untitled

Journal of Rheumatic Diseases Vol. 22, No. 1, February, 2015 http://dx.doi.org/10.4078/jrd.2015.22.1.4 Review Article 메타분석 이영호고려대학교의과대학내과학교실류마티스내과 Meta-analysis Young Ho Lee Division of Rheumatology, Department of Internal Medicine, Korea University College of Medicine, Seoul, Korea Meta-analysis is a statistical tool for combining the results of different studies on the same topic, providing a precise estimate of the effect size and increasing statistical strength, which is particularly important when the strength of the primary study is limited because of a small sample size. Properly conducted meta-analysis provides an invaluable link between past and future studies by quantitatively synthesizing evidence while minimizing bias. Recently, because studies on meta-analysis have been published increasingly, there is a need for rheumatologists to understand meta-analysis. In order to help rheumatologists in use of a meta-analysis, the author describes the basic steps in statistical analysis of a meta-analysis: 1) search for presence of between-study heterogeneity, 2) performing statistical analysis of meta-analysis, 3) checking publication bias, 4) search for causes of heterogeneity, and 5) interpreting and presenting meta-analysis results. (J Rheum Dis 2015;22:4-9) Key Words. Meta-analysis, Statistical analysis 서론 의학연구의발달로매년많은연구논문들이발표되고있으나개별연구에서는통계검정력이낮아서일관성없는결과를보이는경우가많다. 동일한주제에대한많은연구들이있으나결과들이일치하지않거나상이한차이를보이는경우가많아서같은주제에대한연구결과들을통합하여분석할필요가생기게되었다. 메타분석은같은주제에대한독립적인개별연구결과들을통계적으로통합하여객관적으로분석하는연구방법이다 [1]. 메타라는단어는나중에 (after) 라는의미로일차적인연구이후에이런연구들을분석하므로연구들의연구 (study of studies) 라고할수있다. 메타분석은각각의연구결과들을가중평균요약추정치로결합하여연구대상수를증가시켜서통계검정력과정확도를높이고, 개별연구의한계를극복하여일반 적, 포괄적, 객관적인결과를얻을수있는장점이있다 [2,3]. 연구논문발표의증가와더불어메타분석에대한연구도매년꾸준히증가하고있으며류마티스질환을포함한다양한의학연구에서유용하게적용되고활용이증가되고있다. 본종설에서는메타분석의방법, 장점및제한점등에대해기술하여메타분석을이해하는데도움을주고자한다. 본론 메타분석과정메타분석은일반적으로 1) 연구주제선정, 2) 문헌검색, 3) 논문선택및논문의질평가, 4) 논문들간의효과크기 (effect size) 의이질성 (between-study heterogeneity) 검사, 5) 고정효과모델 (fixed effect model) 또는변량효과모델 (random effect model) 을이용한메타분석, 6) 출판편 Received:December 3, 2014, Accepted:December 18, 2014 Corresponding to:young Ho Lee, Division of Rheumatology, Department of Internal Medicine, Korea University Anam Hospital, Korea University College of Medicine, 73 Inchon-ro, Seongbuk-gu, Seoul 136-705, Korea. E-mail:lyhcgh@korea.ac.kr pissn: 2093-940X, eissn: 2233-4718 Copyright c 2015 by The Korean College of Rheumatology. All rights reserved. This is a Free Access article, which permits unrestricted non-commerical use, distribution, and reproduction in any medium, provided the original work is properly cited. 4

Meta-analysis Method 견 (publication bias) 조사, 7) 이질성원인검사, 8) 제한점과장점을포함한결과해석및논문발표의절차를거쳐서시행된다 (Table 1)[1]. 1) 이질성검사 효과크기란여러연구들을비교하거나종합하고자할때중재또는치료효과의크기를서로비교해볼수있는동일한척도로정의된다 [4]. 동일한주제에대한논문들의효과크기 (odds ratio [OR], relative risk [RR] 등 ) 가논문 Table 1. Process of statistical analysis of meta-analysis 1. Search for presence of between-study heterogeneity: Cochran Q test, I 2 2. Performing meta-analysis: fixed or random effect model, forrest plot 3. Checking publication bias: funnel plot, Egger s regression test 4. Search for causes of heterogeneity: subgroup analysis, sensitivity analysis, meta-regression 5. Interpreting and presenting meta-analysis result 들간에의미있는차이가있을수있다. 연구결과들간에의미있는차이가있는경우 ( 이질성이있을때 ) 와없을때 ( 연구결과들이동질일때 ) 의메타분석방법은다르게적용된다 [1]. 따라서연구결과들간의이질성의존재와정도를검사하는것이중요하다. 이질성검사는각연구들의효과크기가산출된효과크기값으로부터얼마나멀리떨어져있는지를검정하는방법으로카이제곱검정인 Cochran Q-검사를이용한다 [5]. Cochran Q-검사는통계검정력이약해서의미있는 p-수치를 0.05가아니라 0.1을기준으로한다 [6]. 따라서 Cochran Q-검사의 p-수치가 0.1보다작을때연구간효과크기의이질성이존재하는것으로간주한다. 연구간이질성을정량적으로분석하는방법으로 I 2 통계량이개발되어사용되고있다 [7]. I 2 통계량은연구의수나효과크기의종류에영향을받지않고이질성을정량적으로표시하며 I 2 =100% (Q df)/q (Q=Cochran Q 이질성값, df=degree of freedom) 로계산된다 (Table 2). I 2 가 0% 인경우는이질성이없는상태이고 I 2 가클수록이질성은증가한다. I 2 의 25%, 50%, 75% 를각각저, 중, 고의이질성으로분류한다 (Table 2)[6-8]. Table 2. Meta-analysis of the associations between the FCRL3-169 C/T polymorphism and rheumatoid arthritis Polymorphism Population No. of studies FCRL3 C vs. T CC vs. CT+TT (recessive) CC+CT vs. TT (dominant) Test of association Test of heterogeneity OR 95% CI p-value Model p-value I 2 Overall 17 1.046 0.997 1.098 0.068 R 0.084 34.1 European 9 1.012 0.962 1.065 0.643 F 0.128 36.2 Asian 7 1.101 1.035 1.171 0.002 F 0.314 15.1 Japanese 3 1.124 1.029 1.227 0.009 F 0.266 24.5 Non-Japanese 4 1.080 0.990 1.177 0.082 F 0.260 25.2 Overall 17 1.069 0.977 1.170 0.146 R 0.052 38.8 European 9 1.004 0.883 1.141 0.955 R 0.040 50.5 Asian 7 1.138 1.014 1.277 0.028 F 0.418 0.75 Japanese 3 1.216 1.027 1.438 0.023 F 0.469 0 Non-Japanese 4 1.074 0.917 1.258 0.375 F 0.330 12.4 Overall 17 1.056 0.996 1.119 0.066 F 0.328 10.7 European 9 1.019 0.941 1.102 0.647 F 0.641 0 Asian 7 1.134 1.037 1.241 0.006 F 0.449 0 Japanese 3 1.144 1.006 1.300 0.040 F 0.264 24.8 Non-Japanese 4 1.125 0.992 1.276 0.067 F 0.379 2.68 CC vs. TT Overall 17 1.100 1.017 1.190 0.018 F 0.105 31.4 European 9 1.032 0.931 1.144 0.549 F 0.129 36.1 Asian 7 1.208 1.063 1.373 0.004 F 0.300 16.9 Japanese 3 1.282 1.064 1.544 0.009 F 0.274 22.7 Non-Japanese 4 1.146 0.961 1.366 0.129 F 0.272 23.1 CC vs. CT Overall 17 1.052 0.960 1.154 0.275 R 0.084 34.0 European 9 1.008 0.918 1.106 0.872 R 0.064 45.7 Asian 7 1.092 0.967 1.234 0.157 F 0.575 0 Japanese 3 1.174 0.983 1.401 0.077 F 0.666 0 Non-Japanese 4 1.024 0.865 1.211 0.785 F 0.472 0 CI: confidence interval, F: fixed effect model, FCRL3: Fc receptor-like 3, OR: odds ratio, R: random effect model. www.jrd.or.kr 5

Young Ho Lee Figure 1. Forrest plot of odds ratios (ORs) and 95% confidence interval (CIs) of individual studies and pooled data for the association between the C allele of the Fc receptor-like 3-169 C/T polymorphism and rheumatoid arthritis (RA) in each ethnic group. NAN: North American Native. 2) 메타분석의통계분석메타분석은개별연구의요약추정치를계산하며, 많은정보를제공하는개별연구에더많은가중치를부여한다 [1]. 일반적으로가중치는표본크기가사용되며, 사건발생률, 분산의역수등이사용되기도한다. 대규모의연구는작은규모의연구에비해우연한차이로인한영향을덜받기때문에대규모연구에상대적으로더많은가중치를부여하는것이다. 메타분석의통계적모델로는고정효과모델과변량효과모델이있다 [9]. 고정효과모델은모든연구의효과크기가유사하다고가정한다. 각각의연구는같은모집단에서얻어졌다는동질성을가정하며, 각연구결과가서로상이한것은표본추출에서생기는표본간의변동때문이라고가정하므로, 연구간의차이는없으며단지연구내변이만을반영한다 [9]. 변량효과모델은모든연구의효과크기가다르다고가정하여각연구가가정하는실제효과크기는각연구별로상이할수있으므로연구내변이뿐만아니라연구간의차이를모두반영한가중치를사용하여분석하는모델이다 [10]. 고정효과모델에서는연구결과의가중치는분산의역수, 즉대상수에따라비례하여많은대상수를포함하여얻은연구결과는적은수를대상으로한연구결과에비해메타분석에더많이반영된다. 반면에변량효과모델에서는대상수뿐만아니라연구간의효과크기차이도가중치에반영되므로효과크기가다른연구들과비교하여현저히차이가있는경우에메타분 석에적게반영된다. 변량효과모델은연구들간에변동을인정하기때문에결합추정치에대해서고정효과모델에비해더넓은신뢰구간을보이며, 소규모연구에상대적으로더많은가중치를부여한다. 만약연구간이질성이없다면고정효과모델이나변량효과모델의결과들사이에차이가나지않으나연구간이질성이있으면두모델로부터나온결과들이차이가있으므로이때에는연구간이질성을반영하여분석하는변량효과모델을사용해야한다 (Table 2)[6,8]. 이분형자료에대한메타분석을위해서는일반적으로각군별전체대상자수와사건발생수가필요하며, 요약통계량으로는 OR, RR, risk difference 등이사용된다. 연속형자료의메타분석을위해서는일반적으로각군의대상자수, 측정변수의평균값및표준편차가필요하다. 효과추정치는모든연구에서동일한척도를사용한경우에평균차이 (mean difference, MD) 를사용하고, 연구마다다른척도를사용하였을경우에표준화된평균차이 (standardized mean difference, SMD) 를사용한다. SMD는각군간의평균값차이를전체참여자결과의표준편차로나누어 (MD/pooled standard deviation) 구한다. 메타분석결과를요약해서제시하는방법으로숲그림 (forest plot) 이사용된다. 이그림은각연구들의효과크기에대한추정치및신뢰구간, 그리고메타분석결합추정치와그신뢰구간을하나의그림으로표시해서보여준다 (Figure 1)[6,11-26]. 6 J Rheum Dis Vol. 22, No. 1, February, 2015

Meta-analysis Method Figure 2. Funnel plot of studies regarding the association between the Fc receptor-like 3-169 C allele and rheumatoid arthritis showed no evidence of asymmetry and Egger s regression test showed no significant p-value (Egger s regression test p-value=0.863), indicating no evidence of publication bias in the meta-analysis. 3) 출판편견조사 출판편견이란의미있는결과를보이는논문이그렇지않은논문보다쉽게그리고빨리논문으로출판되는경향으로인해발생하는오류로 [27], 메타분석은출판된논문을대상으로분석하기때문에이러한출판편견이메타분석결과에영향을미칠수있다. 따라서메타분석과정에서출판편견이있는지에대한조사가필요하다. 출판편견은 x축을효과크기로, y축을정밀도 ( 샘플수나 OR의표준오차등 ) 로하는깔대기점도표 (funnel plot) 를그려서검사한다 (Figure 2)[9,28]. 대상수가적은연구는정확도가떨어지므로다양한효과크기를나타내고, 많은대상수의연구는정확도가높아효과크기의범위가작아서출판편견이없으면깔대기모양이대칭이된다. 출판편견이있는경우에는비대칭성의깔대기모양을보여출판편견존재를판정한다 (Figure 2). 깔대기점도표의단점은대칭판정이주관적이라는것과깔대기모양이되려면다양한대상수의많은연구가있어야한다는것이다. 보완적인방법으로통계적인방법을사용한 Egger 등 [28] 이개발한회귀검사를사용하여깔때기모양의대칭성을 p-수치로나타내기도하므로깔대기점도표방법과보완적으로사용할수있다. 4) 이질성원인검사 의미있는연구간이질성이있는경우에는이질성의원인을찾아야하며, 하위군분석 (subgroup analysis), 민감도분석 (sensitivity analysis), 또는메타회귀분석 (meta-regression) 등의방법이사용된다 [29,30]. 하위군분석에서는메타분석효과추정치에영향을미칠만한환자군이나연구방법의특성과같은요인에따라대상연구들을동질적인 세부집단으로분류해서분석하는것을말한다 (Table 1). 민감도검사란어떤변수에따라메타분석결과가변하는지에대해서검사하는방법으로분석하는변수에따라메타분석결과가민감하지않다면이결과는더신뢰할만한결과를보여주는것이다. 예를들어인종에따라하위군분석을하였을때메타분석결과가변하는지, 또는제일크거나작은효과크기를보이는연구를제외했을때메타분석결과가변하는지를확인하는것이다 (Table 1). 메타회귀분석은결과변수 (outcome variable) 에하나이상의설명변수 (explanatory variables) 가미치는영향을회귀분석하는것으로, 결과변수에는효과크기, 설명변수에는효과크기에영향을미칠수있는연구들의특성이사용된다 [30]. 5) 결과해석및논문발표이와같은방법을통해서나온메타분석결과를분석및해석하여발표한다. 메타분석은후향적연구이고다양한기존의연구를포함하므로출판편견및연구간이질성으로인한오류가메타분석결과에영향을미칠수있으므로메타분석결과해석시이런제한점을고려해야한다. 따라서이에대한분석이필요하며결과해석시이에대한주의및언급이필요하다. 메타분석결과의제한점을포함하여해석하여기술하고메타분석결과를바탕으로향후새로운가설과연구방향을제시할수있다. 논문작성시메타분석및체계적고찰출판가이드라인인 preferred reporting items for systematic reviews and meta-analyses (PRISMA) 가이드라인을따르는것이바람직하다 [31]. 메타분석의제한점및장점메타분석의제한점으로다음과같은사항들이언급되고있다 [3,32,33]. 1) 출판편견등의오류를포함하여메타분석연구결과에영향을미칠수있다 [27]. 2) 기존의연구들을대상으로하는연구이므로기존연구의질에영향을받는다 (gabage in, gabage out). 3) 이질적인연구들을단순하게묶어서분석하는경향이있어서마치사과와배를섞어서연구하는방법이라는비판이있다 (mixing apples with oranges)[34]. 이런제한점들을극복하기위해출판편견유무를확인하고가능한오류를줄이기위한노력이필요하다. 또한양질의연구를포함시키며연구의질이낮은논문이포함된경우에는따로논문의질에따른하위군분석을하고연구결과를비판적으로해석해야한다. 과일의특성을연구하는것이라면사과와배를섞어서분석하는것이도움이될수있으며, 사과와배그룹을각각나누어민감도검사등을통해단점을보완해야한다. 이러한비평에도불구하고메타분석은다음과같은많은장점들을가지고있다 [3,33]. 1) 메타분석은효과크기및효과크기의불확실성을정량적으로측정한다. 2) 연구대상수를증가시켜통계검정력과정밀도가증가하여, 소규모연구에서뚜렷하지않았던결과를보다정확히분석할 www.jrd.or.kr 7

Young Ho Lee 수있다. 3) 연구간의이질성의이유를밝히고, 포괄적이고거시적이고일반적이며객관적인결과를얻어서개별연구결과들간의논란을해결할수있다. 4) 통합된결과를기반으로새로운가설을제공할수있다. 결 론 메타분석은같은주제에대한개별적인연구결과들을통합하여객관적으로분석하는통계적인연구방법으로근거중심의학에서중요한역할을한다. 오류를최소화하고적절히수행된메타분석은통계검정력과연구결과의정밀도를증가시키고개별연구결과들을정량적으로통합하여거시적이고일반적인결과를도출하여개별연구들간의논란을해결할수있으며, 새로운가설을제시할수있어서류마티스질환연구에서중요한연구방법으로활용될수있다. CONFLICT OF INTEREST No potential conflict of interest relevant to this article was reported. REFERENCES 1. Egger M, Smith GD, Phillips AN. Meta-analysis: principles and procedures. BMJ 1997;315:1533-7. 2. Egger M, Smith GD. Meta-analysis. Potentials and promise. BMJ 1997;315:1371-4. 3. Noble JH Jr. Meta-analysis: Methods, strengths, weaknesses, and political uses. J Lab Clin Med 2006;147:7-20. 4. Kelley K, Preacher KJ. On effect size. Psychol Methods 2012;17:137-52. 5. Fleiss JL. The statistical basis of meta-analysis. Stat Methods Med Res 1993;2:121-45. 6. Fleiss JL. Analysis of data from multiclinic trials. Control Clin Trials 1986;7:267-75. 7. Higgins JP, Thompson SG, Deeks JJ, Altman DG. Measuring inconsistency in meta-analyses. BMJ 2003;327:557-60. 8. Lee YH, Woo JH, Choi SJ, Ji JD, Song GG. Fc receptor-like 3-169 C/T polymorphism and RA susceptibility: a metaanalysis. Rheumatol Int 2010;30:947-53. 9. Borenstein M, Hedges LV, Higgins J, Rothstein HR. A basic introduction to fixed effect and random effects models for meta analysis. Res Synth Method 2010;1:97-111. 10. DerSimonian R, Laird N. Meta-analysis in clinical trials. Control Clin Trials 1986;7:177-88. 11. Han SW, Sa KH, Kim SI, Lee SI, Park YW, Lee SS, et al. FCRL3 gene polymorphisms contribute to the radiographic severity rather than susceptibility of rheumatoid arthritis. Hum Immunol 2012;73:537-42. 12. Chen JY, Wang CM, Wu YJ, Kuo SN, Shiu CF, Chang SW, et al. Disease phenotypes and gender association of FCRL3 single-nucleotide polymorphism -169T/C in Taiwanese patients with systemic lupus erythematosus and rheumatoid arthritis. J Rheumatol 2011;38:264-70. 13. Wu H, Yang LH, Zuo J, Liang YL, Li PQ, Liu W, et al. Fc receptor-like 3 gene polymorphisms confer susceptibility to rheumatoid arthritis in a Chinese population. Hum Immunol 2010;71:1203-8. 14. Takata Y, Inoue H, Sato A, Tsugawa K, Miyatake K, Hamada D, et al. Replication of reported genetic associations of PADI4, FCRL3, SLC22A4 and RUNX1 genes with rheumatoid arthritis: results of an independent Japanese population and evidence from meta-analysis of East Asian studies. J Hum Genet 2008;53:163-73. 15. Choi CB, Kang CP, Seong SS, Bae SC, Kang C. The -169C/T polymorphism in FCRL3 is not associated with susceptibility to rheumatoid arthritis or systemic lupus erythematosus in a case-control study of Koreans. Arthritis Rheum 2006;54:3838-41. 16. Ikari K, Momohara S, Nakamura T, Hara M, Yamanaka H, Tomatsu T, et al. Supportive evidence for a genetic association of the FCRL3 promoter polymorphism with rheumatoid arthritis. Ann Rheum Dis 2006;65:671-3. 17. Kochi Y, Yamada R, Suzuki A, Harley JB, Shirasawa S, Sawada T, et al. A functional variant in FCRL3, encoding Fc receptor-like 3, is associated with rheumatoid arthritis and several autoimmunities. Nat Genet 2005;37:478-85. 18. Maehlen MT, Nordang GB, Syversen SW, van der Heijde DM, Kvien TK, Uhlig T, et al. FCRL3-169C/C genotype is associated with anti-citrullinated protein antibody-positive rheumatoid arthritis and with radiographic progression. J Rheumatol 2011;38:2329-35. 19. Owen CJ, Kelly H, Eden JA, Merriman ME, Pearce SH, Merriman TR. Analysis of the Fc receptor-like-3 (FCRL3) locus in Caucasians with autoimmune disorders suggests a complex pattern of disease association. J Clin Endocrinol Metab 2007;92:1106-11. 20. Thabet MM, Wesoly J, Slagboom PE, Toes RE, Huizinga TW. FCRL3 promoter 169 CC homozygosity is associated with susceptibility to rheumatoid arthritis in Dutch Caucasians. Ann Rheum Dis 2007;66:803-6. 21. Newman WG, Zhang Q, Liu X, Walker E, Ternan H, Owen J, et al. Rheumatoid arthritis association with the FCRL3-169C polymorphism is restricted to PTPN22 1858T-homozygous individuals in a Canadian population. Arthritis Rheum 2006;54:3820-7. 22. Eyre S, Bowes J, Potter C, Worthington J, Barton A. Association of the FCRL3 gene with rheumatoid arthritis: a further example of population specificity? Arthritis Res Ther 2006;8:R117. 23. Hu X, Chang M, Saiki RK, Cargill MA, Begovich AB, Ardlie KG, et al. The functional -169T-->C single-nucleotide polymorphism in FCRL3 is not associated with rheumatoid arthritis in white North Americans. Arthritis Rheum 2006;54:1022-5. 24. Martínez A, Sánchez E, Valdivia A, Orozco G, López-Nevot MA, Pascual-Salcedo D, et al. Epistatic interaction between FCRL3 and NFkappaB1 genes in Spanish patients with rheumatoid arthritis. Ann Rheum Dis 2006;65:1188-91. 25. El-Gabalawy HS, Robinson DB, Daha NA, Oen KG, Smolik I, Elias B, et al. Non-HLA genes modulate the risk of rheumatoid arthritis associated with HLA-DRB1 in a susceptible North American Native population. Genes Immun 2011;12:568-74. 26. Lewis S, Clarke M. Forest plots: trying to see the wood and 8 J Rheum Dis Vol. 22, No. 1, February, 2015

Meta-analysis Method the trees. BMJ 2001;322:1479-80. 27. Simes RJ. Confronting publication bias: a cohort design for meta-analysis. Stat Med 1987;6:11-29. 28. Egger M, Davey Smith G, Schneider M, Minder C. Bias in meta-analysis detected by a simple, graphical test. BMJ 1997;315:629-34. 29. Oxman AD, Guyatt GH. A consumer's guide to subgroup analyses. Ann Intern Med 1992;116:78-84. 30. Schmid CH. Exploring heterogeneity in randomized trials via meta-analysis*. Drug Inf J 1999;33: 211-24. 31. Moher D, Liberati A, Tetzlaff J, Altman DG; PRISMA Group. Preferred reporting items for systematic reviews and meta-analyses: the PRISMA statement. Ann Intern Med 2009;151:264-9. 32. Thompson SG, Pocock SJ. Can meta-analyses be trusted? Lancet 1991;338:1127-30. 33. Bailar JC 3rd. The promise and problems of meta-analysis. N Engl J Med 1997;337:559-61. 34. Sharpe D. Of apples and oranges, file drawers and garbage: why validity issues in meta-analysis will not go away. Clin Psychol Rev 1997;17:881-901. www.jrd.or.kr 9