All questions

Does 90% probability to be best mean the same as 90% confidence?

Asked by curtis_n on

Ccurtis_n

I support a marketing team that is used to seeing "significant at 95% confidence" in every results deck. Before Croct I would run their raw numbers through an online significance calculator and report the p-value.

Now the dashboard shows Probability to Be Best, and I want to be careful with terms before I present anything. Is a 90% Probability to Be Best equivalent to 90% confidence in the frequentist sense, or are these different quantities that happen to both be percentages? My stakeholders will absolutely conflate them if I am not precise

Was this helpful?

1 answer

Croct Bot
AI
Accepted answer

They are different quantities, so keep them separate in the deck. Croct's statistics are Bayesian. Probability to Be Best is the direct probability that a variant is the top performer given the data observed so far. A frequentist confidence level is a statement about the procedure, not about the variant, and it answers a two-sided question by default.

In practice, Probability to Be Best is closer to the one-tailed question stakeholders actually ask, which is "how likely is it that B beats A". That makes it easier to present honestly: "there is a 90% probability this variant is the best" is a literally true reading of the number, which you can never say about a confidence level.

Two things to include in your deck framing:

  1. The recommended-winner rule combines two checks: Probability to Be Best above 95% AND potential loss below 0.1%. A high PBB alone does not trigger a recommendation, because a variant can be probably-best while the downside of being wrong is still material.
  2. Bayesian results are readable at any time without a fixed horizon, but the in-progress thresholds (1000 visitors, 25 conversions, and 1 week per variant) still guard against early calls.

There is more depth on the metric itself in this thread on how Probability to Be Best is computed.

Was this helpful?
Still have questions?