Basic Statistics 43. sample size Yosuke Fujii
sample size approach Conventional approach Set sample size statistically before experiments. Adaptive approach Interim analysis (clinical trial) Flexible approach (not recommended) Time, cost Multiple testing
Sample size and confidential interval (CI)
Caution 平均的な誤差範囲は設定した W に等しいは ず。狭い確率も広い確率も 50% 。 実験終了時に必要なサンプル数を計算する ので、除外されるものを考えるともう少し 必要。 designexperimentanalysis
Sample size and statistical test
Example
How to determine effect size? Cohen: preset small, medium, and large effect size. Lenth: consider scientific questions es <- mapply(function(x) cohen.ES("t", x)$effect.size, c("small", "medium", "large")) mapply(function(x) pwr.t.test(d=x, power=0.8)$n, es)
How to determine effect size? 勝者の呪い Setting of effect size is difficult. Refer previous report. Larger effects would be reported but smaller would not. Tend to accept larger effect size. Larger effect size requires small sample size. True size is smaller than reported size. Low power.
Non-parametric test Power analysis requires probabilistic distribution. More 15% samples are required if … Adequate samples. Not abnormal distribution.
Logistic analysis n: samples, m: variables n > 10m n > 20m (not select variables) n > 40m (select variables) n > m (Green)