weighting • 13.2 Estimating the mean outcome via modeling • 13.3 Standardizing the mean outcome to the confounder distribution • 13.4 IP weighting or standardization? • 13.5 How seriously do we take our estimates? • Fine Point 13.1 • Technical Point 13.1 • Fine Point 13.2 • Technical Point 13.2
• 13.2 Estimating the mean outcome via modeling • 13.3 Standardizing the mean outcome to the confounder distribution • 13.4 IP weighting or standardization? • 13.5 How seriously do we take our estimates? • Fine Point 13.1 • Technical Point 13.1 • Fine Point 13.2 • Technical Point 13.2 4
weightingを⽤いて、禁煙が体重増加におよぼすaverage causal effectを推定した。 ※以下の変数Lで調整し、conditional exchangeabilityを仮定した。 >?@,A?B − [>?B,A?B] ※a: treatment c: cencering Average causal effectの定義: Sex Age Race Education intensity and duration of smoking physical activity in daily life recreational exercise weight 6
causal effectを推定する。 データ仕様書(n=1566)※打ち切りがなかった例数 項⽬ 型 詳細 Weight gain Number Smoking cessation Category 0: untreated、1: treated Age Integer Sex Category 0: male、1: female Race Category 0: white、1: other Education Category 5 categories Weight Number kg Intensity of smoking Number number of cigarettes per day duration of smoking Number years of smoking Physical activity in daily life Category 3 categories Recreational exercise Category 3 categories Baseline Characteristic 7
weighting • 13.2 Estimating the mean outcome via modeling • 13.3 Standardizing the mean outcome to the confounder distribution • 13.4 IP weighting or standardization? • 13.5 How seriously do we take our estimates? • Fine Point 13.1 • Technical Point 13.1 • Fine Point 13.2 • Technical Point 13.2
weighting • 13.2 Estimating the mean outcome via modeling • 13.3 Standardizing the mean outcome to the confounder distribution • 13.4 IP weighting or standardization? • 13.5 How seriously do we take our estimates? • Fine Point 13.1 • Technical Point 13.1 • Fine Point 13.2 • Technical Point 13.2
weighting • 13.2 Estimating the mean outcome via modeling • 13.3 Standardizing the mean outcome to the confounder distribution • 13.4 IP weighting or standardization? • 13.5 How seriously do we take our estimates? • Fine Point 13.1 • Technical Point 13.1 • Fine Point 13.2 • Technical Point 13.2
モデルで推定する場合、以下のように利⽤するモデルが異なるため、最終アウトプットは⼀般的に異なる! ▪IP weightingの場合: Pr[A = a, C = 0 | L]をPr[A = a | L]およびPr[C = 0 | A = a, L]から推定 ※Pr[A = a | L]およびPr[C = 0 | A = a, L]はlogistic regression modelを当てはめて推定(Chapter 12) ▪Standardizationの場合: E[Y | A = a, C = 0, L = l]をparemetric linear regression modelを当てはめて推定(Chapter 13)
methodも使⽤する(Fine point 13.2 or Technical point 13.2) 最後に・・・ 今回は全体のaverage causal effectを推定したけど、特定のサブセットに限定したaverage causal effectも 計算可能である。 その際は計算対象のサブセットを限定するだけで、その他は全て同じやり⽅で実施可能である!
weighting • 13.2 Estimating the mean outcome via modeling • 13.3 Standardizing the mean outcome to the confounder distribution • 13.4 IP weighting or standardization? • 13.5 How seriously do we take our estimates? • Fine Point 13.1 • Technical Point 13.1 • Fine Point 13.2 • Technical Point 13.2
weighting • 13.2 Estimating the mean outcome via modeling • 13.3 Standardizing the mean outcome to the confounder distribution • 13.4 IP weighting or standardization? • 13.5 How seriously do we take our estimates? • Fine Point 13.1 • Technical Point 13.1 • Fine Point 13.2 • Technical Point 13.2
L = l] = 0 and Pr[L = l] = 0の場合E[Y | A = a, L = l]がundefinedとなるため ただし、StandardizationとIP weightingではPositivity違反の影響が異なる 理由:parametric modelで外挿すればpositivityの違反を無視できるため ※ただし、推定にbiasを持ち込むので95%信頼区間で真の効果を推定することを⾏う age E[Y |A = 1, C = 0, L = l] conditional relation between age and the mean outcome ^ 48 外挿 注意点: ①外装はあくまでデータが無限にあっても推定できない 区間の推定を⽬的に⾏う ※データ量不⾜を補うものではない =「外挿できるからデータ量は少なくても良い」は成⽴しない ②このことがIP weightingよりStandardizationを推奨 することにならない =標準誤差が⼩さくなるがbiasを含む。Biasが標準誤差 よりも多くの推定の誤りを⽣むことも・・・
weighting • 13.2 Estimating the mean outcome via modeling • 13.3 Standardizing the mean outcome to the confounder distribution • 13.4 IP weighting or standardization? • 13.5 How seriously do we take our estimates? • Fine Point 13.1 • Technical Point 13.1 • Fine Point 13.2 • Technical Point 13.2
weighting • 13.2 Estimating the mean outcome via modeling • 13.3 Standardizing the mean outcome to the confounder distribution • 13.4 IP weighting or standardization? • 13.5 How seriously do we take our estimates? • Fine Point 13.1 • Technical Point 13.1 • Fine Point 13.2 • Technical Point 13.2
weighting • 13.2 Estimating the mean outcome via modeling • 13.3 Standardizing the mean outcome to the confounder distribution • 13.4 IP weighting or standardization? • 13.5 How seriously do we take our estimates? • Fine Point 13.1 • Technical Point 13.1 • Fine Point 13.2 • Technical Point 13.2
outcomeを推定する E[Ya = 1]はE[b(L)]と書くことができる。この時b(L)は以下のいずれかで表現される b(L) = E[Y | A=1 , L] or E pP v = ※ここで = Pr A = 1 L] 本Chapterではplug-in g-formula estimatorである@ m ∑ | b m b?@ について の推定値を研究対象者の数nで 平均を取った値で置き換えられることを説明した ※ は linear regression modelによって推定する ※IP weightingの場合は@ m ∑ p•P• v ‚ =• について、 m b?@ ( で推定)の推定値を 研究対象者の数nで平均を取った値で置き換えられることを説明した