(AGE*SEX*YEAR)导致x2值为8615,概率为0.1964(不小于默认判据0.05),故该效 应项被剔除 第二步,剔除2阶交互效应项,概率均小于0.05,故2阶交互效应项不能剔除。即本例 用2阶交互效应项(同时含1阶主效应项)描述模型已为最佳。 Backward Elimination(p=.050) for DESIGN I with generating class AGE*SEX* YEAR Likelihood ratio chi square 00000DF=0P=1.000 If Deleted Simple Effect is DF L.R. Chisq Change Prob Iter AGE*SEXYEAR 6 8.615 19644 The best model has generating class AGE*SEX AGE*YEAR SEXYEAR Likelihood ratio chi square= 8.61546 DF=6 P=. 196 If Deleted Simple Effect is DF L.R. Chisq Change Prob Iter AGE*SEX 310.816 AGE*YEAR 62.829 0000 SEXYEAR 00462 Step 2 The best model has generating class AGE*SEX AGE*YEAR SEXYEAR Likelihood ratio chi square= 8.61546 DF=6 P=. 196 The final model has generating class AGE*SEX AGE*YEAR SEX YEAR The Iterative Proportional Fit algorithm converged at iteration 0 The maximum difference between observed and fitted marginal totals is 131 and the convergence criterion is 由于剔除了3阶交互效应项,故原全饱和模型变为层次模型,因而期望例数改变,期望 例数与实际例数不同,进而残差、标准化残差均不为0。若标准化残差界于-1.96-1.96范围 内,则表示模型是恰当的。从下面的结果可知,本例的标准化残差均在1.96-196范围内,（AGE*SEX*YEAR）导致χ2 值为 8.615，概率为 0.1964（不小于默认判据 0.05），故该效 应项被剔除。 第二步，剔除 2 阶交互效应项，概率均小于 0.05，故 2 阶交互效应项不能剔除。即本例 用 2 阶交互效应项（同时含 1 阶主效应项）描述模型已为最佳。 Backward Elimination (p = .050) for DESIGN 1 with generating class AGE*SEX*YEAR Likelihood ratio chi square = .00000 DF = 0 P = 1.000 If Deleted Simple Effect is DF L.R. Chisq Change Prob Iter AGE*SEX*YEAR 6 8.615 .1964 4 Step 1 The best model has generating class AGE*SEX AGE*YEAR SEX*YEAR Likelihood ratio chi square = 8.61546 DF = 6 P = .196 If Deleted Simple Effect is DF L.R. Chisq Change Prob Iter AGE*SEX 2 310.816 .0000 2 AGE*YEAR 6 62.829 .0000 2 SEX*YEAR 3 13.024 .0046 2 Step 2 The best model has generating class AGE*SEX AGE*YEAR SEX*YEAR Likelihood ratio chi square = 8.61546 DF = 6 P = .196 The final model has generating class AGE*SEX AGE*YEAR SEX*YEAR The Iterative Proportional Fit algorithm converged at iteration 0. The maximum difference between observed and fitted marginal totals is .131 and the convergence criterion is .278 由于剔除了 3 阶交互效应项，故原全饱和模型变为层次模型，因而期望例数改变，期望 例数与实际例数不同，进而残差、标准化残差均不为 0。若标准化残差界于-1.96—1.96 范围 内，则表示模型是恰当的。从下面的结果可知，本例的标准化残差均在-1.96—1.96 范围内