I think doing this genre of modelling work could be really valuable. After my PhD I applied for a job with GiveWell because I was excited about this (but didn't make it past their first round)

My POV is that scientific best practises on modelling are worth adhering to. And that it'd be possible to do the types of sensitivity and assumption checking analysis GiveWell that values more effectively using best practise statistical methods (e.g. Bayesian workflow)

There has been some work aimed to help spreadsheet users model uncertainty e.g. Tom Adamczewski wrote carlo.app with this in mind

My understanding of this is:

* The domain is characterised by many difficult decisions about assumptions
* GiveWell staff and leadership feel more comfortable working and communicating using spreadsheets

GiveWell have supported development of Bayesian software, though in practise I don't think they are using it very much: www.givewell.org/research/inc... github.com/wwiecek/baggr

Noah Haber wrote a piece for their "change our mind" contest about it (and won). But I'm not sure it changed practise: www.metacausal.com/givewells-un...

I've also looked into this and reached the a similar conclusion. Here are some additional links/thoughts:

My bad!