Multivariate approach to Mixed ANOVA with two Within-Subjects and one Between-Subjects Factors: Howell p. 486, 8th ed. The design is Groups x Cycle x Phase, 3 x 2 x 4, and the dependent variable is a measure of the frequency of bar pressing (for food) for a rat in an experimental box. During the Phase 1 of the experiment, each rat was placed in Box A and, while the animal was bar pressing, a tone (or, in Group L-A-B a light) was presented and paired with shock. During the second phase of the experiment, animals in Group L-A-B and Group A-B were tested in a different box, Box B where the tone stimulus was presented without the shock. It was expected that the tone would initially suppress bar pressing in these animals (but only somewhat in the group for which shock had been paired with a light rather than with a tone). Group A-A rats were tested in Box A in both phases. The training just described was repeated for three more cycles. It was expected that across cycles the rats in Groups A-B and L-A-B would learn that Box B is safe and bar pressing rates would increase, but rats in Group A-A would continue to have low rates of bar pressing. proc format; value gr 1='A-B' 2='A-A' 3='L-A-B'; data sup; infile 'C:\Users\Vati\Documents\StatData\MAN_2W1B.dat'; INPUT group c1p1 c1p2 c2p1 c2p2 c3p1 c3p2 c4p1 c4p2; format group gr.; ****************************************************************************; proc anova; class group; model c1p1 -- c4p2 = group / nouni; repeated cycle 4, phase 2 / printe; The first column of the data file contains the level of the Group variable (between-subjects) followed by subjects’ scores on Cycle-1/Phase-1, Cycle-1/Phase-2, Cycle-2/Phase 1, … Cycle-4/Phase-2. We have 4 x 2 = 8 cells in the matrix of repeated factors, represented by 8 dependent variables, C1P1 through C4P2, in the INPUT and MODEL statements. The REPEATED statement indicates that we have two within-subjects factors, CYCLE with 4 levels, PHASE with 2 levels. The comma separating one within-subject factor from another must be there. The order of the dependent variables is very important: The further to the right a repeated factor is in the REPEATED statement, the more rapidly its index values must change. Phase, the right-more factor, changes more rapidly (1,2,1,2,1,2,1,2) than does Cycle, the left-more factor (1,1,2,2,3,3,4,4). Check the “REPEATED MEASURES LEVEL INFORMATION” output page to assure that the within-subjects factors are properly defined. The ANOVA Procedure Class Level Information Class Levels Values group 3 A-A A-B L-A-B Number of Observations Read 24 Number of Observations Used 24 The ANOVA Procedure Repeated Measures Analysis of Variance Repeated Measures Level Information Dependent Variable c1p1 c1p2 c2p1 c2p2 c3p1 c3p2 c4p1 c4p2 Level of cycle 1 1 2 2 3 3 4 4 Level of phase 1 2 1 2 1 2 1 2 Sphericity Tests Variables DF Mauchly's Criterion Chi-Square Pr > ChiSq Orthogonal Components 5 0.6880092 7.3751845 0.1942 There is no problem with the sphericity assumption for the Cycle Effect. Sphericity is not an issue for the main effect of Phase, since there are only two phases. MANOVA Test Criteria and Exact F Statistics for the Hypothesis of no cycle Effect H = Anova SSCP Matrix for cycle E = Error SSCP Matrix S=1 M=0.5 N=8.5 Statistic Wilks' Lambda Value F Value Num DF Den DF Pr > F 0.33701445 12.46 3 19 <.0001 MANOVA Test Criteria and F Approximations for the Hypothesis of no cycle*group Effect H = Anova SSCP Matrix for cycle*group E = Error SSCP Matrix S=2 M=0 N=8.5 Statistic Value F Value Num DF Den DF Pr > F Wilks' Lambda 0.56393662 2.10 6 38 0.0759 Pillai's Trace 0.47621936 2.08 6 40 0.0767 Hotelling-Lawley Trace 0.70204234 2.17 6 23.636 0.0822 Roy's Greatest Root 0.57907665 3.86 3 20 0.0250 NOTE: F Statistic for Roy's Greatest Root is an upper bound. NOTE: F Statistic for Wilks' Lambda is exact. MANOVA Test Criteria and Exact F Statistics for the Hypothesis of no phase Effect H = Anova SSCP Matrix for phase E = Error SSCP Matrix S=1 M=-0.5 N=9.5 Statistic Wilks' Lambda Value F Value Num DF Den DF Pr > F 0.13920606 129.86 1 21 <.0001 MANOVA Test Criteria and Exact F Statistics for the Hypothesis of no phase*group Effect H = Anova SSCP Matrix for phase*group E = Error SSCP Matrix S=1 M=0 N=9.5 Statistic Wilks' Lambda Value F Value Num DF Den DF Pr > F 0.31824635 22.49 2 21 <.0001 Sphericity Tests Variables DF Mauchly's Criterion Chi-Square Pr > ChiSq Orthogonal Components 5 0.4846221 14.286499 0.0139 Sphericity is an issue for the Cycle x Phase effect (and any effects within which it is contained). While there are only two levels of Phase, there are 8 Cycle x Phase cells. MANOVA Test Criteria and Exact F Statistics for the Hypothesis of no cycle*phase Effect H = Anova SSCP Matrix for cycle*phase E = Error SSCP Matrix S=1 M=0.5 N=8.5 Statistic Wilks' Lambda Value F Value Num DF Den DF Pr > F 0.58602897 4.47 3 19 0.0154 MANOVA Test Criteria and F Approximations for the Hypothesis of no cycle*phase*group Effect H = Anova SSCP Matrix for cycle*phase*group E = Error SSCP Matrix S=2 M=0 N=8.5 Statistic Value F Value Num DF Den DF Pr > F Wilks' Lambda 0.42511798 3.38 6 38 0.0091 Pillai's Trace 0.67842819 3.42 6 40 0.0080 Hotelling-Lawley Trace 1.10871778 3.43 6 23.636 0.0139 Roy's Greatest Root 0.80683333 5.38 3 20 0.0070 NOTE: F Statistic for Roy's Greatest Root is an upper bound. NOTE: F Statistic for Wilks' Lambda is exact. The ANOVA Procedure Repeated Measures Analysis of Variance Tests of Hypotheses for Between Subjects Effects Source DF group Error 2 Anova SS Mean Square F Value Pr > F 4616.76042 2308.38021 21 15723.35938 748.73140 3.08 0.0670 The ANOVA Procedure Repeated Measures Analysis of Variance Univariate Tests of Hypotheses for Within Subject Effects Source DF Anova SS Mean Square F Value Pr > F Adj Pr > F G - G H-F-L cycle 3 2726.973958 908.991319 12.03 <.0001 <.0001 <.0001 cycle*group 6 1047.072917 174.512153 2.31 0.0445 0.0612 0.0522 Error(cycle) 63 4761.328125 75.576637 Greenhouse-Geisser Epsilon 0.7872 Huynh-Feldt-Lecoutre Epsilon 0.8934 Source DF Anova SS Mean Square F Value Pr > F phase 1 11703.13021 11703.13021 phase*group 2 4054.38542 2027.19271 Error(phase) 21 1892.60937 90.12426 Source DF 129.86 <.0001 22.49 <.0001 Anova SS Mean Square F Value Pr > F Adj Pr > F G - G H-F-L cycle*phase 3 741.515625 247.171875 4.04 0.0109 0.0222 0.0181 cycle*phase*group 6 1273.781250 212.296875 3.47 0.0051 0.0133 0.0100 Error(cycle*phase) 63 3859.078125 61.255208 Greenhouse-Geisser Epsilon 0.7124 Huynh-Feldt-Lecoutre Epsilon 0.7955 Excepting the main effect of group, every effect is statistically significant, even the triple interaction. Howell chose to evaluate the simple Cycle x Group interaction at each level of Phase. Look at the program to see how I did this by dropping variables from the left side of the model statement. Simple effects at Phase 1 -- Shock proc anova; class group; model c1p1 c2p1 c3p1 c4p1 = group / nouni; repeated cycle 4; The ANOVA Procedure Class Level Information Class Levels Values group 3 A-A A-B L-A-B Number of Observations Read 24 Number of Observations Used 24 The ANOVA Procedure Repeated Measures Analysis of Variance Repeated Measures Level Information Dependent Variable c1p1 c2p1 c3p1 c4p1 Level of cycle 1 2 3 4 MANOVA Test Criteria and Exact F Statistics for the Hypothesis of no cycle Effect H = Anova SSCP Matrix for cycle E = Error SSCP Matrix S=1 M=0.5 N=8.5 Statistic Wilks' Lambda Value F Value Num DF Den DF Pr > F 0.70304067 2.68 3 19 0.0764 MANOVA Test Criteria and F Approximations for the Hypothesis of no cycle*group Effect H = Anova SSCP Matrix for cycle*group E = Error SSCP Matrix S=2 M=0 N=8.5 Statistic Value F Value Num DF Den DF Pr > F Wilks' Lambda 0.74031448 1.03 6 38 0.4226 Pillai's Trace 0.26196650 1.00 6 40 0.4358 Hotelling-Lawley Trace 0.34769622 1.08 6 23.636 0.4045 Roy's Greatest Root 0.33859661 2.26 3 20 0.1130 NOTE: F Statistic for Roy's Greatest Root is an upper bound. NOTE: F Statistic for Wilks' Lambda is exact. The ANOVA Procedure Repeated Measures Analysis of Variance Tests of Hypotheses for Between Subjects Effects Source DF Anova SS Mean Square F Value Pr > F group 458.39583 229.19792 21 11695.59375 556.93304 Error 2 0.41 0.6679 Repeated Measures Analysis of Variance Univariate Tests of Hypotheses for Within Subject Effects Source DF Anova SS Mean Square F Value Pr > F Adj Pr > F G - G H-F-L cycle 3 403.614583 134.538194 1.74 0.1679 0.1801 0.1735 cycle*group 6 415.604167 69.267361 0.90 0.5036 0.4876 0.4967 Error(cycle) 63 4871.031250 77.317956 Greenhouse-Geisser Epsilon 0.7971 Huynh-Feldt-Lecoutre Epsilon 0.9065 The prediction was that all groups would show high suppression of bar-pressing on all cycles during the shock phase (1), but that during the non-shock phase (2) Group 2 (A-A) should show more suppression (lower scores) than the other groups, with the difference between Group 2 and the other groups increasing across cycles. Given this prediction, one expects no effects at all for the data from the first phase and a Cycle x Group interaction for the data from the second phase -- and that is exactly what we get. As shown above, the prediction holds for Phase 1 (shock). Simple effects at Phase 2 -- No Shock -- and simple, simple main effects of group at each level of Cycle for Phase 2 with pairwise comparisons among Groups. proc anova; class group; model c1p2 c2p2 c3p2 c4p2 = group; repeated cycle 4; means group / lsd lines; The ANOVA Procedure Class Level Information Class Levels Values group 3 A-A A-B L-A-B Number of Observations Read 24 Number of Observations Used 24 Dependent Variable: c1p2 Source DF Sum of Squares Mean Square F Value Pr > F Model 2 1125.750000 562.875000 Error 21 4081.875000 194.375000 Corrected Total 23 5207.625000 The groups do not differ significantly at Cycle 1, Phase 2. 2.90 0.0775 Dependent Variable: c2p2 Source DF Sum of Squares Mean Square F Value Pr > F Model 2 1476.333333 738.166667 Error 21 2466.625000 117.458333 Corrected Total 23 3942.958333 6.28 0.0073 t Tests (LSD) for c2p2 Alpha 0.05 Error Degrees of Freedom 21 Error Mean Square 117.4583 Critical Value of t 2.07961 Least Significant Difference 11.269 Means with the same letter are not significantly different. t Grouping Mean A N group 46.875 8 L-A-B A A 41.125 8 A-B B 28.125 8 A-A The groups do differ significantly at Cycle 1, Phase 2. As predicted, rate of bar pressing was lower in the A-A group than in the other two groups. The group B mean is 15.9 points lower than the mean of the other two groups (hand calculation). Dependent Variable: c3p2 Source DF Sum of Squares Mean Square F Value Pr > F Model 2 4105.583333 2052.791667 Error 21 1700.250000 Corrected Total 23 5805.833333 25.35 <.0001 80.964286 Means with the same letter are not significantly different. t Grouping Mean A N group 50.375 8 L-A-B A A 46.125 8 A-B B 20.750 8 A-A Same pattern of results in Cycle 2, but the difference between the Group B mean and the mean of the other two groups has increased to 27.5, due to the rats in Groups L-A-B and A-B having learned that Box B is safe (and thus bar-pressing more in Box B). Dependent Variable: c4p2 Source DF Sum of Squares Mean Square F Value Pr > F Model 2 3410.333333 1705.166667 Error 21 1421.000000 Corrected Total 23 4831.333333 67.666667 25.20 <.0001 Means with the same letter are not significantly different. t Grouping Mean A N group 51.750 8 A-B A A 46.500 8 L-A-B B 24.250 8 A-A Same pattern of results again, with the difference between the Group B mean and the mean of the other two groups by 24.9 points, about the same as in the previous cycle. Simple, simple main effect of Cycle for each Group during Phase 2 An alternative, perhaps better, way to dissect the significant Group x Cycle interaction for the data from the second phase is to evaluate the simple, simple main effects of Cycle at each level of Group (after sorting the data by group). I have included such an analysis, and it shows exactly what was predicted, a significant increase in bar pressing across cycles in Groups A-B and L-A-B (with the univariate test), but not in Group A-A. I did no pairwise comparisons here, as I decided they would not add anything of value. The ANOVA Procedure group=A-B Number of Observations Read 8 Number of Observations Used 8 Repeated Measures Analysis of Variance group=A-B Repeated Measures Level Information Dependent Variable c1p2 c2p2 c3p2 c4p2 Level of cycle 1 2 3 4 MANOVA Test Criteria and Exact F Statistics for the Hypothesis of no cycle Effect H = Anova SSCP Matrix for cycle E = Error SSCP Matrix S=1 M=0.5 N=1.5 Statistic Wilks' Lambda Value F Value Num DF Den DF Pr > F 0.00642851 257.60 3 5 <.0001 Repeated Measures Analysis of Variance Univariate Tests of Hypotheses for Within Subject Effects group=A-B Source DF Anova SS Mean Square F Value Pr > F Adj Pr > F G-G cycle 3 3789.125000 1263.041667 Error(cycle) 21 378.875000 H-F 70.01 <.0001 <.0001 <.0001 18.041667 Greenhouse-Geisser Epsilon 0.7361 Huynh-Feldt Epsilon 1.0897 MANOVA Test Criteria and Exact F Statistics for the Hypothesis of no cycle Effect H = Anova SSCP Matrix for cycle E = Error SSCP Matrix S=1 M=0.5 N=1.5 Statistic Wilks' Lambda Value F Value Num DF Den DF Pr > F 0.44485179 2.08 3 5 0.2216 Repeated Measures Analysis of Variance Univariate Tests of Hypotheses for Within Subject Effects group=A-A Source DF Anova SS Mean Square F Value Pr > F Adj Pr > F G-G cycle 3 291.250000 97.083333 Error(cycle) 21 2368.250000 112.773810 H-F 0.86 0.4767 0.4451 0.4729 Greenhouse-Geisser Epsilon 0.6759 Huynh-Feldt Epsilon 0.9534 Repeated Measures Analysis of Variance group=L-A-B Repeated Measures Level Information Dependent Variable c1p2 c2p2 c3p2 c4p2 Level of cycle 1 2 3 4 MANOVA Test Criteria and Exact F Statistics for the Hypothesis of no cycle Effect H = Anova SSCP Matrix for cycle E = Error SSCP Matrix S=1 M=0.5 N=1.5 Statistic Wilks' Lambda Value F Value Num DF Den DF Pr > F 0.27183574 4.46 3 5 0.0704 Univariate Tests of Hypotheses for Within Subject Effects group=L-A-B Source DF Anova SS Mean Square F Value Pr > F Adj Pr > F G-G cycle 3 889.750000 296.583333 Error(cycle) 21 1002.250000 47.726190 H-F 6.21 0.0034 0.0140 0.0064 Greenhouse-Geisser Epsilon 0.6179 Huynh-Feldt Epsilon 0.8310 The interaction plot on the next page illustrates the simple interaction well. Youse guys are gonna pay a price for having shocked my friends. One of these days I am going to find you in a dark alley. Sincerely, Mickey Rat Karl L. Wuensch, November, 2013. Fair Use of this Document