Description
Uses a pilot sample's variance and a t quantile to decide how many more observations guarantee a confidence interval of preset width with at least the stated coverage.
With a fixed sample size and unknown variance, the width of a t interval is random and can be large; Dantzig proved in 1940 that no fixed-n procedure can guarantee both a chosen width and coverage. Stein's solution is double sampling. A pilot of m observations estimates the variance S squared. The total sample size N is the larger of m and the smallest integer at least t(m-1, alpha/2) squared times S squared divided by d squared, where d is the desired half-width. After collecting the remaining N - m observations, the interval is the overall mean plus or minus d, with coverage guaranteed to be at least 1 - alpha for every parameter value. The price is oversampling: in simulation, average N exceeded the optimal size by about 100% with m = 5 and about 33% with m = 10, so a larger pilot is more efficient. A bounded-risk variant keeps the mean squared error below a set bound.
Attributes
- Discipline
- Statistics & Modelling
- Loss type
- Unmigrated
- Stage
- Evaluation
- Data
- Quantitative
- Effort
- Days
Steps
- Fix the desired half-width d (for example +/-5 seconds on mean task time) and the confidence level 1 - alpha.
- Choose a pilot size m of at least 2, preferably 10 or more to limit oversampling, and collect the pilot observations.
- Compute the pilot sample variance S squared.
- Compute N as the larger of m and the smallest integer at least t(m-1, alpha/2)^2 x S^2 / d^2.
- If N exceeds m, collect N - m further observations; otherwise stop.
- Report the mean of all N observations plus or minus d as the interval.
Sources
- Probability and Statistical Inference (Mukhopadhyay, 2000)
- Linear Statistical Inference and Its Applications (Rao, 1973)
Also known as
- Stein's two-stage procedure for a fixed-width confidence interval
- Stein's double sampling
- Two-stage sampling for fixed-width interval
- Stein two-stage procedure
- Bounded-risk two-stage estimation
Categories
- Analytics & Behaviour
- Statistics & Modelling