Categorical Data Analysis Using SAS And Stata

Categorical Data AnalysisUsing SAS and StataHsueh-Sheng WuCenter for Family and Demographic ResearchMar 3, 20141

Outline Why do we need to learn categorical data analyses? A summary of different categorical data analyses– Analyses of contingency tables– Regression models Logistic regression Ordered logistic regression Multinomial logistic regression Stata commands SAS commands Interpreting the results Predicted probability Conclusions2

Why Do We Need to Learn CategoricalData Analysis? Four measurement levels––––Nominal (e.g., gender, race)Ordinal (e.g., attitude toward cohabitation)Interval (e.g., temperature)Ratio (e.g., income) Categorical variables are those measured atnominal and ordinal levels. Interval or ratio variables can be transformedinto nominal or ordinal variables, but not theother way around.3

What Is Special about Categorical Variable? The distribution of a categorical variable is described by itsfrequency and proportion rather than by its mean andvariance. Statistical methods (i.e., t-test, correlation, OLS regression)designed for continuous dependent variables are notadequate for analyzing categorical dependent variables. The decision on how to analyze categorical variables is oftenbased on:– The measurement level and number of categories in dependentvariables– The measurement level and number of categories in independentvariables– Sample size– Number of independent variables4

Different Models for CategoricalDependent VariablesCategorical models address three types of questions: Examination of contingency tables– Proportions– Relative risks– Odds ratio How the characteristics of individuals affect the choice– Binary logistic regression– Ordered logistic regression– Multinomial logistic regression5

Analyzing a Two-way Contingency Table Analyzing a 2x2 tableTable. Gender and EmploymentTable. Gender and 200200Maleρ11-ρ 1Female200400Femaleρ21-ρ 2Difference of Two Proportions 1 2 1 2SE 1 (1 1) 2 (1 2)n1 n26

Analyzing a Two-way Contingency Table (Cont.) 1Relative Risk 2Odds RatioOdds1Odds Ratio Odds2 1 2(1 1) 11 (1 2)SE 1n11 12 21 22 1n12 11. 22 12 . 211n 21 1n 227

Example DataP1 200/400 0.5P2 200/600 0.33– Difference of two proportionsP1 - P2 0.17– Relative riskP1/P2 1.51– Odds Ratio(200*400)/(200*200) 28

Analyzing a Three-way Contingency Table A three-way contingency table can be viewedas multiple two-way contingency tablescreated at different levels of a third variable. Example:Table. Relations among Country, Gender, and EmploymentCounty ACountry BEmployedUnemployedEmployed UnemployedMale1801202080Female12080803209

Example– Difference of proportionCountry A: (180/300) – (120/200) 0Country B: (20/100) –(80/320) 0– Relative riskCountry A: (180/300)/(120/200) 0.6/0.6 1Country B: (20/100) –(80/320) 0.2/0.2 1– Odds RatioCountry A: (180*80)/(120*120) 1Country B: (20*320)*(80*80) 110

Models for Examining How Characteristicsof Individuals Affect ChoicesLogistic Regression ( x) 1log( ) log() 1 ( x) 2 exp( )e ( ) 1 exp( ) 1 e Ordered Logistic Regressionp(Y j ) 1 . j,j 1,., J . j ,p(Y j ) 1logit [ p(Y j )] log[] log[], j 1,., J1 p(Y j ) j 1 . J ,11

Models for Examining How Characteristics ofIndividuals Affect Choices (Cont.)Multinomial Logistic Regressionlog(log( j) jJa ) log( aj , j 1,., J 1 J) log( ( ) ( ) ( ) ( ) bb aJ) log( b)JJaa bbabb12

Relations among These Three Models Ordered logistic regression and multinomial logisticregression are an extension of logistic regression. Both ordered and multinomial logistic regression can betreated as models simultaneously estimating a series oflogistic regression. Ordered logistic regression assumes different intercepts,but the same slope for different categories, whilemultinomial logistic regression assumes differentintercept and slope parameters for different categories.13

A List of Variables in the Datavariable namevariable labelLabel ValueLabel Label57101310 99719978aidIDmarriedMarital Status01Not marriedMarriededucEducation1234Less than High SchoolHigh SchoolSome collegecolleges or moreunionUnion FemaleageAgeagesqAge squaredfemaleageInteraction term of female and age24-33576-10890-3314

Data for Logistic Regression, Ordered LogisticRegression, and Multinomial Logistic 282615

Stata CommandsLogistic Regressionlogit married female age femaleagelogit married female age femaleage, orOrdered Logistic Regressionologit educ female age femaleageologit educ female age femaleage, orMultinomial Logistic Regressionmlogit union female age femaleage,base(0)16

SAS CommandsLogistic RegressionProc Logistic data in.annotated 3 2;Format married marriedf. educ educf.;Model married educ female age femaleage;run;17

SAS and Stata CommandsOrdered Logistic RegressionProc Logistic data in.annotated descending;Format educ educf. female femalef.;Model educ female age femaleage;run;PROC QLIM data in.annotated;MODEL educ female agefemaleage/DISCRETE (DIST LOGISTIC);RUN;Multinomial Logistic Regressionmlogit union educ female age femaleage, base(0)18

SAS and Stata CommandsMultinomial Logistic Regressionproc logistic data in.annotated 3 2;class union (ref "0");model union educ female age femaleage/link glogit;run;19

Interpreting the Results The sample sizeThe reference categoryThe regression coefficientsThe odds ratio20

Predicted Probability Predicted probability is useful to describethe results Odds Exp(the sum of coefficients) Predicted Probability Odds/(1 Odds) You can present predicted probability withgraphs21

Predicted Probability (continued)Ta bl e 3. Predi ca ted Proba bi l i ty for Ma l e a nd Fema l e Res pondentsFema l eInterceptAgeAge*Fema l eSum ofcoeffi centsOdds Ra tioPredi ctedProba bi l i tycoeffi centva l uecoeffi centva l uecoeffi centva l uecoeffi 9771822

Predicted Probability (continued)23

Conclusions If you have categorical dependent variables, youneed to choose adequate methods to analyze them. You need to choose the regression models that fityour data and research questions. If you have event counts (e.g., the number ofaccidents), you need to use other models such asPoisson regression, Log-linear model, or Negativebinomial regression for analyses. For additional help with categorical data analysis,feel free to contact me at wuh@bgsu.edu and 3723119.24

Categorical Data Analysis Using SAS and Stata Hsueh-Sheng Wu Center for Family and Demographic Research Mar 3, 2014 . . 12 1 11 1 12. 21 11. 22 22 21 12 11 (1 2) 2 (1 1) 1 Odds2 Odds1 Odds Ratio S S S S S S O S S S S S S S 2 . For additional help with categorical data analysis, feel free to contact me at wuh@bgsu.edu and 372-3119. 24 .