Supplement to Accompany “The Sensitivity of Three Methods to Non-normality and Unequal Variances in Interval Estimation of Effect Sizes” By L.-T. Chen and C.-Y. J. Peng January, 2014 Table of Contents Table A Population Characteristics for CR-3 Design with Coefficient (1, -0.5, -0.5) .................................. 2 Table B Population Characteristics for CR-4 Design with Coefficient (0.5, 0.5, -0.5, -0.5) ........................ 4 Example 1 ..................................................................................................................................................... 6 Example 2 ..................................................................................................................................................... 7 Additional Analyses: Eta-Squared Analyses of the Factors Manipulated .................................................... 8 1 Table A Population Characteristics for CR-3 Design with Coefficient (1, -0.5, -0.5) Population ES (πΏBonett = Variances (skewness, kurtosis) 1 (0, 0): 1 (0, 0): 1 (0, 0) 1 (0, 0): 1 (0, 0): 1 (2,6; chi-squared with df =2) 1 (0, 0): 1 (0, 0): 1 (1.75, 3.75) 1 (0, 0): 1 (0, 0): 1 (1, 1.5) 1 (0, 0): 1 (0, 0): 1 (0.25, ΜΆ 0.75) 1 (0, 0): 1 (0, 0): 1 (0, ΜΆ 1.2; uniform) 1 (2,6; chi-squared with df =2) : 1 (0, 0): 1 (0, 0) 1 (1.75, 3.75): 1 (0, 0): 1 (0, 0) 1 (1, 1.5): 1 (0, 0): 1 (0, 0) 1 (0.25, ΜΆ 0.75): 1 (0, 0): 1 (0, 0) 1 (0, ΜΆ 1.2; uniform): 1 (0, 0): 1 (0, 0) 1 (0, 0): 1 (0, 0): 2.25 (0, 0) 1 (0, 0): 1 (0, 0): 2.25 (2,6; chi-squared with df =2) 1 (0, 0): 1 (0, 0): 2.25 (1.75, 3.75) 1 (0, 0): 1 (0, 0): 2.25 (1, 1.5) 1 (0, 0): 1 (0, 0): 2.25 (0.25, ΜΆ 0.75) 1 (0, 0): 1 (0, 0): 2.25 (0, ΜΆ 1.2; uniform) 1 (2,6; chi-squared with df =2) : 1 (0, 0): 2.25 (0, 0) 1 (1.75, 3.75): 1 (0, 0): 2.25 (0, 0) 1 (1, 1.5): 1 (0, 0): 2.25 (0, 0) 1 (0.25, ΜΆ 0.75): 1 (0, 0): 2.25 (0, 0) 1 (0, ΜΆ 1.2; uniform): 1 (0, 0): 2.25 (0, 0) ∑ππ=1 πj μj ∑π σ2 √ π=1 π π = ψ ) ∑π σ2 √ π=1 π π 0 π1 = π2 = π3 =0 0.2 π1 = 1, π2 = π3 = 0.8 0.5 π1 = 1, π2 = π3 = 0.5 0.8 π1 = 1, π2 = π3 = 0.2 π1 = 0 π2 = 0 π3 = 0 π1 = 1 π2 = 0.762 π3 = 0.762 π1 = 1 π2 = 0.405 π3 = 0.405 π1 = 1 π2 = 0.048 π3 = 0.048 2 1 (0, 0): 1 (0, 0): 4 (0, 0) 1 (0, 0): 1 (0, 0): 4 (2,6; chi-squared with df =2) 1 (0, 0): 1 (0, 0): 4 (1.75, 3.75) 1 (0, 0): 1 (0, 0): 4 (1, 1.5) 1 (0, 0): 1 (0, 0): 4 (0.25, ΜΆ 0.75) 1 (0, 0): 1 (0, 0): 4 (0, ΜΆ 1.2; uniform) 1 (2,6; chi-squared with df =2) : 1 (0, 0): 4 (0, 0) 1 (1.75, 3.75): 1 (0, 0): 4 (0, 0) 1 (1, 1.5): 1 (0, 0): 4 (0, 0) 1 (0.25, ΜΆ 0.75): 1 (0, 0): 4 (0, 0) 1 (0, ΜΆ 1.2; uniform): 1 (0, 0): 4 (0, 0) 1 (0, 0): 1 (0, 0): 8 (0, 0) 1 (0, 0): 1 (0, 0): 8 (2,6; chi-squared with df =2) 1 (0, 0): 1 (0, 0): 8 (1.75, 3.75) 1 (0, 0): 1 (0, 0): 8 (1, 1.5) 1 (0, 0): 1 (0, 0): 8 (0.25, ΜΆ 0.75) 1 (0, 0): 1 (0, 0): 8 (0, ΜΆ 1.2; uniform) 1 (2,6; chi-squared with df =2) : 1 (0, 0): 8 (0, 0) 1 (1.75, 3.75): 1 (0, 0): 8 (0, 0) 1 (1, 1.5): 1 (0, 0): 8 (0, 0) 1 (0.25, ΜΆ 0.75): 1 (0, 0): 8 (0, 0) 1 (0, ΜΆ 1.2; uniform): 1 (0, 0): 8 (0, 0) π1 = 0 π2 = 0 π3 = 0 π1 = 1 π2 = 0.717 π3 = 0.717 π1 = 1 π2 = 0.293 π3 = 0.293 π1 = 1 π2 = −0.131 π3 = −0.131 π1 = 0 π2 = 0 π3 = 0 π1 = 1 π2 = 0.635 π3 = 0.635 π1 = 1 π2 = 0.087 π3 = 0.087 π1 = 1 π2 = −0.461 π3 = −0.461 3 Table B Population Characteristics for CR-4 Design with Coefficient (0.5, 0.5, -0.5, -0.5) Population ES (πΏBonett = Variances (skewness, kurtosis) 1 (0, 0): 1 (0, 0): 1 (0, 0): 1 (0, 0) 1 (0, 0): 1 (0, 0): 1 (0, 0): 1 (2,6; chi-squared with df =2) 1 (0, 0): 1 (0, 0): 1 (0, 0): 1 (1.75, 3.75) 1 (0, 0): 1 (0, 0): 1 (0, 0): 1 (1, 1.5) 1 (0, 0): 1 (0, 0): 1 (0, 0): 1 (0.25, ΜΆ 0.75) 1 (0, 0): 1 (0, 0): 1 (0, 0): 1 (0, ΜΆ 1.2; uniform) 1 (2,6; chi-squared with df =2): 1 (0, 0): 1 (0, 0): 1 (0, 0) 1 (1.75, 3.75): 1 (0, 0): 1 (0, 0): 1 (0, 0) 1 (1, 1.5): 1 (0, 0): 1 (0, 0): 1 (0, 0) 1 (0.25, ΜΆ 0.75): 1 (0, 0): 1 (0, 0): 1 (0, 0) 1 (0, ΜΆ 1.2; uniform): 1 (0, 0): 1 (0, 0): 1 (0, 0) 1 (0, 0): 1 (0, 0): 1 (0, 0): 2.25 (0, 0) 1 (0, 0): 1 (0, 0): 1 (0, 0): 2.25 (2,6; chi-squared with df =2) 1 (0, 0): 1 (0, 0): 1 (0, 0): 2.25 (1.75, 3.75) 1 (0, 0): 1 (0, 0): 1 (0, 0): 2.25 (1, 1.5) 1 (0, 0): 1 (0, 0): 1 (0, 0): 2.25 (0.25, ΜΆ 0.75) 1 (0, 0): 1 (0, 0): 1 (0, 0): 2.25 (0, ΜΆ 1.2; uniform) 1 (2,6; chi-squared with df =2): 1 (0, 0): 1 (0, 0): 2.25 (0, 0) 1 (1.75, 3.75) : 1 (0, 0): 1 (0, 0): 2.25 (0, 0) 1 (1, 1.5) : 1 (0, 0): 1 (0, 0): 2.25 (0, 0) 1 (0.25, ΜΆ 0.75) : 1 (0, 0): 1 (0, 0): 2.25 (0, 0) 1 (0, ΜΆ 1.2; uniform) : 1 (0, 0): 1 (0, 0): 2.25 (0, 0) 0 π1 = π2 = π3 = π4 = 0 π1 π2 π3 π4 =0 =0 =0 =0 4 ∑ππ=1 πj μj ∑π σ2 √ π=1 π π = ψ ) ∑π σ2 √ π=1 π π 0.2 π1 = π2 = 1, π3 = π4 = 0.8 0.5 π1 = π2 = 1, π3 = π 4 = 0.5 0.8 π1 = π2 = 1, π3 = π4 = 0.2 π1 = 1 π2 = 1 π3 = 0.771 π4 = 0.771 π1 = 1 π2 = 1 π3 = 0.427 π4 = 0.427 π1 = 1 π2 = 1 π3 = 0.083 π4 = 0.083 1 (0, 0): 1 (0, 0): 1 (0, 0): 4 (0, 0) 1 (0, 0): 1 (0, 0): 1 (0, 0): 4 (2,6; chi-squared with df =2) 1 (0, 0): 1 (0, 0): 1 (0, 0): 4 (1.75, 3.75) 1 (0, 0): 1 (0, 0): 1 (0, 0): 4 (1, 1.5) 1 (0, 0): 1 (0, 0): 1 (0, 0): 4 (0.25, ΜΆ 0.75) 1 (0, 0): 1 (0, 0): 1 (0, 0): 4 (0, ΜΆ 1.2; uniform) 1 (2,6; chi-squared with df =2): 1 (0, 0): 1 (0, 0): 4 (0, 0) 1 (1.75, 3.75) : 1 (0, 0): 1 (0, 0): 4 (0, 0) 1 (1, 1.5) : 1 (0, 0): 1 (0, 0): 4 (0, 0) 1 (0.25, ΜΆ 0.75) : 1 (0, 0): 1 (0, 0): 4 (0, 0) 1 (0, ΜΆ 1.2; uniform) : 1 (0, 0): 1 (0, 0): 4 (0, 0) 1 (0, 0): 1 (0, 0): 1 (0, 0): 8 (0, 0) 1 (0, 0): 1 (0, 0): 1 (0, 0): 8 (2,6; chi-squared with df =2) 1 (0, 0): 1 (0, 0): 1 (0, 0): 8 (1.75, 3.75) 1 (0, 0): 1 (0, 0): 1 (0, 0): 8 (1, 1.5) 1 (0, 0): 1 (0, 0): 1 (0, 0): 8 (0.25, ΜΆ 0.75) 1 (0, 0): 1 (0, 0): 1 (0, 0): 8 (0, ΜΆ 1.2; uniform) 1 (2,6; chi-squared with df =2): 1 (0, 0): 1 (0, 0): 8 (0, 0) 1 (1.75, 3.75) : 1 (0, 0): 1 (0, 0): 8 (0, 0) 1 (1, 1.5) : 1 (0, 0): 1 (0, 0): 8 (0, 0) 1 (0.25, ΜΆ 0.75) : 1 (0, 0): 1 (0, 0): 8 (0, 0) 1 (0, ΜΆ 1.2; uniform) : 1 (0, 0): 1 (0, 0): 8 (0, 0) π1 π2 π3 π4 =0 =0 =0 =0 π1 = 1 π2 = 1 π3 = 0.735 π4 = 0.735 π1 = 1 π2 = 1 π3 = 0.339 π4 = 0.339 π1 = 1 π2 = 1 π3 = −0.058 π4 = −0.058 π1 π2 π3 π4 =0 =0 =0 =0 π1 = 1 π2 = 1 π3 = 0.668 π4 = 0.668 π1 = 1 π2 = 1 π3 = 0.171 π4 = 0.171 π1 = 1 π2 = 1 π3 = −0.327 π4 = −0.327 5 Example 1 Condition examined: k =3, the first group sampled from a uniform distribution with variance ratio 2.25 and ES = 0.2 The first group sampled from a uniform distribution with variance ratio 2.25 and ES = 0.2, the scores of the first group were generated from the CALL RANUNI function and the scores had .5 subtracted from them, followed by the square root of 1/12 divided from them, and then had 1 added to them. The scores of the second group were generated from the function RAND (‘NORMAL’, .762, 1). The scores from the third group were generated from the function RAND (‘NORMAL’, .762’, 1.5). 6 Example 2 Condition examined: k =4, the last group sampled from a chi-square distribution with df = 2 with variance ratio 8 and ES = 0.2 The scores of the first two groups were generated from the function RAND('NORMAL',1,1). The scores of the third group were generated from the function RAND('NORMAL',.668,1). The scores of the fourth group were generated from RAND (‘CHISQUARE’, 2) and the scores had 2 subtracted from them, followed by 2 divided from them, the square root of 8 multiplied by them, and then had .668 added to them. 7 Additional Analyses: Eta-Squared Analyses of the Factors Manipulated To examine the relative contributions of factors to the variance of coverage probability and of interval width, we computed eta-squared ( ο¨ 2 ) for six main factors, and their interactions. Results are shown in Tables 9 (for k = 3) and 10 (for k = 4). We decided to report eta-squared, instead of F-tests of effects or its p-value, because there is only one observation per cell, resulting in zero degrees of freedom for error term. If an eta-squared was ≥ 8%, its corresponding effect was considered substantial in explaining the variability in coverage probabilities or interval widths. The 8% cutoff was chosen because (1) it falls between a medium effect (.0588) and a large effect (.1379), according to Cohen (1969), and (2) it helped to identify patterns of substantial effects in this simulation study. The substantial eta-squared values are bolded in Tables 9 and 10. The eta squares were obtained from imposing a 3 ο΄ 6 ο΄ 4 ο΄ 4 ο΄ 3 factorial linear model on either the estimated coverage probability or the estimated interval width, as the dependent variable. The independent variables were the CI methods (= 3), population distributions (= 6), population variance ratios (= 4), population ESs (= 4), and sample size patterns (= 3). Because either the first group or the last was sampled from a non-normal distribution, two separate ANOVAs were performed to correspond to these two pairings. The pairing of the first group with a non-normal distribution will be referred to as First Non-normal in subsequent paragraphs. The pairing of the last group with a non-normal distribution will be referred to as Last Nonnormal in subsequent paragraphs. These two configurations will be compared to aid our understanding of the confounding of population distributions and variance ratios with weights (for k = 3), or without weights (for k = 4). When k = 3, the weights assigned to the three groups were (1, −0.5, −0.5). Hence, the confounding of non-normal distributions paired with the first or 8 the last group, and variance ratios was further confounded with the weight, because large variances were always assigned to the last group in the unequal variance conditions. When k = 4, the weights assigned to the four groups were (0.5, 0.5, −0.5, −0.5), all equal in absolute values. Hence, the confounding of non-normal distributions paired with the first or the last group, and variance ratios was not confounded with the weight, even when large variances were always assigned to the last group in the unequal variance conditions. Results are presented in the next four sections. Eta-squared analysis of coverage probabilities for k = 3 (Table 9) First Non-normal conditions. Using 8% as a criterion, we identified five effects that were substantial under the First Non-normal conditions for coverage probabilities: method, sample size pattern, method by sample size pattern, variance ratio by sample size pattern, and method by variance ratio by sample size pattern. Together they explained 77.61% of the total variance. Among these five effects, method by sample size pattern was the most influential effect, accounting for nearly one third of the total variance in coverage probabilities. Figure 1 further illustrates that Bonett’s and the BCa methods performed similarly to each other, yet the noncentral method performed poorly for the sample size patterns of (10 10 10) and (10 10 30). The other four effects each explained about 10% of the total variance. These effects were further elaborated on in Figures A to D. Last Non-normal conditions. Using the same criterion, we identified five effects that were substantial under the Last Non-normal conditions for coverage probabilities: distribution, ES, sample size pattern, method by sample size pattern, and variance ratio by sample size pattern. Together they explained 55.56% of the total variance. Among these five effects, method by sample size pattern was once again the most influential effect, accounting for 15.44% of the total 9 variance in coverage probabilities. Figure 2 further illustrates that Bonett’s and the BCa methods performed similarly to each other, yet the noncentral method performed poorly for the sample size patterns of (10 10 30) and (30 10 10). The other four effects each explained about 10% of the total variance. These effects were further elaborated on in Figures E to H. Eta-squared analysis of interval widths for k =3 (Table 9) Results for both the First Non-normal and the Last Non-normal conditions were similar. Using 8% as a criterion, we identified four effects that were substantial: sample size pattern, method by sample size pattern, variance ratio by sample size pattern, and variance ratio (primarily for the First Non-normal conditions). Together they explained approximately 80% of the total variability in interval widths. Among these four effects, the sample size pattern accounted for the largest eta-squared, about 54%. The narrowest to the widest intervals were given by the sample size patterns of (30 10 10), (10 10 30), and (10 10 10), respectively (see Figures 3 and 4). Because the sample size pattern (30 10 10) was always paired the largest sample size (30) with the largest weight (1), it therefore yielded the narrowest interval, regardless of the methods, distributions, variance ratios, population ESs, or the pairing of the first or the last group with a non-normal distribution. The interaction of variance ratio by sample size pattern (Figures I and J) modified the main effect of sample size pattern and the main effect of variance ratio under the First Non-Normal conditions only (Figure K). Likewise, the interaction of method by sample size pattern (Figures L and M) modified the main effect of sample size pattern. Eta-squared analysis of coverage probabilities for k = 4 (Table 10) As stated before, the weights assigned to the four groups were (0.5, 0.5, −0.5, −0.5) for k = 4. Weights were not confounded with distributions and variance ratios, even when large variances were always assigned to the last group in the unequal variance conditions. 10 First Non-normal conditions. Using 8% as a criterion, we identified five effects that were substantial under the First Non-normal conditions for coverage probabilities: ES, sample size pattern, method by ES, method by sample size pattern, and method by variance ratio by sample size pattern. Together they explained 78.34% of the total variance. Among these five effects, method by sample size pattern was the most influential effect, accounting for a little more than a quarter of the total variance in coverage probabilities. Figure 5 further illustrates that Bonett’s and the BCa methods performed similarly to each other, yet the noncentral method performed poorly for the sample size patterns of (10 10 10 30) and (30 10 10 10). In other words, the unequal sample sizes affected the performance of the noncentral method, but not Bonett’s method or the BCa method. The other four effects explained 8.22% to 17.43% of the total variance. These effects were further elaborated on in Figures N to Q. Last Non-normal conditions. Using the same criterion, we identified five effects that were substantial under the Last Non-normal conditions for coverage probabilities: distribution, variance ratio, ES, sample size pattern, and method by sample size pattern. Together they explained 55.09% of the total variance. Among these five effects, the distribution was the most influential effect, accounting for 13.13% of the total variance in coverage probabilities. Figure 6 further illustrates the impact of distributions on coverage probabilities. The coverage probabilities obtained from normal distributions were similar to those obtained from uniform and the slightly non-normal distribution with skewness = 0.25 and kurtosis = ο 0.75 , than the other three distributions. The other four effects each explained about 10% of the total variance. These effects were further elaborated on in Figures R to U. Eta-squared analysis of interval widths for k = 4 (Table 10) 11 Results for both the First Non-normal and the Last Non-normal conditions were similar. Using 8% as a criterion, we identified three effects that were substantial: sample size pattern, method by sample size pattern, and variance ratio by sample size pattern. Together they explained 87.86% of the total variability in interval widths under the First Non-normal conditions and 74.39% of the total variability under the Last Non-normal conditions. Among these three effects, the sample size pattern accounted for the largest eta-squared, more than 50%. The narrowest to the widest intervals were given by the sample size patterns of (10 10 10 30), (30 10 10 10), (10 10 10 10), respectively (Figures 7 and 8). Because the width of a CI is inversely related to SE which is determined from the total sample size. The two unequal sample size patterns therefore resulted in narrower intervals than the equal sample size pattern, regardless of the methods, distributions, variance ratios, population ESs, or the pairing of the first or the last group with a non-normal distribution. The interactions of method or variance ratio by sample size pattern modified the main effect of sample size pattern (Figures 9 to 12). Regardless of the pairing of the first or the last group with a non-normal distribution, the CIs obtained from the noncentral method were similar and narrower for the two unequal sample size patterns than those obtained under equal sample sizes (Figures 9 and 10). The narrowest to the widest CIs of Bonett’s and the BCa’s were obtained from (10 10 10 30), (30 10 10 10), and (10 10 10 10) patterns respectively. Furthermore, the intervals widened as the variance ratio increased for the sample size patterns of (10 10 10 10) and (30 10 10 10), whereas they narrowed as the variance ratio increased for the sample size pattern of (10 10 10 30) (see Figures 11 and 12). This phenomenon was observed under both the First Non-normal and the Last Non-normal conditions. 12 Table 9 Eta-squared Analysis of Coverage Probabilities and Interval Widths for k =3: Normal Distributions and the First or the Last Group Sampled from a Non-normal Distribution Eta-squared in Percentage Coverage Probability Interval Width First Last First Last NonNonNonNonnormal normal normal normal Simulation Main and Interaction Effects method 2.6292 4.5135 2.0721 12.8979 distribution 0.0836 0.5608 1.5079 11.7111 variance ratio 0.7823 4.1192 9.5077 7.0713 ES 3.0590 1.1842 3.7592 8.3606 sample size pattern 9.5146 11.5478 54.2837 53.5709 method*distribution 1.3540 1.3024 0.4664 1.6978 method*variance ratio 1.0899 0.5670 4.9144 4.2434 distribution*variance ratio 0.0869 4.0755 0.1018 0.2478 method*ES 5.0448 1.4207 0.3620 0.5277 distribution*ES 0.8258 5.2493 0.0919 0.4361 variance ratio*ES 1.6482 2.7950 0.1446 0.2471 method*sample size pattern 33.9598 15.4382 9.6285 8.9634 distribution*sample size pattern 0.4279 3.0740 0.1446 0.0535 variance ratio*sample size pattern 10.1565 8.5072 9.1724 9.4594 ES*sample size pattern 1.6214 2.7770 0.1373 0.2063 method*distribution*variance ratio 0.2206 0.3484 0.0937 0.3170 method*distribution*ES 0.0903 0.7804 0.1581 0.6418 method*variance ratio*ES 0.5550 0.2558 0.0659 0.1398 distribution*variance ratio*ES 0.3611 1.6495 0.0223 0.1107 method*distribution*sample size pattern 0.2073 0.0810 0.0963 0.0368 method*variance ratio*sample size pattern 4.0937 4.2321 11.0826 6.7060 distribution*variance ratio*sample size pattern 0.2407 1.3217 0.0267 0.0251 method*ES*sample size pattern 1.9424 0.9450 0.0242 0.0147 distribution*ES*sample size pattern 0.2808 1.2484 0.0139 0.0395 variance ratio*ES*sample size pattern 0.4781 0.7881 0.0231 0.0482 Note. Percentage of main or interaction effects greater than 8 is bolded. Since no four-way or five-way interaction is greater than 1%, they are not presented in this table. 13 Table 10 Eta-squared Analysis of Coverage Probabilities and Interval Widths for k =4: Normal Distributions and the First or the Last Group Sampled from a Non-normal Distribution Eta-squared in Percentage Coverage Probability Interval Width First Last First Last NonNonNonNonnormal normal normal normal Simulation Main and Interaction Effects method 0.6806 0.9492 1.9842 5.3726 distribution 0.1342 0.0459 1.1261 13.1343 variance ratio 3.5490 0.2139 0.1272 8.4433 ES 2.6004 5.4840 8.2166 10.3457 sample size pattern 17.4257 12.2309 64.3490 53.7211 method*distribution 0.0827 1.0433 0.0898 1.7593 method*variance ratio 1.6374 1.0380 0.4449 1.3670 distribution*variance ratio 0.1830 4.9492 0.0119 0.3171 method*ES 4.0626 0.3022 0.7896 11.2437 distribution*ES 0.1854 5.9127 0.0224 0.7394 variance ratio*ES 0.8641 2.4371 0.1207 0.3438 method*sample size pattern 27.9892 10.9392 11.2921 9.5761 distribution*sample size pattern 0.1811 1.7964 0.0321 0.0377 variance ratio*sample size pattern 7.8313 6.6847 12.2142 11.0974 ES*sample size pattern 1.1263 0.9454 0.1479 0.2337 method*distribution*variance ratio 0.0719 0.3339 0.0241 0.5290 method*distribution*ES 0.0669 0.8766 0.0452 1.1646 method*variance ratio*ES 0.2277 0.2207 0.0590 0.2555 distribution*variance ratio*ES 0.3738 1.9016 0.0225 0.1695 method*distribution*sample size pattern 0.0773 0.1111 0.0407 0.0785 method*variance ratio*sample size pattern 5.4825 5.0270 13.4683 7.1443 distribution*variance ratio*sample size pattern 0.2676 1.1382 0.0316 0.0168 method*ES*sample size pattern 1.4167 0.3488 0.0218 0.0338 distribution*ES*sample size pattern 0.1707 0.7510 0.0273 0.0393 variance ratio*ES*sample size pattern 0.4672 0.4575 0.0395 0.0560 Note. Percentage of main or interaction effects greater than 8 is bolded. Since no four-way or five-way interaction is greater than 1%, they are not presented in this table. 14 Figure 1. The interaction of method by sample size pattern on coverage probability under First Non-normal for k =3 15 Figure 2. The interaction of method by sample size pattern on coverage probability under Last Non-normal for k =3 16 Figure 3. The main effect of sample size pattern on interval width under First Non-normal for k =3 17 Figure 4. The main effect of sample size pattern on interval width under Last Non-normal for k =3 18 Figure 5. The interaction of method by sample size pattern on coverage probability under First Non-normal for k =4 19 Figure 6. The main effect of population distribution on coverage probability under Last Nonnormal for k =4 20 Figure 7. The main effect of sample size pattern on interval width under First Non-normal for k =4 21 Figure 8. The main effect of sample size pattern on interval width under Last Non-normal for k =4 22 Figure 9. The interaction of method by sample size pattern on interval width under First Nonnormal for k =4 The eta-squared for the interaction of method by sample size pattern under the First Non-normal conditions was 11.29% demonstrating the impact of sample size patterns on interval widths to be different across three methods. Figure 9 shows that the interval widths obtained from the noncentral method were similar under the sample size patterns of (10 10 10 30) and (30 10 10 10); the CIs obtained from these two sample size patterns were narrower than the CIs obtained from equal sample sizes. The interval widths obtained from Bonett’s method and the BCa method from the narrowest to the widest was in the order of (30 10 10 10), (10 10 10 30), and (10 10 10 10). 23 Figure 10. The interaction of method by sample size pattern on interval width under Last Nonnormal for k =4 The eta-squared for the interaction of method by sample size pattern under the Last Non-normal conditions was 9.58% demonstrating the impact of sample size patterns on interval widths to be different across three methods. Figure 10 shows that the interval widths obtained from the noncentral method were similar under the sample size patterns of (10 10 10 30) and (30 10 10 10); the CIs obtained from these two sample size patterns were narrower than the CIs obtained from the equal sample sizes. The CIs obtained from Bonett’s method and the BCa method from the narrowest to the widest was in the order of (30 10 10 10), (10 10 10 30), and (10 10 10 10). 24 Figure 11. The interaction of variance ratio by sample size pattern on interval width under First Non-normal for k =4 The eta-squared for the interaction of variance ratio by sample size pattern under the First Nonnormal conditions was 12.21% demonstrating the impact of variance ratios on interval widths to be different across three sample size patters. Figure 11 shows that the CIs obtained from the sample size patterns of (10 10 10 10) and (30 10 10 10) became wider as the variance ratio increased, whereas the CIs obtained from the sample size pattern of (10 10 10 30) became narrower as the variance ratio increased. 25 Figure 12. The interaction of variance ratio by sample size pattern on interval width under Last Non-normal for k =4 The eta-squared for the interaction of variance ratio by sample size pattern under the Last Nonnormal conditions was 11.10% demonstrating the impact of variance ratios on interval widths to be different across three sample size patters. Figure 12 shows that the CIs obtained from the sample size patterns of (10 10 10 10) and (30 10 10 10) became wider as the variance ratio increased; whereas the CIs obtained from the sample size pattern of (10 10 10 30) became narrower as the variance ratio increased. 26 Figure A. The method main effect on coverage probability under First Non-normal for k =3 The eta-squared for the main effect of method under the First Non-normal conditions was 12.90% demonstrating the impact of methods on coverage probabilities. Figure A shows that Bonett’s and the BCa methods performed similarly to each other and they performed better than the noncentral method. 27 Figure B. The main effect of sample size pattern on coverage probability under First Non-normal for k =3 The eta-squared for the main effect of sample size pattern under the First Non-normal conditions was 9.51% demonstrating the impact of sample size patterns on coverage probabilities. Figure B shows when the sample size pattern was (30 10 10), the mean of the coverage probabilities was closer to the nominal CI, than when the sample size patterns were (10 10 10) or (10 10 30). Yet, when the sample size pattern was (30 10 10), the range of the coverage probabilities was the largest. 28 Figure C. The interaction of variance ratio by sample size pattern on coverage probability under First Non-normal for k =3 The eta-squared for the interaction of variance ratio by sample size pattern under the First Nonnormal conditions was 10.16% demonstrating the impact of variance ratio on coverage probabilities to be different across three sample size patterns. Figure C shows when the sample size patterns were (10 10 10) and (10 10 30), the mean coverage probabilities were higher when the variance ratio was 4 or 8, than when the variance ratio was 1 or 2.25. When the sample size pattern was (30 10 10), the coverage probabilities decreased as the variance ratio increased. 29 Figure D. The interaction of method by variance ratio by sample size pattern on coverage probability under First Non-normal for k =3 30 The eta-squared for the interaction of method by variance ratio by sample size pattern under the First Non-normal conditions was 11.08% demonstrating the interaction of variance ratio by sample size pattern on coverage probability to be different across three methods. Figure D shows for the noncentral method and the sample size patterns of (10 10 10) and (10 10 30), the coverage probabilities increased as the variance ratio increased. When the sample size pattern was (30 10 10), the coverage probabilities decreased as the variance ratio increased for the noncental method. For Bonett’s and the BCa methods, their coverage probabilities were relatively consistent across different variance ratios and sample size patterns, compared to the corresponding coverage probabilities produced by the noncentral method. 31 Figure E. The main distribution effect on coverage probability under Last Non-normal for k =3 The eta-squared for the main effect of distributions under the Last Non-normal conditions was 11.71% demonstrating the impact of distributions on coverage probabilities. Figure E shows that the coverage probabilities obtained from normal distributions were similar to those obtained from uniform and slightly non-normal distribution (skewness = .25 and kurtosis = −.75) than the other three distributions. 32 Figure F. The main effect of ES on coverage probability under Last Non-normal for k =3 The eta-squared for the main effect of ES under the Last Non-normal conditions was 8.36% demonstrating the impact of ES on coverage probability. Figure F shows that the coverage probabilities obtained from ES = 0 were more similar to those obtained from ES = 0.2 or 0.5 than ES = 0.8. 33 Figure G. The main effect of sample size pattern on coverage probability under Last Non-normal for k =3 The eta-squared for the main effect of sample size pattern under the Last Non-normal conditions was 11.55% demonstrating the impact of sample size patterns on coverage probabilities. Figure G shows that the coverage probabilities obtained from the sample size pattern of (10 10 10) was more similar to the sample size pattern of (10 10 30) than to (30 10 10). 34 Figure H. The interaction of variance ratio by sample size pattern on coverage probability under Last Non-normal for k =3 The eta-squared for the interaction of variance ratio by sample size pattern under the Last Nonnormal conditions was 8.51% demonstrating the impact of variance ratios on coverage probabilities to be different across three sample size patterns. Figure H shows that when the sample size pattern was (10 10 30), the coverage probabilities increased as the variance ratio increased. When the sample size pattern was (30 10 10), the coverage probabilities decreased as the variance ratio increased. When the sample size pattern was (10 10 10), the coverage probabilities obtained from equal variances were more similar to those obtained from variance ratios of 2.25 or 4, than from the variance ratio of 8. 35 Figure I. The interaction of variance ratio by sample size pattern on interval width under First Non-normal for k =3 The eta-squared for the interaction of variance ratio by sample size pattern under the First Nonnormal conditions was 9.17% demonstrating the impact of variance ratios on interval widths to be different across three sample size patterns. Figure I shows that the interval widths became narrower as the variance ratio increased under sample size patterns of (10 10 10) and (10 10 30). The interval widths did not change much across different variance ratios under the sample size pattern of (30 10 10). 36 Figure J. The interaction of variance ratio by sample size pattern on interval width under Last Non-normal for k =3 The eta-squared for the interaction of variance ratio by sample size pattern under the Last Nonnormal conditions was 9.46% demonstrating the impact of variance ratios on interval widths to be different across three sample size patterns. Figure J shows that the interval widths became narrower as the variance ratio increased under sample size patterns of (10 10 10) and (10 10 30). The interval widths did not change much across different variance ratios under the sample size pattern of (30 10 10). 37 Figure K. The main effect of variance ratio on interval width under First Non-normal for k =3 The eta-squared for the main effect of variance ratio under the First Non-normal conditions was 9.51% demonstrating the impact of variance ratios on interval widths. Figure K shows that the interval widths became narrower as the variance ratio increased. 38 Figure L. The interaction of method by sample size pattern on interval width under First Nonnormal for k =3 The eta-squared for the interaction of method by sample size pattern under the First Non-normal conditions was 9.63% demonstrating the impact of sample size patterns on interval widths to be different across three methods. Figure L shows that the interval widths obtained from the noncentral method from the narrowest to the widest was in the order of (30 10 10), (10 10 30), and (10 10 10). The interval widths obtained from Bonett’s method and the BCa method were similar under the sample size patterns of (10 10 30) and (30 10 10); the interval widths obtained from these two sample size patterns were narrower than the interval widths obtained from the equal sample size pattern. 39 Figure M. The interaction of method by sample size pattern on interval width under Last Nonnormal for k =3 The eta-squared for the interaction of method by sample size pattern under the Last Non-normal conditions was 8.96% demonstrating the impact of sample size patterns on interval widths to be different across three methods. Figure M shows that the interval widths obtained from the noncentral method from the narrowest to the widest was in the order of (30 10 10), (10 10 30), and (10 10 10). The interval widths obtained from Bonett’s method and the BCa method were similar under the sample size patterns of (10 10 30) and (30 10 10) and the interval widths from these two sample size patterns were narrower than those obtained from the equal sample size pattern. 40 Figure N. The main effect of ES on coverage probability under First Non-normal for k =4 The eta-squared for the main effect of ES under the First Non-normal conditions was 8.22% demonstrating the impact of ES on coverage probabilities. Figure N shows that the coverage probabilities were closer to the nominal CI when ES = 0 or 0.8 than when ES = 0.2 or 0.5. 41 Figure O. The main effect of sample size pattern on coverage probability under First Non-normal for k =4 The eta-squared for the main effect of sample size pattern under the First Non-normal conditions was 17.43% demonstrating the impact of sample size patterns on coverage probabilities. Figure O shows that when the sample size pattern was (30 10 10 10), the mean of the coverage probabilities was closer to the nominal CI, than when the sample size patterns were (10 10 10 10) or (10 10 10 30). Yet, when the sample size pattern was (30 10 10 10), the range of the coverage probabilities was the largest. 42 Figure P. The interaction of method by ES on coverage probability under First Non-normal for k =4 The eta-squared for the interaction of method by ES under the First Non-normal conditions was 11.24% demonstrating the impact of ES on coverage probabilities to be different across three methods. Figure P shows that coverage probabilities produced by the noncentral method were closer to the nominal CI when ES = 0 or 0.8 than when ES = 0.2 or 0.5. The coverage probabilities produced by Bonett’s method and the BCa method were consistently close to the nominal CI regardless of ESs. 43 Figure Q. The interaction of method by variance ratio by sample size pattern on coverage probability under First Non-normal for k =4 44 The eta-squared for the interaction of method by variance ratio by sample size pattern under the First Non-normal conditions was 13.47% demonstrating the interaction of variance ratio by sample size pattern on coverage probability to be different across three methods. Figure Q shows that for the noncentral method and the sample size pattern of (10 10 10 30), the coverage probabilities increased as the variance ratio increased. When the sample size patterns were (10, 10, 10, 10) and (30, 10, 10, 10), the coverage probabilities decreased as the variance ratio increased for the noncentral method. For Bonett’s and the BCa method, their coverage probabilities were relatively consistent across different variance ratios and sample size patterns, compared to the corresponding coverage probabilities produced by the noncentral method. 45 Figure R. The interaction of method by sample size pattern on coverage probability under Last Non-normal for k =4 The eta-squared for the interaction of method by sample size pattern under the Last Non-normal conditions was 10.94% demonstrating the impact of sample size patterns on coverage probabilities to be different across three methods. Figure R further illustrates that Bonett’s and the BCa methods performed similarly to each other, yet the noncentral method performed poorly for the sample size patterns of (10 10 10 30) and (30 10 10 10). 46 Figure S. The main effect of variance ratio on coverage probability under Last Non-normal for k =4 The eta-squared for the main effect of variance ratio under the Last Non-normal conditions was 8.44% demonstrating the impact of variance ratio on coverage probability. Figure S shows that the coverage probabilities obtained from ES = 0 were more similar to those obtained from ES = 0.2 or 0.5 than ES = 0.8. 47 Figure T. The main effect of ES on coverage probability under Last Non-normal for k =4 The eta-squared for the main effect of ES under the Last Non-normal conditions was 10.35% demonstrating the impact of ES on coverage probabilities. Figure T shows that the coverage probabilities obtained from ES =0 were similar to those obtained from ES = 0.2 or 0.5 than ES = 0.8. 48 Figure U. The main effect of sample size pattern on coverage probability under Last Non-normal for k = 4 The eta-squared for the main effect of sample size pattern under the Last Non-normal conditions was 12.23% demonstrating the impact of sample size pattern on coverage probabilities. Figure U shows that the mean coverage probabilities from the closest to the farthest from the nominal CI were in the order of sample size patterns of (10 10 10 10), (10 10 10 30), and (30 10 10 10). 49