cuBERT | Fast implementation of BERT inference | Machine Learning library

I thought 'x' was being defined in line 3?

Line 3 is lambda<-with(bc, x[which.max(y)]). It doesn't define x, it defines lambda. It does use x, which it looks for within the bc environment. If you're using boxcox() from the MASS package, bc should indeed include x and y components, so bc$x shouldn't give you the same error message. I'd expect an error about the replacement lengths. Because...

bc$x are the potential lambda values tried by boxcox - you're using the default seq(-2, 2, 1/10), and it would be an unlikely coincidence if your data had a multiple of 41 rows needed to not give an error when assigning 41 values to a new column.

Line 3 picks out the lambda value that maximizes the likelihood, so you shouldn't need the rest of the values in bc ever again. I'd expect you to use that lambda values to transform your response variable, as that's what the Box Cox transformation is for. ((x^lambda)-1/lambda) doesn't make any statistical or programmatic sense. Use this instead: