Unquoted `weights` argument

[kylebutts]

Okay, here's a bit of a strange one. Trying to use fixest::feglm with ggplot2 for estimation (so I can fit an exponential curve and have standard errors on the fitted values).
But internally, they pass the argument weights = weight where weight is a vector in the data argument.
I.e. non-standard evaluation.
This creates problems since this errors in feols.
The simplest solution would be to patch fixest to allow for NSE for weights argument.

I was trying to patch this using method.args = list(family = "poisson", weights = NULL) but I get an error: ! formal argument "weights" matched by multiple actual arguments.
@teunbrand is this something that could be addressed on the ggplot2 side?

library(fixest)
library(ggplot2)
base <- mtcars
base$w <- runif(nrow(base), 0.5, 1.5)

lm(mpg ~ hp, base, weights = w)
#> 
#> Call:
#> lm(formula = mpg ~ hp, data = base, weights = w)
#> 
#> Coefficients:
#> (Intercept)           hp  
#>    29.66891     -0.06503
feols(mpg ~ hp, base, weights = w)
#> Error in feols(mpg ~ hp, base, weights = w): Argument `weights` could not be evaluated: object 'w' not found

ggplot(aes(x = hp, y = mpg), data = mtcars) + 
  geom_point() + 
  geom_smooth(method = "lm")
#> `geom_smooth()` using formula = 'y ~ x'

ggplot(aes(x = hp, y = mpg), data = mtcars) + 
  geom_point() + 
  geom_smooth(method = fixest::feols)
#> `geom_smooth()` using formula = 'y ~ x'
#> Warning: Failed to fit group -1.
#> Caused by error in `method()`:
#> ! Argument `weights` could not be evaluated: object 'weight' not found

[lrberge]

Hi Kyle, adding NSE to fixest is a no go, this is a design choice. It would be a breaking change and be unaligned with other options.

From ggplot2 code I see (https://github.com/tidyverse/ggplot2/blob/6a447ba747f14942c2ff3ea00a1e4d74f23ec1f2/R/stat-smooth.R#L82):

    if (is.null(data$weight)) data$weight <- 1

    # [...]

    prediction <- try_fetch(
      {
        model <- inject(method(
          formula,
          data = data,
          weights = weight,
          !!!method.args
        ))
        predictdf(model, xseq, se, level)
      },

Maybe using a "real vector" with a strange name (that wouldn't be caught in NSE) could solve the problem:

    weights__internal__ <- data[["weight"]]
    if (is.null(weights__internal__)) weights__internal__ <- 1

    # [...]

    prediction <- try_fetch(
      {
        model <- inject(method(
          formula,
          data = data,
          weights = weights__internal__,
          !!!method.args
        ))
        predictdf(model, xseq, se, level)
      },

Given that the weight is only used locally within this function, it could work.

[teunbrand]

@kylebutts Is this is a new problem since ggplot2 4.0.0 or a longer standing one?
Generally speaking, when geom_smooth() doesn't suffice for a paricular model, our recommendation is to build the model outside of ggplot2 and visualise the results from scratch.

[lrberge]

OK, I just checked and the fix I propose does not work since it leads to another problem down the road. Maybe it's just incompatible.

[kylebutts]

@teunbrand @lrberge this could be fixed with a very simple patch. Changing https://github.com/tidyverse/ggplot2/blob/6a447ba747f14942c2ff3ea00a1e4d74f23ec1f2/R/stat-smooth.R#L106 to data$weight works. I think this is a better default anyways since it passes the actual vector instead of an unquoted name.

Reprex:

library(ggplot2) 
library(fixest)
predictdf.fixest <- function(model, xseq, se, level) {
  pred <- stats::predict(
    model,
    newdata = data_frame0(x = xseq),
    se.fit = se,
    level = level,
    interval = if (se) "confidence" else "none"
  )

  if (se) {
    fit <- pred
    names(fit) <- c("y", "se", "ymin", "ymax")
    base::data.frame(x = xseq, fit)
  } else {
    base::data.frame(x = xseq, y = as.vector(pred))
  }
}

ggplot(aes(x = hp, y = mpg), data = mtcars) +
  geom_point() +
  geom_smooth(
    method = "lm",
    se = FALSE
  )
ggplot(aes(x = hp, y = mpg), data = mtcars) +
  geom_point() +
  geom_smooth(
    method = "glm",
    method.args = list(family = "poisson"),
    se = FALSE
  )
ggplot(aes(x = hp, y = mpg), data = mtcars) +
  geom_point() +
  geom_smooth(
    method = "loess",
    se = FALSE
  )
ggplot(aes(x = hp, y = mpg), data = mtcars) +
  geom_point() +
  geom_smooth(
    method = "gam",
    se = FALSE
  )

ggplot(aes(x = hp, y = mpg), data = mtcars) +
  geom_point() +
  geom_smooth(
    method = fixest::feols, 
    se = FALSE
  )

ggplot(aes(x = hp, y = mpg), data = mtcars) +
  geom_point() + 
  geom_smooth(
    method = fixest::feglm, 
    method.args = list(family = "poisson"),
    se = FALSE
  )

[teunbrand]

We'd be happy to consider a PR on this. If that is inconvenient for whatever reason, you can redirect this issue to ggplot2 so it doesn't fall off the radar.