1、临床试验常用统计分析方法临床试验常用统计分析方法多因素分析多因素分析HEPATOLOGY 2015;61:1809-1820Clinical Gastroenterology and Hepatology.2015;动物喂养实验动物喂养实验(评价不同的膳食组合评价不同的膳食组合)2 组兔子组兔子抓一组放在新笼子：新膳食组合抓一组放在新笼子：新膳食组合剩余留在旧笼子中：标准膳食组合剩余留在旧笼子中：标准膳食组合3个月后评价组间个月后评价组间体重增加体重增加 的差异的差异动物喂养实验动物喂养实验-3个月后结果个月后结果试验组：平均体重比常规膳食组试验组：平均体重比常规膳食组显著增加显著增加问题问题
2、：新膳食组基线的平均体重，比常规：新膳食组基线的平均体重，比常规组重组重(有显著性差异有显著性差异)原因原因?!?DES(n=11516)BMS(n=6210)p valueAge yrs64.4 12.365.3 13.00.001Female(%)32.0 32.2 0.85Diabetes Mellitus(%)28.927.40.04 Insulin Dependent Diabetes9.78.80.07Hyperlipidemia(%)77.171.90.001Hypertension(%)76.172.60.001Current Smoker(%)19.623.20.001Pri
3、or PCI(%)22.520.70.004Prior MI(%)26.928.70.01Prior CABG(%)13.016.10.001Laura Mauri,N Engl J Med 2008;359:1330-42.12.4%6.7%MortalityP 0.0001 RevascularizationP 0.000124.0%20.0%MIP 0.000112.1%9.1%11.9%9.4%MortalityP 0.0001 RevascularizationP 0.000123.9%20.1%MIP=0.1111.8%10.8%Before AdjustmentAfter Adj
4、ustment预后因素预后因素结局指标结局指标混杂混杂年年龄龄60岁岁5%15%试验组试验组对对照照组组10%20%试验组试验组对对照照组组指标指标试验组试验组(N=100)对照组对照组(N=100)p基线：基线：age6050501.0000主要终点：死亡主要终点：死亡9190.0396指标指标试验组试验组(N=100)对照组对照组(N=100)p基线：基线：age606733 死死不掉河里不掉河里=死活不知死活不知怎么办：怎么办：尝试过河尝试过河关心的问题：关心的问题：成功的可能性？成功的可能性？删失删失生存时间生存时间生存函数生存函数12345678910开始过河！开始过河！INTERV
5、ALSTARTED INTERVALDIEDCENSORED0-120221-21612-315413-410114-5815-676-77117-858-9519-10420只兔子过河的结果只兔子过河的结果INTERVALSTARTED INTERVALDIEDCENSOREDP(SURV)S(t)0-12022.900.9001-2161.937.8442-31541.733.6193-41011.900.5574-581.875.4875-671.00.4876-7711.857.4187-851.00.4188-9518.00.3349-1041.00.334生存函数的估计生存函数的估
6、计石头编号石头编号生存函数生存函数Kaplan-Meier曲线曲线12345678910过河不易过河不易Eur Heart J 2004;25:201318.PRAIS-UK研究研究注册登记注册登记Non-STEMI患者患者 653例例随访随访4年年研究与死亡相关的危险因素研究与死亡相关的危险因素HEPATOLOGY 2015;61:1809-1820Factors for Transplant-Free Survival036072005101520No.at RiskPCICABG300300272276236239Days Since RandomizationCumulative In
7、cidence,%8.712.26.78.1PCICABGN Engl J Med 2011;364:1718-27.PROCOMBAT study8,6888,763010203086420Cumulative incidence(%)ClopidogrelTicagrelor4.775.43HR 0.88(95%CI 0.771.00),p=0.045No.at riskClopidogrelTicagrelor9,2919,3338,8758,9428,7638,827Days after randomisation319015021027033086420ClopidogrelTica
8、grelor5.286.608,6888,7638,2868,3976,3796,480Days after randomisation*HR 0.80(95%CI 0.700.91),p0.0018,4378,5436,9457,0284,7514,822Cumulative incidence(%)Landmark analysis:Primary efficacy endpoint(composite of CV death,MI or stroke)*Excludes patients with any primary event during the first 30 days线性回
9、归线性回归Logistic回归回归Cox回归回归结局指标结局指标类型类型连续连续二分二分事件时间事件时间删失删失不允许不允许不允许不允许允许允许建模对象建模对象结局的均值结局的均值Log(结局的结局的odds)Log(结局的结局的Hazard)预后变量预后变量连续连续x一个单位一个单位:y的均数改变的均数改变x一个单位一个单位:y的的OR改变改变x一个单位一个单位:y的的HR改变改变二分二分x两水平间两水平间:y的均数改变的均数改变x两水平间两水平间:y的的OR改变改变x两水平间两水平间:y的的HR改变改变多分类多分类x各水平比参照各水平比参照:y的均数改变的均数改变x各水平比参照各水平比参照
10、:y的的OR改变改变x各水平比参照各水平比参照:y的的HR改变改变BMJ 2017;356:j273Eur J Epidemiol(2009)24:733736常用方法占比：常用方法占比：28%for PK&20%for SW值得注意的数字值得注意的数字0%vs.35%根根据既往临床经验指定变量据既往临床经验指定变量N Engl J Med 2014;371:612-23.Forward selection algorithm(i)Fit all possible simple linear models of the response variable against each separa
11、te covariate.Select the covariate with the lowest P-value and include it in the models of the subsequent steps.(ii)Fit all possible models with the covariate(s)selected in the preceding step(s)plus one other of the remaining covariates.Select the new covariate that has the lowest P-value and add it
12、to all subsequent models.(iii)Repeat step(ii)to add additional variables,one variable at a time.Continue this process until either none of the remaining covariates has a P-value less than some threshold or all of the covariates have been selected.BMC Medical Research Methodology 2010,10:8704,06和和08年
13、发表的年发表的1882篇篇，仅仅287(15%)谈及谈及UnivariteMultivariate*ItemOR(95%CI)POR(95%CI)pNutrition support0.46(0.25-0.82)0.0090.28(0.14-0.57)0.003Severity of disease score4.91(2.08-11.57)0.00032.65(0.99-7.00)0.051Nutritional status score1.02(0.77-1.34)0.9011.66(1.21-2.28)0.002Age 1.02(1.00-1.04)0.0481.02(1.00-1.04
14、)0.038Sex 1.13(0.66-1.94)0.6481.02(0.68-1.51)0.939Medical versus Surgical admissions 0.19(0.08-0.45)0.00020.28(0.10-0.79)0.017Insurance2.26(0.79-6.48)0.1312.35(0.57-9.78)0.239现有方法的局限性现有方法的局限性全模型全模型 样本量不足导致过度拟合样本量不足导致过度拟合临床经验设定临床经验设定 遗漏有显著影响的变量遗漏有显著影响的变量Stepwise 假阳性增加、外推性减低假阳性增加、外推性减低多多因因素分析的目的？素分析的目
15、的？评价评价“因果关联因果关联”时，时，控制混杂控制混杂建建立预测模型立预测模型发生终点发生终点OR值值仅仅SBP仅仅DBP全模型全模型逐步回归逐步回归SBP1.04(1.02,1.07)-1.07(0.98,1.16)1.04(1.02,1.07)DBP-1.03(1.01,1.06)0.98(0.91,1.05)-c统计量统计量0.7190.6870.7280.719发生终点发生终点OR值值仅仅SBP仅仅DBP全模型全模型逐步回归逐步回归SBP1.04(1.01,1.06)-0.97(0.90,1.04)-DBP-1.05(1.02,1.08)1.08(1.01,1.17)1.05(1.0
16、2,1.08)c统计量统计量0.6600.6860.7020.686模拟模拟100例患者的血压与终点事件发生的数据例患者的血压与终点事件发生的数据识别共线性的方法识别共线性的方法相关分析相关分析容差容差(tolerance),VIF,条件指数条件指数判定标准判定标准?消除共线性的方法消除共线性的方法逐步回归逐步回归?对存在共线性的指标进行敏感性分析对存在共线性的指标进行敏感性分析扩大样本，减小标准误，减弱共线性影响扩大样本，减小标准误，减弱共线性影响VariableUnivariateP valueMultivariteP valueOR Value95%CIOR Value95%CIAge1
17、.071.04-1.110.00011.061.03-1.100.0001Family history of CHD2.101.20-3.680.0101.630.86-3.090.133Diabetes2.831.48-5.420.0022.231.10-4.520.027Hypertension2.151.20-3.850.0101.300.66-2.600.449Dyslipidemia1.931.07-3.480.0301.110.55-2.240.762RDW1.971.28-3.030.0022.001.24-3.220.004hs-CRP1.191.05-1.340.0051.1
18、41.01-1.290.029Table X Logistic regression analysis of independent predictor for CHD*Cohort study,consecutive recruit 233 patients.Their multivariate statistical models are performed within only 233 patients,therefore a power analysis could be helpful in view of over fitting problems(RDW).回答审稿人的问题回答
19、审稿人的问题把握度分析把握度分析(事后事后)Over fitting 指的是什么指的是什么?VariableCHD group(N=147)Control(N=86)P valueAge(years)63.08.757.39.60.0001BMI(kg/m2)25.43.625.13.10.518Smoking history11(7.5)7(8.1)1.000Diabetes55(37.4)15(17.4)0.002Hypertension114(77.6)53(61.6)0.013Dyslipidemia115(78.2)56(65.1)0.031Family history of CHD
20、70(47.6)26(30.2)0.011Hemoglobin(g/L)129.311.2127.77.80.221RDW(%)12.90.712.60.60.001hs-CRP1.94(0.85,3.35)1.58(0.69,2.76)0.139VariableUnivariateP valueMultivariteP valueOR Value95%CIOR Value95%CIAge1.071.04-1.110.00011.061.03-1.100.0001Family history of CHD2.101.20-3.680.0101.630.86-3.090.133Diabetes2
21、.831.48-5.420.0022.231.10-4.520.027Hypertension2.151.20-3.850.0101.300.66-2.600.449Dyslipidemia1.931.07-3.480.0301.110.55-2.240.762RDW1.971.28-3.030.0022.001.24-3.220.004hs-CRP1.191.05-1.340.0051.141.01-1.290.029OR=exp(beta),beta=log(OR),ORadjusted=exp(beta/10)ORadjusted=1.07,95%CI:1.02-1.12 生活质量评分生
22、活质量评分1=非常差非常差,3=一般一般,5=感觉很好感觉很好取平均得分意味着取平均得分意味着1和和2分之间的差别与分之间的差别与4和和5分之间一致分之间一致(1+5)/2=(2+4)/2如果如果1=死亡将如何？死亡将如何？地区地区 赋值赋值东部东部1中部中部2西部西部3地区地区赋值赋值哑变量哑变量1哑变量哑变量2东部东部110中部中部201西部西部300 通常，哑变量的数量等于原始变量的分类数减通常，哑变量的数量等于原始变量的分类数减 1评价某实验室指标与终点事件发生率的关系评价某实验室指标与终点事件发生率的关系按四分位数分组按四分位数分组分组变量对应的取值分组变量对应的取值1-4Logis
23、tic回归回归不同策略的比较不同策略的比较直接采用分组变量直接采用分组变量哑变量处理哑变量处理变量变量p值值OR值值95%置信区间置信区间Group0.16091.1160.957,1.301变量变量p值值OR值值95%置信区间置信区间Group 1 vs 40.39400.7850.498,1.238Group 2 vs 40.01430.4590.277,0.758Group 3 vs 40.57990.6340.396,1.016直接采用分组变量直接采用分组变量哑变量处理哑变量处理指标指标Quintile1(N=200)Quintile2(N=200)Quintile3(N=200)Qu
24、intile4(N=200)事件数事件数(%)45(22.5%)29(14.5%)38(19.0%)54(27.0%)缺缺失失缺缺失失缺缺失失缺缺失失缺缺失失Missing54CoCr(%)PtCr(%)Relative Risk(95%CI)Relative Risk(95%CI)P InteractionAll randomized(n=1507)7.1%5.9%0.84 0.56,1.26Age 65 yrs(n=779)8.0%5.3%0.67 0.38,1.160.29Age 65 yrs(n=728)6.8%6.9%1.02 0.58,1.79Male(n=1074)6.3%6.3
25、%1.00 0.62,1.610.17Female(n=433)10.4%5.8%0.56 0.28,1.11Diabetic(n=351)12.7%8.0%0.63 0.32,1.240.30Nondiabetic(n=1156)5.8%5.6%0.97 0.60,1.58Single vessel Tx(n=1399)7.7%5.9%0.77 0.51,1.160.25Dual vessel Tx(n=108)4.3%8.8%2.02 0.41,9.92BMI 2.62 mm(n=728)6.2%4.9%0.80 0.43,1.50Lesion 13.0 mm(n=870)7.2%6.0%
26、0.84 0.50,1.410.90Lesion 13.0 mm(n=636)7.9%6.3%0.80 0.44,1.44TLF at 3 YearsPtCr betterCoCr betterBinary RatesGood OutcomeIntermediate OutcomeBad OutcomeOutcomes from an RCTMean Treatment EffectFalse-negative mortality effect in a subgroup defined only by the birth sign from ISIS-2Astrological Birth
27、signNo.of 1-month deaths(Aspirin vs placebo)Statistical significanceLibra or GeminiAll other signsAny birth sign150 vs 147654 vs 869804(9.4%)vs 1016(11.8%)NS2p0.0000012p0.05P for interaction 0.05P for interaction 0.05总样本总样本=18,000,相对风险相对风险=0.84(95%CI 0.77-0.92),P=0.0003Circulation.2011;124:544-554.Primary Efficacy OutcomeASA Maintenance DoseDays from Randomization0 060120180240300360161412108 86 64 42 20 0Tic:ASA HighClop:ASA HighClop:ASA LowTic:ASA LowASA Low(300mg):ASA High(300mg):HR(95%CI),0.79(0.71,0.88)HR(95%CI),1.45(1.01,2.09)Kaplan-Meier Estimated Rate(%)
