Avoid estimation for a model in tidymodels

no_estimation.R

library(tidymodels)

actual_fit <- glm(mpg ~ ., data = mtcars)

not_fit <- glm(
  mpg ~ .,
  data = mtcars,
  start = coef(actual_fit),
  control = glm.control(maxit = 1, trace = TRUE)
)

# Can also trick tidymodels if we can initialize a model known to
no_parsnip_fit <-
  linear_reg() |>
  set_engine(
    "glm",
    start = coef(actual_fit),
    control = glm.control(maxit = 1)
  ) |>
  fit(mpg ~ ., data = mtcars)

waldo::compare(
  coef(actual_fit),
  no_parsnip_fit |> extract_fit_engine() |> coef()
)

This doesn't seem to work if using other (non-estimated) coefficients:

library(tidymodels)

fixed_coef <- c("(Intercept)" = 30, cyl = -1.5, disp = -0.02)

not_fit <- glm(
  mpg ~ cyl + disp,
  data = mtcars,
  start = fixed_coef,
  control = glm.control(maxit = 1, trace = TRUE)
)
#> Deviance = 270.7403 Iterations - 1
#> Warning: glm.fit: algorithm did not converge

# Can also trick tidymodels if we can initialize a model known to
no_parsnip_fit <-
  linear_reg() |>
  set_engine(
    "glm",
    start = fixed_coef,
    control = glm.control(maxit = 1)
  ) |>
  fit(mpg ~ cyl + disp, data = mtcars)
#> Warning: glm.fit: algorithm did not converge

waldo::compare(
  fixed_coef,
  not_fit |> coef()
)
#> `old`: 30.00000 -1.50000 -0.02000
#> `new`: 34.66099 -1.58728 -0.02058

waldo::compare(
  fixed_coef,
  no_parsnip_fit |> extract_fit_engine() |> coef()
)
#> `old`: 30.00000 -1.50000 -0.02000
#> `new`: 34.66099 -1.58728 -0.02058

^{Created on 2025-11-06 with reprex v2.1.1}