Ability to ignore "not enough finite observations" error

[ymer]

correlation calls cor.test.default which gives a not enough finite observation error when applicable. I think there could be a value in correlation() such as ignore_errors that sets those to NA instead.

[DominiqueMakowski]

Hi, would you have a reproducible example so that we can resolve that? Thanks!

[ymer]

library(tidyverse)
library(correlation)

tibble(v2 = c(2, 1, 1, 2), v3 = c(1,2, NA, NA)) %>% 
    correlation()

[DominiqueMakowski]

Yes, that's because you can't compute a correlation with only 2 observations (2 complete rows). If you add one row it works:

library(tidyverse)
#> Warning: package 'tibble' was built under R version 4.0.2
library(correlation)

tibble(v2 = c(2, 1, 1, 2, 2), v3 = c(1,2,  NA, NA, 3)) %>% 
  correlation()
#> Parameter1 | Parameter2 |    r |    t | df |      p |  Method | n_Obs
#> ---------------------------------------------------------------------
#> v2         |         v3 | 0.00 | 0.00 |  1 | > .999 | Pearson |     3

^{Created on 2020-09-22 by the reprex package (v0.3.0)}

[ymer]

I understand why it can't be calculated. I am interested in the ability to ignore it.

For example, say I have a table with 1000 elements and would like to calculate all correlations. It would be very easy to use correlation::correlation() for that. But if some pairs have too few observations, it gives this error, so the method can't be used.

[DominiqueMakowski]

I see! Indeed, it would be better to return NA with a warning than throw an error that stops the process. Let me see what I can do.