Download as PDF, PPTX
Uploaded bySayuri Shimizu
「傾向スコア分析」 報告事例
傾向スコア(propensity score)を用いた論文の珍妙な事例紹介です。2012年に小規模の研究会向けに作ったスライドなので、最後の方はもう募集してないのですが。まぁいいか。
「傾向スコア分析」 報告事例
- 1.
- 2. 報告の流れ ■ 研究紹介 ２報 (30分) -「無作為化比較試験」ではなく「コホート研究」が対象 ■ 質疑 (10分)
- 3. Count Medlineに登録された傾向スコアを用いた論文数 Calendar Year 『Some journalsand their reviewers seem to be ‘enamored’ with propensity scores and may request such analyses as a ‘better tool’ to control for confounding.』 2012 John Seegerを改編して引用
- 4. Exposure propensity score Patientsnever treated with study drug % of subjects Patients always treated with study drug 0.5 0 = treated with study drug 1 = treated with comparison drug 2012 John Seegerを改編して引用
- 5. PS distribution aftermatching Patients never treated with study drug % of subjects Patients always treated with study drug 0.5 0 = treated with study drug 1 = treated with comparison drug 2012 John Seegerを改編して引用
- 6. 傾向スコアを用いた２論文 1. Survival amonghigh-risk patients after bariatric surgery. Maciejewski ML LE. JAMA. 2011; 305: 2419–26. doi:10.1001/jama.2011.817. 2. Resource Utilization and Costs of Schizophrenia Patients Treated with Olanzapine versus Quetiapine in a Medicaid Population. Yu AP, Atanasov P, Ben‐Hamadi R, Birnbaum H, Stensland MD, Philips G. Value in Health. 2009; 12: 708– 15.
- 7. Survival among high-riskpatients after bariatric surgery. Maciejewski ML LE. JAMA. 2011; 305: 2419–26. http://www.webmd.com/diet/weight-loss-surgery/slideshow-weight-loss-surgery#webmd_rf_photo_of_body_mass_index.jpg
- 8. ■背景 肥満手術の生存に関するエビデンスは、肥満関連 死亡リスクの低い若い女性の研究に基づいている。 高齢の男性における肥満手術と死亡の関連に関す るエビデンスは定かではない。 ■目的 高齢の男性の肥満手術の実施は、死亡率低下に関 係するだろうか？ Diagram of SurgicalOptions. Image credit: Walter Pories, M.D. FACS. NIDDK
- 9. ■研究デザイン Retrospective cohort studyin 退役軍人医療センター Patients 850人 2000-2006年に外科手術の質改善プロ グラムに登録された肥満手術実施患者 VA indicates Veterans Affairs; BMI, body mass index; and DCG, diagnostic cost group (a measure of comorbidity burden). Nonsurgical controls 41244人 2000年に外来受診した180万人のうち、BMI が35-80で肥満手術を実施していない患者 BMI indicates body mass index; FY, fiscal year; VA, Veterans Affairs; and DCG, diagnostic cost group (a measure of comorbidity burden).
- 10.
- 11. ■統計解析 ① Unadjusted difference 『Unadjusteddifferences between patients undergoing or not undergoing bariatric surgery were compared with χ2 tests for categorical variables and 2-tailed unpaired t tests for continuous variables in the unmatched cohorts, McNemar tests and paired t tests in the matched cohorts, and standardized differences to enable comparison of covariate imbalance between the matched and unmatched cohorts. The association between bariatric surgery and all-cause mortality was examined in the unmatched cohorts with crude mortality rate comparisons and unadjusted and multivariable Cox proportional hazards regression models.』
- 12. ■統計解析 ② Multivariable Coxmodel 年齢,性別,自己申告の人種,婚姻状況,BMI,併存症,VAの地域 ※併存症の評価方法 Diagnostic cost group scores 『The diagnostic cost group score aggregates inpatient and outpatient diagnoses in the year before baseline, with scores greater than 1.0 implying above-average expected expenditures and scores less than 1.0 implying below-average expected expenditures. Diagnostic cost group scores were as predictive of veterans‘ 1-year mortality as other comorbidity scores and were highly predictive of mortality, use, and expenditures in bariatric surgery.』
- 13. ■統計解析 ③ 傾向スコアの推定 『In athird analysis, we accounted for the nonequivalence (eg, selection bias) of the nonsurgical control cohort via propensity score matching with logistic regression…. A propensity score represents the predicted probability that a given patient will undergo bariatric surgery, and patients who had bariatric surgery procedures were matched to controls with a greedy algorithm. The propensity score model included interaction of age, age squared, diagnostic cost group, BMI, BMI squared, BMI cubed, sex, race, marital status, and Veterans Integrated Service Networks, as well as numerous 2-way interactions, and had a concordance index of 0.85.』
- 14. ■統計解析 ④ マッチング法・バランスの確認 『 Despitethe large (n = 41 244) sample of controls, we conducted one-to-one matching to avoid the possible bias of many-to-one matching. Each surgical case patient was matched to a single nonsurgical control if their predicted propensity scores were identical to 8 digits. If such a match was not found, the case patient was matched to a control on the basis of a 7-, 6-, 5-, 4-, 3-, 2-, or 1-digit match. This process matched 847 surgical case patients (of 850 possible; 99.6%) to 847 nonsurgical controls, and covariate balance between matched surgical case patients and nonsurgical controls was assessed via McNemar tests, 2-tailed paired t tests, and standardized differences. We then conducted an unadjusted Cox regression stratified on the matched pairs to account for the lack of independence between cohorts induced by matching and a Cox regression adjusted for differences in the year indicating zero time. Alternative propensity score analyses from many-to-one matching generated similar results, so we present the 1:1 matching results here. 』
- 15.
- 16.
- 17.
- 18.
- 19. ■結論 超肥満者に対する肥満手術は死亡率の減少 に寄与しなかった ／(^o^)＼ 過去の観察研究とは逆の結果 →調査対象,フォローアップ期間,共変量の違い VA病院の肥満手術前後の死亡率が高すぎる →30日死亡率が1.3%(他のコホートは若い,女性,手技も異なる) コホートのcharacteristicsの影響 …etc 『この研究のいいところ：統計的な調整の重要性がわかった』 →過去の研究は、リスク調整やマッチングに用いた共変量がイマイチ →regressionの結果に見られた手術と生存期間の関係は、 傾向スコアによる調整によって影響が減少した
- 20. Resource Utilization andCosts of Schizophrenia Patients Treated with Olanzapine versus Quetiapine in a Medicaid Population. AP, Atanasov P, Ben‐Hamadi R, Birnbaum H, Stensland MD, Philips G. Value in Health. 2009; 12: 708–15.
- 21.
- 22.
- 23. ■対象 1999-2003のペンシルベニア州のMedicaid claim databaseに登録された大人の統合失調症患者 90日O orQ処方無し オランザピンかクエチアピン を30日以上単剤処方 オーバーラップ可 2000/6 2002/6 1年間の総コスト Indexでオランザピン群とクエチアピン群に２分 併存症を評価
- 24. 傾向スコアの記載法 傾向スコアの説明がくどくど・・・ 『The following processwas applied to implement the propensity score method with optimal matching. First, patients’ baseline characteristics were profiled during a 12-month preindex period. These characteristics were compared between the two study cohorts. Second, propensity scores were generated using logistic regressions and given patient baseline covariates, which were selected using backward selection. Third, optimal matching was applied to match olanzapine and quetiapine patients based on propensity score using a SAS …. Finally, baseline characteristics were compared for the full and the matched cohorts over the 12-month preindex period. Once the propensity score matching process was completed, outcomes were compared using paired t tests for continuous variables, and McNemar’s tests for categorical variables for matched pairs. Each cost component was studied separately. 』
- 25. ■傾向スコアの推定 -ロジスティック回帰分析 ■マッチングのアルゴリズム -Optimal matching (メイヨー・クリニックのSASマクロ) One-to-onepair matching ■効果の推定 -対応のあるt検定 -マクマネー検定 ＋ ■２つの感度分析 -単剤治療の期間を’少なくとも60日以上’に拡張 -DID(Difference-in-Difference)分析 (傾向スコア＋共変量調整 のコンビネーションは推定の有効性を増加)
- 26.
- 27.
- 28.
- 29.
- 30. 処方から１年間の医療サービスコストと薬 剤費は、オランザピン群が安かった １年間の総医療費は1204ドル安かった ■リミテーション Claim dataを利用したので、重症度補正が十分でない 傾向スコアの利用でセレクションバイアスをコントロールしたけれど… 一般化可能性 １つの州のデータ 統合失調症患者のうちメディケイド被保険者は32％ 1999年から2003年のデータ 処方は変わっていく 多くのオランザピン患者のデータを除外している コストに与える共変量の影響は見ていない 12ヶ月間のコストデータ ２つの薬剤治療の因果効果は推定していない 治療開始がO or Qの時の経済的アウトカムを推計 Bipolarの患者に対する経済効果は不明
- 31.
- 32.
- 33.
- 34.