Chapter 3 Analysis of Variance (ANOVA; 變異 數 分 析 ) 許湘伶 Applied Linear Regression Models (Kutner, Nachtsheim, Neter, Li) Chapter 16 Design and Analysis of Experiments (Douglas C. Montgomery) hsuhl (NUK) DAE Chap. 3 1 / 46 Part I Supplement hsuhl (NUK) DAE Chap. 3 2 / 46 Relation between Regression and Analysis of Variance Regression model: yi = β0 + β1 X1i + · · · + βk Xki + εi , i = 1, . . . , n ANOVA Model or One-way model: yij = µi + ij = µ + τi + ij , hsuhl (NUK) DAE Chap. 3 i = 1, . . . , a j = 1, . . . , n 3 / 46 Relation between Regression and Analysis of Variance (cont.) Analysis of variance models differ from ordinary regression models in two key respects: 1 The explanatory or predictor variables in ANOVA models may be qualitative. 2 If the predictor variables are quantitative, no assumption is made in ANOVA models about the nature of the statistical relation between Xs and Y. hsuhl (NUK) DAE Chap. 3 4 / 46 Relation between Regression and Analysis of Variance (cont.) hsuhl (NUK) DAE Chap. 3 5 / 46 Relation between Regression and Analysis of Variance (cont.) When indicator variables are so used with regression models, the regression results will be identical to those obtained with ANOVA models. ANOVA models and regression models with indicator variables will lead to identical results. hsuhl (NUK) DAE Chap. 3 6 / 46 Relation between Regression and Analysis of Variance (cont.) Figure : Regression model Figure : Figure 16.4 Illustration of Partitioning of Total Deviations Yij − Ȳ·· hsuhl (NUK) DAE Chap. 3 7 / 46 Part II Chapter 3 The Analysis of Variance hsuhl (NUK) DAE Chap. 3 8 / 46 Outline 1 Example 2 The analysis of variance 3 Analysis of the fixed effects model 4 Model adequacy checking 5 Practical interpretation of results 6 Sample computer output 7 Determining sample size 8 Other example of single-factor experiments 9 The random effect model 10 The regression approach to the ANOVA 11 Nonparametric methods in the ANOVA hsuhl (NUK) DAE Chap. 3 9 / 46 Example methods for the design and analysis of single-factor experiments with a levels of the factor (or a treatments) Assume: completely randomized wafer(晶片) Relationship: RF power setting vs. the etch rate(蝕刻 速率) I I RF power: 4 levels: 160, 180, 200, 220 W 蝕刻速率: 測量物質從晶圓 表面被移除的的速率有多快 n = 5 replicates- 20 runs in random order hsuhl (NUK) DAE Chap. 3 10 / 46 Example (cont.) hsuhl (NUK) DAE Chap. 3 11 / 46 Example (cont.) no strong evidence to suggest that the variability in the etch rate around the average depends on the power setting Test: differences between the mean etch rates at a = 4 levels of RF power 1 2 t-test for all six possible pairs of means: inflates the type I error the analysis of variance hsuhl (NUK) DAE Chap. 3 12 / 46 ANOVA a treatments of a single factor yij : the jth observation taken under treatment i means model: yij = µi + ij hsuhl (NUK) DAE Chap. 3 i = 1, 2, . . . , a j = 1, 2, . . . , n 13 / 46 ANOVA (cont.) Model: yij = µi + ij i = 1, 2, . . . , a j = 1, 2, . . . , n = µ + τi + ij mean model effect model yij : the ij observation µi : the mean of the ith factor level µ: overall mean τi : the ith treatment effect ij : the random error component; sources of variability I I I I measurement variability from uncontrolled factors differences between the experimental unit noise in the process hsuhl (NUK) DAE Chap. 3 14 / 46 ANOVA (cont.) yij = µi + ij i = 1, 2, . . . , a j = 1, 2, . . . , n = µ + τi + ij mean model effect model linear statistical models one-way or single-factor analysis of variance model (單因子變異 數分析) the effect model is more widely encountered in the experimental design literature object: I I test hypotheses about the treatment means estimate model parameters: (µ, τi , σ 2 ) ij ∼ NID(0, σ 2 ) ⇒ yij ∼ N(µ + τi , σ 2 ) hsuhl (NUK) DAE Chap. 3 15 / 46 ANOVA (cont.) fixed effects model (固定效應模型): chosen by experimenter random effects model (隨機效應模型; components of variance model變異數成分模型): (Chap. 3.9; Chap. 13) a treatment could be a random sample from a larger population of treatments hsuhl (NUK) DAE Chap. 3 16 / 46 Notation ȳi· : the average of the observations under the ith treatment y·· : the grand total of all the observations ȳ·· : the grand average of all the observations yi· = y·· = n X ȳi· = yi· /n yij j=1 a X n X yij ȳ·· = y·· /N, i = 1, 2, . . . , a N = an i=1 j=1 hsuhl (NUK) DAE Chap. 3 17 / 46 Testing Testing the equality of the a treatment means E(yij ) = µ + τi = µi : Hypothesis: H 0 : µ1 = µ2 = · · · = µa H a : µi = 6 µj for at least one pair (i, j) H0 : τ1 = τ2 = · · · = τa = 0 Ha : τi = 6 0 for at least one i Pa a X i=1 µi ∵ =µ ⇔ τi = 0 a i=1 ⇔ The appropriate procedure for testing the equality of a treatment means is the analysis of variance. hsuhl (NUK) DAE Chap. 3 18 / 46 Testing (cont.) hsuhl (NUK) DAE Chap. 3 19 / 46 Decomposition of the Total Sum of Squares ANOVA: derived from a partitioning of total variability into its component parts 1 2 3 SST : the total corrected sum of squares SSTreatment : the sum of squares due to treatments (between treatment) SSE : the sum of squares due to error (within treatments) a X n X SST = (yij − ȳ·· )2 (N−1) i=1 j=1 a X 2 (ȳi· − ȳ·· ) + =n i=1 a X n X (yij − ȳi· )2 i=1 j=1 = SSTreatment + SSE (a−1) hsuhl (NUK) (N−a) DAE Chap. 3 20 / 46 Decomposition of the Total Sum of Squares (cont.) hsuhl (NUK) DAE Chap. 3 21 / 46 Decomposition of the Total Sum of Squares (cont.) Total variability: can be partitioned into 1 the total corrected sum of squares a X n a X n X X y2 2 SST = (yij − ȳ·· ) = y2ij − ·· N i=1 j=1 2 a sum of squares of the differences between the treatment average and the grand average SSTreatment = n a X i=1 3 i=1 j=1 a 1 X 2 y2·· (ȳi· − ȳ·· ) = yi· − n N 2 i=1 a sum of squares of the differences of observation within treatments from the treatment average SSE = a X n X (yij − ȳi· )2 = SST − SSTreatment i=1 j=1 hsuhl (NUK) DAE Chap. 3 22 / 46 Decomposition of the Total Sum of Squares (cont.) 1 2 a pooled estimate of the common variance within each of the a treatments SSE N−a an estimate of σ 2 if µi s are all equal SSTreatment a−1 3 ANOVA identity: provide two estimated of σ 2 hsuhl (NUK) DAE Chap. 3 23 / 46 Decomposition of the Total Sum of Squares (cont.) Error mean square (MSE; 誤差均方): 1 2 SSE N−a E(MSE) = σ 2 MSE = Treatment mean square (處理均方): 1 2 3 SSTreatment a−1 P n ai=1 τi2 E(MSTreatment ) = σ 2 + a−1 if there are no differences in treatment means (i.e. τi = 0), MSTreatment also estimate σ 2 MSTreatment = A test of hypothesis of no difference in treatment means can be performed by comparing METreatment and MSE hsuhl (NUK) DAE Chap. 3 24 / 46 Statistical Analysis Assumptions ij ∼ NID(0, σ 2 ) ⇒ yij ∼ NID(µ + τi , σ 2 ) Cochran’s Theorem SST : a sum of squares in normally distributed r.v. 1 2 3 4 SST /σ 2 ∼ χ2N−1 SSTreatment /σ 2 ∼ χ2a−1 if H0 : τi = 0 is true SSE /σ 2 ∼ χ2N−a SSTreatment /σ 2 and SSE /σ 2 are independent χ2 r.v. ⇒ test statistic: F0 = hsuhl (NUK) SSTreatment /(a − 1) MSTreatment H0 = ∼ Fa−1,N−a SSE /(N − a) MSE DAE Chap. 3 25 / 46 Statistical Analysis (cont.) Cochran’s Theorem Let Zi ∼ NID(0, 1), i = 1, . . . , ν, and ν X Zi2 = i=1 s X Qi , i=1 where s ≤ ν and Qi has νi d.f. (i=1,. . . ,s). Then Qi , ı = 1, . . . , s are independent χ2νi r.v., if and only if ν= s X νi i=1 hsuhl (NUK) DAE Chap. 3 26 / 46 Statistical Analysis (cont.) If H0 is false, MSTreatment > MSE ⇒ reject H0 if F0 is too large, i.e., F0 > Fα,a−1,N−a ANOVA table: hsuhl (NUK) DAE Chap. 3 27 / 46 Statistical Analysis (cont.) The Plasma Etching Experiment H0 : µ1 = µ2 = µ3 = µ4 vs. H1 : some means are different hsuhl (NUK) DAE Chap. 3 28 / 46 Statistical Analysis (cont.) ## ANOVA table etch$FRF <- as.factor(etch$RF) etch.aov <- aov(rate˜FRF,data=etch) summary(etch.aov) FRF Residuals Df 3 16 Sum Sq Mean Sq F value Pr(>F) 66870.55 22290.18 66.80 0.0000 5339.20 333.70 F0 > F(0.99, 3, 16) = 5.29 hsuhl (NUK) DAE Chap. 3 29 / 46 Estimation of the Model Parameters Model: yij = µ + τi + ij i = 1, . . . , a j = 1, . . . , n Parameter: µ, τi , σ 2 Estimates: I I I I overall mean: µ̂ = ȳ·· treatment effect: τ̂i = ȳi· − ȳ·· , i = 1, . . . , a µi : µ̂i = µ̂ + τ̂i = ȳi· σ 2 : σ̂ 2 = MSE hsuhl (NUK) DAE Chap. 3 30 / 46 Estimation of the Model Parameters (cont.) ij ∼ NID(0, σ 2 ) ⇒ ȳi· ∼ N(µi , σ 2 /n) 100(1 − α)% Confidence interval: r r MSE MSE ȳi· − tα/2,N−a ≤µi ≤ ȳi· + tα/2,N−a n n r r 2MSE 2MSE ȳi· − ȳj· − tα/2,N−a ≤ µi −µj ≤ ȳi· − ȳj· + tα/2,N−a n n hsuhl (NUK) DAE Chap. 3 31 / 46 Estimation of the Model Parameters (cont.) Ex 3.3 overall mean: µ̂ = 617.75 treatment effect: i 1 2 3 4 RF power 160 180 200 220 τ̂i -66.55 -30.35 7.65 89.25 95% confidence interval for µ4 : (one-at-a-time) 689.6815 ≤ µ4 ≤ 724.3185 Bonferroni method: correct level α/2r hsuhl (NUK) DAE Chap. 3 32 / 46 Unbalanced data ni observations under treatment i (i = 1, . . . , a) P N = ai=1 ni : total sample size SST = SSTreatment = ni a X X i=1 j=1 a X y2i· i=1 hsuhl (NUK) DAE Chap. 3 ni y2ij − − y2·· N y2·· N 33 / 46 Model Adequacy Checking ŷij : estimate of yij ŷij = µ̂ + τ̂i = ȳi· residual eij : investigating violations of the basic assumptions and model adequacy eij = yij − ŷij I I I I The checking should be automatic Model is adequate ⇒ eij s should be structureless graphical analysis how to deal with commonly occurring abnormalities standardized residual: dij = √ hsuhl (NUK) eij MSE DAE Chap. 3 34 / 46 Model Adequacy Checking (cont.) Residual plot ## Residual plot opar <- par(mfrow=c(2,2),cex=.8) plot(etch.aov) par(opar) hsuhl (NUK) DAE Chap. 3 35 / 46 Model Adequacy Checking (cont.) hsuhl (NUK) DAE Chap. 3 36 / 46 Model Adequacy Checking (cont.) eij vs. time: independence assumption eij vs. ŷij : nonconstant variance-variance-stabilizing transformation hsuhl (NUK) DAE Chap. 3 37 / 46 Statistical Tests for Equality of Variance Bartlett’s test: H0 :σ12 = σ22 = · · · = σa2 Ha : above not true for at least on σi2 Test statistic: χ20 = 2.3026 q H0 2 ∼ χa−1 c q = (N − a) log10 Sp2 − a X (ni − 1) log10 Si2 i=1 a X 1 c=1+ (ni − 1)−a − (N − a)−1 3(a − 1) i=1 Pa 2 (ni − 1)Si Sp2 = i=1 N−a hsuhl (NUK) DAE Chap. 3 ! 38 / 46 Statistical Tests for Equality of Variance (cont.) Reject H0 : χ20 > χα,a−1 very sensitive to the normality assumption > bartlett.test(rate˜RF,data=etch) Bartlett test of homogeneity of variances data: rate by RF Bartlett’s K-squared = 0.4335, df = 3, p-value = 0.9332 > qchisq(0.95,3) [1] 7.814728 hsuhl (NUK) DAE Chap. 3 39 / 46 Statistical Tests for Equality of Variance (cont.) Modified Levene test: robust to departures from normality considering the absolute deviation of yij from the treatment median ỹi· : dij = |yij − ỹi· | i = 1, 2, . . . , a j = 1, 2, . . . , n The test statistic for Levene’s test is simply the usual ANOVA F statistic for testing equality of means applied to the absolute deviations hsuhl (NUK) DAE Chap. 3 40 / 46 Statistical Tests for Equality of Variance (cont.) Peak Discharge Data hsuhl (NUK) DAE Chap. 3 41 / 46 Statistical Tests for Equality of Variance (cont.) hsuhl (NUK) DAE Chap. 3 42 / 46 Statistical Tests for Equality of Variance (cont.) > > > library(lawstat) peak.aov<-aov(Observ˜as.factor(Method),data=peak) summary(peak.aov) Df Sum Sq Mean Sq F value Pr(>F) as.factor(Method) 3 708.3 236.1 76.07 4.11e-11 *** Residuals 20 62.1 3.1 --Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1 > leveneTest(peak$Observ,as.factor(peak$Method)) Levene’s Test for Homogeneity of Variance (center = median) Df F value Pr(>F) group 3 4.5684 0.01357 * 20 --Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1 hsuhl (NUK) DAE Chap. 3 43 / 46 Statistical Tests for Equality of Variance (cont.) Transformation: y∗ij = hsuhl (NUK) √ yij DAE Chap. 3 44 / 46 Statistical Tests for Equality of Variance (cont.) Formal method: Box-Cox Method ## Box-Cox Method library(MASS) boxcox(Observ ˜ Method, data = peak,lambda = seq(-1, 1, length = 10)) hsuhl (NUK) DAE Chap. 3 45 / 46 Comparing Among Treatment Means ANOVA: reject H0 ⇒ differences between the treatment means which means differ is not specified multiple comparison methods ȳi· ∼ N(µi , σ 2 /n), σ̂ 2 = MSE ⇒µ1 6= µ2 6= µ3 6= µ4 hsuhl (NUK) DAE Chap. 3 46 / 46