2 for Random Effects, aka the Intraclass Correlation Coefficient Lori Foster Thompson provided me with data from 40 teams, each consisting of 4 persons. Twenty of the teams were in an experimental group referred to as “face to face” (FTF) and the other twenty were in the “computer-mediated” (CM) group. The dependent variable was the score on a five item scale designed to measure satisfaction with the team process. Lori compared the two groups using individual persons as the experimental unit -- that is, she had 4 x 40 = 160 cases, 80 in each of two groups. I replicated this analysis using a pooled variances t (equivalent to a one-way ANOVA). This analysis showed that satisfaction in the FTF group (M = 4.25, s = 0.70) was significantly greater than in the CM group (M = 3.33, SD = 0.83), t(158) = 7.54, p < .001. If this comparison was made with ANOVA rather than t, the F would be the square of the t -- F(1, 158) = 58.86, p < .001. The editor of the journal pointed out to Lori that participants were nested within teams and asked her to report an intraclass correlation to indicate the portion of the variance due to teams and to report both individual-level and group-level comparisons between FTF and CM teams. As noted by Howell (2010, page 439), some authors use 2 with both fixed and random effects, but others use the symbol 2 with fixed effects and the symbol ρ2 (the intraclass correlation coefficient) with random effects. Statistically, there are three classification variables (“factors” in the language of ANOVA) in Lori’s design: Using the variable names in her SPSS data file, they are CONDTN (FTF or CM), TnNo2 (team number), and Subjects. The Subjects variable is nested within the Team variable and the Team variable is nested within the Condtn variable -- each subject served on only one team and each team served in only one of the experimental conditions. Lori concluded, correctly I believe, that the editor wanted her to conduct a one-way random effects ANOVA using process satisfaction (PrSat) as the dependent variable and teams (TnNo2) as the classification variable, and then report 2 as an estimate of the proportion of the variance in process satisfaction that is explained by differences among the teams. Do note that in treating Teams as a random rather than a fixed factor we are asserting that our teams represent a random sample from a population of teams to which we wish to generalize our results. Put another way, we are not just interested in the 40 teams on which we have data but rather on the entire population of teams from which our sample of teams could be considered to be a random sample. Of course, we are also treating Subjects as a random factor, pretending that they really represent a random sample from the population of subjects that is of interest to us. The experimental factor (Condtn) is a fixed factor -- we did not randomly choose two experimental conditions from a population of conditions, we deliberately chose these two conditions, the only two conditions in which we are interested -- that is, on this factor we have the entire population of interest. One-Way Random Effects ANOVA, 2 = .425, 2 = .562 Copyright 2010, Karl L. Wuensch - All rights reserved. ICC-OmegaSq.doc Page 2 David Howell (2007, page 328) points out that “one version of the intraclass correlation coefficient is nothing but 2 for the random model.” On pages 414 through 415, he shows how to compute 2 for one-way through three-way designs, including the one-way random design, which I describe here: (MSA MSE ) / n , where MSA is the mean square among teams, MSE is the MSE (MSA MSE ) / n error (within teams) mean square, and n is the number of scores in each team. In the seventh edition of Howell’s Methods text only two-way ANOVA is included, but he provides a link to a table of expected mean squares for three-way designs. 2 Now I obtain the random effects ANOVA, using my preferred statistical program, SAS. Here is the procedural code I used, followed by the output: proc glm; class tm_no2; model prsat = tm_no2; random tm_no2 / test; run; -----------------------------------------------------------------------------------------------The SAS System 2 The GLM Procedure Dependent Variable: PRSAT Source DF Sum of Squares Mean Square F Value Pr > F Model 39 71.1840000 1.8252308 3.96 <.0001 Error 120 55.3600000 0.4613333 Corrected Total 159 126.5440000 R-Square Coeff Var Root MSE PRSAT Mean 0.562524 17.92125 0.679215 3.790000 Source DF Type III SS Mean Square F Value Pr > F TM_NO2 39 71.18400000 1.82523077 3.96 <.0001 -----------------------------------------------------------------------------------------------The SAS System 4 The GLM Procedure Tests of Hypotheses for Random Model Analysis of Variance Dependent Variable: PRSAT Source DF Type III SS Mean Square F Value Pr > F TM_NO2 39 71.184000 1.825231 3.96 <.0001 Error: MS(Error) 120 55.360000 0.461333 ------------------------------------------------------------------------------------------------ For a one-way ANOVA, the basic analysis for the random effect model is identical to that for the fixed effect model. The computation of the 2 does differ, Page 3 however, between the fixed effect model and the random effect model. Computation of the 2 for this analysis is: 2 (MSA MSE ) / n (1.825 0.461) / 4 0.341 0.425 . MSE (MSA MSE ) / n 0.461 (1.825 0.461) / 4 0.802 Omega-squared is considered to be superior to eta-squared for estimating the proportion of variance accounted for by an ANOVA factor. Eta-squared tends to overestimate that proportion -- but eta-squared is certainly easier to compute: 2 SSEffect 71.184 0.563 . Notice that this is the R2 reported by SAS. SSTotal 126.544 One-Way Fixed Effects ANOVA, 2 = .419, 2 = .563 For pedagogical purposes, I’ll show the computation of 2 treating teams as a fixed variable. The estimated treatment variance is (a 1)(MSA MSE ) (40 1)(1.825 0.461) 0.332. The term “a” is the number of (n)(a) (4)( 40) levels of the team variable. The estimated total variance is equal to the estimated treatment variance plus the estimated error variance, .332 MSE .332 .461 .793. .332 .419 . Accordingly, 2 .793 Two-Way Mixed Effects ANOVA One should keep in mind that the 2 for teams, as computed above, includes the effect of the experimental treatment, Condtn -- that is, we have estimated the variance among the teams, but part of that variance is due to the fact that the two experimental groups of teams were treated differently, and the other part of it is due to other differences among teams (error, effects of extraneous variables, reflected in differences among subjects’ scores within teams). If one wanted to determine the variance of teams after excluding variance due to the experimental condition, a mixed factorial ANOVA, Condtn x Teams (nested within Condtn), could be conducted. Here is the SAS code and output for such an analysis, treating Condtn as fixed and Teams as random: proc glm; class condtn tm_no2; model prsat = condtn tm_no2(condtn); random tm_no2(condtn) / test; run; -----------------------------------------------------------------------------------------------The SAS System 6 The GLM Procedure Dependent Variable: PRSAT Source DF Sum of Squares Mean Square F Value Pr > F Model 39 71.1840000 1.8252308 3.96 <.0001 Error 120 55.3600000 0.4613333 Page 4 Corrected Total 159 126.5440000 R-Square Coeff Var Root MSE PRSAT Mean 0.562524 17.92125 0.679215 3.790000 Source DF Type III SS Mean Square F Value Pr > F CONDTN TM_NO2(CONDTN) 1 38 33.48900000 37.69500000 33.48900000 0.99197368 72.59 2.15 <.0001 0.0009 -----------------------------------------------------------------------------------------------The SAS System 8 The GLM Procedure Tests of Hypotheses for Mixed Model Analysis of Variance Dependent Variable: PRSAT PRSAT Source DF Type III SS Mean Square F Value Pr > F CONDTN 1 33.489000 33.489000 33.76 <.0001 Error Error: MS(TM_NO2(CONDTN)) 38 37.695000 0.991974 Source DF Type III SS Mean Square F Value Pr > F TM_NO2(CONDTN) 38 37.695000 0.991974 2.15 0.0009 120 55.360000 0.461333 Error: MS(Error) Note that the SS for Teams from the previous analysis, 71.184, has been partitioned into a SS for Condtn, 33.489, and a SS for Teams within conditions, 37.695. The Total SS (126.544) is comprised of the SS for Condtn, 33.489, plus the SS for Teams within conditions, 37.695, plus the SS for Subjects within teams within conditions, 55.36. Do note that the analysis above uses the variance for teams within conditions as the error variance for testing the effect of Condtn. The 2 for the entire effect of teams is 71.184/126.544 = .563. That part due to the experimental manipulation is 33.489/126.544 = .265. References Howell, D. C. (2007). Statistical methods for psychology (6th ed.). Belmont, CA: Thompson Wadsworth. Howell, D. C. (2010). Statistical methods for psychology (7th ed.). Belmont, CA: Cengage Wadsworth. Return to Wuensch’s Statistics Lessons Page Copyright 2010, Karl L. Wuensch - All rights reserved.