In lwjohnst86/mason: Build Data Structures for Common Statistical Analysis

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)

This vignette goes over typical 'construction' projects for each of the analysis designs. First, let's load up mason!

library(mason)

T-tests

swiss %>% 
    design('t.test') %>% 
    add_settings() %>% 
    add_variables('yvars', c('Fertility', 'Agriculture')) %>% 
    add_variables('xvars', c('Examination', 'Education')) %>% 
    construct() %>% 
    scrub()

Correlations

swiss %>% 
    design('cor') %>% 
    add_settings() %>% 
    add_variables('yvars', c('Fertility', 'Agriculture')) %>% 
    add_variables('xvars', c('Examination', 'Education')) %>% 
    construct() %>% 
    scrub()

Linear regression

swiss %>% 
    design('glm') %>% 
    add_settings() %>% 
    add_variables('yvars', c('Fertility', 'Agriculture')) %>% 
    add_variables('xvars', c('Examination', 'Education')) %>% 
    add_variables('covariates', 'Catholic') %>% 
    construct() %>% 
    scrub()

Generalized estimating equations

data.frame(state.x77, state.region) %>% 
    design('gee') %>% 
    add_settings(cluster.id = 'state.region') %>% 
    add_variables('yvars', c('Income', 'Frost')) %>% 
    add_variables('xvars', c('Population', 'Murder')) %>% 
    add_variables('covariates', c('Life.Exp', 'Area')) %>% 
    add_variables('interaction', 'Area') %>% 
    construct() %>% 
    add_variables('xvars', c('Illiteracy')) %>%
    construct() %>%
    scrub()

Partial least squares (PLS) regression

PLS is implemented through several different algorithm, which are often fairly similar in their results. The default for mason is the default [pls::plsr()] kernel algorithm. Eventually mason will allow using other algorithms (like SIMPLS)... but baby steps. The kernel algorithm is good for "tall matrices", i.e. many observations and few variables (the type of data I often use).

Resources include this CrossValidated post and this website, both of which provide a lot of detail on interpreting the results of PLS. And of course there is the vignette from the pls package itself: vignette("pls-manual", package = "pls").

There are several outputs/results from PLS models:

Scores: Calculated value from each component from the PLS model for each observation. Use these to look for patterns between components.
Loadings: Very similar to principal component analysis loadings, including in their interpretation. Strongly correlated variables will have similar loadings. Loadings also include relationship with Y.
Coefficients: Are calculated by regressing the T scores on X. They are only useful for understanding the model, but not really for interpretation.
Y scores are used for mainly interpretation purposes, but the X scores are more important.

For more detail of what outputs are "scrubbed", see the [tidy_up()] function documentation. Below is an example usage:

pls_model <- swiss %>% 
    design('pls') %>% 
    add_settings(validation = 'CV', cv.data = TRUE) %>% 
    add_variables('yvars', c('Agriculture')) %>% 
    add_variables('xvars', c('Examination', 'Education', 'Fertility')) %>% 
    construct() 

scrub(pls_model) 
scrub(pls_model, "default") 
scrub(pls_model, "scores") 
scrub(pls_model, "loadings")