Error in `fit_xy()` with GAM model

[akrm1283]

I get the following message when trying to run a workflow.

Error in fit_xy():
! fit() must be used with GAM models (due to its use of formulas).

Here is the complete workflow; note that I call "fit()", not "fit_xy():

 gam_formula <- as.formula(str_c(dep_var, "~", treat_var, "+", paste(unlist(all_indep_vars), collapse = " + ") ))           
                        
 gam_recipe <- recipe(modl_df, formula = gam_formula) %>%
   update_role(all_of(!!ID_vars), new_role = "ID") %>%
   step_intercept() %>%
   step_naomit(!!dep_var) %>% 
   step_zv(all_predictors()) %>%
   step_nzv(all_predictors()) %>%
   step_dummy(all_nominal_predictors()) %>%
   step_ns(all_of(!!treat_var), deg_free = 2) 
 
 gam_model <- gen_additive_mod(select_features = TRUE, engine = "mgcv") %>%
   set_engine("mgcv") %>%
   set_mode("regression")
 
 gam_wkflo <- workflow() %>% 
   add_model(gam_model) %>%        
   add_recipe(gam_recipe) 
 
 response_gam <- gam_wkflo %>% fit(data = modl_df)    

Is "fit_xy()" a default process? Is there way to force "fit()" as the workflow processor?

[simoncouch]

Thanks for the post!

Can you please provide a minimal reprex (reproducible example)? A reprex will help us troubleshoot the issue more quickly + thoroughly.

[akrm1283]

library(tidyverse, quietly = TRUE)
library(dplyr, quietly = TRUE)
library(sf, quietly = TRUE)
#> Linking to GEOS 3.9.1, GDAL 3.4.3, PROJ 7.2.1; sf_use_s2() is TRUE
library(tidymodels, quietly = TRUE)
library(EnvStats, quietly = TRUE)
#> 
#> Attaching package: 'EnvStats'
#> The following objects are masked from 'package:stats':
#> 
#>     predict, predict.lm
#> The following object is masked from 'package:base':
#> 
#>     print.default


#### PARAMETERS
#.. set dependent variable, treatment variable & ID variables
dep_var <- "y"
treat_var <- "xt"
ID_vars <- c("sampleID", "replicateID")
#.. define sets of potential independent vars
  quant_indep <- c("X1", "X2", "X3")
  qual_indep <- c("N1", "N2")
all_indep_vars <- c(ID_vars, quant_indep, qual_indep)
####          end parameters        ####


####  create artificial dataset
txt <- c("A", "B", "C")
samp_ids <- c("sample1", "sample2")
repl_ids <- c("rep1", "rep2", "rep3", "rep4", "rep5")

modl_df <- data.frame(y = EnvStats::rnormTrunc(250, mean = 0, sd = 0.7, min = -1.5, max = 1.5),
                      sampleID = array(sample(samp_ids), 250),
                      replicateID = array(sample(repl_ids),250),
                      xt = rnorm(250, mean = 0, sd = 1),
                      x = matrix(sample(runif(5000, min = -0.5, max = 1.5), 750), 250, 3),
                      c = matrix(sample(txt, 500, replace = TRUE), 250, 2))
modl_df <- modl_df %>% dplyr::rename(X1 = x.1, X2 = x.2, X3 = x.3, N1 = c.1, N2 = c.2)


####  Workflow for GAM modeling
gam_formula <- as.formula(str_c(dep_var, "~", treat_var, "+", paste(unlist(all_indep_vars), collapse = " + ") ))           

gam_recipe <- recipe(modl_df, formula = gam_formula) %>%
  update_role(all_of(!!ID_vars), new_role = "ID") %>%
  step_intercept() %>%
  step_naomit(!!dep_var) %>% 
  step_zv(all_predictors()) %>%
  step_nzv(all_predictors()) %>%
  step_dummy(all_nominal_predictors()) %>%
  step_ns(all_of(!!treat_var), deg_free = 2) 

##** View outcome of applying 'gam_recipe' to data 'modl_df' via prep() and bake()
##*        prep_modl_df <- prep(gam_recipe, retain = TRUE, verbose = TRUE)
##*        bakd_modl_df <- bake(prep_modl_df, new_data = NULL)     # -OR- bakd_modl_df <- bake(prep_modl_df, modl_df)
##*        View(bakd_modl_df)
##**

gam_model <- gen_additive_mod(select_features = TRUE, engine = "mgcv") %>%
  set_engine("mgcv") %>%
  set_mode("regression")

gam_wkflo <- workflow() %>% 
  add_model(gam_model) %>%        
  add_recipe(gam_recipe) 


response_gam <- gam_wkflo %>% parsnip::fit(data = modl_df) 

#> Error in `fit_xy()`:
#> ! `fit()` must be used with GAM models (due to its use of formulas).
Created on 2022-07-26 by the reprex package (v2.0.1)

[simoncouch]

Ah, thanks for the additional context!

You'll need to change your line:

add_model(gam_model) %>%

to

add_model(gam_model, formula = gam_formula) %>%

The formula needs to be dropped in additionally at the add_model step in order to make sure we dispatch to the formula fit method first.

Another error is triggered related to the treat_var you define not being in the data, but the debugging should be a bit more straightforward for you from here.

There's currently an active issue that proposes either improving our documentation or interface here. We'll hopefully get to this sooner than later so others don't run into this hitch.