bayesplain.teach.precision_planning¶
-
bayesplain.teach.precision_planning(target_width, expected_rate=
0.5, level=0.95, prior='uninformed', max_n=1000000)[source]¶ Find the sample size that makes an interval narrow enough to act on.
Computed exactly rather than through a normal approximation: for each candidate
nthe actual Beta posterior interval is measured, and the search stops at the firstnthat meets the target.- Parameters:¶
- target_width : float¶
The widest interval you could still act on, on the rate’s own scale.
0.05means five percentage points.- expected_rate : float, default 0.5¶
The rate to plan around. 0.5 is the conservative choice, since a rate near one half is the hardest to pin down; plan around a rate closer to what you expect if you have grounds for one.
- level : float, default 0.95¶
Credible level.
- prior : str, tuple, or BetaPrior, default 'uninformed'¶
Prior on the rate. Note that strength alone does not help: a prior centred near
expected_ratereduces the data required, while an equally strong prior centred elsewhere increases it, since it pulls the posterior toward a region of higher variance. Making that distinction visible is much of the point of this method.- max_n : int, default 1_000_000¶
Give up beyond this, rather than searching forever for an unattainable precision.
- Returns:¶
Call
.summary().- Return type:¶
Examples
>>> import bayesplain as bp >>> plan = bp.teach.precision_planning(target_width=0.10, expected_rate=0.15) >>> plan.required_n 195What matters is not how strong the prior is but whether it agrees with what you expect to find. A prior centred on the rate you are planning around halves the data you need; one of similar strength centred somewhere else increases it, because it drags the posterior toward the middle where a rate is hardest to pin down:
>>> agrees = bp.priors.from_previous_study(successes=15, n=100) >>> bp.teach.precision_planning(0.10, 0.15, prior=agrees).required_n 95 >>> bp.teach.precision_planning(0.10, 0.15, prior="skeptical").required_n 207