Meta Analysis(mata分析)大量资料汇总3

合集下载

Meta-分析

Meta-分析
• 方差可用于可信区间和假设检验的计算。
3.异质性检验与异质性分析
• 按统计原理,只有同质的资料才能进行合并或比较等统 计分析,反之,则不能。
• 因此,Meta分析过程需要对多个研究的结果进行异质 性分析,尽可能地消除导致异质的原因,使之达到同质。
异质性检验
• 异 质 性 检 验 (tests for heterogeneity) 又 称 同 质 性 检 验 (tests for homogeneity)
大小合适: 选题太大,纳入研究的文献太多,而且问题也 不明确,研究难以完成;选题太少,缺乏推广应用的代表 性,而且纳入研究的文献也太少,达不到汇总的效果。 一般而言,纳入研究的文献以10-30篇比较适合做meta 分析。
Meta分析的步骤和方法
• 提出问题,制定研究计划 • 检索资料 • 选择符合纳入标准的研究 • 纳入研究的质量评价 • 提取纳入文献的数据信息 • 资料的统计学处理 • 敏感性分析 • 形成结果报告
报告失访原因; (7)是否采用ITT(意向分析法)分析结果; (8)患者的依从性(compliance)如何。
Jadad量表
记分为1~5分,1或2分:低质量,3~5分:高质量: • 随机化方法:
– 恰当-如计算机产生的随机数字或类似的方法 (2 分); – 不清楚-试验描述为随机试验,但没有告知随机分配产生的
计分析的一种方法(Glass) • 是对先前研究结果进行统计合并和评述的一种新方法(Sack) • 是用以汇总众多研究结果的各种定量分析(Hedge) • 是一类统计方法,用来比较和综合针对同一科学问题所取得的研究
结果。比较和综合的结论是否有意义,取决于这些研究是否满足特 定的条件(Fleiss & Gross)

元分析(Meta-analysis)方法

元分析(Meta-analysis)方法

2

df为元分析所研究的各个研究样本的自由度。要求:每个研究样本的 容量大于或等于10,当df≥10时,接近正态分布。 缺点 是,不能对样 本容量10的测验进行统合,而通常情况下,小于10的测验样本很少, 因此,其缺点就不易体现出来了。

Stouffer统合(把 p值转换为z值,而非t值): Z c
解释的问题
• 关于d值的无偏估计
– d值是一个有些微偏差的效应估计值,应对其进 行校正( Hedges, 1982),公式如下:
解释的问题
选取的研究是否同质?

齐性检验,公式(之一):
d
wd w
w
2N 8 d
x2 w(d d)2
2


w
2 N 8 d
2
d为未加权的效应值,w指元分析中每个研究的权重
常常可以得出更有力的结论,引发对某一问题的激 烈争论
尤其存在研究结果相悖的情况时,能够给出一个是
集中于整体效应,对中间变量或交互效应没有给 予充分的重视
有把“苹果”和“桔子”混杂之嫌 由于对研究进行组织处理的标准不同,偏差仍有
值以确定变量之间关系性质及其大小,更尊重客观, 结论更具推广意义
二、概括性分析
向心性( central tendency)
指概括化的结果,可由效应值的统合值、显著性水 平的统合值以及平均效应值的置信区间来衡量
解释的问题(仅以d值为例)
• 关于d值
• 建立平均效应值的.99或.95的置信区间,看是否包围 0(Sdx为d分布的标准差;SEx为d 分布的标准误;n 为进入元分析的研究个数):
元分析还可以运用于非文献分析的研究之中 即使有时不能给出一个具体的量值,元分析的思想

Meta-分析过程的解读

Meta-分析过程的解读

结果的不一致性。
Meta-分析是系统评价(Systematic Reviews) 中一个可选择的部分。
系统评价
Meta-分析
循证医学的分级
所有RCT的 系统评价/Meta-分析 单个样本量 足够大的RCT 设有对照组但未用 随机方法分组的研究
无对照的系列病例观察
专家意见
(二)Meta-分析的原理与基本思想
原理:
统计学用抽样分布的理论来描述样本统计量的变 化规律。 在用样本信息推断总体参数时,是存在抽样误差的, 并且抽样误差的大小与样本量的大小有关。
基本思想:
理想状态:
我们把不同作者对相同问题进行的研究可以看作从同一总体 中进行抽样得到的一个随机样本,如果他们都是按照相同的设计 得到的研究结果,并且我们可以找到每一项研究的结果,这样就 可以根据上面的原理得到一个更为可靠的结果。 实际情况: 不同作者所使用的设计方案会有一定的差别,选择的实验对 象有所不同,研究结果不一定都能发表到专业杂志上。 因此实际能够得到的资料可能是不完整的,甚至是有偏性的, 如阳性结果的文章,以及和目前大家普遍能够接受的观点一致的 文章可能更容易发表在专业杂志上。 重复发表。
Risk difference (RD)
Risk on control - risk on treatment For the example before:
130 119 164 164
=0.793-0.726=0.067 Usually expressed as a %, so:6.7% 治疗减少发生事件的危险性约7个百分点
连续变量的效应值
Cohen’d=(M1-M2)/Qpooled=2t/√df=2r/ √(1-r2) Hedges g=(M1-M2)/Spooled Glass’s ∆=(M1-M2)/Scontrol

Meta分析

Meta分析

传统文献综述的主要问题
传统文献评价的结果必然存在两个问题: 一是多个研究的质量不相同 二是各个研究的样本含量的大小(权重) 不相等。 因此,传统文献综述的方法很难保证研 究结果的真实性、可靠性和科学性,尤其当 多个研究的结果不一致时,让人容易产生困 惑或误解。
Meta分析的统计目的
对多个同类独立研究的结果进行汇总和
1)二分类资料:CI包括1,无统计学意义 2)连续性资料:CI包括0,无统计学意义
连续性资料的Meta-分析
1)条件:需正态分布(偏态分布,中位数等不适合)
2)所有试验测量单位相同时,选加权均数差(WMD) 3)各试验测量单位不同时选标准化均数差(SMD) 4)方法学不及二分类资料成熟
Meta-分析图(森林图)解读
合并分析,以达到增大样本含量,提高检验 效能的目的,尤其是当多个研究结果不一致 或都没有统计意义时,采用Meta分析可得到 更加接近真实情况的统计分析结果。
Meta分析与系统评价(一)
在系统评价(systematic review)中, 当数据资料适合使用Meta分析时,用Meta分 析可以克服传统文献综述的两大问题,其分 析结果的可靠性更高;当数据资料不适合做 Meta分析时,系统评价只能解决文献评价的 问题,不能解决样本含量的问题,因此,对 其分析结论应慎重。
任何研究间的变异都可称为异质性。
在作资料综合前进行同质性检验是十分重要 的。通过采用统计学的方法对治疗效应的变 异性(异质性)进行检验,了解异质性的大小。
异质性的种类

方法学异质性
研究的类型 RCT vs 非随机对照研究, 不同质量的研究

异质性的种类


临床异质性
观察对象 年龄,性别,人种,疾病程度,病程长短,研究 间纳入/排除标准差异 研究设计 随机、盲法、样本大小 干预措施 剂量,给药途径,疗程,辨证论治,复杂干预 结果测量指标 量度,测量时间,测量方法

“meta分析”资料汇整

“meta分析”资料汇整

“meta分析”资料汇整目录一、儿童重症肺炎支原体肺炎危险因素的Meta分析二、中医药治疗月经后期的Meta分析三、10种口服中成药联合化疗治疗胃癌的网状Meta分析四、5种保肝药物治疗药物性肝损伤的网状meta分析五、中国脑卒中患者并发肺炎的危险因素的Meta分析儿童重症肺炎支原体肺炎危险因素的Meta分析儿童重症肺炎支原体肺炎是一种常见的儿科疾病,其危险因素一直是研究的重点。

本文旨在通过Meta分析的方法,综合分析儿童重症肺炎支原体肺炎的危险因素,为预防和治疗该疾病提供依据。

本文采用Meta分析的方法,收集了近年来关于儿童重症肺炎支原体肺炎危险因素的研究,对相关文献进行了筛选和评价,并采用统计软件进行了数据分析。

经过筛选和评价,最终纳入10篇相关文献进行分析。

Meta分析结果显示,儿童重症肺炎支原体肺炎的危险因素主要包括以下几个方面:年龄:年龄小于3岁的儿童是重症肺炎支原体肺炎的高危人群,可能与儿童免疫系统发育不完全有关。

性别:男性儿童相对于女性儿童更容易患重症肺炎支原体肺炎,可能与男性儿童免疫系统发育较为薄弱有关。

季节:冬季和春季是重症肺炎支原体肺炎的高发季节,可能与气温变化和空气质量有关。

家族史:有家族史的儿童更容易患重症肺炎支原体肺炎,可能与遗传因素有关。

营养不良:营养不良的儿童免疫系统发育不良,容易感染病原菌,增加重症肺炎支原体肺炎的患病风险。

其他疾病:患有其他呼吸系统疾病的儿童容易并发重症肺炎支原体肺炎,如哮喘、慢性咳嗽等。

本文通过Meta分析的方法综合分析了儿童重症肺炎支原体肺炎的危险因素,结果显示年龄、性别、季节、家族史、营养不良和其他疾病是该疾病的主要危险因素。

因此,针对这些危险因素,家长和医生应该加强儿童的预防和保健工作,减少重症肺炎支原体肺炎的发生。

对于已经患病的儿童,应该积极治疗,控制病情的发展,提高治愈率。

中医药治疗月经后期的Meta分析月经后期是妇科常见病,中医药治疗本病具有独特的优势。

meta分析简介

meta分析简介

VS
评价文献质量
对筛选后的文献进行质量评价,包括研究 设计、样本大小、数据分析方法等。
提取数据
提取数据
根据meta分析的需要,从筛选后的文献中 提取相关数据,包括样本特征、结局指标等 。
数据整理
对提取的数据进行整理,确保数据的准确性 和完整性。
统计分析
选择统计分析方法
根据研究问题和数据特征,选择合适的统计 分析方法,如随机效应模型、固定效应模型 等。
meta分析的起源与发展
起源
Meta分析起源于20世纪70年代,最初由美国学者Glass和Smith提出,用于整合和分析多个独立研究 的结果。
发展
随着统计技术和计算机技术的不断发展,Meta分析的方法不断完善和扩展,应用领域也日益广泛。目 前,Meta分析已广泛应用于医学、社会科学、心理学、生物学等多个领域。
06
meta分析案例介绍
研究背景及目的
要点一
背景
在医学和心理学领域,为了评估某种治疗或干预措施的效 果,经常需要进行大量的研究。然而,由于各种原因,不 同的研究结果之间往往存在差异。为了更全面地评估治疗 或干预措施的效果,meta分析应运而生。
要点二
目的
meta分析旨在汇总和分析多个相关研究的结果,以获得更 准确、全面的结论,为决策提供依据。
5
meta分析的应用与前 景
meta分析的应用领域
临床医学
meta分析被广泛应用于 临床医学研究,通过对多 项研究的合并和分析,提 供更全面、准确的医学证 据,指导临床实践。
社会科学
在社会科学领域,meta 分析可用于研究社会现象 、历史事件、政治经济等 问题,提供更深入、客观 的分析结果。
生物医学

meta分析

meta分析
是以0R及其95%可信区间作图的方法,可 用于描述每个原始研究效应量分布及其特征,展 示研究间结果的差异情况。形式如图。
0.4
0.6
0.8
1
1.2
1.4
1.6
二、异质性检验
(一)异质性检验常用方法是Q检验,统计公式:
Q Wi Ti T

,其中
2
T
2 W T Q WT i i ,则 W i 1
(二)确定效应量的表达形式
ቤተ መጻሕፍቲ ባይዱ
效应量(effect size,ES) 是指临床上有意义或实际 价值的数值或观察指标改变量。 观察指标为分 类变量资料,可采用RR、OR、 ARR 、NNT等; 观察指标为数值变量资料时,效应量采用加权均 数差值(WMD) 、标准化差值(SMD) 。
(三)Forest图(森林图):

3.SAS for windows:国际权威的统 计软件,可完成各种Meta分析(包括 数值、分类资料及固定效应、随机效 应模型)的统计工作。 4.SPSS for Windows :该软件是一个 统计专用软件,在其“Crosstable”菜 单中,可完成四格表资料Fleiss法的计 算工作。
4.Microsoft-Excel: 由于Meta分析的计算公式较为简 单,因此,可利用Excel的公式计算与 表格制做的功能,完成各种Meta分析 的计算工作。

2.STATA该软件

美国Computer Resource Center研制的统计软 件,从1985年起,连续推出了多个版本。 该软件可完成二分类变量和连续性变量的Meta 分析,也可以进行Meta回归分析,还可以绘制 Meta分析的相关图型,如森林图(Forest)、 漏斗图(Funnel)和L’Abbe图。

meta-analysis指南

meta-analysis指南

meta-analysis指南Meta - Analysis指南。

一、Meta - Analysis简介。

Meta - analysis(元分析)是一种对多个独立研究结果进行综合统计分析的方法。

它旨在通过整合相关研究的数据,增大样本量,提高统计效能,从而更精确地估计研究效应,解决单个研究可能存在的样本量小、结果不稳定等问题。

二、Meta - Analysis的步骤。

(一)提出研究问题。

1. 明确研究目的。

- 确定想要探究的总体效应,例如某种治疗方法对特定疾病的疗效、某个风险因素与疾病发生的关联等。

- 问题应该具有明确的研究对象、干预措施(如果有)、对照(如果有)和结局指标。

例如:“不同类型的运动干预对肥胖青少年体重减轻的效果比较”。

2. 检索相关研究。

- 选择数据库。

- 常用的数据库包括PubMed、Embase、Web of Science等。

根据研究领域的不同,可能还需要检索专业数据库,如Cochrane图书馆(在循证医学领域非常重要)、PsycINFO(心理学领域)等。

- 制定检索策略。

- 确定关键词和检索词的组合。

例如,对于上述运动干预的研究问题,可以使用“运动干预”、“肥胖青少年”、“体重减轻”等关键词,通过逻辑运算符(如“AND”、“OR”)构建检索式。

同时,要注意不同数据库的检索语法可能有所差异。

- 检索的全面性。

- 除了电子数据库,还应考虑检索灰色文献(如未发表的研究报告、学位论文等),以减少发表偏倚。

可以通过搜索特定机构的知识库、联系相关领域的专家获取未发表的研究。

(二)文献筛选。

1. 初筛。

- 根据题目和摘要,排除明显不相关的文献。

例如,如果研究题目中未涉及研究问题中的关键要素,如运动干预和肥胖青少年,就可以初步排除。

2. 复筛。

- 获取初筛后可能相关文献的全文,仔细阅读并根据预先设定的纳入和排除标准进行筛选。

纳入标准可能包括研究类型(如随机对照试验、队列研究等)、研究对象的特征(如年龄范围、疾病严重程度等)、干预措施的具体细节、结局指标的测量方法等。

系统评价Meta分析详细介绍

系统评价Meta分析详细介绍

系统评价Meta分析详细介绍目录一、系统评价Meta分析的基本概念 (2)1.1 系统评价的定义 (3)1.2 Meta分析的定义 (4)二、系统评价Meta分析的目的和意义 (4)三、系统评价Meta分析的流程 (5)3.1 明确研究问题 (6)3.2 检索文献 (7)3.3 筛选文献 (8)3.4 数据提取 (9)3.5 整理数据 (10)3.6 进行Meta分析 (11)3.7 结果解释 (12)3.8 评估偏倚风险 (13)3.9 结果的综合评价 (14)四、系统评价Meta分析中的统计方法 (15)4.1 基本统计方法 (16)4.2 元分析统计方法 (17)五、系统评价Meta分析的质量评价 (19)5.1 文献质量评价 (20)5.2 结果的一致性评价 (21)5.3 可靠性评价 (22)六、系统评价Meta分析的结果解释和应用 (24)6.1 结果的解释 (25)6.2 结果的应用 (26)6.3 对未来研究的启示 (27)七、系统评价Meta分析的局限性 (28)7.1 样本选择偏差 (29)7.2 数据质量问题 (31)7.3 不同研究结果间的异质性 (32)八、系统评价Meta分析的伦理问题 (33)8.1 保护受试者隐私 (35)8.2 避免学术不端行为 (36)九、系统评价Meta分析的未来发展趋势 (37)9.1 技术的发展 (38)9.2 方法学的创新 (39)一、系统评价Meta分析的基本概念系统评价(Systematic Review,简称SR)是一种多学科研究方法,旨在通过收集、整理和分析大量关于某一主题的独立研究结果,以便得出全面、准确和可靠的结论。

Meta分析(Metaanalysis)是系统评价的一种扩展和深化,它通过对多个独立研究的统计分析,对原始研究结果进行加权汇总,以提高研究结果的可靠性和推广性。

系统评价的目的是对现有的研究进行全面、客观和公正的评估,从而为实践提供有价值的指导。

Meta分析概述

Meta分析概述

Meta分析概述3.1Meta分析(Meta-analysis)原理Meta分析(Meta-analysis)中文翻译为“荟萃分析”。

其在英文中的定义是“The statistical analysis of a large collection of analysis results from individual studies for the purpose of integrating the findings.”中文翻译为:一种综合性强的统计方法,并且是同一课题研究的内容,而且在特定条件下对研究结果进行分析和整合。

也有国内的学者将Meta分析翻译为“综合性分析,单元分析,共性分析”等,但本文统一翻译为Meta分析。

Meta分析思想不是一蹴而就的,而且有一个比较漫长的发展过程。

最开始是1920年由Fisher统计学家做的Beecher HK.,1955)得到了确定。

到了19世纪50年代由Beecher正式提出了Meta的分析概念。

后来美国心里学家又把这种思想进行扩大。

3.2Meta分析在国外的发展状况以及历史据历料记载Meta分析是在实践中提出的。

1904年英国的统计学家把统计好的五个数学进行平均,再根据统计结果对当时英国所使用的疫苗与当时英国人死亡率之间的关系进行分析,即检验疫苗的有效与否(PearsonK,1904)。

Meta分析真正兴起的时间在70年代。

而且当时英国还把这种统计分析方法运用到军事实验,对实验结果进行科学的综合分析。

这是Meta分析开始形成的邹型。

而真正意义上Meta分析的提出还应该算是美国教学专家兼心理学家在统计心理治疗效果时把这种实用的定量分析法命名为“Meta-analysis”。

在学术界普遍认为这才是真正意义上的Meta分析。

之外,Glass(1976)又提出了EffectSize(效应值)的概念。

19世纪90年代在生态学领域有几篇有关Meta分析引起了专家的关注,所以Meta分析一直到上世纪90年代才真正应用于生态领域。

Meta分析(例子较详细)教材

Meta分析(例子较详细)教材
值( Sd 2 )
d widi wi
Sd 2

wi (di d )2 wi

widi2 wi
d2
d 114.18 0.39 294
Sd
2

60.72 294

(0.39)2

0.0544
21
3.齐性检验 H0:各研究效应量的总体均数相等 H1:各研究效应量的总体均数不全相等
由于Meta分析本质上属于观察性研究,在解释分 析结果时尤其要谨慎,主要考虑齐性及其对结果 的影响,各种偏倚的识别与控制,分析结果不能 脱离专业知识背景,要具有实际意义
15
第四节 Meta分析方法
Meta统计分析方法很多,各方法的主要步骤有两个: 一是对各个研究的效应量进行齐性检验; 二是对各个研究的效应量进行合并及总体区间估计
7
第三节 Meta分析的基本步骤
三、文献检索
检索时必须查全、查准,最好能找到所有相关文 献(包括未发表的),以减少发表偏倚对结论的 影响,这是十分重要的环节 从立题入手确定检索词,制定检索策略和检索范围
8
用计算机检索时,常用的医学网站或数据库有 ki.nt、 、Medline、 中国医院数字图书馆、中国学术期刊全文数据库、 中国生物医学文献数据库等
3
5.疾病预防干预措施的评价 6.疾病防治的成本效益分析、卫生经济学
研究 7.卫生服务评价 8.卫生决策、卫生管理评价
4
二、Meta分析的特点
1.提高统计检验效能 2.评价各研究结果的不一致或矛盾 3.发现单个研究的不足 4.具有处理大量文献的能力且不受研究数目的限
制 5.Meta分析是更高一级的统计分析,被称为“分 析

(医学课件)meta分析简介

(医学课件)meta分析简介
• 目的:对多个研究结果进行统计 分析,得出合并后的效应值,以 此对研究结果进行全面评估。
meta分析的历史与发

• 起源:Meta分析起源于20世纪70 年代,由美国Glass和Smith提出 。
• 发展:Meta分析经历了三个阶段 :定性、定量和综合评价阶段。
meta分析与系统评价
、综述的关系
• 关系:Meta分析是系统评价和综 述的高级形式。
使用数据提取表格
为了准确、全面地提取数据,可以使用数据提取表格进行记录和整理。
统计分析
选择合适的统计分析方法
01
meta分析需要选择合适的统计分析方法,如随机效应模型、Leabharlann 固定效应模型等。进行统计分析
02
根据选定的统计分析方法,对提取的数据进行统计分析,得出
合并后的效应值及其可信区间。
结果解释与报告
03
根据统计分析结果,解释合并效应的意义和临床价值,并撰写
meta分析报告。
05
meta分析的局限性及挑战
数据汇总和解释的限制
纳入研究间的异质性
在将不同研究结果进行合并时,可能存在 研究设计、样本大小、研究人群、结局评 估等方面的差异,影响结果的稳定性和可 解释性。
数据缺失和错误
纳入的研究中可能存在数据缺失和错误, 例如缺失关键结局指标、数据错误或数据 报告不完全等,导致数据分析不完整或不 准确。
原始研究设计和方法学异质性
由于meta分析纳入的研究可能来自不同的研究设计和方法学,这些差异可能导致合并后的结果存在偏 差。
结局指标和测量方法的异质性
不同的研究可能采用不同的结局指标和测量方法,这可能导致合并后的结果存在偏差。
06
meta分析的实例演示

Meta分析的简单介绍---文本资料

Meta分析的简单介绍---文本资料
Meta分析的统计学过程
吉林大学白求恩第一医院 结直肠肛门外科 张婷
“Meta”一词源于希腊文,意为“more comprehensive”,即“更广泛、更全面”
上个世纪60年代开始,在医学文献中,陆续出 现许多对多个独立研究的统计量进行合并的报 道。 英国心理学家G.V.Glass,在1976年首先将这 种多个同类研究的统计量合并方法称为 “Meta-Analysis”。该方法现在已广泛应用于 医学和健康领域,尤其是针对疾病的诊断、治 疗、预防和病因等问题的综合评价。
RR和OR的森林图



RR和OR的森林图(forest plots) 无效线竖线的横轴尺度为1; 每条横线为该研究的95%可信区间上下限的 连线; 线条长短直观地表示了可信区间范围的大小; 线条中央的小方块为RR或OR值的位置,其 方块大小为该研究权重大小。 若某个研究的95%可信区间的线条横跨过无 效竖线,即该研究无统计学意义,反之,若 该横线落在无效竖线的左侧或右侧,该研究 有统计学意义。
异质性定义
广义上用于描述试验的参与者、试验 的干预措施和多个研究测量结果的变 异,即各研究的内在真实性的变异。
种类: 临床异质性 方法异质性 统计学异质性
统计学异质性




是指干预效果的评价在不同试验间的变异, 它是研究间的临床和方法学上变异联合作用 的结果。 通常将Meta分析的统计学异质性简称为“异 质性”,它是以各研究之间可信区间(CI) 的重合程度来度量异质性的大小; 多个研究间的CI重合程度越大,存在统计学 异质性的可能性就越小,反之,各研究间存 在统计学异质性的可能性就越大。 异质性分析的意义:Meta分析的核心计算是 合并(相加),按统计原理,只有同质的资 料才能进行合并或比较等统计分析,反之则 不能。

META分析

META分析

第一节定义Meta分析,又称“荟萃分析”,“元分析”、“综合分析”,也有人翻译为“分析的分析”、“资料的再分析”等。

Meta分析可简单归纳为定量的系统评价。

Glass把Meta分析定义为“以综合现有的发现为目的,对单个研究结果的集合的统计分析方法”。

Meta分析解释如:对具有共同研究目的相互独立的多个研究结果给予定量分析,合并分析,剖析研究间差异特征,综合评价研究结果。

英国心理学家G1ass认为Meta分析是为达到统一研究目的,对收集到的多个研究进行的综合统计分析,是数据收集和相关信息处理的一系列统计原则和过程,而不是一个简单的方法。

Finney则把对不同来源科学技术信息的定量化汇总分析,统称为Meta分析。

Meta分析是汇总多个研究的结果并分析评价其合并效应量的一系列过程,包括提出研究问题、制定纳入和排除标准、检索相关研究、汇总基本信息、综合分析并报告结果等。

G1ass最早在教育学研究中使用了Meta分析。

二十世纪八十年代中期开始被引入到临床随机对照试验以及观察性的流行病学研究中。

在过去的15年内,有大约几百篇有关Meta分析的文章出现在医学杂志上。

Meta分析结果能够帮助解决重要的公共健康问题或使个体直接受益,同时能作为可靠的证据指导临床实践及卫生决策的科学化。

Meta分析可以用于分析危险因素较弱,但为公众所关心的重要健康问题(如被动吸烟与肺癌、低剂量辐射与白血病、避孕药与乳腺癌等);可以得到危险因素定量化的综合效应(如标准化死亡比、相对危险比);还可用于较复杂的剂量反应关系研究及诊断试验研究的综合分析。

第二节Meta分析能解决的问题一、放大统计功效在临床研究中,如果样本量小,则结果受偶然因素的影响就大,且难以明确肯定或排除某些相对较弱的药物作用,而这些作用对临床来说可能又是重要的。

如果要从统计学上来肯定或排除这些作用,研究所需要的样本量可能较大。

Meta分析通过整合大量的临床研究报告,增加了样本量,增加了结论的统计功效。

Meta-analysis经典

Meta-analysis经典

Introduction to Meta-AnalysisMichael BorensteinLarry HedgesHannah RothsteinJuly 1, 2007 (C) M Borenstein, L Hedges, H Rothstein 2007DedicationDedicated in honor of Sir Iain Chalmers and to the memory of Dr. Thomas Chalmers.Dedication (2)Acknowledgements (9)Preface (11)An ethical imperative (11)From narrative reviews to systematic reviews (12)The systematic review and meta-analysis (13)Meta-Analysis as part of the research process (14)Role in informing policy (14)Role in planning new studies (14)Role in papers that report results for primary studies (15)A new perspective (15)The intended audience for this book (16)An outline of this book’s contents (16)What this book does not cover (17)Web site (19)Introduction (20)Introduction (20)The streptokinase meta-analysis (20)What we know now (23)The treatment effect (24)The combined effect (24)Heterogeneity in effects (25)The Swiss patient (27)Systematic reviews and meta-analyses (28)Key points (29)Section: Treatment effects and effect sizes (31)July 1, 2007 (C) M Borenstein, L Hedges, H Rothstein 2007Overview (31)Effect size and precision in primary studies and in meta-analysis (32)Treatment effects, effect sizes, and point estimates (33)Treatment effects (34)Effect sizes (34)Point estimates (34)Standard error, confidence interval, and variance (34)Standard error (34)Confidence interval (35)Variance (35)Effect sizes rather than p-values (35)Effect size and precision (36)The significance test (37)Comparing significance tests and confidence intervals (38)How the two approaches differ (41)Implications for primary studies (43)Implications for meta-analysis (45)The analysis should focus on effect sizes (48)The irony of type I and type II errors being switched (48)Key points (49)Section: Effect size indices (50)Overview (51)Data based on means and standard deviations (52)Raw Mean difference (52)Standardized mean difference (52)Variants of the standardized mean difference (53)Computational formulas (53)Studies that used independent groups (54)Standard error (assuming a common SD) (55)Cohen’s d (56)Hedges’s G (56)Studies that used pre-post scores or matched pairs (60)Raw Mean Difference (61)Standardized mean difference, d (62)Hedges’s G (64)Other indices based on means in two groups (66)Binary data in two groups (2x2 tables) (67)Risk ratios (67)Odds ratios (67)Risk Difference (67)Computational formulas (68)Means in one-group studies (79)July 1, 2007 (C) M Borenstein, L Hedges, H Rothstein 2007Factors that affect precision (81)Sample size (81)Study design (82)Section: Fixed effect vs. random effects models (85)Section: Fixed effect vs. random effects models (86)Overview (86)Fixed effect model (88)Definition of a combined effect (88)Assigning weights to the studies (88)Random effects model (94)Definition of a combined effect (94)Assigning weights to the studies (95)Fixed effect vs. random effects models (105)The concept (105)Definition of a combined effect (105)Computing the combined effect (106)Extreme effect size in large study (106)Extreme effect size in small study (108)Confidence interval width (109)Which model should we use? (114)Mistakes to avoid in selecting a model (115)Section: Heterogeneity within groups (116)Overview (116)Heterogeneity in effect sizes – Part 1 (117)Testing the null hypothesis that the studies are homogeneous (123)Benchmarks for I-squared (126)Limitations of I-squared (131)Confidence intervals for I-squared (138)Choosing the appropriate model (139)Summary (140)Section: Heterogeneity across groups (142)Analysis of variance (143)Fixed or random effects within groups (154)Pooling effects across groups (167)Meta-regression (170)Section: Other issues (190)July 1, 2007 (C) M Borenstein, L Hedges, H Rothstein 2007Cumulative meta analysis (192)Complex data structures (198)Independent subgroups within a study (201)Subgroup as the unit of analysis (Option 1) (201)Study as the unit of analysis (Option 2) (205)Comparing options 1 and 2 (206)Computing effect across subgroups (Option 3) (209)Multiple outcomes, time-points, or comparison groups within a study (212)Multiple outcomes within a study (212)Multiple time-points within a study (223)Multiple comparisons within a study (224)Comparing outcomes, time-points, or comparison groups within a study (225)Multiple outcomes within a study (227)Multiple time-points within a study (237)Multiple comparisons within a study (238)Forest plots (242)Sensitivity analyses (243)Publication bias (243)Evidence that publication bias exists (244)A short history of publication bias (245)Impact of publication bias (246)Overview of computational methods for addressing bias (246)Getting a sense of the data (247)Is there evidence of bias? (247)The funnel plot (247)Tests for asymmetry (248)Is it possible that the observed effect is solely a function of bias (248)The failsafe N (248)Orwin’s failsafe N (249)What is the effect size after we adjust for bias? (250)Duval and Tweedie’s trim and fill (250)Illustrative example (252)What is the effect size after we adjust for bias? (255)Prospective registration of trials (259)July 1, 2007 (C) M Borenstein, L Hedges, H Rothstein 2007Publication bias in perspective (261)Publication bias over the life span of a research question (262)Other kinds of suppression bias (262)Working with data in more than one format (263)Alternative weighting schemes (264)Mantel-Haenszel Weights (264)Odds ratio (264)Risk ratio (270)Risk difference (274)Peto weights (279)Odds ratio (279)Criticisms of meta-analysis (282)Overview (282)Meta-analysis focuses on the combined effect (282)Apples and oranges (283)Meta-analysis and RCT’s may not agree (284)Publication bias (288)Many meta-analyses are performed poorly (289)Is the narrative review a better choice? (289)Is the meta-analysis inherently problematic? (291)Other Approaches to Meta-Analysis (293)Psychometric Meta-Analysis (Hunter-Schmidt) (293)Overview (293)Sample Size Weighting (293)Use of correlations (293)Corrections for Artifacts (293)Sampling error variance (293)Error of measurement (293)Restriction of range (293)Assessment of heterogeneity (293)Bayesian Meta-Analysis (293)Worked examples (294)Medical examples using binary data (294)Social science examples using continuous data (294)Social science examples using psychometric meta-analysis (294)July 1, 2007 (C) M Borenstein, L Hedges, H Rothstein 2007Resources for meta-analysis (296)Software (296)Computer programs (296)Professional organizations (300)Cochrane and Campbell (300)Books on meta-analysis (300)Web sites (301)References (303)July 1, 2007 (C) M Borenstein, L Hedges, H Rothstein 2007Section: Fixed effect vs. random effects models OverviewOne goal of a meta-analysis will often be to estimate the overall, or combined effect.If all studies in the analysis were equally precise we could simply compute the mean of the effect sizes. However, if some studies were more precise than others we would want to assign more weight to the studies that carried more information. This is what we do in a meta-analysis. Rather than compute a simple mean of the effect sizes we compute a weighted mean, with more weight given to some studies and less weight given to others.The question that we need to address, then, is how the weights are assigned. It turns out that this depends on what we mean by a “combined effect”. There are two models used in meta-analysis, the fixed effect model and the random effects model. The two make different assumptions about the nature of the studies, and these assumptions lead to different definitions for the combined effect, and different mechanisms for assigning weights.Definition of the combined effectUnder the fixed effect model we assume that there is one true effect size which is shared by all the included studies. It follows that the combined effect is our estimate of this common effect size.By contrast, under the random effects model we allow that the true effect could vary from study to study. For example, the effect size might be a little higher if the subjects are older, or more educated, or healthier, and so on. The studies included in the meta-analysis are assumed to be a random sample of the relevant distribution of effects, and the combined effect estimates the mean effect in this distribution.Computing a combined effectUnder the fixed effect model all studies are estimating the same effect size, and so we can assign weights to all studies based entirely on the amount of information captured by that study. A large study would be given the lion’s share of the weight, and a small study could be largely ignored.July 1, 2007 (C) M Borenstein, L Hedges, H Rothstein 2007By contrast, under the random effects model we are trying to estimate the mean of a distribution of true effects. Large studies may yield more precise estimates than small studies, but each study is estimating a different effect size, and we want to be sure that all of these effect sizes are included in our estimate of the mean. Therefore, as compared with the fixed effect model, the weights assigned under random effects are more balanced. Large studies are less likely to dominate the analysis and small studies are less likely to be trivialized. Precision of the combined effectUnder the fixed effect model the only source of error in our estimate of the combined effect is the random error within studies. Therefore, with a large enough sample size the error will tend toward zero. This holds true whether the large sample size is confined to one study or distributed across many studies. By contrast, under the random effects model there are two levels of sampling and two levels of error. First, each study is used to estimate the true effect in a specific population. Second, all of the true effects are used to estimate the mean of the true effects. Therefore, our ability to estimate the combined effect precisely will depend on both the number of subjects within studies (which addresses the first source of error) and also the total number of studies (which addresses the second).How this section is organizedThe two chapters that follow provide detail on the fixed effect model and the random effects model. These chapters include computational details and worked examples for each model. Then, a chapter highlights the differences between the two.July 1, 2007 (C) M Borenstein, L Hedges, H Rothstein 2007July 1, 2007 (C) M Borenstein, L Hedges, H Rothstein 2007Fixed effect modelDefinition of a combined effectIn a fixed effect analysis we assume that all the included studies share acommon effect size, μ. The observed effects will be distributed about μ, with avariance σ2 that depends primarily on the sample size for each study.In this schematic the observed effect in Study 1, T 1, is a determined by thecommon effect μ plus the within-study error ε1. More generally, for any observedeffect T i ,i i T μe =+ (2.2)Assigning weights to the studiesIn the fixed effect model there is only one level of sampling, since all studies aresampled from a population with effect size μ. Therefore, we need to deal withonly one source of sampling error – within studies (e).July 1, 2007 (C) M Borenstein, L Hedges, H Rothstein 2007Since our goal is to assign more weight to the studies that carry moreinformation, we might propose to weight each study by its sample size, so that a study with 1000 subjects would get 10 times the weight of a study with 100 subjects. This is basically the approach used, except that we assign weights based on the inverse of the variance rather than sample size. The inversevariance is roughly proportional to sample size, but is a more nuanced measure (see notes), and serves to minimize the variance of the combined effect.Concretely, the weight assigned to each study is1i iw v = (2.3)where v i is the within-study variance for study (i ). The weighted mean (T •) is then computed as11ki ii kii w TT w=•==∑∑ (2.4)that is, the sum of the products w i T i (effect size multiplied by weight) divided by the sum of the weights. The variance of the combined effect is defined as the reciprocal of the sum of the weights, or11kii v w •==∑ (2.5)and the standard error of the combined effect is then the square root of the variance,()SE T •= (2.6)The 95% confidence interval for the combined effect would be computed as1.96*()Lower Limit T SE T ••=− (2.7)1.96*()Upper Limit T SE T ••=+ (2.8)July 1, 2007 (C) M Borenstein, L Hedges, H Rothstein 2007Finally, if one were so inclined, the Z -value could be computed using()T Z SE T ••= (2.9)For a one-tailed test the p-value would be given by1Φ(||)p Z =− (2.10)(assuming that the effect is in the hypothesized direction), and for a two-tailed test by()()21Φ||p Z ⎡⎤=−⎣⎦ (2.11)Where Φ is the standard normal cumulative distribution function.Illustrative exampleThe following figure is the forest plot of a fictional meta-analysis that looked at the impact of an intervention on reading scores in children.July 1, 2007 (C) M Borenstein, L Hedges, H Rothstein 2007In this example the Carroll study has a variance of 0.03. The weight for that study would computed as1133.333(0.03)w ==and so on for the other studies. Then,101.8330.3968256.667T •==10.0039256.667v •==()0.0624SE T •== 0.3968 1.96*0.06240.2744Lower Limit =−= 0.3968 1.96*0.06240.5191Upper Limit =−=0.39686.35630.0624Z ==()()11Φ|6.3563|.0001T p =−<()()221Φ|6.3563|.0001Tp ⎡⎤=−<⎣⎦The fixed effect computations are shown in this spreadsheetJuly 1, 2007 (C) M Borenstein, L Hedges, H Rothstein 2007Column (Cell) LabelContentExcel Formula* See formula(Section 1) Effect size and weights for each studyA Study name EnteredB Effect size EnteredC Variance Entered(Section 2) Compute WT and WT*ES for each studyD Variance within study =$C3E Weight =1/D3 (2.3)F ES*WT =$B3*E3 Sum the columns E9 Sum of WT =SUM(E3:E8) F9Sum of WT*ES=SUM(F3:F8)(Section 3) Compute combined effect and related statisticsF13 Effect size =F9/E9 (2.4)F14 Variance =1/E9 (2.5) F15 Standard error =SQRT(F14) (2.6) F16 95% lower limit =F13-1.96*F15 (2.7) F17 95% upper limit =F13+1.96*F15 (2.8)F18 Z-value =F13/F15 (2.9) F19 p-value (1-tailed) =(1-(NORMDIST(ABS(F18),0,1,TRUE))) (2.10) F20 p-value (2-tailed) =(1-(NORMDIST(ABS(F18),0,1,TRUE)))*2(2.11)CommentsSome formulas include a “$”. In Excel this means that the reference is to a specific column. These are not needed here, but will be needed when we expand this spreadsheet in the next chapter to allow for other computational models.Inverse variance vs. sample size.As noted, weights are based on the inverse variance rather than the sample size. The inverse variance is determined primarily by the sample size, but it is a more nuanced measure. For example, the variance of a mean difference takes account not only of the total N, but also the sample size in each group. Similarly, the variance of an odds ratio is based not only on the total N but also on the number of subjects in each cell.July 1, 2007 (C) M Borenstein, L Hedges, H Rothstein 2007Random effects modelThe fixed effect model, discussed above, starts with the assumption that the true effect is the same in all studies. However, this is a difficult assumption to make in many (or most) systematic reviews. When we decide to incorporate a group of studies in a meta-analysis we assume that the studies have enough in common that it makes sense to synthesize the information. However, there is generally no reason to assume that they are “identical” in the sense that the true effect size is exactly the same in all the studies.For example, assume that we’re working with studies that compare the proportion of patients developing a disease in two groups (vaccination vs. placebo). If the treatment works we would expect the effect size (say, the risk ratio) to be similar but not identical across studies. The impact of the treatment impact might be more pronounced in studies where the patients were older, or where they had less natural immunity.Or, assume that we’re working with studies that assess the impact of an educational intervention. The magnitude of the impact might vary depending on the other resources available to the children, the class size, the age, and other factors, which are likely to vary from study to study.We might not have assessed these covariates in each study. Indeed, we might not even know what covariates actually are related to the size of the effect. Nevertheless, experience says that such factors exist and may lead to variations in the magnitude of the effect.Definition of a combined effectRather than assume that there is one true effect, we allow that there is a distribution of true effect sizes. The combined effect therefore cannot represent the one common effect, but instead represents the mean of the population of true effects.July 1, 2007 (C) M Borenstein, L Hedges, H Rothstein 2007July 1, 2007 (C) M Borenstein, L Hedges, H Rothstein 2007In this schematic the observed effect in Study 1, T 1, is a determined by the true effect θ1 plus the within-study error ε1. In turn, θ1, is determined by the mean of all true effects, μ and the between-study error ξ1. More generally, for any observed effect T i , i i i i i T θe μεe =+=++ (3.1)Assigning weights to the studiesUnder the random effects model we need to take account of two levels of sampling, and two source of error. First, the true effect sizes θ are distributed about μ with a variance τ2 that reflects the actual distribution of the true effects about their mean. Second, the observed effect T for any given θ will bedistributed about that θ with a variance σ2 that depends primarily on the sample size for that study. Therefore, in assigning weights to estimate μ, we need to deal with both sources of sampling error – within studies (e ), and between studies (ε).Decomposing the varianceThe approach of a random effects analysis is to decompose the observed variance into its two component parts, within-studies and between-studies, and then use both parts when assigning the weights. The goal will be to reduce both sources of imprecision.The mechanism used to decompose the variance is to compute the total variance (which is observed) and then to isolate the within-studies variance. The difference between these two values will give us the variance between-studies, which is called Tau-squared (τ2). Consider the three graphs in the following figure.In (A), the studies all line up pretty much in a row. There is no variance between studies, and therefore tau-squared is low (or zero).In (B) there is variance between studies, but it is fully explained by the variance within studies. Put another way, given the imprecision of the studies, we would July 1, 2007 (C) M Borenstein, L Hedges, H Rothstein 2007July 1, 2007 (C) M Borenstein, L Hedges, H Rothstein 2007expect the effect size to vary somewhat from one study to the next. Therefore, the between-studies variance is again low (or zero).In (C ) there is variance between studies. And, it cannot be fully explained by the variance within studies, since the within-study variance is minimal. The excess variation (between-studies variance), will be reflected in the value of tau-squared.It follows that tau-squared will increase as either the variance within-studies decreases and/or the observed variance increases.This logic is operationalized in a series of formulas. We will compute Q , which represents the total variance, and df , which represents the expected variance if all studies have the same true effect. The difference, Q - df , will give us the excess variance. Finally, this value will be transformed, to put it into the same scale as the within-study variance. This last value is called Tau-squared.The Q statistic represents the total variance and is defined as()2.1ki i i Q w T T ==−∑ (3.2)that is, the sum of the squared deviations of each study (T i ) from the combined mean (.T ). Note the “w i ” in the formula, which indicates that each of the squared deviations is weighted by the study’s inverse variance. A large study that falls far from the mean will have more impact on Q than would a small study in the same location. An equivalent formula, useful for computations, is21211k i i ki i i ki ii w T Q w T w ===⎛⎞⎜⎟⎝⎠=−∑∑∑ (3.3)Since Q reflects the total variance, it must now be broken down into itscomponent parts. If the only source of variance was within-study error, then the expected value of Q would be the degrees of freedom for the meta-analysis (df ) where()1df Number Studies =− (3.4)This allows us to compute the between-studies variance, τ2, asJuly 1, 2007 (C) M Borenstein, L Hedges, H Rothstein 2007if Q > df 0 if Q dfQ dfτC 2−⎧⎪=⎨⎪≤⎩ (3.5)where2i iiw C w w=−∑∑∑ (3.6)The numerator, Q - df , is the excess (observed minus expected) variance. The denominator, C , is a scaling factor that has to do with the fact that Q is a weighted sum of squares. By applying this scaling factor we ensure that tau-squared is in the same metric as the variance within-studies.In the running example,2101.83353.20812.8056256.667Q ⎛⎞=−=⎜⎟⎝⎠(61)5df =−=15522.222256.667196.1905256.667C ⎛⎞=−=⎜⎟⎝⎠212.805650.0398196.1905τ−==Assigning weights under the random effects modelIn the fixed effect analysis each study was weighted by the inverse of itsvariance. In the random effects analysis, too, each study will be weighted by the inverse of its variance. The difference is that the variance now includes theoriginal (within-studies) variance plus the between-studies variance, tau-squared.Note the correspondence between the formulas here and those in the previous chapter. We use the same notations, but add a (*) to represent the randomeffects version. Concretely, under the random effects model the weight assigned to each study isJuly 1, 2007 (C) M Borenstein, L Hedges, H Rothstein 2007**1i iw v =(3.7)where v*i is the within-study variance for study (i ) plus the between-studies variance, tau-squared. The weighted mean (T •*) is then computed as*1*1*ki i i kii w TT w=•==∑∑ (3.8)that is, the sum of the products (effect size multiplied by weight) divided by the sum of the weights.The variance of the combined effect is defined as the reciprocal of the sum of the weights, or*.*11ki i v w ==∑ (3.9)and the standard error of the combined effect is then the square root of the variance,(*)SE T •=The 95% confidence interval for the combined effect would be computed as** 1.96*(*)Lower Limit T SE T ••=− (3.11)** 1.96*(*)Upper Limit T SE T ••=+ (3.12)Finally, if one were so inclined, the Z -value could be computed using**(*)T Z SE ••= (3.13)July 1, 2007 (C) M Borenstein, L Hedges, H Rothstein 2007The one-tailed p -value (assuming an effect in the hypothesized direction) is given by ()**1Φ||p Z =− (3.14)and the two-tailed p -value by()**21Φ||p Z ⎡⎤=−⎣⎦(3.15)Where Φ is the standard normal cumulative distribution function.Illustrative exampleThe following figure is based on the same studies we used for the fixed effect example.Note the differences from the fixed effect model.• The weights are more balanced. The boxes for the large studies such as Donat have decreased in size while those for the small studies such as Peck have increase in size.July 1, 2007 (C) M Borenstein, L Hedges, H Rothstein 2007• The combined effect has moved toward the left, from 0.40 to 0.34. This reflects the fact that the impact of Donat (on the right) has been reduced. • The confidence interval for the combined effect has increased in width.In the running example the weight for the Carroll study would be computed as*1114.330(0.0300.040)(0.070)i w ===+and so on for the other studies. Then,30.207*0.344287.747T •==1*0.011487.747v •=(*)0.1068SE T •== *0.3442 1.96*0.10680.1350Lower Limit =−=*0.3968 1.96*0.10680.5535Upper Limit =−=0.3442* 3.22470.1068Z ==()()11Φ(3.2247)0.0006T P ABS =−=()()21Φ(3.2247)*20.0013T P ABS ⎡⎤=−=⎣⎦These formulas are incorporated in the following spreadsheetThis spreadsheet builds on the spreadsheet for a fixed effect analysis. Columns A-F are identical to those in that spreadsheet. Here, we add columns for tau-squared (columns G-H) and random effects analysis (columns I-M).Note that the formulas for fixed effect and random effects analyses are identical, the only difference being the definition of the variance. For the fixed effect analysis the variance (Column D) is defined as the variance within-studies (for example D3=C3). For the random effects analysis the variance is defined as the variance within-studies plus the variance between-studies (for example,K3=I3+J3).July 1, 2007 (C) M Borenstein, L Hedges, H Rothstein 2007July 1, 2007 (C) M Borenstein, L Hedges, H Rothstein 2007Column (Cell) LabelContentExcel Formula*See formula(Section 1) Effect size and weights for each studyA Study name EnteredB Effect size EnteredC Variance Entered(Section 2) Compute fixed effect WT and WT*ES for each studyD Variance within study =$C3E Weight =1/D3(2.3)F ES*WT =$B3*E3Sum the columnsE9 Sum of WT =SUM(E3:E8)F9Sum of WT*ES =SUM(F3:F8) (Section 3) Compute combined effect and related statistics for fixed effect modelF13 Effect size =F9/E9(2.4) F14 Variance =1/E9 (2.5)F15 Standard error =SQRT(F14) (2.6)F16 95% lower limit =F13-1.96*F15 (2.7) F17 95% upper limit =F13+1.96*F15 (2.8) F18 Z-value =F13/F15 (2.9) F19 p-value (1-tailed) =(1-(NORMDIST(ABS(F18),0,1,TRUE))) (2.10) F20 p-value (2-tailed) =(1-(NORMDIST(ABS(F18),0,1,TRUE)))*2 (2.11) (Section 4) Compute values needed for Tau-squared G3 ES^2*WT =B3^2*E3H3 WT^2 =E3^2 Sum the columns G9 Sum of ES^2*WT =SUM(G3:G8) H9 Sum of WT^2 =SUM(H3:H8)(Section 5) Compute Tau-squaredH13 Q =G9-F9^2/E9 (3.3) H14 Df=COUNT(B3:B8)-1 (3.4)H15 Numerator =MAX(H13-H14,0) H16 C=E9-H9/E9 (3.6) H17 Tau-sq =H15/H16 (3.5)(Section 6) Compute random effects WT and WT*ES for each studyI3 Variance within =$C3J3 Variance between =$H$17 K3 Variance total =I3+J3 L3 WT =1/K3 (3.7)M3 ES*WT =$B3*L3Sum the columnsL9 Sum of WT =SUM(L3:L8) M9 Sum of ES*WT =SUM(M3:M8)(Section 7) Compute combined effect and related statistics for random effects modelJuly 1, 2007 (C) M Borenstein, L Hedges, H Rothstein 2007M13 Effect size =M9/L9 (3.8) M14 Variance =1/L9 (3.9) M15 Standard error =SQRT(M14) (3.10) M16 95% lower limit =M13-1.96*M15 (3.11) M17 95% upper limit =M13+1.96*M15 (3.12) M18 Z-value =M13/M15 (3.13) M19 p-value (1-tailed) =(1-(NORMDIST(ABS(M18),0,1,TRUE))) (3.14) M20 p-value (2-tailed) =(1-(NORMDIST(ABS(M18),0,1,TRUE)))*2 (3.15)。

Meta分析

Meta分析

LCPULA),相比于安慰剂对照,能否预防或延缓肾病
的进展? 入选标准:1)RCT;2)使用n-3 LCPUFAs,包括EPA和 /或DHA或鱼油作为干预措施;3)安慰剂作为对照;4) 报告了肾功能的结果。 排除器官移植和终末期肾病患者。
• 根据入选标准选择合格的研究
文献的筛选

初筛:题目、摘要阅读
• 提出临床问题-PICO原则

P population/patients 人群/病人群
– 关心的是哪一类病人群或对象
• I intervention/exposure 干预/暴露 – 我们所感兴趣的治疗策略,诊断试验或暴露是什么?如一种药物
、食物,外科手术方式,诊断试验或暴露于一种化学物质?
• C comparison or control 比较物或对照 – 我们感兴趣的干预措施相比较的对照物、处理策略、试验或暴露 是什么? • O outcome 结果 – 病人经干预处理后得到的结果是什么?
案例: 皮肤感染 一个28岁男性,在过去8个月中反复发生 皮肤疖肿,应用过几个疗程的抗生素。 能否预防复发?
问题 • P opulation :人群
– 复发性皮肤感染的病人


I
ndicator (intervention, test, etc):干预
– 抗生素治疗
C omparator:对照
– 对照不治疗
不清楚
法(1分) 不恰当
试验描述为随机试验,但没有告知随机分配产生的方
如采用交替分配或类似方法的半随机化(0分)
盲法:
恰当 不祥 使用完全一致的安慰剂或类似的方法(2分) 试验称为双盲法,但未交代具体的方法(1分)
非盲法 未采用双盲法或盲的方法不恰当(0分)
  1. 1、下载文档前请自行甄别文档内容的完整性,平台不提供额外的编辑、内容补充、找答案等附加服务。
  2. 2、"仅部分预览"的文档,不可在线预览部分如存在完整性等问题,可反馈申请退款(可完整预览的文档不适用该条件!)。
  3. 3、如文档侵犯您的权益,请联系客服反馈,我们会尽快为您处理(人工客服工作时间:9:00-18:30)。

万方数据
生堡放射堂盘壶垫!Q生!圣旦筮丝鲞筮!!魍g丛!』娶璺堂型:坠婴!出!垫!Q:!吐丝:盟垒!兰
probabilities,GCP)。
公式中验前概率为临床医师根据患者病史、症状和 体征等临床指标所估计的患病概率,验后概率指根 据某项检查结果(如阳性或阴性)所获得的患病 概率。
验前比=黼

(1) (2)
金标准为CAG;(5)所有纳入研究对象均经DSCT 和CAG检查;(6)以冠状动脉直径狭窄I>50%作为
者中诊断显著性冠状动脉狭窄的价值,为临床决策 提供依据。 资料与方法 一、文献检索
根据汇总敏感度和特异度值计算似然比;(3)建立
分层汇总受试者工作特征曲线(hierarchical
receiver operating characteristic summary
curve,HSROC),曲线
计算机检索Medline数据库、中文期刊全文数
分析结果的临床应用价值。结果共纳入10篇英文文献,研究对象127l例,平均5.6%(33/590)患
者和2.3%(271/11 745)节段不可评价;95%可信区问的汇总敏感度、特异度分别为99%(97%~ 99%)和86%(79%~90%),阳性和阴性似然比分别为6.84和0.01。患者验前概率<84%,诊断为 阴性时,患病概率<5%;验前概率>13%,诊断为阳性时,患病概率>50%。结论DSCT冠状动脉成 像诊断CAD有很高的准确性和临床庇用价值,彳只.仍不能取代传统冠状动脉造影。 【关键词】冠状动脉疾病;体层摄影术,x线计算机;诊断;Meta分析
performance
in the detection of CAD;however,it could
completely
replace conventional coronary angiography.
【Key words】Coronary
artery
disease;Tomography,X-ray computed;Diagnosis;Meta—analysis
Total of
ten
articles enrolled
were
in this study.included 1271
mean
rate
of nonevaluable patients and set目ments
5.6%
(33/590)and 2.3%(271/1l 745),respectively.The pooled statistical results were as follows:the sensitivity and specificity were 99%(97%—一99%)and 86%(79%—-90%),respectively;the positive and
of conditional
disease”或“coronary
computed
tomography”或
angiography”、
CT”、“coronary
specificity”。中文文献检索关键词
为:双源CT、冠状动脉。为避免漏查文献,对所纳入 文献和相关综述中提供的参考文献进行二次检索。 二、文献筛选 由本研究的2名作者独立进行筛选,如有不一 致时经讨论确定。 1.纳入标准:(1)中文或英文文献;(2)研究对 象为可疑CAD患者;(3)采用DSCT诊断CAD;(4)
tomography(DSCT)in
January 2006
China National
the diagnosis of coronary artery
disease(CAD).Methods
or
Literatures publicated Medline,
July 2009,including in Enhe (95%CI)were acquired
calculated based
on
Was enrolled if it:(1)used DSCT angiography鹳the diagnostic test for the detection of stenosis(≥50%diameter stenosis)in patients with suspected CAD;(2)used coronary reference standard.The pooled sensitivity and specificity of the 95%confidence interval
Chinese languages,were searched in
Knowledge study
Infrastructure(CNKI),and Chinese Medical Assosiation Digital Periodicsis
(CMADP).A
significant coronary
期或低质量文献;(5)未采用盲法。 三、数据提取
一、文献检索及筛选 共查得文献309篇,其中英文47篇,中文 262篇。根据文题和(或)摘要排除不相关文献 286篇;在剩余23篇中,阅读全文后根据筛选标准
由2名本研究文献筛选者独立迸行提取,如有 不一致经讨论确定。
1.一般信息:(1)研究作者、发表时间、研究国 家;(2)研究对象总数和不可评价数、平均年龄、男 性数量、证实为CAD患者数;(3)接受CTA检查时 使用降心率药物对象数量,CTA扫描时平均心率; (4)CTA扫描辐射量;(5)节段评价基础及不可评价 节段数;(6)平均钙化积分。
心血管疾病是我国居民首要致死因素,其中冠
并且患病率和病死率呈逐年上升的趋势…。近半 个世纪以来,冠状动脉造影(coronary
angiography,
状动脉疾病(coronary
artery
disease,CAD)占第2位,
CAG)一直是诊断CAD的金标准,同时还可起到治
DOI:10.3760/cma.j.issn.1005-1201.2010.12.014 作者单位:710061西安交通大学医学院第一附属医院影像科 (李敏、麻少辉、张明);西安交通大学医学院第二附属医院影像科 (张晓娜) 通信作者:张明,Email:zmmri@163.corn
阳性标准。 2.排除标准:(1)研究对象不足20例;(2)纳
验后概率=怒(3)
验后比=验前比×似然比
6.统计学软件:Stata 10.0,P<0.05为差异有
统计学意义。

入了已知CAD患者,且未提供基于可疑CAD患者
分析数据;(3)无法获得诊断结果绝对值;(4)数据
重叠(以Email形式联系通信作者确认),小样本、早
生生越蕴堂盘查垫!Q生!!旦筮丝盏筮!!魍垦堕!』壁型!!:旦塑!出!垫!Q:!堕竺:盟些!兰
・1285・
.胸部放射学.
双源CT诊断冠状动脉疾病的Meta分析
李敏张晓娜麻少辉张明
【摘要】
目的运用Meta分析方法评价双源cT(DSCT)诊断冠状动脉疾病(CAD)的价值。
方法采用Medline数据库、中文期刊全文数据库以及中华医学会数字化期刊数据库检索2006年 1月至2009年7月国内外公开发表的中英文文献:(1)在町疑CAD患者中以DSCT血管成像诊断显 著性冠状动脉狭窄(狭窄卣径≥50%);(2)以冠状动脉造影作为诊断金标准。基于患者水平采用双 变量随机效应模型和分层综合受试者工作特征曲线(HSROC)模型分析数据;根据贝叶斯原理,评价
Dual・翻flqlPce CT in the detection of coronary artery
msease:a Meta analysis
ⅡMin’,Z鼠4ⅣG
Xiao一础,MA
Xi’aR Corresponding
Shao—hui,ZHANG
Ming.+Department ofImaging,First
疗的作用,但缺点是有创伤性、费用高。因此,寻求
一种无创或微创、准确性高、费用低的检查手段,使 阴性患者不必接受CAG检查,具有一定的临床应用 价值。
万方数据
生生照射堂盈盍!Q!Q生!!旦箜丝鲞筮!;翅鱼也』曼型丛:望婴!些旦!Q!Q,!尘:丝:坠!兰 目前,CTA是应用最广泛的无创性CAD检查方 法,尤其是64层螺旋CT的问世,真正开启了CAD 无创性检查的新纪元,但仍有高达2l%的患者不能 获得评价…。双源CT(dual—source CT,DSCT)单扇 区重建时间分辨率达83 ms,即使在不控制心率的 情况下也可高质量成像冠状动脉,甚至可用于心房 颤动患者。笔者的目的是汇总国内外公开发表的文 2.诊断结果数据:基于患者水平的真阳性值、 假阳性值、假阴性值、真阴性值。 四、文献质量评估及数据分析 1.采用诊断准确性研究质量评价(quality
based
Oil
the bivariate random—effects mod01.Hierarchical also estimated.The pooled
wei【ghted symmetric ratios were
summary receiver.operating
curve(HSROC)WaS Was estimated according
patients.,11le
Iikelihood
the pooled sensitivity and specificity.Combined with the pooled likelihood ratios.the
to
clinical utility of the results
Bayes’theory.Results
排除13篇,其中ll篇纳入已知CAD患者,2篇不能
提取诊断性数据。最后纳入文献10篇HJ 二、纳入文献基本信息 Weustink等【l到的研究在不同时期对2组患者 采用不同的扫描方式,故将2组患者独立分析,最后 实际纳入研究11项(表1)。共有1271例患者,
相关文档
最新文档