FYI, I discovered I’d been reading the model incorrectly (I’d had the covariate matrices transposed in my head) and there’s far less opportunity for this trick to make a difference than I thought, at least with the simple approach from that case study example. There’s still a lot of redundancy in that for a given covariate and country, there’s only a couple unique values being multiplied by that covariate’s alpha, and I’ve begun to think about approaches to eliminating that redundancy as well, but it’ll take me a while to work out something general/flexible.