Combining Feature selection and mixed effects models: Error in eval(predvars, data, env): object ' ' not found

[AlbertoImg]

Combining Feature selection and mixed effects models: Error in eval(predvars, data, env): object ' ' not found

[hfrick]

Thanks for including an example! I can't run it, though, so this is a bit of a general reply: if your feature selection step changes the names of the predictors, e.g., through transformations, you can't use the original predictor name in the model formula (that's the one that you use in add_model()) because they will not be there anymore after the preprocessing.

You could use the dot notation in the formula, i.e., something similar to Disease ~ . + (1|ID) - ID. You'd need to ensure only variables you want to use in the model are left after preprocessing and remove the fixed effect for the ID variable. Here is an illustration of that idea

library(tidymodels)
library(multilevelmod)

data(sleepstudy, package = "lme4")
# we want to use the formula Reaction ~ Days + (1|Subject)

lmer_spec <- 
  linear_reg() %>% 
  set_engine("lmer")

# recipe here without any further preprocessing/feature engineering
# because the data already only contains the 3 variables we are going to use
rec <- recipe(Reaction ~ ., sleepstudy) 

wflow <- workflow() %>% 
  add_recipe(rec) %>% 
  add_model(lmer_spec, formula = Reaction ~ . -Subject + (1|Subject))

fit(wflow, data = sleepstudy)
#> ══ Workflow [trained] ══════════════════════════════════════════════════════════
#> Preprocessor: Recipe
#> Model: linear_reg()
#> 
#> ── Preprocessor ────────────────────────────────────────────────────────────────
#> 0 Recipe Steps
#> 
#> ── Model ───────────────────────────────────────────────────────────────────────
#> Linear mixed model fit by REML ['lmerMod']
#> Formula: Reaction ~ . - Subject + (1 | Subject)
#>    Data: data
#> REML criterion at convergence: 1786.465
#> Random effects:
#>  Groups   Name        Std.Dev.
#>  Subject  (Intercept) 37.12   
#>  Residual             30.99   
#> Number of obs: 180, groups:  Subject, 18
#> Fixed Effects:
#> (Intercept)         Days  
#>      251.41        10.47

# same fit as with Reaction ~ Days + (1|Subject)
lmer_spec %>% 
  fit(Reaction ~ Days + (1|Subject), data = sleepstudy)
#> parsnip model object
#> 
#> Linear mixed model fit by REML ['lmerMod']
#> Formula: Reaction ~ Days + (1 | Subject)
#>    Data: data
#> REML criterion at convergence: 1786.465
#> Random effects:
#>  Groups   Name        Std.Dev.
#>  Subject  (Intercept) 37.12   
#>  Residual             30.99   
#> Number of obs: 180, groups:  Subject, 18
#> Fixed Effects:
#> (Intercept)         Days  
#>      251.41        10.47

^{Created on 2024-01-24 with reprex v2.0.2}

[AlbertoImg]

Hi hfrick,
Thanks for your answer!
I tried it and the fit function is working with my dataset and workflow (as your example).
However, then when I use augment or predict I got the following error:
fitted_model <-
fit(
current_workflow,
data = data_training
)
prediction_fold <- fitted_model %>%
augment(new_data_fold)

Error in terms.formula(ff) : '.' in formula and no 'data' argument

I tried this but I get the same: fitted_model %>%
augment(data=data_training, new_data=new_data_fold)

Thanks again
Best
Alberto

[hfrick]

I can't help you with this one without a proper reprex. Please check out the reprex package for easily making those and the article on dos and donts. Github issues are best used for bug reports and feature requests; for general help in how to get a piece of code to run, Posit Community is the best place, also because more people see your question there and can chime in.

[AlbertoImg]

Thank a lot for your support! I will see if in Posit Community I can get some tips as well.
Best
Alberto