隨機分配對照臨床試驗（RCT）的多重比較

家族錯誤率的控制

m 個比較：H1, H2…Hm, 第 i 個比較是 Hi

目的：維持家族錯誤率（整體第一型錯誤率 α）0.05、維持足夠的統計檢定力（1-第二型錯誤率 β）

保守性由高至低（統計檢定力由低至高）：

1. Bonferroni 法：Hi ≤ 臨界值 α/m，缺點是太保守了。

2. Holm 法：將 p 值由低至高排序 1…m，從最低的 p 值看是否 ≤ 臨界值 α/(m-i+1)，適用於不相關或是相關的比較

3. Hochberg 法：將 p 值由低至高排序 1…m，從最高的 p 值看是否 ≤ 臨界值 α/(m-i+1)，適用於不相關或是某些正相關的比較

固定次序法：若H1 ≤ 臨界值 α 則看 H2，若H2 ≤ 臨界值 α 則看 H3，以此類推，缺點是若前面的試驗陰性，則後面的試驗就沒有機會陽性了（無論 p 值多少）。

倒退（fallback）法：將 α 分配給 α1, α2, α3（不一定是均分）。若H1 ≤ 臨界值 α1 則看 H2，若H2 ≤ 臨界值 α1+α2 則看 H3 是否 ≤ 臨界值 α1+α2+α3，以此類推。

序列把關（gatekeeping）法：所有的主要終點（家族一）都有意義才能測試次要終點（家族二）

平行把關法：只要有一個主要終點（家族一）有意義就能測試次要終點（家族二）

家族一用截斷法（Holm, Hochberg）：看 p 值是否 ≤ 臨界值（γ 是截斷比例）α x [γ/(m-i+1)+(1-γ)/m]，當 γ=0 時等於 Bonferroni，當 γ=1 時等於 Holm, Hochberg。好處是讓 Holm, Hochberg 有意義時未使用的 α 可以被下一個家族重複使用，α ᵢ₊₁ =αᵢ-αᵢ[γ+(1-γ)(ai/m)], αi: number of accepted H。

Dunnett 法：各組跟同一個對照組比較

假發現率（FDR）：適用於多維的比較，例如：體學（基因體、蛋白質體、代謝體）

不需要校正：事先決定的少數比較、副作用/傷害/壞處的比較、假陰性的壞處比假陽性的壞處大

需要校正：驗證性研究、假陽性的壞處比假陰性的壞處大

探索性研究：不需要校正或是用假發現率校正

留言

張貼留言

這個網誌中的熱門文章

Bayesian reanalysis of ORBITA

import pymc3 as pm import numpy as np import arviz as az # Percutaneous coronary intervention in stable angina (ORBITA): a double-blind, randomised controlled trial 18 https://www.sciencedirect.com/science/article/abs/pii/S0140673617327149 diff = 16.6 diff_lower = -8.9 diff_upper = 42 # The scale for the Cauchy distribution cauchy_scale = 5 # Model setup with pm.Model() as model:     # Prior distribution for the difference     diff_prior = pm.Cauchy("diff", alpha=0, beta=cauchy_scale)      # Likelihood (normal approximation of the distribution of the difference)     se_diff = (diff_upper - diff_lower) / 3.92     likelihood = pm.Normal("likelihood", mu=diff_prior, sigma=se_diff, observed=np.array([diff]))     # Sample from the posterior distribution     trace = pm.sample(500000, cores=1, tune=200000, chains=4) # Calculate the posterior probability that the difference is negative post_prob_diff_negative = np.mean(t...
閱讀完整內容

Bayesian reanalysis with pymc3

 # Empagliflozin after Acute Myocardial Infarction 24 https://www.nejm.org/doi/full/10.1056/NEJMoa2314051 Do a Bayesian reanalysis of a RCT for log(hazard ratio) with 500000 samplings with a weakly informative prior, calculate the probability of hazard ratio less than 1 and a sensitivity analysis using pymc3 and only show the results in traditional Chinese instead of the code: A total of 3260 patients were assigned to receive empagliflozin and 3262 to receive placebo. During a median follow-up of 17.9 months, a first hospitalization for heart failure or death from any cause occurred in 267 patients (8.2%) in the empagliflozin group and in 298 patients (9.1%) in the placebo group, with incidence rates of 5.9 and 6.6 events, respectively, per 100 patient-years (hazard ratio, 0.90; 95% confidence interval [CI], 0.76 to 1.06; P=0.21). 使用 pymc3 對一項 RCT 進行貝氏再分析，在弱信息先驗條件下抽樣 500000 次，計算危害比小於 1 的概率，並進行敏感性分析，只用繁體中文顯示結果，而不顯示代碼:  Do a Bayesian reanalysis of a RCT with 500000 samplings wit...
閱讀完整內容

Bayesian reanalysis of Apolipoprotein A1 Infusions

Apolipoprotein A1 Infusions and Cardiovascular Outcomes after Acute Myocardial Infarction 24 https://www.nejm.org/doi/full/10.1056/NEJMoa2400969 P(θ | D) = (P(D | θ) * P(θ)) / P(D) marginal likelihood P(D) = ∫ P(D | θ) * P(θ) dθ Likelihood: P(D | θ) ~ N(μ_l, σ_l^2) Likelihood ratio: (P(D | θ1))/(P(D | θ2)*) Bayes factor: (P(D | θ1)* P(θ1))/(P(D | θ2)* P(θ2)) Prior: P(θ) ~ N(μ_p, σ_p^2) P(θ | D) ∝ P(D | θ) * P(θ) When we normalize the product of the two PDFs by dividing by the marginal likelihood p(D), the constant term cancels out, and we are left with a normal distribution with mean μ_new and variance σ_new^2. Prior mean 0, sd 0.5 log(HR) = log(0.93) = -0.0726 SE = (log(1.05) - log(0.81)) / (2 * 1.96) = 0.0644 Prior precision = 1 / (0.35^2) = 8.16 Likelihood precision = 1 / (0.0644^2) = 241.18 Posterior precision = 8.16 + 241.18 = 249.34 Posterior mean = (8.16 * 0 + 241.18 * -0.0726) / 249.34 = -0.0701 Posterior SD = sqrt(1 / 249.34) = 0.0634 Posterior 95% credible interval for the lo...
閱讀完整內容