ANOVA problems with revoScaleR::rxGlm() in R

Question

I build lots of GLMs. Usually on large data sets with many model parameters. This means that base R's glm() function isn't really useful because it won't cope with the size/complexity, so I usually use revoScaleR::rxGlm() instead.

However I'd like to be able to do ANOVA tests on pairs of nested models, and I haven't found a way to do this with the model objects that rxGlm() creates, because R's anova() function won't work with them. revoScaleR provides an as.glm() function which converts an rxGlm() object to a glm() object - sort of - but it doesn't work here.

For example:

library(dplyr)

data(mtcars)

# don't like having named rows

mtcars <- mtcars %>% 
  mutate(veh_name = rownames(.)) %>%
  select(veh_name, everything())

# fit a GLM: mpg ~ everything else

glm_a1 <- glm(mpg ~ cyl + disp + hp + drat + wt + qsec + vs + am + gear + carb,
             data = mtcars,
             family = gaussian(link = "identity"),
             trace = TRUE)

summary(glm_a1)

# fit another GLM where gear is removed

glm_a2 <- glm(mpg ~ cyl + disp + hp + drat + wt + qsec + vs + am + carb,
             data = mtcars,
             family = gaussian(link = "identity"),
             trace = TRUE)

summary(glm_a2)

# F test on difference

anova(glm_a1, glm_a2, test = "F")

works fine, but if instead I do:

library(dplyr)

data(mtcars)

# don't like having named rows

mtcars <- mtcars %>% 
  mutate(veh_name = rownames(.)) %>%
  select(veh_name, everything())

glm_b1 <- rxGlm(mpg ~ cyl + disp + hp + drat + wt + qsec + vs + am + gear + carb,
               data = mtcars,
               family = gaussian(link = "identity"),
               verbose = 1)

summary(glm_b1)

# fit another GLM where gear is removed

glm_b2 <- rxGlm(mpg ~ cyl + disp + hp + drat + wt + qsec + vs + am + carb,
               data = mtcars,
               family = gaussian(link = "identity"),
               verbose = 1)

summary(glm_b2)

# F test on difference

anova(as.glm(glm_b1), as.glm(glm_b2), test = "F")

I see the error message:

Error in qr.lm(object) : lm object does not have a proper 'qr'
component. Rank zero or should not have used lm(.., qr=FALSE)

The same problem cropped up on a previous SO posting: Error converting rxGlm to GLM but doesn't seem to have been solved.

Can anyone help please? if as.glm() isn't going to help here, is there some other way? Could I write a custom function to do this (stretching my coding abilities to their limit I suspect?)?

Also, is SO the best forum, or would one of the other StackExchange forums be a better place to look for guidance?

Thank you.

Answer 1

Partial solution...

my_anova <- function (model_1, model_2, test_type)
{
  
  # only applies for nested GLMs. How do I test for this?
  
  cat("\n")
  
  if(test_type != "F")
  {
    cat("Invalid function call")
  }
  else
  {
    # display model formulae
    
    cat("Model 1:", format(glm_b1$formula), "\n")
    cat("Model 2:", format(glm_b2$formula), "\n")
    
    if(test_type == "F") 
    {
      
      if (model_1$df[2] < model_2$df[2]) # model 1 is big, model 2 is small
        
      {
        dev_s <- model_2$deviance
        df_s  <- model_2$df[2]
        dev_b <- model_1$deviance
        df_b  <- model_1$df[2]
      }
      else # model 2 is big, model 1 is small
      {
        dev_s <- model_1$deviance
        df_s  <- model_1$df[2]
        dev_b <- model_2$deviance
        df_b  <- model_2$df[2]
      }
      
      F <- (dev_s - dev_b) / ((df_s - df_b) * dev_b / df_b)
      
    }
    
    # still need to calculate the F tail probability however
    
    # df of F: numerator: df_s - df_b
    # df of F: denominator: df_b
    
    F_test <- pf(F, df_s - df_b, df_b, lower.tail = FALSE)
    
    cat("\n")
    cat("F:     ", round(F, 4), "\n")
    cat("Pr(>F):", round(F_test, 4))
  }
}