...

メタ・アナリシス - Researchmap

by user

on
Category: Documents
54

views

Report

Comments

Transcript

メタ・アナリシス - Researchmap
「メタ・アナリシス」の書き方
東京医科歯科大学大学院 医歯学総合研究科
心療・緩和医療学分野 博士課程 市倉加奈子
メタ・アナリシスの最新動向
①メタ・アナリシスの処理や報告に関する知識が広がり、
Cochraneのエントリーが11,000件を超過 (2009年3月)
②バイアスのリスクを評価する指標として,「アウトカムレ
ベル」などの新しい概念が進展
③無作為化比較試験 (Randomized controlled trial: RCT)
以外のエビデンスを統合する手法としても発展
QUOROM: Quality of Reporting of Meta-analysis
PRISMA: Preferred Reporting Items for Systematic
reviews and Meta-analysis
2
メタ・アナリシス実施の前に
プロトコルの立案と登録
http://www.crd.york.ac.uk/prospero/
会員手続き後,
無料で登録可
3
メタ・アナリシス概要
1.問題の定式化
2.文献検索
3.研究の選択・データベース化
4.個々の研究のバイアス評価
5.結果の統合
6.バイアスの検討
4
メタ・アナリシス概要
1.問題の定式化
2.文献検索
3.研究の選択・データベース化
4.個々の研究のバイアス評価
5.結果の統合
6.バイアスの検討
5
本日のMENU
個々の研究の特性とバイアス評価
結果の統合
①指標の決定
②事前の計算処理
③分析実施の検討
④モデルと手法の決定
⑤効果量と信頼区間の算出
バイアスの検討
アディショナル・アナリシス
メタ・アナリシスのまとめ
6
メタ・アナリシスとRパッケージ
metafor
meta
rmeta
固定効果モデル
○
○
○
変量効果モデル
○
○
○
various
DL
DL
Mantel-Haenzel法
○
○
○
Peto法
○
○
×
Forest plots
○
○
○
Funnel plots
○
○
○
Radial plots
○
○
×
Q-Q normal plots
○
×
×
質変数の調整分析
multiple
single
×
量変数の調整分析
multiple
×
×
○
×
×
異質性の推定
混合効果モデル
7
使用するデータセット
(Viechtbauer., J Stat Software., 2010, Conducting Meta-Analysis in R with the metafor package.)
8
本日のMENU
個々の研究の特性とバイアス評価
結果の統合
①指標の決定
②事前の計算処理
③分析実施の検討
④モデルと手法の決定
⑤効果量と信頼区間の算出
バイアスの検討
アディショナル・アナリシス
メタ・アナリシスのまとめ
9
個々の研究の特性
Methods
研究デザイン/研究期間
Participants
背景/研究に関連した健康状態の詳細/年齢/性別/国
Intervention
薬物→薬名/用量/頻度/投与方法/期間
非薬物→介入の留意点と構成内容
Outcomes
評価時期/評価指標
Notes
上記以外の必要事項,資金源,など
10
個々の研究の特性
≪Cochraneでの記載≫
(Cipriani et al., The Cochrane Collaboration., 2009, Escitalopram versus other antidepressive agents for
depression.)
11
個々の研究の特性
≪結果の書き方:薬物≫
患者
背景
疾患
状態
介入群の
内容とN数
統制群の
内容とN数
評価
指標
【推奨されない例】
評価指標の欄に結果を併記すべきではない
(Hauser et al., Rheumatology., 2011, Comparative efficacy and acceptability of amitriptyline, duloxetine
and milnacipran in fibromyalgia syndrome: a systematic review with meta-analysis.)
12
個々の研究の特性
≪結果の書き方:非薬物≫
疾患
名
各群の
N数
介入群の
内容
統制群の
内容
評価
時期
(Smit et al., Clin Psychol Rev., 2012, The effectiveness of long-term psychoanalytic psychotherapy--a metaanalysis of randomized controlled trials.)
13
NEW
個々の研究のバイアス評価
質 (Quality)
個々の研究で研究者が除去可能な限りのバイアス
に対する評価
バイアスのリスク (Risk of Bias)
客観的なバイアスの有無に対する評価
例)外科的手術で介入者に盲検化することは実質不可
Quality
→ 被験者への盲検化で質は十分と評価
Risk of Bias → 介入者に盲検化されていないことで結果に
バイアスが生じる可能性ありと評価
14
個々の研究のバイアス評価
バイアスのリスクをどのように評価?
盲検化の有無?
2人以上の評価者による評価?
評価者同士の独立した評価?
評価者チーム内で実施したキャリブレーション試験?
評価したリスクのデータ統合時にサブグループ分析
でどのように活用?
除外する研究の有無?根拠?
評価したリスクによる感度分析とサブグループ分析
の計画?
15
個々の研究のバイアス評価
(1) Cochrane推奨のバイアスのリスク評価法
①Sequence generation
②Allocation concealment
③Blinding of participants, personnel and outcome assessors
④Incomplete outcome data
⑤Selective outcome reporting
⑥Other potential threats to validity
・6領域についに,”Yes”/”No”/”Unclear”で評価
・”Yes”である方が,バイアスのリスクが低い
16
個々の研究のバイアス評価
①Sequence generation
割付けの列が適切に生成されているか?
【”Yes”の例】
乱数表の活用/コンピューター上の乱数生成器の利用/コイント
ス/カードや封筒のシャッフル/サイコロ振り/くじ引き
【”No”の例】
誕生月や日の偶数・奇数/参加の日や曜日/病院のID/臨床
医の判断による割付け/参加者の好みによる割付け/実験室試
験や一連の研究結果に基づく割付け/介入の利用しやすさによ
る割付け
17
個々の研究のバイアス評価
②Allocation concealment
対象者・調査者に割付けが十分に隠されているか?
【”Yes”の例】
中心部からの割付け(電話,ウェブ,調剤上のランダム化)/連続
的にナンバリングされた同じ形の薬剤の袋/連続的にナンバリン
グされ,不透明で,封がされた封筒
【”No”の例】
オープンなランダム割付けスケジュール(乱数リスト)/十分な保
護なしの封筒の割付け(封がされていない,不透明でない,連続
的にナンバリングされていない)/交互または循環的/誕生日/
ID番号
18
個々の研究のバイアス評価
③Blinding of participants, personnel and outcome
assessors
対象者・介入者・評価者に盲検化されているか?
【”Yes”の例】
非盲検化だが効果は盲検化に影響されないと判断可/対象者と
介入者が盲検化/対象者・介入者いずれかが非盲検化だが,効
果評価は盲検化され,他評価は盲検化によるバイアスの可能性
が低い
【”No”の例】
非盲検化または不十分だが,アウトカムは盲検化に影響される/
対象者・介入者に盲検化されているが,分かる可能性/対象者・
介入者のいずれかが非盲検化でバイアスの可能性あり
19
個々の研究のバイアス評価
④Incomplete outcome data
不十分なアウトカムデータに適切に対処しているか?
【”Yes”の例】
欠損なし/欠損の理由が真値と関連する可能性が低い(生存
データなど)/各群の欠損理由が同様で介入群の欠損数が均一
/観察されたイベントと欠損の比率が介入効果の推定に意味を
なさない/連続変数のため,欠損の妥当な効果量が観察された
効果量に意味をなさない/適切な方法で欠損が置換されている
【”No”の例】
欠損の理由が真値と関連し,介入群の欠損数または理由が不均
一/観察されたイベントと欠損の比率が介入効果の推定に意味
をもつ/連続変数のため,欠損の妥当な効果量が観察された効
果量でバイアスを引き起こしうる/単純な置換の不適切な利用
20
個々の研究のバイアス評価
⑤Selective outcome reporting
報告するアウトカムの選択について報告されている?
【”Yes”の例】
研究プロトコルが入手可能で,初めに設定されたアウトカム全て
が既定の方法で報告されている/研究プロトコルは入手不可能
だが,初めに設定されたものを含めた全ての予測されたアウトカ
ムが報告されていることが明確
【”No”の例】
初めに設定されたアウトカム全てが報告されていない/プライマ
リアウトカムは,初めに設定されていない分析で用いられている
/プライマリアウトカムが初めに設定されていない/レビューされ
た関心のあるプライマリアウトカムの報告が不十分なためにメタ
アナリシスに入れられない/アウトカムが報告されていない
21
個々の研究のバイアス評価
⑥Other potential threats to validity
バイアスのリスクになり得るものが他にないか?
【”Yes”の例】
他にバイアスの原因となるものはない
【”No”の例】
用いられた研究デザインに関連したバイアスが潜在的に存在して
いる/データ依存性のプロセスのために早期に終了(正式な中断
ルール)/ベースラインの不均衡が著しい
22
個々の研究のバイアス評価
≪方法の書き方≫
Risk of bias was evaluated in accordance with the
Cochrane Handbook for Systematic Reviews of
interventions, using the following parameters: adequacy of
sequence generation; allocation concealment; blinding of
participants, personnel and outcome assessors; incomplete
outcome data; and selective outcome reporting
(Apler., BMJ., 2011, Citalopram for major depressive disorder in adults: a systematic review and metaanalysis of published placebo-controlled trials.)
23
個々の研究のバイアス評価
≪結果の書き方≫
このように「バイアスのリスク」について評価する表は,
個々の研究の特性の表とは分けて作成すべき
(Apler., BMJ., 2011, Citalopram for major depressive disorder in adults: a systematic review and metaanalysis of published placebo-controlled trials.)
24
個々の研究のバイアス評価
Ⅱ.その他のバイアスのリスク評価法
①スケール評価
例1) Jadad score (ハダッド・スコア)
①ランダム化 (明示され,適切である=2点,明示されているが適切でない
=1点,明示されていない=0点),②ブラインド化 (明示され,適切である=2
点,明示されているが適切でない=1点,明示されていない=0点),③中断
と脱落 (記載されている=1点,明示されていない=0点) の5点満点で評価
例2) Clinical Trials Assessment Measure (CTAM)
サンプル2問 (最大10点),割り付け3問 (最大16点),アセスメント5問 (最
大32点),統制群1問 (最大16点),分析2問 (最大15点),積極的治療2問
(最大11点) の計15問100点満点で評価
②チェックリスト評価
25
個々の研究のバイアス評価
≪方法の書き方:①≫
A quality review of each RCT was done to include details of
randomization, generation of random numbers, details of doubleblinding procedure, information on withdrawals, and allocation
concealment. One point was awarded for the specification of
each criterion, with a maximum score of 5. High quality RCTs
scored ≥ 3 points whereas low-quality RCTs scored < 2 points,
according to a modified Jadad score.
(Chen et al., Cardiovasc Drugs Ther., Effects of omega-3 fatty acid for sudden cardiac death prevention in
patients with cardiovascular disease: a contemporary meta-analysis of randomized, controlled trials.)
We used a modified Jadad scale to evaluate the quality of the
randomized controlled trials.
(Ghanbari et al., Arch Intern Med., Effectiveness of implantable cardioverter-defibrillators for the primary
prevention of sudden cardiac death in women with advanced heart failure: a meta-analysis of randomized
controlled trials.)
26
個々の研究のバイアス評価
≪結果の書き方:①≫
【推奨されない例】
研究の特性の表にまとめるべきではない
(Chen et al., Cardiovasc Drugs Ther., Effects of omega-3 fatty acid for sudden cardiac death prevention in
patients with cardiovascular disease: a contemporary meta-analysis of randomized, controlled trials.)
27
個々の研究のバイアス評価
≪結果の書き方:①≫
(Ghanbari et al., Arch Intern Med., Effectiveness of implantable cardioverter-defibrillators for the primary
prevention of sudden cardiac death in women with advanced heart failure: a meta-analysis of randomized
controlled trials.)
28
個々の研究のバイアス評価
≪方法の書き方:②≫
We assessed the quality of each study using the Maastricht
Amsterdam criteria (van Tulder et al., 1997) and eight criteria
proposed by Cuijpers et al. (2009). The inter-rater agreement was
80% overall (77% for the Maastricht Amsterdam criteria and 87%
for the criteria used by Cuijpers et al.). Consensus on the quality
rating was achieved through discussion and the final ratings for
the criteria used by Cuijpers et al. are depicted in Table 2. The
quality of selected studies was variable. Most importantly, the
quality of internal validity was・・・
(Smit et al., Clin Psychol Rev., 2012, The effectiveness of long-term psychoanalytic psychotherapy--a metaanalysis of randomized controlled trials..)
29
個々の研究のバイアス評価
≪結果の書き方:②≫
(Smit et al., Clin Psychol Rev., The effectiveness of long-term psychoanalytic psychotherapy--a meta-analysis
of randomized controlled trials..)
30
本日のMENU
個々の研究の特性とバイアス評価
結果の統合
①指標の決定
②事前の計算処理
③分析実施の検討
④モデルと手法の決定
⑤効果量と信頼区間の算出
バイアスの検討
アディショナル・アナリシス
メタ・アナリシスのまとめ
31
メタ・アナリシスの実施
①要約指標の決定
(1) 2値変数
例)死亡,再入院,糖尿病の発症,がんの再発
(2) 連続変数
例)うつの重症度,生活の質 (QOL),日常生活動作 (ADL)
(3) 生存時間変数
例)時間×死亡,時間×再入院,時間×発症
32
(1) 2値変数の要約指標
イベント発生数
介入群 (n1)
a
統制群 (n2)
リスク比
a/n1
c/n2
非イベント発生数
b
c
オッズ比
a/b
c/d
a/c
b/d
d
リスク差
(a/n1) – (c/n2)
 リスク比とオッズ比:非介入時の発生リスクに左右されにくい
 リスク差:群間の発生リスクの差が小さい時に有用
33
(1) 2値変数の要約指標
≪方法の書き方≫
心疾患患者に対するオメガ3脂肪酸(悪玉コレステロールを分解
し,血液を浄化)の心臓突然死予防効果
Treatment effects are presented as risk ratio (RR) with 95%
confidence interval (CI).
(Chen et al., Cardiovasc Drugs Ther., Effects of omega-3 fatty acid for sudden cardiac death prevention in
patients with cardiovascular disease: a contemporary meta-analysis of randomized, controlled trials.)
慢性尋常性乾癬(赤みとうろこ状の皮膚が特徴の慢性疾患)患者
に対する生物学的治療と心血管性のイベントとの関連
Absolute risk differences were used as an effect measure,
measuring the excess probability of MACEs in those
receiving active treatment compared with those receiving
placebo.
(Ryan et al., JAMA., 2011, Association between biologic therapies for chronic plaque psoriasis and
cardiovascular events: a meta-analysis of randomized controlled trials..)
34
(2) 連続変数の要約指標
【対応なし】
平均得点
介入群 (n1)
X1
平均得点
統制群 (n2)
Swithin
X2
群間の標準偏差
の平均値
Cohen’s d =
X1 - X2
Swithin
35
(2) 連続変数の要約指標
【対応あり】
統制期 (n2)
介入期 (n1)
平均得点
平均得点
Y2
Y1
Swithin
平均値差の
標準偏差
Cohen’s d =
Y1 - Y2
Swithin
36
(2) 連続変数の要約指標
• Cohen’s dはサンプルサイズが小さいと,過大
評価につながりやすい。
→自由度で調整したHedges’ gが推奨される
Hedges’ g =
(1 -
3
4df - 1
)
×d
37
(2) 連続変数の要約指標
≪方法の書き方≫
大うつ病性障害患者に対するシタロプラム投与治療と
うつ症状改善効果
I carried out a meta-analysis of the change-from baseline scores
on the 17, 21 and 24-item HAM-D scales for participants
included in outcome evaluation. I applied a random effects
model to calculate Hedges’ g for standardized mean differences
between citalopram and placebo groups. Standard deviations
(SD) were computed from the p values, ・・・・・・
(Apler., BMJ., 2011, Citalopram for major depressive disorder in adults: a systematic review and metaanalysis of published placebo-controlled trials.)
38
(3) 生存時間変数の要約指標
≪方法の書き方≫
心不全患者に対する植込み型除細動器治療の心臓
突然死一次予防効果に対するメタ分析
Hazard ratio (HR) was chosen as the principal measure of effect.
The HR from each included trial was pooled by the use of fixed
effects and random-effects models that used weighting based on
inverse variance calculated according to the methods of Der
Simonian and Laird. The Q test and I2 ・・・・・・
生存時間変数を用いるときは、
計算方法を丁寧に記載すべき
※測定時期が研究によって様々な場合に1時点に区切って2値
変数として統合する場合もある
(Chen et al., Cardiovasc Drugs Ther., Effects of omega-3 fatty acid for sudden cardiac death prevention in
patients with cardiovascular disease: a contemporary meta-analysis of randomized, controlled trials.)
39
本日のMENU
個々の研究の特性とバイアス評価
結果の統合
①指標の決定
②事前の計算処理
③分析実施の検討
④モデルと手法の決定
⑤効果量と信頼区間の算出
バイアスの検討
アディショナル・アナリシス
メタ・アナリシスのまとめ
40
メタ・アナリシスの実施
②事前の計算処理
メタ・アナリシスを実施するうえで必要な事前の計算
処理の過程を明記する
• 2つ以上の介入群がある場合に、2群以上の結果を一
体化しておく手続き
• 適格基準を一致させるためにサブセットのみを抽出し
ておく手続き
• 連続変数において測定指標が研究間で異なっている
時に,アウトカムを標準化して揃える手続き
• 標準偏差や標準誤差の記載がない場合に,論文に
記載されているデータから計算しなおす手続き
41
事前の計算処理
≪方法の書き方≫
Standard deviations (SD) were computed from the p values, taken
at the upper limit and converted into a t-statistic. I used the
formula SD. SE/O(1/Ne+1/Nc), where SE. difference in means of the
two change from baseline scores divided by the t-statistic, and Ne
and Nc are the sample sizes in the experimental and control groups
respectively. I multiplied the result by 1 to convert a measure of
symptom reduction into an improvement score.
(Apler., BMJ., 2012, Citalopram for major depressive disorder in adults: a systematic review and metaanalysis of published placebo-controlled trials.)
42
本日のMENU
個々の研究の特性とバイアス評価
結果の統合
①指標の決定
②事前の計算処理
③分析実施の検討
④モデルと手法の決定
⑤効果量と信頼区間の算出
バイアスの検討
アディショナル・アナリシス
メタ・アナリシスのまとめ
43
メタ・アナリシスの実施
③分析実施の検討
(1) 臨床的検討
(2) 方法論的検討
(3) 統計学的検討
※(3) 以外は,明確な基準はない。
→研究者が主観的意図を持ち,根拠を記す!
44
メタ・アナリシスの実施
③分析実施の検討
(1) 臨床的検討
 対象患者の性質が異なっている
 介入内容の性質が異なっている
(2) 方法論的検討
 主観的指標を用いている
 盲検化されている研究とされていない研究が
混在している
あり
(3) 統計学的検討
あり
実施困難
なし
(3) 統計学的検討
なし
あり
要検討
なし
実施可
45
(3) 統計学的検討 (異質性の検定)
k
(
Q = iΣ= 1
N (研究数) に
左右されないように
自由度で調整
Yi - M
Si
N (研究数) に
左右されないように
自由度で調整
Q - df
Q - df
T2 =
)
×100%
I2 =
C
Q
用いる指標による分散の違いに左右されない
ように,変量効果モデルで算出される重みづけ
が用いられた値のCを調整
46
(3) 統計学的検討 (異質性の検定)
値の範囲
研究数に
指標に
左右されるか? 左右されるか?
Q
0≤Q
Yes
No
T2
0 ≤ T2
No
Yes
I2
0% ≤ I2 <100%
No
No
※I2は研究数や指標に左右されにくく,値の範囲も分
かりやすいため,最も好まれる。
→ただし,研究数が小さい場合に信頼区間が広がる
と,不確定な値になるため,注意が必要
47
(3) 統計学的検討
≪Rを使った出力≫
統計学的に異質性ありと判断
48
(3) 統計学的検討
≪Rを使った出力≫
無視できない異質性の大きさありと判断
49
分析実施の検討
≪方法の書き方≫
Interstudy heterogeneity was tested by using the Cochran Q
statistic (chi-square value), with the significance level set at a P
value less than 0.10, and was quantified by using the I2 statistic,
where a value of 50% or greater indicates substantial
heterogeneity.
(Sievenpiper., Ann Intern Med., 2012, Effect of Fructose on Body Weight in Controlled Feeding Trials - A
Systematic Review and Meta-analysis)
50
分析実施の検討
≪方法の書き方≫
We investigated potential sources of clinical and
methodological heterogeneity by using sensitivity analyses and
a priori subgroup analyses of comparator, fructose form (solid,
liquid, or mixed), dose (American Heart Association threshold
37.5 g/d or 37.5 g/d [150 kcal/d or 150 kcal/d]), follow-up (4
weeks or 4 weeks), study quality (Heyland MQS 8 or 8),
randomization (yes or no), and baseline body weight (70 kg or 70
kg).
(Sievenpiper., Ann Intern Med., 2012, Effect of Fructose on Body Weight in Controlled Feeding Trials - A
Systematic Review and Meta-analysis)
51
分析実施の検討
≪結果の書き方≫
The effect was to eliminate evidence of interstudy heterogeneity
(I2 value) and increase the relative weighting of effects detected
in parallel trials so that the body weight– decreasing effect of
fructose then became statistically significant in the overall
analysis; for example, with a correlation coefficient of 0.66, the
MD was -0.36 kg (CI, -0.60 to -0.12 kg) and the I2 value was 0%.
However, a formal test of interaction of weight change by study
design was not statistically significant.
(Sievenpiper., Ann Intern Med., 2012, Effect of Fructose on Body Weight in Controlled Feeding Trials - A
Systematic Review and Meta-analysis)
52
分析実施の検討
≪結果の書き方≫
Appendix Figure 1 shows subgroup analyses in the isocaloric
trials. No evidence of effect modification was seen in any of the
subgroup analyses on meta-regression analyses except for
fructose form. The weight-decreasing effect of fructose in solid
(fruit) and fluid form differed statistically from the weightincreasing effect of fructose in mixed form (p< 0.05). Statistically
significant unexplained interstudy heterogeneity, however,
remained in at least 1 level of all subgroup analyses.
(Sievenpiper., Ann Intern Med., 2012, Effect of Fructose on Body Weight in Controlled Feeding Trials - A
Systematic Review and Meta-analysis)
53
本日のMENU
個々の研究の特性とバイアス評価
結果の統合
①指標の決定
②事前の計算処理
③分析実施の検討
④モデルと手法の決定
⑤効果量と信頼区間の算出
バイアスの検討
アディショナル・アナリシス
メタ・アナリシスのまとめ
54
メタ・アナリシスの実施
④モデルと手法の決定
(1) 固定効果モデル (fixed-effect model)
• Mantel-Haenzel法
• Peto法
• General variance-based method
(2) 変量効果モデル (random-effect model)
• DerSimonian-Laired法 (DL法)
• restriced maximum likelihood method
• Bayesian method
55
(1) 固定効果モデル
→統合する研究の母集団が同一と仮定したモデル
→効果のばらつきは偶然誤差のみが原因
56
(2) 変量効果モデル
→統合する研究の母集団が異なると仮定したモデル
→効果のばらつきは偶然誤差と各研究の偏りが原因
57
モデルと手法の決定
(1) 固定効果モデル (fixed-effect model)
※信頼区間が狭くなり,効果が示されやすくなる
※被験者内の比較のみ使用できる
※異質性がない場合には,変量効果モデルと同じ結果
になる
(2) 変量効果モデル (fixed-effect model)
※研究間の異質性が認められても,メタ・アナリシスが
実施できる可能性がある
※被験者間/被験者内の比較,いずれも使用できる
※サンプルサイズの小さい研究の重みづけが劣性にな
58
るため,出版バイアスが出やすい
モデルと手法の決定
≪方法の書き方≫
【固定効果】
If neither clinical nor statistical heterogeneity was found, we
pooled results using a fixed effect model. If significant statistical
heterogeneity was found (p < 0.05), random-effect model was
considered to do the sensitivity analysis.
(Liu & Wang., Int J Clin Pract., 2012, Ramelteon in the treatment of chronic insomnia:
systematic review and meta-analysis.)
【変量効果】精神分析的心理療法のうつ病治療効果
All analyses used the random effects model because we expected
that the data could be heterogeneous, given the large diversity in
populations (mental disorders) and control treatments (Higgins,
Thompson, & Spiegelhalter, 2009; Lau, Ioannidis, & Schmid, 1997)
(Smit et al., Clin Psychol Rev., 2012, The effectiveness of long-term psychoanalytic psychotherapy--a metaanalysis of randomized controlled trials.)
59
モデルと手法の決定
≪Rを使った出力≫
推定効果量
95%信頼区間
60
本日のMENU
個々の研究の特性とバイアス評価
結果の統合
①指標の決定
②事前の計算処理
③分析実施の検討
④モデルと手法の決定
⑤効果量と信頼区間の算出
バイアスの検討
アディショナル・アナリシス
メタ・アナリシスのまとめ
61
効果量と信頼区間の算出
介入群・統制群それぞれのデータ一覧
個々の研究における指標の推定効果量
個々の研究における指標の信頼区間
要約した指標の推定効果量
要約した指標の信頼区間
Forest plots (個々の研究の効果量と信頼区
間/各群のN数/統合された効果量と信頼
区間/各研究の重みづけのパーセンテージ)
※通常は信頼区間として,95%信頼区間 (95%CI) を用いる
62
効果量と信頼区間の算出
≪結果の書き方≫
個々の研究の
効果量と信頼区間
統合された
効果量と信頼区間
各群のN数
各研究の重みづけの
パーセンテージ
(Apler., BMJ., 2012, Citalopram for major depressive disorder in adults: a systematic review and metaanalysis of published placebo-controlled trials.)
63
効果量と信頼区間の算出
≪結果の書き方≫
■の大きさが個々の研究の
重みづけに対応
■の中心が個々の研究の
効果量の点推定
横棒が個々の研究の
信頼区間の幅
信頼区間がまたいでいると
有意ではないという基準線
(null value)
◆の中心が統合された
効果量の点推定
(Apler., BMJ., 2012, Citalopram for major depressive disorder in adults: a systematic review and metaanalysis of published placebo-controlled trials.)
64
効果量と信頼区間の算出
≪Rを使った出力≫
65
効果量と信頼区間の算出
≪Rを使った出力≫
66
本日のMENU
個々の研究の特性とバイアス評価
結果の統合
①指標の決定
②事前の計算処理
③分析実施の検討
④モデルと手法の決定
⑤効果量と信頼区間の算出
バイアスの検討
アディショナル・アナリシス
メタ・アナリシスのまとめ
67
NEW
バイアスの検討
①出版バイアス (Publication Bias)
・ポジティブな結果が報告されやすいことによる
バイアス
・研究の欠如
②選択バイアス (Selective Reporting Bias)
・研究が恣意的に収集されることによるバイアス
・アウトカムの欠如
68
バイアスの検討
①出版バイアスについて
Funnel Plotが報告されているなら、効果の推定量と
精度をそれぞれ記載
Funnel Plotの非対称性について、統計的検討を
行ったかどうか、どのように行ったかを記載
②報告バイアスについて
プロトコルが入手可能なら出版されたデータを比較。
不可能なら、方法と結果の報告を比較。
③総合的に
研究を統合する際のバイアスを検討するために想
定していた分析がなされたか、その理由を記載
69
バイアスの検討
≪方法&結果の書き方≫
【Methods】We used funnel plot and Orwin’s fail safe N
analysis to assess this problem. The funnel plot appears
asymmetrical in the presence of potential file drawer
effect, which requires a further search for missing studies;
Orwin’s fail safe N analysis provides the number of
nonincluded studies that may bring the overall effect size
down to below trivial value and thus overthrow the study.
【Results】Every outcome variable showed a symmetric
funnel plot.
【推奨されない例】
できる限りFunnel Plotの図を入れる
(Feng et al., Worldviews on Evidence-Based Nursing., 2011, The Effect of Cognitive Behavioral Group
Therapy for Depression: A Meta-Analysis 2000–2010.)
70
バイアスの検討
≪方法の書き方≫
To increase the statistical power of subgroup analyses,
we combined the effects of fructose on body weight
across categories of weight and diabetes. We used metaregression analyses to assess the significance of
subgroup effects. We investigated publication bias by
inspection of funnel plots and by Egger and Begg tests.
(Sievenpiper., Ann Intern Med., 2012, Effect of Fructose on Body Weight in Controlled Feeding Trials - A
Systematic Review and Meta-analysis)
71
バイアスの検討
≪結果の書き方≫
We inspected funnel plots for evidence of publication
bias. No asymmetry or small-study effects were
detected among the isocaloric and hypercaloric
trials; the Egger and Begg tests were not statistically
significant (p < 0.05). However, 5 of the hypercaloric
trials were conducted by the same group of
investigators (with the same senior author).
(Sievenpiper., Ann Intern Med., 2012, Effect of Fructose on Body Weight in Controlled Feeding Trials - A
Systematic Review and Meta-analysis)
72
バイアスの検討
≪結果の書き方≫
Sievenpiper., Ann Intern Med., 2012, Effect of Fructose on Body Weight in Controlled Feeding Trials - A
Systematic Review and Meta-analysis)
73
バイアスの検討
≪Rを使った出力≫
視覚的に非対称性はない
=出版バイアスはないと判断
74
バイアスの検討
≪Rを使った出力≫
統計的に非対称でない
=出版バイアスはないと判断
75
本日のMENU
個々の研究の特性とバイアス評価
結果の統合
①指標の決定
②事前の計算処理
③分析実施の検討
④モデルと手法の決定
⑤効果量と信頼区間の算出
バイアスの検討
アディショナル・アナリシス
メタ・アナリシスのまとめ
76
アディショナル・アナリシスの実施
①感度分析
統合された主要な結果が個々の研究の方法や
データの変化に影響されている程度を検討
②サブグループ分析
要約効果量が研究やその参加者の固有の特性
に関連して変化するかどうかを検討
③メタ回帰分析
研究の特性が要約効果量に与える量的な影響
性を検討(サブグループ分析の考え方を拡大)
77
アディショナル・アナリシスの実施
分析を実施するかどうか、その根拠は何か、あらか
じめ想定されていた分析はどれか、について記載
メインのメタアナリシスの結果と異なるかどうか記載
(特に①感度分析において重要な情報)
結果が出ない分析もすべて報告
②サブグループ分析
交互作用のための検定について報告
サブグループごとに統合された推定量と信頼区間を
報告
③メタ回帰分析
p値だけでなく、効果量と信頼区間も報告
78
アディショナル・アナリシスの実施
≪方法の書き方≫
We intended subgroup-analyses (per type of disorder,
type of treatment and control etc.) and meta-regression
to explore heterogeneity, but the small number of
studies precluded many of the intended subgroup
analyses. We explored the few subgroup analyses that
were possible. We considered the proxy session ratio
(number of sessions in the intervention group/number of
sessions in the control group) to be a covariate that
might explain part of the heterogeneity in effect sizes
and explored this by meta-regression.
(Smit et al., Clin Psychol Rev., 2012, The effectiveness of long-term psychoanalytic psychotherapy--a metaanalysis of randomized controlled trials..)
79
アディショナル・アナリシスの実施
≪方法の書き方≫
To increase the statistical power of subgroup analyses,
we combined the effects of fructose on body weight
across categories of weight and diabetes. We used metaregression analyses to assess the significance of
subgroup effects. We investigated publication bias by
inspection of funnel plots and by Egger and Begg tests.
アディショナル・アナリシスの目的を
明記することが重要
(Sievenpiper., Ann Intern Med., 2012, Effect of Fructose on Body Weight in Controlled Feeding Trials - A
Systematic Review and Meta-analysis)
80
アディショナル・アナリシスの実施
≪結果の書き方≫
Appendix Figure 2 shows subgroup analyses in the
hypercaloric trials. None of the subgroups was
statistically significant in meta-regression analyses.
Statistically significant unexplained interstudy
heterogeneity, however, remained in at least 1 level of
most of the subgroup analyses.
(Sievenpiper., Ann Intern Med., 2012, Effect of Fructose on Body Weight in Controlled Feeding Trials - A
Systematic Review and Meta-analysis)
81
本日のMENU
個々の研究の特性とバイアス評価
結果の統合
①指標の決定
②事前の計算処理
③分析実施の検討
④モデルと手法の決定
⑤効果量と信頼区間の算出
バイアスの検討
アディショナル・アナリシス
メタ・アナリシスのまとめ
82
メタ・アナリシスのまとめ
重要なアウトカム全てのリスト
各群でイベントが生じるリスク
効果の絶対的および相対的な影響の大きさ
対象者のサンプル数
アウトカムごとのエビデンス全体の質の評価
コメント
※Cochraneのデータベース → “背景”の前に記載
※その他の論文 → “結果”と”考察”の間に記載
83
メタ・アナリシスのまとめ
①Table title / Header
②Outcomes
③Illustrative comparative risks1: Assumed risk
④Illustrative comparative risks2: Corresponding risk
⑤Relative effect (95%CI)
⑥Number of participants (studies)
⑦Quality of the evidence (GRADE)
⑧Comments
・上記の内容について,”Summary of findings”の表に
記載する
84
メタ・アナリシスのまとめ
①Table title / Header
以下の項目が含まれるように,1行にまとめる
 Patients of population
対象者の特性や属性
 Setting
研究上の限界領域としての環境
 Intervention
介入群の介入内容
 Comparison
統制群の介入内容
85
メタ・アナリシスのまとめ
②Outcomes
 重要な順に全てのアウトカムを記載 (最大7個)
 評価指標と評価時期の詳細を記載
・プロトコル作成時点で重要と考えたアウトカムを優先
(後から重要なアウトカムに気づくこともあるので注意)
・重篤な有害事象は含める
(軽度の有害事象は包含できる可能性がある)
・複数時点で測定している指標は,必要不可欠な指標
のみ残し,残りは共通の1時点を記載
・連続変数は単位が明確で説明可能なものを記載
(うつ症状・QOLなど基準が不明確な場合は,50%改善などの2
値変数に置換)
86
メタ・アナリシスのまとめ
③Illustrative comparative risks1: Assumed risk
 統制群でイベントが生じるリスク
・自然な状態で1000人中,イベントを経験する人数
・頻度が低いなら10000人中,頻度が高いなら100人中
と分母を変えても良い.
・脚注には,上記の根拠と資料を記載
87
メタ・アナリシスのまとめ
④Illustrative comparative risks1: Corresponding risk
 介入群でイベントが生じるリスク
・要約指標が2値変数の場合,メタ分析の結果,1000人
中,イベントを経験する人数 (絶対数) と信頼区間
・要約指標が連続変数の場合,平均値差・標準偏差と
信頼区間
88
メタ・アナリシスのまとめ
⑤Relative effect (95%CI)
 統制群と比較した介入群の相対的効果
・リスク比やオッズ比などの相対指標と信頼区間
・2値変数のメタ・アナリシスの統合で算出された結果
89
メタ・アナリシスのまとめ
⑥Number of participants (studies)
 統合された研究全ての対象者数の合計
 それぞれのアウトカムごとの研究数
90
メタ・アナリシスのまとめ
⑦Quality of the evidence (GRADE)
 メタ・アナリシスによるエビデンスの質の評価
・アウトカムごとに評価を行う
・GRADEシステムによる評価が推奨されている
根本的な方法論
質の評価
RCT / 2ランクアップ観察研究
高い
1ランクダウンRCT / 1ランクアップ観察研究
中程度
2ランクダウンRCT / 観察研究
3ランクダウンRCT / 1ランクダウン観察研究
連載 / ケースレポート
低い
非常に低い
91
メタ・アナリシスのまとめ
≪結果の書き方≫
①
②
③
④
⑤
⑥
⑦
⑧
(Cochrane Handbook for Systematic Reviews of Interventions)
92
参考文献
Cochrane Handbook for Systematic Reviews of Interventions
(Wiley)
93
参考文献
Introduction to Meta-Analysis -Statistics in Practice-
(Wiley)
94
参考文献
メタ・アナリシス入門
(丹後俊郎/朝倉書店)
95
参考文献
はじめてのメタアナリシス
(野口善令/臨床研究デザイン塾)
The PRISMA Statement for Reporting Systematic
Reviews and Meta-Analyses of Studies That Evaluate
Health Care Interventions: Explanation and Elaboration
(PLoS Medicine/www.plosmedicine.org)
Conducting Meta-Analysis in R with the metafor Package
(J Stat Software/www.jstatsoft.org)
市倉加奈子(東京医科歯科大学 心療・緩和医療学)
[email protected]
96
Fly UP