| Generally, using one's brain. This is especially important when metrics would not be expected to be available -- for example, if you're designing a nuclear reactor, you need to think hard about ways to prevent a meltdown in advance, rather than collecting meltdown statistics and then fixing the piping problems that correlated with the most nuclear meltdowns. This is also necessary when the true metric that matters is very hard to evaluate counterfactually. For example, perhaps your real task is "maximize profit for the company", but you can't actually evaluate how your actions have influenced that metric, even though you can see the number going up and down. And necessary as well when a goal is too abstract to directly capture by metrics, resulting in bad surrogate metrics: for example, "improve user experience" is hard to measure directly, so "increase time spent interacting with website" might be measured as a substitute, with predictable outcomes that bad UI design can force users to waste more time on a page trying to find what they came for. All of these problems are faced by metric designers, who need to pick directly-measurable metric B (UX design metric) in order to maximize metric A (long-term profits) that the shareholders actually care about, but they cannot evaluate the quality of their own metrics by a metric, for the same reason that they were not using metric A directly to begin with. (See also the McNamara fallacy, which parent comment is a splendid example of: https://en.m.wikipedia.org/wiki/McNamara_fallacy ) |