1、Univariate Analysis and Table 1Mark Drazner,MD,MScUniversity of Texas Southwestern Medical CenterJune 11,2004Risk Factor OutcomeExposureExplanatory variableIndependent variableDependent variableResponseDiseaseTypes of Variables in Your ExperimentObjective Is the risk factor associated with the outco
2、me?Is there a pattern?Validity and Generalizability Internal Validity I find a pattern,is it real(reproducible)?ChanceBiasConfounding External Validity Does it apply to other patients?GeneralizabilityTools Against Invalidity Chance:Statistics Bias:Study design(Was it a fair study?)Confounding:Multiv
3、ariable analysisStatistical Analyses Univariate 1 Risk Factor 1 Outcome MultivariableMultiple risk factors 1 OutcomeWhich Statistical Test Do I Use?It depends on type of data being analyzed Types of data in clinical research Categorical:Nominal vs.Ordinal Dichotomous 2 categories Continuous Survival
4、 Paired data Limited types of data means limited number of statistical testsHypertensionConcentric LVHDilated CardiomyopathyParadigm of Hypertensive Heart Disease“Transition to Failure”Rame et.al,AJC,2004Retrospective Study:Parkland Echocardiography Database(1992 1994)Low EF Normal EF n=28 n=131 (18
5、%)(82%)159 subjects Concentric LVH(1.3 cm)Normal LVEF Follow-up echo.1 year later 4 year f/uRame,AJC,2004Some Purposes of Table 1 Provide baseline characteristics What patients were studied?Are they applicable to the readers patients(generalizability)?Is the study believable e.g.,authors say they st
6、udied advanced heart failure,do patients have characteristics of advanced heart failure?What are differences at baseline between 2 groups(e.g.,case-control)?Randomized clinical trial:Did randomization work?How Did We Make Table 1?Methods Section “Data are expressed as proportions,means SD,or medians
7、(25th and 75th percentiles).Differences in characteristics between patients who did or did not progress to a depressed EF were compared by Fishers exact test,Students t test with equal or unequal variances,or Wilcoxon rank-sum test,where appropriate.”Rame,AJC,2004What Type of Data are You Analyzing?
8、Dichotomous OutcomeEither OrDichotomous Risk FactorsEpidemiologists Love 2 x 2 TablesDisease/OutcomeRisk Factor/Exposure+-+-Statistical Tests for Dichotomous Risk Factor+Dichotomous Outcome Chi-square(approximation)Fishers Exact TestChi-SquareObserved effect(what we see)Expected(under null)Expected(
9、under null)DiseaseExposure+-+-A BC Dn1n2c1 c2X2=N(AD-BC)2n1n2c1c2A or D=patternDichotomous Risk Factors18/28 who develop low EF are AA104/131 who do not develop low EF are AAP=0.14Dichotomous Risk FactorsOnly 129 subjects(not 159)=missing data10(45%)who develop low EF had CAD19(18%)who did not devel
10、op low EF had CADP=0.01Learn a Statistical Package!SAS SPSS STATAContinuous Risk FactorsContinuous Risk FactorDichotomous Outcome Students T test Equal variance Unequal variance Wilcoxon Rank Sum(Mann-Whitney-U)T test vs.Wilcoxon Rank Sum:Are data normally distributed?How Do You Tell if Data are Nor
11、mally Distributed?Look for characteristics of normality Is it bell-shaped(histogram)?Mean=median=mode Symmetry of data(skewness)Length of tails/how common are values near and far from mean(kurtosis)Statistical tests of normality Normal probability plot Wilk-Shapiro testNormal DistributionLeft skewRi
12、ght skewHigh kurtosisLot of tailsLow kurtosis“Boxy”Table 1 A series of statistical tests Dichotomous risk factors+outcome Chi square Fishers exact test Continuous risk factors and dichotomous outcome T test Wilcoxon rank sum Have not addressed categorical risk factors or outcomes with 2 levels,survi
13、val analysis,paired data Confounder 3rd factor that is associated with exposure/risk factor(height)Independent determinant of outcome(wealth)How to handle in analysis:StratificationMultivariable analysisUnivariate Analysis Its a start but Need to move to Multivariable Analysis to know if your risk factor is independently associated with outcome(“unconfounded association”)Conclusions Guard against chance,bias,confounding Learn a statistical package Use the right statistical test,based on what type of data