Add function for Weighted Effect Coding?

[mattansb]

Is this something we can add here?
(After I built the function, I saw it was implemented in the {wec} package....)

contr.wsum <- function(x, ref, ...) {
  x <- as.factor(x)
  lvls <- levels(x)
  n <- nlevels(x)
  
  if (!missing(ref)) {
    if (!ref %in% lvls) stop("")
    lvls <- c(setdiff(lvls, ref), ref)
    x <- factor(x, levels = lvls)
  } else {
    ref <- lvls[n]
  }
  
  M <- contr.sum(n)
  rownames(M) <- lvls
  
  tab <- proportions(table(x))
  M[ref,] <- -unname(tab[-n] / tab[n])
  M
}

contr.wsum(mtcars$cyl)
#>         [,1] [,2]
#> 4  1.0000000  0.0
#> 6  0.0000000  1.0
#> 8 -0.7857143 -0.5

# same as:  
wec::contr.wec(factor(mtcars$cyl), "8")
#>            4    6
#> 1  1.0000000  0.0
#> 2  0.0000000  1.0
#> 3 -0.7857143 -0.5

Usage:

mtcars$cyl_f <- factor(mtcars$cyl)
contrasts(mtcars$cyl_f) <- contr.wsum(mtcars$cyl_f)
m <- lm(mpg ~ cyl_f, mtcars)
coef(m)[1]
#> (Intercept) 
#>    20.09062
mean(mtcars$mpg)
#> [1] 20.09062

^{Created on 2023-02-08 with reprex v2.0.2}

[etiennebacher]

@easystats/core-team WDYT? (I have no idea what this is)

[bwiernik]

Don't contrasts need to sum to 0?or is this for something like different levels of dosage? What are these used for?

[bwiernik]

Oh, it's so that the parameters are deviations from the sample mean, rather than the grand mean of the group means

https://journal.r-project.org/archive/2017/RJ-2017-017/RJ-2017-017.pdf

I'm good with adding this, but we should be sure to include some of the language from the wec paper clearly describing what this is for and how it differs from unweighted effect coding

In effect coding (also known as deviation contrast or ANOVA coding), parameters represent the deviation of each category from the grand mean across all categories (i.e., the sum of arithmetic means in all categories divided by the number of categories). To achieve this, the sum of all parameters is constrained to 0. This implies that the possibly different numbers of observations in categories is not taken into account. In weighted effect coding, the parameters represent the deviation of each category from the sample mean, corresponding to a constraint in which the weighted sum of all parameters is equal to zero. The weights are equal to the number of observations per category.

[etiennebacher]

bump