Knowing when a number is good enough

Analysis can always be refined. More data, more controls, more scenarios, a better method — each improves the estimate and each costs time. Knowing when to stop is a skill and it is rarely taught.

The stopping rule comes from the decision rather than from the analysis.

What is the stopping rule?

That further precision would not change what is done. If the decision is the same across the plausible range of the estimate, the estimate is good enough and additional work is spent for comfort.

That is an easy test to apply and it requires the decision to be named, which is why it is so often skipped.

How do you apply it in practice?

Take the current estimate and its range, and check the decision at both ends. If both ends point the same way, stop. If they point different ways, the remaining uncertainty is decision-relevant and more work is justified.

The same test tells you what kind of work is needed: narrowing the range that matters rather than refining everything.

What are the costs of over-refinement?

Time, the delay to the decision, and a false impression of certainty. An analysis refined for three weeks arrives with more precision and an implicit claim to authority that the underlying data may not support.

There is also an opportunity cost that is rarely counted. The analyst refining a sufficient answer is not answering the next question.

What are the costs of stopping too early?

A decision made on an estimate whose range crosses the threshold, which is a decision made effectively at random. That is the failure the stopping rule is designed to prevent, and it is why the rule is about the decision rather than about effort.

How does this interact with presentation?

A good-enough estimate should be presented as such, with its range. Presenting a deliberately rough answer with precise-looking figures invites it to be treated as more than it is, and it will then be quoted in contexts the roughness cannot support.

Rounding appropriately is part of this. A figure quoted to several significant digits from an estimate with a wide range is making a claim the analysis does not.

What about numbers that will be reused?

They need more care, because a figure produced for one decision has a habit of being quoted in another where the tolerance is different. Anything likely to circulate deserves the range and the caveat attached to the number itself.

What should an analyst say when asked for more precision?

What it would cost and what it would change. If the answer is that it would not change the decision, that is worth saying plainly, and it is frequently accepted.