GForce optimize weighted.mean

[renkun-ken]

I'm not sure if it is possible to GForce optimize weighted.mean? It's quite different from existing GForce functions because it takes two variables.

[MichaelChirico]

I think it should be doable. Several other g functions are taking several arguments (though none taking several vector arguments...)

[renkun-ken]

Yes, that's exactly what I was trying to say. It'd be great if it can be supported since I use a lot, one weighted.mean in the call makes everything broken, sometimes I have to single out this and merge later.

[myoung3]

+1 for this. The current workaround is to first take a product and assign it a new column, then doing the sum in a subsequent by-group call.

[ben-schwen]

I took a look into this and we have basically two options.

1. Write everything as a new gforce function
2. Implement it like this gweighted.mean = function(x, w, ..., na.rm=FALSE) { if (is.null(w)) gmean(x, na.rm) else gsum(x*w, na.rm) / gsum(w, na.rm)}

Advantages / Disadvantages

Option 2 is easy maintenance since it dispatches everything onto gsum/gmean but is less verbose and might return confusing error messages/warnings.

Benchmarking both, it turns out that option 2 is also faster than option 1 (at least with a naive implementation similar to gvar/gsd) since gsum is heavily optimized. However, option 1 is numerically more stable.