Dplyr: mutate_all & rowwise

[Enuma]

Suppose you have an EEG dataset where every column is an electrode E1....E129.
First, you want to compute the spatial standard deviation GFP. How would you rewrite rowSD using dplyr rowwise? Next you want to normalize each electrode by dividing it with GFP. What goes into the mutate_all? Thanks!

library(dplyr)
library(RFunctionsSN)
tibble(E1 = rnorm(10), E2 = rnorm(10), E3 = rnorm(10), E4 = rnorm(10)) %>% 
  mutate(GFP = rowSD(.)) %>% 
  mutate_all()

[technocrat]

Is this the rowSD() function from https://github.com/snandi/RFunctionsSN/blob/master/R/rowSD.R

A reproducible example, called a reprex is really helpful.

And is GFP global field power, or what?

[mattwarkentin]

Hi @Enuma,

I'm not actually that familiar with rowwise, but here is one way to do it by pivot_longer(), doing your calculations, and then pivot_wider() to get back to the same original shape.

tibble(E1 = rnorm(10), E2 = rnorm(10), E3 = rnorm(10), E4 = rnorm(10)) %>% 
  mutate(id = 1:n()) %>% 
  pivot_longer(E1:E4) %>% 
  group_by(id) %>% 
  mutate(row_sd = sd(value),
         normalized = value / row_sd) %>% 
  pivot_wider(names_from = name, values_from = normalized, id_cols = id)

[Enuma]

Sorry, I thought it is exhaustive enough. Yes, rowwSD is from RFunctionsSN, that's why I call it at the beginning. GFP is global field power, yes.

[Enuma]

Thanks, this is also good. I wanted to learn rowwise on this particular example, because I struggled with other online examples.

[mattwarkentin]

Okay, here is another more row-wise approach. Let me know if this is more in line with what you hoped, @Enuma.

tibble(E1 = rnorm(10), E2 = rnorm(10), E3 = rnorm(10), E4 = rnorm(10)) %>% 
  mutate(row_sd = pmap_dbl(., lift_vd(sd))) %>% 
  mutate_at(vars(E1:E4), list(norm = ~ . / row_sd))

Or if you want to remove the original E1:E4 variables:

tibble(E1 = rnorm(10), E2 = rnorm(10), E3 = rnorm(10), E4 = rnorm(10)) %>% 
  mutate(row_sd = pmap_dbl(., lift_vd(sd))) %>% 
  transmute_at(vars(E1:E4), list(norm = ~ . / row_sd))

[technocrat]

The reason I asked is that RFunctionSN isn't a package, as such, just a collection of functions, which makes them hard to track down.

For an N x 19 data frame, rowwise operations become extremely inefficient as N increases. Converted to a matrix object, applying a function to rows is much faster and dividing all terms in a matrix also. Converting back to a tibble is possible with enframe, although you'll probably need to add back the original colnames

[Enuma]

Thanks, that is a useful info.

[Enuma]

Nice, seems quite elegant! Thank you.