Meta-analysis - PDF Free Download

체계적문헌고찰논문작성 : 연구평가와메타분석을중심으로 고려대학교의과대학근거중심의학연구소 김현정

순서 체계적고찰은무엇인가? 포함된연구설계의질평가 연구결과의합산 적용가능성

체계적고찰은무엇인가?

Systematic Review or meta-analysis? A Systematic Review 는명백하게구조화된질문을가진종설로, 체계적이고엄격한방법에따라관련된연구를모두찾고선택하여연구에포함된자료를수집분석한것으로정의됨. 통계적인방법 (meta-analysis) 은포함된연구의결과를요약하기위해사용되거나사용하지않을수있음

질문 1 상자안에무엇이있을까? 질문 2 동일한상자안에다른표본이나오는이유는? 표본 1 모집? 표본 4 표본 2 표본 3

표본 1 체계적고찰의본질 : 모아서합친다 표본 4? Systematic Review!!! 모집단 표본 2 표본 3

가정은. 표본 1 표본 4 모집단 표본 2 표본 3

체계적고찰의본질적인한계 : 이질성 & 합칠수있는가? 진실은!!! 표본 1 표본 4 모집단 표본 2 표본 3

Systematic Review 1 질문에대한구조화 2 사전에정의된명백한기준 3 엄격하고재생산가능한방법론적용 4 체계적인문헌검색 5 선택된문헌에대한비평적평가 6 근거의요약과결과해석

APPRAISE & select studies 좋은질의연구는무엇인가?

biased? OK? 1. 모든긍정적인결과의연구 2. 대상환자가 100 명이상인연구 3. BMJ, Lancet, JAMA or NEJM 등의저널에등재된연구 4. 모든사전등록연구

근거의수준이란무엇인가? How can we judge our confidence that adherence to specific intervention will do benefit?

근거수준의평가 두단계에대한평가가필요함 첫번째 : 개별연구의단계 두번째 : 개별연구가모인핵심질문의단계 ROB (Risk Of Bias) LOE(Level Of Evidence)

ROB : 개별연구단계 Bias 연구대상인구집단에존재하는사실로부터체계적 ( 계통적 ) 으로차이가있는결과를생산하게경향 Selection bias( 선택비뚤림 ) o Avoid with randomisation and concealment Performance bias( 수행비뚤림 ) o Avoid with standardisation of care and blinding Attrition bias( 탈락비뚤림 ) o Avoid using all subjects intention to treat analysis Detection or Measurement bias( 결과확인비뚤림 ) o Avoid with blinding of outcome assessors and patients

연구대상선정 대상의할당 Selection bias Randomisation Concealment 대조군 중재 실험군 중재 Performance bias Standardisation of care protocol Blinding of care providers and patients 추적조사 추적조사 Attrition bias Drop-outs? Cross-over? Is everyone accounted for? 결과측정 결과측정 Measurement bias Blinding of outcome assessors and patients

문헌평가방법 연구설계에따라문헌평가방법을적용 무작위대조군연구 : Cochrane Risk of Bias(ROB 2.0) o 무작위방법 o 배정은폐 o 맹검적용 o 불완전한보고 비무작위대조군연구 : ROBINS-I(The Risk Of Bias In Non-randomized Studies of Interventions) o 교란변수, 환자의선택, 중재의분류, 의도된중재로부터의분리, 결측자료, 결과변수의측정, 보고의선택로인한비뚤림 진단방법 : QUADAS Ⅱ 적절한도구를선택 ( 점수화시키는도구는사용을권고하지않음 )

보다근본적으로.. 과일과야채를먹으면간암발생의위험을감소시킬것인가? 중재는? 대조군은? 연구기간은? 연구결과는?

Vegetables and fruit intake and cancer mortality in the Hiroshima/ Nagasaki Life Span Study British Journal of Cancer (2003) The Life Span Study (LSS) is a prospective study of 120 321 subjects including atomic-bomb survivors and nonexposed controls. The Atomic Bomb Casualty Commission (ABCC), followed by the Radiation Effects Research Foundation (RERF), initiated the follow-up in 1950, and regularly monitors the causes of death among the participants through death certificates and other vital status surveys (Pierce et al, 1996). A mail survey was carried out among the 55 650 LSS subjects who were alive as of 1 September 1978 (Radiation Effects Research Foundation, 1978), and of whom 40 349 persons completed the questionnaire (response rate of 72.5%). Completion of the mail survey was effective as from 1 January 1980 for men, and 1 February 1981 for women, during which time 525 persons had died. After exclusion of 1284 cancer cases at baseline based on the information of incidence cases from the Hiroshima and Nagasaki tumour registries, the study population consisted of 38 540

중재및결과확인 중재에대한확인 : The mailed lifestyle questionnaire included questions on past medical history, smoking and drinking habits, marital status, reproductive history, occupation, education, and 22 dietary items, which assessed the average frequency of intake over the previous year. 용량에대한확인 : For the analyses, green-yellow vegetables and fruit consumption were classified into three frequency categories: once per week or less, 2 4 times per week, and daily or almost daily. 암으로인한사망확인 : Follow-up for mortality was linked with the Japanese nationwide family registration system (Koseki). The Koseki provides complete mortality ascertainment for the LSS cohort members residing in Japan. The cause of death was coded according to the International Classification of Diseases, Ninth and Tenth Revisions (ICD-9, ICD-10). The end of the follow-up was defined as the date of death, or 31 March 1998, whichever came first.

무엇을평가해야할까? 만약 ROBINS-I 로평가한다면? 교란변수의확인 환자의선택 : 환자의선택이중재이후의특성에영향을받았는가? 중재의분류 : 중재는명확하게정의되었는가?, 중재의시점과정보의제공시점이동일한가? 의도된중재의분리 : 일반적으로기대되는것외에의도된중재로부터의분리가있었는가? 탈락된데이터 : 모든대상자에게결과자료가가능한가? 결과변수의측정 : 중재에대한지식이결과측정에영향을미치는가?.. 이것으로충분한가?

과일과야채를많이먹으면간암을예방하는가? 과일과야채의많이 ( 어느정도의용량인가?) 어느정도의기간을추적해야하는가? 기간의문제인가? 용량의문제인가? 무엇이비교군인가? 과일과야채를먹지않는사람? 과일과야채를먹지않는사람이존재하는가? 과일과야채대신육식을먹는다면이는과일과야채의효과인가? 육식의효과인가? 10 년이상연구를지속할수있는가?

개별연구의결과가비뚤리는이유 AUSTRALASIAN COCHRANE CENTRE

Internal Validity Intervention A vs B Outcome Important Prognostic Characteristics

Statin 사용과암발생과의관련성 Statin use and the risk of breast cancer;journal of Clinical Epidemiology 2003,56(3),280-285 Exposure time Nonexposed Breast cancer cases Follow-up time (PY) a 0 days 566 165,930 3.41 Statin users Breast cancer rate (per 1,000 PY) Relative rate (95% CI) b 1 90 days 38 7,017 5.42 1.59 (1.14, 2.20) 91 180 days 32 4,633 6.91 2.02 (1.42, 2.89) 181 365 days 17 5,428 3.13 0.92 (0.57, 1.49) 1 2 years 37 7,311 5.06 1.48 (1.06, 2.07) 2 3 years 19 5,204 3.65 1.07 (0.68, 1.69) 3 4 years 13 4,002 3.25 0.95 (0.55, 1.65) 4 years 7 7,920 0.88 0.26 (0.12, 0.55)

동일하게사건이발생하였다해도사건발생수 1 년 8 명 2 년 2 명 3 년 1 명 발생까지의시간이길어서오래약물을복용한효과???

동일한시간을서로추적한경우라고해도 대상자선정에따라추적기간의차이 Time-varying (actual data) Time-invarying approach 약물시작시점 약물의사용을사용자 / 비사용자로구분한경우비뚤림발생

투약시점으로환자를구분한경우사건발생수 비복용군 6 명 복용시기진단시기가다른경우 복용시기와진단시기가같은경우 6 명

복용여부로환자를구분한경우사건발생수 비복용군 4 명 복용군 복용군 8 명

이전시기에대한정의 (2005 년유병환자대상 ) 1 년을이전시기로정의한경우 : 2004 년의료내역이없다면신환자로정의 883( 참값 ) vs 1021(1 년 wash out) 2005 년기준 2002년이력없음만 2003 년만 2002~ 2003 년 2004 년만 2003~ 2004 년 2002 년,2 004 년 2002~ 2004 년 환자수 1,093 883 18 39 81 11 8 28 25 구성비 100.0 80.8 1.6 3.6 7.4 1.0 0.7 2.6 2.3 신환자로분류 (11.6%) 잘못된분류의가능성 유병환자로분류됨 (6.6%)

이전시기에대한정의 (2005 년유병환자대상 ) 1 년을이전시기로정의한경우 정상으로분류 (40.1%) 잘못된분류의가능성 배제? 기존환자군? 전체 이력없음 2002 년만 2003 년만 2002~ 2003 년 2004 년만 2003~ 2004 년 2002 년, 2004 년 2002~ 2004 년 환자수 7733 3363 1605 1353 139 1121 90 36 26 구성비 100.0 43.5 20.8 17.5 1.8 14.5 1.2 0.5 0.3 환자군 환자군 환자군 순응도가높은환자군 환자군 정상군 환자군 순응도가낮은환자군

Back up period 와사망 2005 2006 2007 2008 2009 2010 2011 2012 2013 전체 빈도 빈도 빈도 빈도 빈도 빈도 빈도 빈도 빈도 행 N % Ontariow1 0 1 6 1 0 1 0 0 2 18.2 Ontariow2 0 1 5 1 0 1 0 0 1 11.1 Ontariow3 0 1 5 1 0 1 0 0 1 11.1 이전기간에대한추적이짧을수록질병을오래가진대상자가신환자로분류될가능성으로인한사망률증가

무작위대조군연구의문제 이상적으로 : 중재의대상이되는상태를가진모든환자가포함되어야함 실제로는 : 윤리적문제, 의료적문제로인한환자의배제 부작용의위험을가진대상 이전연구에서이미효과가입증된대상 Benefit 이명백하지않거나작음 추가적으로 : 관리와과학적측면에서의배제 정확성의증가 ( 효과를통계적으로입증하기위해 base line risk 가높은집단에대한선택, 변이를감소시키기위한대상군선정기준의축소 ) 비뚤림의배제 (ITT analysis 등을위한탈락군최소화 ) 결과적으로 : 전체환자군중일부의상태만을가진대상군이포함됨

연구에서의선택 매우동질한환자군만을포함시킴 - 정확한연구결과 - 일반화에대한문제제기 Versus 대부분의환자군을포함시킴 - 동일하지않은결과로인한넓은신뢰구간 - 보다직접적인일반화된추정치

BHAT Study: Beta-Blocker Therapy after MI Choosing between randomised and nonrandomised studies:a systematic review (Health Technology Assessment 1998; Vol. 2: No. 13) BHAT 연구와두개의다른코호트대상군에대한비교 코호트군 : 1979 년 1982 년까지 Yale-New Haven Hospital 에 AMI 로입원하거나치료를받은대상자 2497 명을대상으로함 넓은대상 : 일반적조건 제한된대상 : BHAT 연구의배제조건

무작위대조군연구에서의환자의선택 Baseline BHAT 코호트 ( 제한된기준 ) 코호트 ( 제한하지않음 ) Beta No Beta Beta No Beta Beta No Beta Blocker Blocker Blocker Blocker Blocker Blocker (n-1916) (n=1921) (n=417) (n=205) (n=626) (n=433) Age (yr) 55 55 57 60 58 60 Male (%) 84 85 75 73 74 68 Severity (% mild) Outcome(2 yr Mortality) 58 61 72 68 64 51 위험도가높은환자군 Crude 7.3 9.2 7.2 10.7 9.3 16.4 Adjusted 7.3 9.2 7.6 9.7 10.2 14.4

Nifedipine in Patients with Cardiovascular Disease ( Risk of Mortality) 9 RCTs with dosages 30-60 mg/day (n > 8000) 1 large, high quality non-randomized study with detailed risk adjustment (n=11,575) with dosages 30-60 mg/day

결론적으로 비무작위대조군연구와무작위대조군연구는각기다른면에서의연구결과의비뚤릴가능성과일반화의문제를가짐 연구결과는개별연구에서의환자의선택과배제에따른선택비뚤림이가장중요한문제이며, 무작위대조군의경우일반화의가능성의문제로 비무작위대조군의경우교란변수와의상호작용을통해연구결과의신뢰성의문제를야기시킴 주의깊은평가와판단이필요함!!

Synthesis: 연구결과에대한합산혹은요약 어디까지합산할것인가?

meta-analysis? Egger, Davey Smith and Altman (2001) Statistical analysis of the results of independent studies, which generally aims to produce a single effect estimate Cochrane handbook v. 5.0.1 (2008) Sec. 9.1.2 The statistical combination of results from two or more separate studies Glass G.V. (1976) use the term meta-analysis Meta-analysis refers to the analysis of analyses the statistical analysis of a large collection of analysis results from individual studies for the purpose of integrating findings. It connotes a rigorous alternative to the casual, narrative discussions of research studies which typify our attempts to make sense of the rapidly expanding literature o Glass GV (1976). Primary, secondary, and meta-analysis of research. Edu. Researcher 5:3-8 18

메타분석정의와목적 체계적문헌고찰 (Systematic Review, SR) 수행과정에서 2 개이상의개별연구들의요약추정치를합성 해당 intervention 의통합된가중평균요약추정치를정량적으로산출하여임상적효과성을평가하기위해사용하는통계적기법 정량적고찰및독립적인연구결과의합성 몇몇연구결과의정보를합산함을통해연구의파워와일반화가능성을증가시킴 여러연구결과에거치혼합된효과혹은평균효과를추정

메타분석의원리 개별연구의요약추정치 (Summary estimate) 를계산하며, 각각의연구에가중치줌 가중치는보다많은정보를제공하는개별연구에더많은가중치를부여함. 일반적으로가중치는표본크기를사용하며, 사건발생률, 분산의역수, 연구의질등이사용되기도함. 가중치가부여된평균을계산하여통합요약통계량을산출함

Forest plot : 메타분석결과표 ( 이분형자료 ) 개별연구의효과범위 이질성평가 효과의추정치 효과의추정치

Fixed effect( 고정효과모형 ) vs Random effect Model( 변량효과모형 )

Fixed effect vs Random effect Model 우리가결과를추정하고자하는모집단에따라결정 효과 (effect) 에대한가정에따라 A data-driven approach to the decision whether to use fixed- vs random-effects models is really not recommended. The correct decision is based on theory: I.e. Is the study sample and intervention mostly the same across studies (fixed model OK) or not (most common situation - random model better)?

Fixed effect Model 표본 1~4 중에가장진실한결과 ( 모집단 ) 에가까운것은? 표본 1 표본 4 모집단 표본 2 표본 3

Random effect Model 표본 1~4 중가장진실한결과 ( 모집단 ) 에가까운것은? 표본 1 표본 3 모집단 1 모집단 2 표본 4 표본 4

Fixed effect Model 동일한모집단 ( 인구집단 ) 에서추출되었다고가정함 따라서동일한상태에서의동일한 true effect(real value) 를공유한다고가정함 연구결과간변이가있다면이것은대상인구집단에서표본추출을하는것으로부터의변이 ( 표본추출변동 (sampling variation, random error)) 임, 따라서각연구내의인구집단의변이로인한것이지연구자체의성격의차이가아님 연구결과의정확성은연구대상자의수에의존함

Random effect Model 대상인구집단의차이가있음을가정함 연구는각인구집단의 true effect 를가지는것으로가정됨, 즉각연구들은어떤평균적인처리효과를중심으로퍼져있는모집단내연구들로부터무작위로추출된연구들임 치료혹은노출의효과에따른변이는연구내뿐만아니라연구간의차이로기인함 연구의표본추출변동 ( 연구내차이 ) 연구간의차이 ( 중재방법의차이, 약의용량등 )

FE methods 비교 MH method is preferable Pooling 할 study 수는많지만 within-study sample size 는작은경우 IVW method is preferable Pooling 할 study 수는작지만 within-study sample size 가큰경우 Peto s method is under strong criticism May produce seriously biased OR and SE o 특별히두집단의수가 severely imbalance 되어있을때 Possibly biased when the estimated OR is far from 1 Trial arm 들에 zero events 가있는경우 For MH, a study with zero total events is completely excluded o 그러나 a continuity correction (add.5 to each cell) 을사용할수도있음 Peto method outperforms MH or IVW o 2 2 tables 의하나이상의 cell 들에서 event 수가작을때 Important to report precisely what methods we used 35

RE model Weighted least squares (WLS) method DerSimonian and Laird, Controlled Clin Trials 1986;7:177-188 Called Dersimonian-Laired (D-L) method o Comparability 나 simplicity 등을고려할때일반적으로 D-L method 사용추천 Unweighted least squares (UWLS) Maximum likelihood (ML) Restricted maximum likelihood (REML) Bayesian method 36

Pooling Summary 언제 random-effects model 사용하는가? RE model 은연구간 (between studies) 이질성을인정하면서 study 들을 combining 하는방법 따라서 study 들간에설명되어지지않는이질성 (heterogeneity) 이존재한다는근거가있을때사용 또한, 비록 test of heterogeneity 결과는 nonsignificant 하지만 study 들이 truly homogeneity 하다는가정을할수없을때사용 o Because the heterogeneity test lacks power (i.e., studies may be regarded as homogeneous when in fact there is a degree of heterogeneity) 47

이질성 (heterogeneity) AUSTRALASIAN COCHRANE CENTRE

Heterogeneity 는무엇인가? Heterogeneity is variation between the results of a set of studies Key point : Judgements as to whether the primary studies are similar enough to be combined in a systematic review is based on whether there is a plausible rational to expect a similar treatment effect (or result) across the range of participants, interventions, outcomes and methodologies of the primary studies.

pressure-controlled ventilation vs volume-controlled ventilation 8 studies : thoracotomy for lung surgery - pneumectomy, lobectomy, wedge resection. 1 study : thoracotomy for robotic-assisted esophagectomy, 1 study : coronary artery bypass 1 study : mitral valve repair respectively

Review comment A data-driven approach to the decision whether to use fixed- vs random-effects models is really not recommended. The correct decision is based on theory: I.e. Is the study sample and intervention mostly the same across studies (fixed model OK) or not (most common situation - random model better)? In this case, there is a lot of heterogeneity in the indication for OLV and the patients in each of the studies. The assumptions of a fixed effects model are therefore likely *not* met, irrespective of measures of statistical heterogeneity like I 2 value

Heterogeneity 의발생원인 : 임상적이질성 명백히동일한연구임에도불구하고이질성이존재하는원인 포함기준과배제기준의차이 동일한선택기준을사용한경우에도포함된연구대상의 risk 의차이가있는경우 중재의차이 ( 용량, 횟수, 주입방법등 ) Management 의차이 ( 환자관리, setting 의차이, cointervention 등 ) 결과측정방법의차이 ( 추적관찰기간의차이, 결과의정의의차이등 ) 분석에서의변이 (withdrawals, dropouts, crossovers 의관리 )

Heterogeneity 원인 : 방법론적 연구설계 e.g. randomised vs non-randomised, parallel group vs crossover vs cluster randomised, length 개별연구에서의방법론적제약 e.g. allocation concealment, blinding, approach to analysis, imputation methods for missing data

Identifying heterogeneity 1. graphically the eyeball test 2. numerically the I 2 test

Heterogeneity 에대한 graphical informal test( 눈으로보기 ) Forest plot 메타분석결과를제시하는가장일반적인방법 신뢰구간에대한겹침정도평가 메타분석의 exploratory stage 에서도매우효과적으로사용될수있음 study estimate 들간의 variablility 에대한 idea 를얻을수있음

Heterogeneity 존재여부에대한통계적검정 Cochran s Q-test (chi-square test) 가정 : 결합할 study 들은모두하나의 population parameter 를추정하는 study 들이라고볼수있다. (study estimates 의 variation 은완전히 random error 에의한것이다 ) 귀무가설 : 이질적이지않다 vs 대립가설 : 이질적이다 Q 를사용하여 heterogeneity 를평가하는경우의문제점 Low statistical power o 실제로 heterogeneity 가존재하는상황에서도 Q 값은일반적인유의수준 ( 예 : 0.05) 하에서유의하지않은결과를제시하게됨. o 따라서메타분석의경우, 유의수준 0.1 을사용하는것이일반적임. 각 study 들의 sample size 가크다면, 개별 effect size 추정값들이서로다르지않은경우에도가정은기각될수있음. Publication bias 가있거나 primary study 에 design flaw 가있는경우, Q 를이용한검정결과를해석하기가어려움.

Heterogeneity 존재여부에대한통계적검정 I 2 통계량 (Higgins)= 총관찰된 variation 중 true heterogeneity 의분율을의미함 I 2 = (Q df) x 100% (df = the number of studies minus 1) Q I 2 값이작을수록 heterogeneity 적음 일반적으로 25%, 50%, 75% 는낮은, 중등도, 높은이질성을반영함 (Higgins et al 2003, BMJ)

Forest plot A Forest plot B I 2 =0 I 2 =98%

Heterogeneity 처리 Lau, J. et. al. Ann Intern Med 1997;127:820-826

Option 1: Ignore heterogeneity

Option 2: Don t pool studies

Option 3: Incorporate heterogeneity : Random effect Fixed=0.98(0.71-1.36) Random=0.90(0.52-1.54)

Random effects meta-analyses fixed effect analyses 를수행한이후명백한지않은이질성 ( 이질성의원인을찾을수없을때 ) 이있다면사용을고려 : 단이모형은연구간변이를인정 - 즉, 연구마다해당 underlying effect 는다를수도있다 ) 더넓은신뢰구간을가짐 ( 이질성이존재하는경우 ) Random effects analyses : 상대적으로연구대상자의수가작은연구에가중값을더많이주게됨 ( 각각이고유한 true effect 임으로 - 따라서정확성이낮은연구에보다많은가중치를주게되는문제를발생 )

Fixed versus random effects almost identical

Fixed versus random effects similar, but wider CIs

Fixed versus random effects very different results source: with thanks to Julian Higgins

Option 4: heterogeneity 에대한연구 통계적으로설명되지않은이질성에대한추적연구가필요

Heterogeneity 처리방법들 exclude studies 가장큰변이를제공하는연구들을제거할수있음 o 즉, 자료분석초기단계에서 outliers 나 extreme results 제거하는것 o 그러나이는 biased estimates 를초래할가능성도있음에유의해야함. o sensitivity analysis 사용을고려 groups of studies 각각분석 (subgroup analysis 와는다름 ) 합산할수없을정도로큰이질성이존재하나유산한그룹의연구들이존재하면, 이들을묶어보는것이타당할수도 메타분석실시전에발생가능한하위그룹에대한정의필요

Heterogeneity 에대한연구방법 subgroup analysis 연구대상, 중재에관한추가적인해답가능 하지만연구결과가잘못된결과를추정될수있음 meta-regression 치료의효과와각연구 ( 환자가아닌 ) 의특성과의관련성을설명할수있음 o not available in RevMan individual patient data (IPD) meta-analysis 참가한환자수준의특성을고려할수있음 많은시간과비용이필요

Cumulative meta-analysis Use a fixed effects model is used in the cumulative meta Although there was great heterogeneity among trials! Reason Using fixed effect analysis aids o the identification of influential studies when this plot is being used in an exploratory manner Using random effect analysis results in o smoothing effect (i.e., more equal weight), and o restrict the impact of individual studies When sources of heterogeneity have been identified Performing a cumulative meta-analysis on subgroups defined by the sources may be more informative o For example, by study quality, baseline risk, etc.

연구간 Heterogeneity 원인에대한탐색 메타분석을수행하는경우, 이질성을평가하기위한방법들을사용해보고, 가능한원인들에대해탐구필요 아직상당한크기의설명불가능한 heterogeneity 가여전히남아있다면, overall estimate 를계산하기위해 results 들을결합하는것이과연적절한지에대한판단을해야할것 이를위해서는상당한양의 subjectivity 가필요할수밖에

subgroup analysis 나 regression method 들의단점 단지탐색적 (exploratory) 인방법일뿐 study characteristic 들과 outcome 들사이의관련성이단지우연 ( 혹은교란변수 ) 이존재하기때문에발생한것인지등에대한탐색적인방법일뿐 회귀분석방법 : aggregation bias(ecologic bias) 발생가능성높음 aggregation bias: patient characteristic 에관한 study 들의평균과 outcome 들간의관계가각개인별 value 들과개인별 outcome 들간의관계를직접적으로반영하지못할때발생하는편향 각 study 들이제공하는분석에필요한자료들이제한적일수있음 subgroup analysis 와 regression method: 민감도분석 (sensitivity analysis) 에서사용될수도있음

Sensitivity analysis ( 민감도분석 ) 메타분석결과가분석에사용된특정방법들이나각연구에서얻어진결론들에얼마나 influencial / robust 한지를평가 민감도분석을통해 결과가쉽게변화되지않는다 고판단되면해당메타분석결과에더자신감을가질수있을것 메타분석각단계에서행해진모든결정들이반영되도록실시 Oxman (1996) Review 결과가 review 를수행하는과정중행해진 key decisions or assumptions 에대해얼마나 robust 한지검사 Reviewer 는문제가될소지가있고, 따라서결과에영향을미칠수있는 key decisions or assumptions 를 identify 해야

Sensitivity analysis Cochrane s advise 1 2 3 4 5 6 7 8 9 Inclusion criteria 를변화시켜줌 Inclusion criteria 를만족하는지애매모호한 study 들의 including/excluding 을시도 Unpublished study 들 including/excluding 시도 Methodological low quality study 들 impact 평가 Extract 된자료와관련해서 uncertainty 가존재하는경우 data 를재분석 Publication bias 평가 Missing information 들을 imputing Extra trial 들에대한모의실험 (simulation) 실시 Cumulative meta-analysis

Subgroup vs sensitivity analysis A sensitivity analysis asks : Are the findings robust to the decisions made in the process of obtaining them? Subgroup analyses can be used to answer specific questions about particular groups of participants, types of intervention or types of study.

sensitivity vs subgroup analysis Effect of homeopathy vs. placebo effect Linde, et al., Lancet 1997:350;834-843 71

메타분석프로그램들 Software info. http://www.prw.le.ac.uk/epidemio/personal/ajs22/meta/index.html Commercial STATA (http://www.stata.com), SAS, S+, R Comprehensive meta-analysis (http://www.meta-analysis.com) Metawin (http://www.metawinsoft.com) WEasyMA (http://www.weasyma.com) Freeware Review-Manager (RevMan v.5.1) (http://www.cc-ims.net/revman) MetaDiSc (http://www.hrc.es/investigacion/metadisc_en.htm) Sinergy (http://www.e-biometria.com/ebiometria/sinergy/sinergy.htm) Epi-meta (http://www.cdc.gov/epo/dpram/epimeta/epimeta.htm): DOS ver. Meta, Meta-Analyst, Meta-Test R

연구결과의합산 모든연구가메타분석이가능하지는않음 메타분석이가능하지않은연구의경우연구결과의범위, 정도등을요약하여제시 무작위대조군연구의경우 연구대상이무작위에의한배정을수행하였음으로두군의차이는중재의차이임 비무작위대조군연구의경우 각군간의위험요인이다름으로두군의차이가중재의차이만이아닌다른교란변수의효과가잠재되어있음. 따라서이러한교란변수의효과를보정한상태에서의결과를비교하여야함

연구결과의합산 모든연구가메타분석이가능하지는않음 메타분석이가능하지않은연구의경우연구결과의범위, 정도등을요약하여제시 연구설계가무작위대조군다른연구의연구는경우서로다른비뚤림위험이있음으로합산할수없음 연구대상이무작위에의한배정을수행하였음으로두군의차이는중재의차이임 연구가비무작위대조군서로다르다면연구의 ( 동질경우하지않은경우 ) 합산 ( 메타분석 ) 하지말아야함 각군간의위험요인이다름으로두군의차이가중재의차이만이아닌다른교란변수의효과가잠재되어있음. 따라서이러한교란변수의효과를보정한상태에서의결과를비교하여야함

THE VALIDITY OF THE MOOD DISORDER QUESTIONNAIRE FOR SCREENING BIPOLAR DISORDER: A META-ANALYSIS- THE VALIDITY OF THE MOOD DISORDER QUESTIONNAIRE FOR SCREENING BIPOLAR DISORDER: A META-ANALYSIS

Background the early detection of BD among patients diagnosed as MDD is very important. One of the most commonly used is the Mood Disorder Questionnaire (MDQ) developed by Hirschfeld et. al. In the original study by Hirschfeld, the authors suggested a standard cutoff value of 7

Questions According to previous studies, the MDQ seems to have relatively good sensitivity and specificity in detecting BD among psychiatric outpatients. However, it seems to have lower sensitivity in detecting BD in the general population. The MDQ has been translated into many other languages, and the validity of these versions has been investigated among patients with mood disorders there is a tendency for studies conducted in Eastern countries to demonstrate relatively lower optimal cutoff values than those conducted in Western countries.

cutoff of 7 Optimal cutoff

The summary sensitivity and specificity of the MDQ according to different inclusion/exclusion criteria of included studies SENSITIVITY/SPECIFICITY standard or modified cutoff of 7 0.62 (95% CI =0.51 0.72) 0.85 (95% CI =0.79 0.89) optimal cutoff values 0.78 (95% CI =0.74 0.81) 0.76 (95% CI =0.71 0.81) Including subjects with both MDD and BD without using a broadened definition of the DSM-IV for BD Excluding subjects with known BD or using the broadened definition of the DSM-IV for BD 0.76 (95% CI =0.69 0.82) 0.81 (95% CI =0.75 0.86) 0.37 (95% CI =0.29 0.45) 0.88 (95% CI =0.80 0.93) Excluding subjects with known BD 0.37 (95% CI =0.22 0.54) 0.88 (95% CI =0.79 0.94) 20 21 11 9 6

cross-cultural differences? First, when pooling studies excluding patients with known BD from the screened population, the summary sensitivity became significantly lower. Second, we found that when studies used a broadened definition of the DSM-IV for BD or when they used a final diagnosis elicited by re-interviewing 1 year later, the summary sensitivity became significantly lower. There have been reports that the optimal cutoff seemed to be lower in Eastern countries in comparison to their Western counterparts. However, after adjusting for various clinical correlates, these cross-cultural differences disappeared.

Subgroup Analysis: streptokinase versus placebo Second International Study of Infarct Survival (ISIS-2) (I) Reprinted from ISIS-2 Collaborative Group. Lancet 1988;2:349-60, with permission from Elsevier.

Subgroup Analysis: aspirin versus placebo Second International Study of Infarct Survival (ISIS-2) (II) Reprinted from ISIS-2 Collaborative Group. Lancet 1988;2:349-60, with permission from Elsevier.

Transferability : 적용가능성

알고자하는것? 우리가아는것 대상상태를 가진환자군 우리가모르는것 연구에포함 연구에포함 되지않음 중재 A 중재 B 중재 A,B 결과 결과 결과 AUSTRALASIAN COCHRANE CENTRE

알고자하는것은반영하는가? 얼마나영향을미치는가? 대상상태를 가진환자군 얼마나같은가? 연구에포함? 연구에포함 되지않음 중재 A 중재 B 중재 A,B 결과 결과 결과 AUSTRALASIAN COCHRANE CENTRE

Transferable? 나의환자에게적용가능한가 연구들의대상과나의환자는유사한가? 만약아니라면, 이유는? 연구설계의차이 연구대상 ( 중재, 결과측정방법등등 ) 의차이 연구와실제대상의 Baseline risk 의차이 만약그렇다면, 추가고려해야할사항 연구대상군과다른대상에게적용가능성

말라리아유행지역에서의사회경제적수준과사망

evidence gap 1 우리가알고싶은것 말라이아유행지역에경제를성장시킨다면장기적으로말라이아유행을감소시키거나사망률을줄일것이다. 우리가알게된것 말라이아유행지역의부모의경제적수준이말라리아에걸린아이의사망을줄인다

systematic reviews 득과실 Advantages 많은대상자수와사건수로인한검정력 전체연구에대한체계적인결과수용 Disadvantages 실제효과가비뚤릴가능성

포함된문헌에따라 garbage in garbage out 체계적고찰에서포함된연구의추정된결과값은포함된연구가바이어스된연구라면개별연구자체보다더바이어스된결과를가져올수있음 출판혹은보고비뚫림을가진체계적고찰은적합하지않은추정값을가져올수있음 모두모았는가? 무엇을모았는가?

체계적고찰과메타분석 사과와오랜지의혼합 사과에대해알기를원할때적합하지않음 과일에대해알기를원할때보다적합함 연구는동일한주제를가지고있어야함 분석에사용된연구가너무다양하다면진실된효과를구별하기에적합하지앟을수있으며결과가무의미할수있음

체계적고찰결과에대한신뢰성에대한불확실성 체계적고찰은현재까지진행된 ( 보고된 ) 연구결과에대한비뚤리지않는포괄적인고찰이지, 모든 ( 보고되지않았거나, 이루어지지않은 ) 사건 ( 효과 ) 에대한고찰이아니다. 모든사건 ( 효과 ) 과체계적고찰의근거

체계적고찰결과에대한신뢰성에대한불확실성 체계적고찰은현재까지진행된 ( 보고된 ) 연구결과에대한비뚤리지않는포괄적인고찰이지, 모든 ( 보고되지않았거나, 이루어지지않은 ) 사건 ( 효과 ) 에대한고찰이아니다. 현재까지진행된연구가적거나없다. 연구에포함된대상자의수나효과를평가하기에적절하지않은사건수를가진다 대상연구들의대상과중재, 결과지표등의이질성이존재한다 포함된연구들의연구설계에따른연구수행이적절하지않아비뚤어질가능성이높다 긍정적인연구결과만이보고되었을가능성이존재한다

경청해주셔서감사합니다 고려대학교의과대학근거중심의학연구소 E-mail : moole02@naver.com 김현정