Stratified Analysis: Mantel-Haenszel Techniques Instructor: 李奕慧 yihwei@mail.tcu.edu.tw 1 Lecture Overview 1. 2. 3. 4. Review example: ”Risk factors associated with lung cancer in Hong Kong” Mantel-Haenszel Technique for Stratified Tables Modification effect (Interaction effect) Application: Meta-Analysis 2 Confounding factors (干擾因素) Confounder: Variable is associated with both the disease and the exposure variable. 3 Method for control for confounders 1. 2. 3. Study design: restriction/ matching/ randomization Statistical adjustment: Standardization; e.g. age standardized (where age is a confounder) Stratified by confounder (Mantel-Haenszel test) Incorporate the confounder into a regression analysis as a covariate. (logistic regression approach) 4 Restriction Example 研究主旨:二手煙(ETS, exposure)與罹患肺癌 (disease)的關係 confounder: 研究對象本身是否抽煙 為了避免干擾只分析ETS對nonsmoker的影響 5 Stratified Analysis 6 將性別當作分層(stratum)的因子 smoking * case * sex Crosstabulation Count sex male smoking ex- and current smoker nonsmoker female Total smoking ex- and current smoker nonsmoker Total case case control 160 116 Total 276 52 212 13 96 212 6 148 424 19 106 119 113 119 219 238 Lung cancer2.sav 7 Sex-Specific OR for smoking Risk Estimate sex male female 95% Confidence Interval Value Lower Upper 2.55 1.68 3.85 Odds Ratio for smoking (ex- and current smoker / nonsmoker) N of Valid Cases Odds Ratio for smoking (ex- and current smoker / nonsmoker) 424 2.31 N of Valid Cases 238 0.85 6.30 可以將男士的OR與女士的OR合併嗎? 怎麼併? Lung cancer2.sav 8 Don’t do! 完全忽略性別(confounder) OR=1.88 距離2.31或2.55 都很遠, smoking * case Crosstabulation smoking Total case case control 173 122 52.3% 36.9% 158 209 47.7% 63.1% 331 331 100.0 100.0 % % ex- and current Count smoker % within case nonsmoker Count % within case Count % within case Total 295 44.6% 367 55.4% 662 100.0 % Risk Estimate 95% Confidence Interval Value Odds Ratio for smoking (ex- and current smoker / nonsmoker) N of Valid Cases Lower 1.88 Upper 1.38 2.56 662 9 男、女的OR很接近嗎?可以將 男女的OR整合嗎? H0: ORm = ORf = OR (common odds ratio) 抽煙對男、女性罹癌的風險是否有差異? Test of the Homogeneity of Odds Ratio (OR的同質性檢定) Tests of Homogeneity of the Odds Ratio Breslow-Day Tarone's Chi-Squared .031 .031 df Asymp. Sig. (2-sided) 1 .860 1 .860 10 整合後的OR如何? Mantel-Haenszel Common Odds Ratio Estimate Estimate 介於2.31~2.55之間 ln(Estimate) ln(2.51)=0.92 Std. Error of ln(Estimate) 標準誤 Asymp. Sig. (2-sided) p-value Common Odds Ratio Lower Bound Asymp. 95% Upper Bound Confidence Interval ln(Common Odds Ratio) Lower Bound Upper Bound 2.509 .920 .195 .000 1.711 3.678 .537 1.302 The Mantel-Haenszel common odds ratio estimate is asymptotically normally distributed under the common odds ratio of 1.000 assumption. So is the natural log of the estimate. 11 Confidence Interval and Testing for common OR 1. Obtain confidence interval for ln(OR) ln(OR) 1.96*SE 0.92 1.96*0.195 (0.38) (0.92-0.38, 0.92+0.38)=(0.54, 1.3) 2. 3. 4. Exponentiate these limits (e0.54, e1.3)=(1.71, 3.68) 當控制性別後,抽煙者罹患肺癌的風險是不抽 煙者的1.7~3.7倍 M-H test for common OR=1: p-value< 0.001 12 Sex-Specific OR for smoking Risk Estimate sex male female 95% Confidence Interval Value Lower Upper 2.55 1.68 3.85 Odds Ratio for smoking (ex- and current smoker / nonsmoker) N of Valid Cases Odds Ratio for smoking (ex- and current smoker / nonsmoker) 424 2.31 N of Valid Cases 238 0.85 6.30 男性OR信賴區間較窄,標準誤較小,給予較大 的權重。女性的CI較寬,標準誤較大,給予較 小的權重。 Common OR=2.51 Lung cancer2.sav 13 M-H分析的應用:Forest Plot Odds ratio No. of events Study (95% CI) Treatment Control male 2.55 ( 1.68, 3.85) 160/212 116/212 female 2.31 ( 0.85, 6.30) 13/119 6/119 2.51 ( 1.71, 3.68) 173/331 122/331 Overall .1 .5 1 2 Odds ratio smoking better non-smoking better 10 Sex-specific OR Common OR 14 Layer:分層 Mantel-Haenszel Statistics 15 如果不能整合,怎麼辦? Table 4: Impact of fatty food consumption on lung cancer risk by Gender Male Female 16 Stratified Tables fat * lungcancer * sex Crosstabulation Count sex male fat Total female fat lungcancer yes no 161 130 51 80 212 210 69 73 50 43 119 116 moderate/high fat low fat moderate/high fat low fat Total Total 291 131 422 142 93 235 Risk Estimate sex male Odds Ratio for fat (moderate/high fat / low fat) N of Valid Cases female Odds Ratio for fat (moderate/high fat / low fat) N of Valid Cases 95% Confidence Interval Value Lower Upper 1.943 1.276 2.958 422 .813 235 .481 1.373 Lung cancer3.sav 17 可以將男女的OR整合嗎? H0: ORm = ORf = OR (common odds ratio) 脂肪攝取對男、女性罹癌的風險是否有差異? 如有差異,則表示此危險因子,在男女性的表 現是不一樣的,不能將兩者整合。 Tests of Homogeneity of the Odds Ratio Chi-Squared df Asymp. Sig. (2-sided) Breslow-Day 6.498 1 .011 Tarone's 6.497 1 .011 18 Interaction or modification If the stratum-specific odds ratios ( say lung cancer) are different across the 2 (or g) strata, then there is an interaction between Exposure (fat consumption) and Confounder (gender), and the Confounder is an effect modifier (修飾因子). 脂肪攝取與性別會交互影響肺癌的發生風 險 19 Multiple 2 X 2 Tables No interaction With interaction 20 M-H分析的應用: Meta-Analysis Odds ratio No. of events Study (95% CI) Treatment Control Ip (1989) 0.12 ( 0.04, 0.36) 7/35 23/34 Liu (1987) 0.03 ( 0.01, 0.14) 3/27 21/26 Xu (1955) 0.20 ( 0.07, 0.58) 7/60 12/30 Xu (1995) 0.46 ( 0.18, 1.17) 14/60 12/30 0.17 ( 0.10, 0.30) 31/182 68/120 Overall .01 .1 1 10 100 Odds ratio Vaccine better Placebo better Hepatitis B.sav 21 Outcome * Vaccine * study Crosstabulation Count study Ip 1989 Outco me Liu 1987 Total Outco me Xu 1995a Total Outco me Xu 1995b Total Outco me Total Vaccine vaccin place e bo Total infected 7 23 30 28 11 39 not infected 35 34 69 infected 3 21 24 24 5 29 not infected 27 26 53 infected 7 12 19 53 18 71 not infected 60 30 90 infected 14 12 26 46 18 64 not infected 60 30 90 Risk Estimate 95% Confidence Interval study Ip 1989 Value Lower Upper OR .120 .040 .358 OR .030 .006 .140 OR .198 .068 .580 OR .457 .178 1.174 Liu 1987 Xu 1995a Xu 1995b 22 H0: OR1=OR2=OR3=OR4 檢定4個研究的OR是否相同 P=0.019 表示這4個OR差異很大 Tests of Homogeneity of the Odds Ratio Chi-Squared Breslow-Day Tarone's Asymp. Sig. (2sided) df 10.003 3 .019 9.967 3 .019 23 M-H分析的應用 Odds ratio No. of events Study (95% CI) Treatment Control Ip (1989) 0.12 ( 0.04, 0.36) 7/35 23/34 Liu (1987) 0.03 ( 0.01, 0.14) 3/27 21/26 Xu (1955) 0.20 ( 0.07, 0.58) 7/60 12/30 Xu (1995) 0.46 ( 0.18, 1.17) 14/60 12/30 0.17 ( 0.10, 0.30) 31/182 68/120 Overall .01 .1 1 10 100 Odds ratio Vaccine better Placebo better Hepatitis B.sav 24 Common OR: 整合後的OR =0.18, 95%CI (0.10- 0.30) 檢定整合後的OR=1, p=0.000 Mantel-Haenszel Common Odds Ratio Estimate Estimate ln(Estimate) Std. Error of ln(Estimate) Asymp. Sig. (2-sided) Asymp. 95% Confidence Interval Common Odds Ratio ln(Common Odds Ratio) .175 -1.744 .269 .000 Lower Bound Upper Bound Lower Bound Upper Bound .103 .296 -2.271 -1.218 The Mantel-Haenszel common odds ratio estimate is asymptotically normally distributed under the common odds ratio of 1.000 assumption. So is the natural log of the estimate. 25 Fig 2 Effect of hepatitis B vaccine on occurrence of hepatitis B in newborn infants. Test for heterogeneity 檢定RR1=RR2=RR3=RR4是否相等 Test for overall effect 檢定整合後的RR是否等於1 BMJ 2006;332:328-336 26 Thank you! 27