 0.86  |    1.00    |    0.26   
#> 
#> The unique latent factor accounted for 83.55% of the total variance of the original data.
#>