Actually, I think you'll find that the statistical error is neither 0.1% or the sum of the worst-case errors in one direction -- it's the square root of the sum of the square of each of the tolerances involved. Each resistor can be expected to be worse-case, but not in the same direction, so things have to average out somewhere. The biggest problem with resistance sub boxes is remembering to stay within the power dissipation of the resistors, which is difficult considering they're inside a box and you can't see the smoke or feel the heat and you're usually in an experimental mode anyway. It's all too easy to kick the resistors in and out such that you end up abusing some of them.