Statistics for EES Factorial analysis of variance

Size: px

Start display at page:

Download "Statistics for EES Factorial analysis of variance"

Aldous Lee
6 years ago
Views:

1 Statistics for EES Factorial analysis of variance Dirk Metzler 1. July 2013

2 1 ANOVA and F-Test 2 Pairwise comparisons and multiple testing 3 Non-parametric: The Kruskal-Wallis Test

3 Contents ANOVA and F-Test 1 ANOVA and F-Test 2 Pairwise comparisons and multiple testing 3 Non-parametric: The Kruskal-Wallis Test

4 ANOVA and F-Test Are the group means significantly different? group 1 group 2 group observed value

5 ANOVA and F-Test Are the group means significantly different? group 1 group 2 group observed value group 1 group 2 group observed value Or does this look like random deviations?

6 ANOVA and F-Test Are the group means significantly different? group 1 group 2 group observed value group 1 group 2 group observed value Or does this look like random deviations? This depends on the ratio of the varibility of group means and the variability within the groups.

7 ANOVA and F-Test Are the group means significantly different? observed value observed value group 1 group 2 group 3 Large variability within groups group 1 group 2 group 3 Small variability with groups Or does this look like random deviations? This depends on the ratio of the varibility of group means and the variability within the groups.

8 ANOVA and F-Test Are the group means significantly different? observed value observed value group 1 group 2 group 3 Large variability within groups group 1 group 2 group 3 Small variability with groups Or does this look like random deviations? This depends on the ratio of the varibility of group means and the variability within the groups. The analysis of variance (ANOVA) quantifies this ratio and its significance.

9 ANOVA and F-Test Example Blood-clotting time in rats under 4 different treatments group observation

10 ANOVA and F-Test Example Blood-clotting time in rats under 4 different treatments group observation global mean x = 64, group means x 1 = 61, x 2 = 66, x 3 = 68, x 4 = 61.

11 ANOVA and F-Test Example Blood-clotting time in rats under 4 different treatments group observation global mean x = 64, group means x 1 = 61, x 2 = 66, x 3 = 68, x 4 = 61. Caution: In this simple example, the global mean is the mean of the group means. This is not always the case!

12 ANOVA and F-Test Example Blood-clotting times in rats under 4 different treatments gr. x i observations (62 61) 2 (60 61) 2 (63 61) 2 (59 61) (63 66) 2 (67 66) 2 (71 66) 2 (64 66) 2 (65 66) 2 (66 66) (68 68) 2 (66 68) 2 (71 68) 2 (67 68) 2 (68 68) 2 (68 68) (56 61) 2 (62 61) 2 (60 61) 2 (61 61) 2 (63 61) 2 (64 61) 2 (63 61) 2 (59 61) 2 global mean x = 64, group means x 1 = 61, x 2 = 66, x 3 = 68, x 4 = 61.

13 ANOVA and F-Test Example Blood-clotting times in rats under 4 different treatments gr. x i observations (62 61) 2 (60 61) 2 (63 61) 2 (59 61) (63 66) 2 (67 66) 2 (71 66) 2 (64 66) 2 (65 66) 2 (66 66) (68 68) 2 (66 68) 2 (71 68) 2 (67 68) 2 (68 68) 2 (68 68) (56 61) 2 (62 61) 2 (60 61) 2 (61 61) 2 (63 61) 2 (64 61) 2 (63 61) 2 (59 61) 2 global mean x = 64, group means x 1 = 61, x 2 = 66, x 3 = 68, x 4 = 61. The red Differences (unsquared) are the residuals: they are the residual variability which is not explained by the model.

14 ANOVA and F-Test Example Blood-clotting times in rats under 4 different treatments gr. x i observations (62 61) 2 (60 61) 2 (63 61) 2 (59 61) (63 66) 2 (67 66) 2 (71 66) 2 (64 66) 2 (65 66) 2 (66 66) (68 68) 2 (66 68) 2 (71 68) 2 (67 68) 2 (68 68) 2 (68 68) (56 61) 2 (62 61) 2 (60 61) 2 (61 61) 2 (63 61) 2 (64 61) 2 (63 61) 2 (59 61) 2 global mean x = 64, group means x 1 = 61, x 2 = 66, x 3 = 68, x 4 = 61. The red Differences (unsquared) are the residuals: they are the residual variability which is not explained by the model. Sums of squares within groups: ss within = 112,

15 ANOVA and F-Test Example Blood-clotting times in rats under 4 different treatments gr. x i observations (62 61) 2 (60 61) 2 (63 61) 2 (59 61) (63 66) 2 (67 66) 2 (71 66) 2 (64 66) 2 (65 66) 2 (66 66) (68 68) 2 (66 68) 2 (71 68) 2 (67 68) 2 (68 68) 2 (68 68) (56 61) 2 (62 61) 2 (60 61) 2 (61 61) 2 (63 61) 2 (64 61) 2 (63 61) 2 (59 61) 2 global mean x = 64, group means x 1 = 61, x 2 = 66, x 3 = 68, x 4 = 61. The red Differences (unsquared) are the residuals: they are the residual variability which is not explained by the model. Sums of squares within groups: ss within = 112, 20 degrees of freedom (df)

16 ANOVA and F-Test Example Blood-clotting times in rats under 4 different treatments gr. x i observations (62 61) 2 (60 61) 2 (63 61) 2 (59 61) (63 66) 2 (67 66) 2 (71 66) 2 (64 66) 2 (65 66) 2 (66 66) (68 68) 2 (66 68) 2 (71 68) 2 (67 68) 2 (68 68) 2 (68 68) (56 61) 2 (62 61) 2 (60 61) 2 (61 61) 2 (63 61) 2 (64 61) 2 (63 61) 2 (59 61) 2 global mean x = 64, group means x 1 = 61, x 2 = 66, x 3 = 68, x 4 = 61. The red Differences (unsquared) are the residuals: they are the residual variability which is not explained by the model. Sums of squares within groups: ss within = 112, 20 degrees of freedom (df) Sums of squares between groups: ss betw = 4 (61 64) (66 64) (68 64) (61 64) 2 = 228,

17 ANOVA and F-Test Example Blood-clotting times in rats under 4 different treatments gr. x i observations (62 61) 2 (60 61) 2 (63 61) 2 (59 61) (63 66) 2 (67 66) 2 (71 66) 2 (64 66) 2 (65 66) 2 (66 66) (68 68) 2 (66 68) 2 (71 68) 2 (67 68) 2 (68 68) 2 (68 68) (56 61) 2 (62 61) 2 (60 61) 2 (61 61) 2 (63 61) 2 (64 61) 2 (63 61) 2 (59 61) 2 global mean x = 64, group means x 1 = 61, x 2 = 66, x 3 = 68, x 4 = 61. The red Differences (unsquared) are the residuals: they are the residual variability which is not explained by the model. Sums of squares within groups: ss within = 112, 20 degrees of freedom (df) Sums of squares between groups: ss betw = 4 (61 64) (66 64) (68 64) (61 64) 2 = 228, 3 degrees of freedom (df)

18 Example ANOVA and F-Test Blood-clotting times in rats under 4 different treatments gr. x i observations (62 61) 2 (60 61) 2 (63 61) 2 (59 61) (63 66) 2 (67 66) 2 (71 66) 2 (64 66) 2 (65 66) 2 (66 66) (68 68) 2 (66 68) 2 (71 68) 2 (67 68) 2 (68 68) 2 (68 68) (56 61) 2 (62 61) 2 (60 61) 2 (61 61) 2 (63 61) 2 (64 61) 2 (63 61) 2 (59 61) 2 global mean x = 64, group means x 1 = 61, x 2 = 66, x 3 = 68, x 4 = 61. The red Differences (unsquared) are the residuals: they are the residual variability which is not explained by the model. Sums of squares within groups: ss within = 112, 20 degrees of freedom (df) Sums of squares between groups: ss betw = 4 (61 64) (66 64) (68 64) (61 64) 2 = 228, 3 degrees of freedom (df) F = ss betw/3 ss within /20 = = 13.57

19 ANOVA and F-Test Example: Blood-clotting times in rats under 4 different treatments. ANOVA table (ANalysis Of VAriance) sum of mean sum df squares of squares F value (ss) (ss/df) groups residuals

20 ANOVA and F-Test Example: Blood-clotting times in rats under 4 different treatments. ANOVA table (ANalysis Of VAriance) sum of mean sum df squares of squares F value (ss) (ss/df) groups residuals Under the Null-Hypothesis H 0 the group means are equal (and assuming independent, normally distributed observations) is F Fisher-distributed with 3 and 20 degrees of freedom, and p = Fisher 3,20 ([13.57, ))

21 ANOVA and F-Test Example: Blood-clotting times in rats under 4 different treatments. ANOVA table (ANalysis Of VAriance) sum of mean sum df squares of squares F value (ss) (ss/df) groups residuals Under the Null-Hypothesis H 0 the group means are equal (and assuming independent, normally distributed observations) is F Fisher-distributed with 3 and 20 degrees of freedom, and p = Fisher 3,20 ([13.57, )) Thus, we can reject H 0.

22 ANOVA and F-Test Sir Ronald Aylmer Fisher,

23 ANOVA and F-Test F-Test n = n 1 + n n I obersvations in I groups, X ij = j-th observation in i-th group, j = 1,..., n i.

24 ANOVA and F-Test F-Test n = n 1 + n n I obersvations in I groups, X ij = j-th observation in i-th group, j = 1,..., n i. Model assumption: X ij = µ i + ε ij, with independent, normally distributed ε ij, E[ε ij ] = 0, Var[ε ij ] = σ 2

25 ANOVA and F-Test F-Test n = n 1 + n n I obersvations in I groups, X ij = j-th observation in i-th group, j = 1,..., n i. Model assumption: X ij = µ i + ε ij, with independent, normally distributed ε ij, E[ε ij ] = 0, Var[ε ij ] = σ 2 (µ i is the true mean within group i.)

26 ANOVA and F-Test F-Test n = n 1 + n n I obersvations in I groups, X ij = j-th observation in i-th group, j = 1,..., n i. Model assumption: X ij = µ i + ε ij, with independent, normally distributed ε ij, E[ε ij ] = 0, Var[ε ij ] = σ 2 (µ i is the true mean within group i.) X = 1 n I i=1 ni j=1 X ij (empirical) global mean X i = 1 n i ni j=1 X ij (empirical) mean of group i

27 ANOVA and F-Test F-Test n = n 1 + n n I obersvations in I groups, X ij = j-th observation in i-th group, j = 1,..., n i. Model assumption: X ij = µ i + ε ij, with independent, normally distributed ε ij, E[ε ij ] = 0, Var[ε ij ] = σ 2 (µ i is the true mean within group i.) X = 1 n I i=1 ni j=1 X ij (empirical) global mean X i = 1 n i ni j=1 X ij (empirical) mean of group i SS within = I n i (X ij X i ) 2 sum of squares within the groups, n I degrees of freedom i=1 j=1 SS betw = I n i (X i X ) 2 i=1 sum of squares between the groups, I 1 degrees of freedom

28 ANOVA and F-Test F-Test n = n 1 + n n I obersvations in I groups, X ij = j-th observation in i-th group, j = 1,..., n i. Model assumption: X ij = µ i + ε ij, with independent, normally distributed ε ij, E[ε ij ] = 0, Var[ε ij ] = σ 2 (µ i is the true mean within group i.) X = 1 n I i=1 ni j=1 X ij (empirical) global mean X i = 1 n i ni j=1 X ij (empirical) mean of group i SS within = I n i (X ij X i ) 2 sum of squares within the groups, n I degrees of freedom i=1 j=1 SS betw = I n i (X i X ) 2 i=1 F = SS betw/(i 1) SS /(n I) sum of squares between the groups, I 1 degrees of freedom

29 ANOVA and F-Test F-Test X ij = j-th observation i-th group, j = 1,..., n i, Model assumption: X ij = µ i + ε ij. E[ε ij ] = 0, Var[ε ij ] = σ 2 SS within = I n i (X ij X i ) 2 sum of squares within groups, n I degrees of feedom i=1 j=1 SS betw = I n i (X i X ) 2 i=1 sum of squares between groups, I 1 degrees of freedom F = SS betw/(i 1) SS within /(n I) Under the hypothesis H 0 : µ 1 = = µ I ( all µ i are equal ) F is Fisher-distributed with I 1 and n I degrees of freedom (no matter what the true joint value of µ i is). F-Test: We reject H 0 on the level of significance α if F q α, whereas q α is the (1 α)-quantile of the Fisher-distribution with I 1 and n I degrees of freedom.

30 ANOVA and F-Test Computing the p-value with R How to choose q such that Pr(F q) = 0.95 for Fisher(6,63)-distributed F? > qf(0.95,df1=6,df2=63) [1]

31 ANOVA and F-Test Computing the p-value with R How to choose q such that Pr(F q) = 0.95 for Fisher(6,63)-distributed F? > qf(0.95,df1=6,df2=63) [1] computation of p-value: How probable is it that a Fisher(3,20)-distributed random variable takes a value 13.57? > pf(13.57, df1=3, df2=20, lower.tail=false) [1] e-05

32 ANOVA and F-Test Table of 95%-quantiles of the F distribution This table shows the (rounded) 95%-quantiles of the Fisher distribution with k 1 and k 2 degrees of freedom (k 1 : numerator, k 2 : denominator) k 2 k

33 ANOVA and F-Test Anova completely in R The text file clotting.txt contains a column time with the clotting times and a column treat with the treatmetns A,B,C,D. > rat<-read.table("clotting.txt",header=true) > rat.aov <- aov(time~treat,data=rat) > summary(rat.aov) Df Sum Sq Mean Sq F value Pr(>F) treat e-05 *** Residuals Signif. codes: 0 *** ** 0.01 *

34 ANOVA and F-Test 7 verschiedene Labors haben jeweils 10 Messungen des Chlorpheniraminmaleat Gehalts von Medikamentenproben vorgenommen: Gehalt an Chlorpheniraminmaleat [mg] Labor

35 ANOVA and F-Test 7 verschiedene Labors haben jeweils 10 Messungen des Chlorpheniraminmaleat Gehalts von Medikamentenproben vorgenommen: Mittelwerte +/ Standardfehler Gehalt an Chlorpheniraminmaleat [mg] Labor

36 ANOVA and F-Test Caution: The labs are labeled with numbers. To tell R that these numbers are just names and have no numeric meaning, we have to convert them into a variable of type factor : > chlor <- read.table("chlorpheniraminmaleat.txt") > str(chlor) data.frame : 70 obs. of 2 variables: $ content: num $ Lab : int > chlor$labor <- as.factor(chlor$labor) > str(chlor) data.frame : 70 obs. of 2 variables: $ content: num $ Lab : Factor w/ 7 levels "1","2","3","4",..:

37 ANOVA and F-Test Now we can perform the anova: > chlor.aov <- aov(content~lab,data=chlor) > summary(chlor.aov) Df Sum Sq Mean Sq F value Pr(>F) Lab e-05 *** Residuals Signif. codes: 0 *** ** 0.01 *

38 Contents Pairwise comparisons and multiple testing 1 ANOVA and F-Test 2 Pairwise comparisons and multiple testing 3 Non-parametric: The Kruskal-Wallis Test

39 Pairwise comparisons and multiple testing The anova shows that there is a significant difference between the labs.

40 Pairwise comparisons and multiple testing The anova shows that there is a significant difference between the labs. But which labs are significantly different?

41 Pairwise comparisons and multiple testing The anova shows that there is a significant difference between the labs. But which labs are significantly different? p-values from pairwise comparisons with t-tests: Lab2 Lab3 Lab4 Lab5 Lab6 Lab7 Lab Lab Lab Lab Lab Lab

42 Pairwise comparisons and multiple testing We see 21 pair-wise comparisons. On the 5% level, some of them are significant.

43 Pairwise comparisons and multiple testing We see 21 pair-wise comparisons. On the 5% level, some of them are significant. Problem of multiple testing: If the null hypothesis ( only random deviations ) is true, we falsly reject it with a probability of 5% in each test.

44 Pairwise comparisons and multiple testing We see 21 pair-wise comparisons. On the 5% level, some of them are significant. Problem of multiple testing: If the null hypothesis ( only random deviations ) is true, we falsly reject it with a probability of 5% in each test.if we then apply 20 or more test, there will be on average one or more test in which we falsely reject the null hypothesis.

45 Pairwise comparisons and multiple testing We see 21 pair-wise comparisons. On the 5% level, some of them are significant. Problem of multiple testing: If the null hypothesis ( only random deviations ) is true, we falsly reject it with a probability of 5% in each test.if we then apply 20 or more test, there will be on average one or more test in which we falsely reject the null hypothesis. Therefore, we should apply a correction to the p-values for multiple testing.

46 Pairwise comparisons and multiple testing We see 21 pair-wise comparisons. On the 5% level, some of them are significant. Problem of multiple testing: If the null hypothesis ( only random deviations ) is true, we falsly reject it with a probability of 5% in each test.if we then apply 20 or more test, there will be on average one or more test in which we falsely reject the null hypothesis. Therefore, we should apply a correction to the p-values for multiple testing. One possibility in the anova case: Tukey s Honest Significant Differences (HSD).

47 Pairwise comparisons and multiple testing > TukeyHSD(chlor.aov) Tukey multiple comparisons of means 95% family-wise confidence level Fit: aov(formula = content ~ lab, data = chlor) $Labor diff lwr upr p adj

48 Pairwise comparisons and multiple testing > TukeyHSD(chlor.aov) Tukey multiple comparisons of means 95% family-wise confidence level Fit: aov(formula = content ~ lab, data = chlor) $Labor diff lwr upr p adj Wie get confidence intervals [lwr,upr] for the differences between the labs and p-values for the null hypotheses that these differences are 0. Both are already corrected for multiple testig.

49 Pairwise comparisons and multiple testing Visualization of HSD confidence intervals with plot(tukeyhsd(chlor.aov)): 95% family wise confidence level Differences in mean levels of Labor

50 Pairwise comparisons and multiple testing Restriction: Tukeys HSD-Methode is only valid for balanced designs, i.e. the sample sizes must be equal across the groups. (And variancen must be equal, which is already required in the anova.)

51 Pairwise comparisons and multiple testing Restriction: Tukeys HSD-Methode is only valid for balanced designs, i.e. the sample sizes must be equal across the groups. (And variancen must be equal, which is already required in the anova.) This is the case for the lab comparisons but not for the rat clotting times dataset.

52 Pairwise comparisons and multiple testing What to use when Tukey s HSD is not applicable?

53 Pairwise comparisons and multiple testing What to use when Tukey s HSD is not applicable? Very general correction for multiple testing: Bonferroni Method: Multiply each p value with the number n of tests performed.

54 Pairwise comparisons and multiple testing Example: pair-wise comparisons (with t-test) for the clotting times, first without correction: B C D A B C

55 Pairwise comparisons and multiple testing Example: pair-wise comparisons (with t-test) for the clotting times, first without correction: B C D A B C Now with Bonferroni correction (multiply each value with 6): B C D A B C

56 Pairwise comparisons and multiple testing Example: pair-wise comparisons (with t-test) for the clotting times, first without correction: B C D A B C Now with Bonferroni correction (multiply each value with 6): B C D A B C Significant differences after Bonferroni correction: A/C, B/D, C/D.

57 Pairwise comparisons and multiple testing Example: pair-wise comparisons (with t-test) for the clotting times, first without correction: B C D A B C Now with Bonferroni correction (multiply each value with 6): B C D A B C Significant differences after Bonferroni correction: A/C, B/D, C/D. (The value of 6.0 is not a proper p-value and should be replaced by 1.0)

58 Pairwise comparisons and multiple testing The Bonferroni method is very conservative, i.e. to be on the safe side, it tends to loose some of the significances.

59 Pairwise comparisons and multiple testing The Bonferroni method is very conservative, i.e. to be on the safe side, it tends to loose some of the significances. An improvement is the Bonferroni-Holm Method: Let k be the number of test. Multiply the smallest p value with k, the second smallest with k 1, the third smallest with k 2 etc.

60 Pairwise comparisons and multiple testing The Bonferroni method is very conservative, i.e. to be on the safe side, it tends to loose some of the significances. An improvement is the Bonferroni-Holm Method: Let k be the number of test. Multiply the smallest p value with k, the second smallest with k 1, the third smallest with k 2 etc. The R command p.adjust corrects p values for multiple testing, using Bonferroni-Holm by default: > pv <- c( , , , , , ) > p.adjust(pv) [1] > p.adjust(pv, method="bonferroni") [1]

61 Pairwise comparisons and multiple testing There is a special R function for pair-wise t-tests that also uses Bonferroni-Holm correction by default: > pairwise.t.test(rat$time,rat$treat,pool.sd=false) Pairwise comparisons using t tests with non-pooled SD data: rat$time and rat$treat A B C B C D P value adjustment method: holm

62 Contents Non-parametric: The Kruskal-Wallis Test 1 ANOVA and F-Test 2 Pairwise comparisons and multiple testing 3 Non-parametric: The Kruskal-Wallis Test

63 Non-parametric: The Kruskal-Wallis Test The one-factor anova is based on the assumption that all values are independent and normally distributed. The group means µ 1, µ 2,..., µ m may be different. (It is the aim of the test to find this out.) The variances within the groups must be equal.

64 Non-parametric: The Kruskal-Wallis Test The one-factor anova is based on the assumption that all values are independent and normally distributed. The group means µ 1, µ 2,..., µ m may be different. (It is the aim of the test to find this out.) The variances within the groups must be equal. In formulae: If Y ij is the jth value in group i, it is assumed that Y ij = µ i + ε ij, whereas all ε ij are independently N (0, σ 2 ) distributed with the same σ 2 for all groups!

65 Non-parametric: The Kruskal-Wallis Test The one-factor anova is based on the assumption that all values are independent and normally distributed. The group means µ 1, µ 2,..., µ m may be different. (It is the aim of the test to find this out.) The variances within the groups must be equal. In formulae: If Y ij is the jth value in group i, it is assumed that Y ij = µ i + ε ij, whereas all ε ij are independently N (0, σ 2 ) distributed with the same σ 2 for all groups! The null hypothesis is µ 1 = µ 2 = = µ m.

66 Non-parametric: The Kruskal-Wallis Test Not all deviations from neutrality cause problems.

67 Non-parametric: The Kruskal-Wallis Test Not all deviations from neutrality cause problems. However, anovas are not robust against outliers or distributions that may somtimes produce exceptionally huge values.

68 Non-parametric: The Kruskal-Wallis Test Not all deviations from neutrality cause problems. However, anovas are not robust against outliers or distributions that may somtimes produce exceptionally huge values. In such cases we can apply thekruskal-wallis Test. Like the Wilcoxon-Test it uses the Ranks instead of the actual values. It is a non-parameteric test, no particular probability distribution is assumed.

69 Non-parametric: The Kruskal-Wallis Test Not all deviations from neutrality cause problems. However, anovas are not robust against outliers or distributions that may somtimes produce exceptionally huge values. In such cases we can apply thekruskal-wallis Test. Like the Wilcoxon-Test it uses the Ranks instead of the actual values. It is a non-parameteric test, no particular probability distribution is assumed. Null hypothesis of the Kruskal-Wallis test: all values of Y ij come from the same distribution, independent of their group.

70 Non-parametric: The Kruskal-Wallis Test Not all deviations from neutrality cause problems. However, anovas are not robust against outliers or distributions that may somtimes produce exceptionally huge values. In such cases we can apply thekruskal-wallis Test. Like the Wilcoxon-Test it uses the Ranks instead of the actual values. It is a non-parameteric test, no particular probability distribution is assumed. Null hypothesis of the Kruskal-Wallis test: all values of Y ij come from the same distribution, independent of their group. Like in the anova, we also have to assume for the Kruskal-Wallis test that the data are independent of each other.

71 Non-parametric: The Kruskal-Wallis Test Let R ij be the rank of Y ij within the total sample.

72 Non-parametric: The Kruskal-Wallis Test Let R ij be the rank of Y ij within the total sample. Let R i. = 1 J i J i R ij be the average rank in group i, with J i the sample size in group i. j=1

73 Non-parametric: The Kruskal-Wallis Test Let R ij be the rank of Y ij within the total sample. Let R i. = 1 J i J i R ij be the average rank in group i, with J i the sample size in group i. j=1 The mean rank in the total sample is R.. = 1 N J I i R ij = N + 1 2, i=1 j=1 with I being the number of groups and N being the total sample size.

74 Non-parametric: The Kruskal-Wallis Test Let R ij be the rank of Y ij within the total sample. Let R i. = 1 J i J i R ij be the average rank in group i, with J i the sample size in group i. j=1 The mean rank in the total sample is R.. = 1 N J I i R ij = N + 1 2, i=1 j=1 with I being the number of groups and N being the total sample size. Under the null hypothesis all ranks have expectation value R...

75 Non-parametric: The Kruskal-Wallis Test We measure the deviation from this expectation by S = I J i (R i. R.. ) 2. i=1

76 Non-parametric: The Kruskal-Wallis Test We measure the deviation from this expectation by S = I J i (R i. R.. ) 2. i=1 To get a p value from S we have to know the probability distribution of S under the null hypothesis. For small values of I and J I, the latter can be found in tables.

77 Non-parametric: The Kruskal-Wallis Test We measure the deviation from this expectation by S = I J i (R i. R.. ) 2. i=1 To get a p value from S we have to know the probability distribution of S under the null hypothesis. For small values of I and J I, the latter can be found in tables. For I 3 und J i 5, or I > 3 and J i 4 we can use that the following scaling K of S is approximately χ 2 distributed with I 1 degrees of freedom: ( I ) 12 K = N (N + 1) S = 12 N (N + 1) J i R 2 i. 3 (N +1) i=1

78 Non-parametric: The Kruskal-Wallis Test Kruskal-Wallis-Test with R > kruskal.test(time~treat,data=rat) Kruskal-Wallis rank sum test data: time by treat Kruskal-Wallis chi-squared = , df = 3, p-value = > kruskal.test(content~lab,data=chlor) Kruskal-Wallis rank sum test data: content by lab Kruskal-Wallis chi-squared = , df = 6, p-value = 4.67e-05

Math Section MW 1-2:30pm SR 117. Bekki George 206 PGH

Math Section MW 1-2:30pm SR 117. Bekki George 206 PGH Math 3339 Section 21155 MW 1-2:30pm SR 117 Bekki George bekki@math.uh.edu 206 PGH Office Hours: M 11-12:30pm & T,TH 10:00 11:00 am and by appointment More than Two Independent Samples: Single Factor Analysis