目次

P値関数

One of the most persistent misconceptions about p-values is that they are necessarily calculated assuming a null hypothesis of no effect is true. Instead, p-values can and should be calculated for multiple hypothesized values for the effect size. 「p値に関する最も根強い誤解の1つは,p値は必ず「効果がない」という帰無仮説を仮定して計算されるというものである.そうではなく,効果量に関する複数の仮説に対してp値を計算することができ,また計算すべきである」

Nullismからの脱却

We view the practice of testing only the null hypothesis of no effect as a manifestation of a cognitive and procedural bias called ‘nullism’, in which excessive focus is made on a ‘no-effect’ hypothesis at the expense of analysing reasonable alternatives as well. It has been argued that nullism is a pervasive source of distorted scientific reporting (Greenland, 2017)

モデルと適合性のよい推定値を一覧する

Unless the point estimate has a value exactly corresponding to no effect, the hypothesis of no effect (and thus that any observed association is pure random error) is not the hypothesis that is most compatible with our data. Any other hypothesis with a larger p-value would be more compatible with the same data given the same background model. We can see this by plotting the p-values of all hypotheses in a compatibility graph or p-value function (also known as a confidence distribution or consonance curve; Berner and Amrhein, 2022; Birnbaum, 1961; Cox, 1958; Infanger and Schmidt-Trucksäss, 2019; Poole, 1987; Rafi and Greenland, 2020; Rothman et al., 2008, Ch: 10).

P値関数を要約する値

具体例

P2002~2010年にカナダのオンタリオ州で公的処方を受けた母親から生まれた子供
E母のセロトニン作動系の抗うつ薬内服への曝露
C上記への非曝露,または非曝露の兄弟
O自閉症スペクトラム障害の発生
結果自閉症スペクトラム障害の発生率は,抗うつ薬に曝露された小児では1000人年当たり4.51人,曝露されていない小児では1000人年当たり2.03人(群間差 1000人年当たり2.48 [95% CI, 2.33-2.62]; HR 2.16 [95% CI, 1.64-2.86]).
高次元傾向スコア propensity score に基づく治療の逆確率重みづけ IPTW 後の関連は統計学的に有意ではなかった(HR 1.61[95%CI、0.997-2.59])
2)
“Scientists rise up against statistical significance.” Nature 567, 305-307 (2019). DOI
3) , 7)
【CC-BY】Journal of Information Technology Vol37, Issue 3, September 2022,316-320 DOI
4)
American Journal of Epidemiology, Volume 186, Issue 6, 15 September 2017, Pages 639–645 DOI
5)
compatible intervalsへの改称が提案されている
6)
JAMA. 2017;317(15):1544-1552. DOI