How to avoid/work around "Downdated VtV is not positive definite" error when using lme4::lmer() with purrr::map() or for loop?

Question

I have a function that runs a variable through a specific lme4::lmer() model syntax as such:

run_my_model <- function(i)
{
        
f <- formula(paste("y_variable ~", i, "+", i, "*interaction_variable + ", i, "*another_interaction_variable + (1|date) + (1|location)"))  

result <- tidy_lmer(lmer(f, data=my_dataset))
  
print(result)

}

When I run it on one specific variable in my dataset it works fine:

run_my_model("one_of_my_variables")

returns:

                  term      estimate   std.error          df   statistic   p.value
1            (Intercept)  1.977634e-04 0.001376695    24.21203  0.14365088 0.8869651
2            one_of_my_variables -2.805953e-03 0.001810723 69367.30600 -1.54963171 0.1212345
3            interaction_variableTRUE -8.648828e-04 0.001038360    20.06112 -0.83293141 0.4146903
4            another_interaction_variableTRUE  7.352286e-05 0.001539428    20.02418  0.04775987 0.9623808
5            one_of_my_variables:interaction_variableTRUE  -1.383274e-04 0.010860708 67714.86149 -0.01273650 0.9898381
6            sd_(Intercept).location 2.358986e-02          NA          NA          NA        NA
7            sd_(Intercept).date  2.098956e-03          NA          NA          NA        NA
8            sd_Observation.Residual  4.812015e-02          NA          NA          NA        NA
                 group
1                fixed
2                fixed
3                fixed
4                fixed
5                fixed
6                location
7                date
8                Residual

but when I try to use it with map() so it works on all variables in the dataset:

test <- map(names(my_dataset), run_my_model)

returns

 Error in eval_f(x, ...) : Downdated VtV is not positive definite

it also returns the same error if I use a for loop e.g.

models <- list()
        
for (i in names(my_dataset)) 
{
          
     f <- formula(paste("y_variable ~", i, "+", i, "*interaction_variable + ", i, "*another_interaction_variable + (1|date) + (1|location)"))  
    
    models[[i]] <- lmer(f, data=my_dataset)

}

I know that this model syntax works for some of the variables (as shown above) so clearly it is some specific ones that are generating this error. I tried looking at the traceback but it didn't provide any information of interest as to at which point the error first occurred. I have also tried nesting the lmer() call inside try() but the error still occurs.

How can I get the map() statement or loop to run successfully for the variables that do not generate this error message and just return NA or something similar for the ones that generate the error, without just writing out the syntax over and over again for each variable?

score 1 · Answer 1 · answered Aug 25 '20 at 13:30

You can try 2 things.

First, you can print the name before evaluating your function. The last variable that get printed will be the one with an error.

test <- map(names(my_dataset), ~{
  print(.x)
  run_my_model(.x)
})

Alternatively, you could use tryCatch to return the error directly:

test <- map(names(my_dataset), ~{
  tryCatch(run_my_model(.x), error=function(e) return(e))
})

You will end up with a list containing either your model or the error that your model threw.

If this doesn't help, I suggest you add a reproducible example to your question.

How to avoid/work around "Downdated VtV is not positive definite" error when using lme4::lmer() with purrr::map() or for loop?

1 Answers1