[baddox]

Ideally, by recording (as best you can, and estimating after that) the cost per unit time of using a non-standard keyboard layout (including the cost of learning it and the cost of occasionally needing to use computers that don’t use it) and the benefit per unit time of using that layout.

Either the benefit per unit time eventually surpasses the cost per unit time (since some costs are fixed, namely the initial learning), or it doesn’t. Only switch if benefit surpasses cost in a time window you’re comfortable with.