Notes on Expected Mean Squares • There are two approaches or models for random effects: The “restricted” model and the “unrestricted” model. We’ll focus on the “unrestricted” model, which is the default in Minitab, and which makes the most sense. The restricted model basically incorporates the assumption that a random effect-by-fixed effect interaction sums to 0 over the levels of the fixed effect. I cannot imagine why, in real life, one would assume this to be so. • The general idea is that each Mean Square arising from the model (appearing in the ANOVA table) has an expected value of the form “part 1 + part 2” where “part 2” is 0 if the associated null hypothesis is true. One looks to find another term in the model whose expected mean square is “part 1”. One then uses this other term as the denominator of the F-statistic. Then it follows that F ≈ part 1 + part 2 part 1 F ≈ part 1 + 0 =1 part 1 and if H0 is true then and if H0 is false then F tends to be greater than 1. • There is a set of not-too-complicated rules for determining the expected mean squares (for both the unrestricted and restricted models) as long as the design is balanced. • These are not-too-complicated but a bit awkward to state. Minitab knows how to work out the expected mean squares and will do so if asked. • Most of the time you won’t actually need to use these rules, since Minitab does the work for you. However my firm belief is that you are less likely to make mistakes over-all if you are capable of handling these rules yourself. In order to state the rules, we first need to establish an elephant-ful of notation: - Let U denote any “effect” — main effect or interaction. Think of U as a sequence of letters such as ABC where A, B, and C represent main effects. (Of course U could consist of a single letter such as A.) - If a main effect C is nested within an effect U , then we represent C as U C . E.g if C is nested within AB (i.e. within each combination of levels of A and B, then we represent C as ABC; note that there will be no AC or BC (and of course no actual C) terms in the model. A bit of useful jargon: In such cases, i.e. where ABC appears in the model to represent the effect C which is nested within AB, we say that A and B appear in ABC “by convention”. - Let the number of levels of each main effect A, B, C, . . . , be denoted by the corresponding lower case letter, a, b, c, . . . . Let the number of replicates be r . - With each effect U in the model there is associated a non-negative quantity denoted by σU2 . If U is a random effect then σU2 is indeed its variance. If U is a fixed effect then σU2 is just a “quadratic form” (sum of squares) involving the parameters associated with that effect. In any case the effect U is actually present if and only if σU2 6= 0 . - The variance of the error term in the model is denoted by σ 2 . • THE RULES: 1. The expected mean square for U always contains a σ 2 and a σU2 . 2. It also contains a σ 2 term for any interaction in which all all the letters of U appear and at least one letter represents a random effect. (For the restricted model the proviso is that all other letters must represent random effects. Note however that this allows sequences of letters used to represent nested random effects, even if some of the individual letters on their own represent fixed effects.) 3. The coefficient of σ2 is always 1. 4. The coefficient of the σ 2 for any other term is r times the product of all the lower case letters a, b, c, . . . whose corresponding upper case letters DO NOT appear in that term. • E.g. # 1: A fixed, B random, crossed with each other, so the model is Y = mean + A + B + AB + Error Source A B AB Error Expected Mean Square 2 + rσ 2 σ2 + rbσA AB 2 + rσ 2 σ2 + raσB AB 2 σ2 + rσAB σ2 From this we see that the proper denominator to test for the A and B effects is the mean square for AB, and the proper denominator to test for the AB effect is the mean square for Error. • E.g. # 2: A fixed, B random, crossed with each other, C random and nested within AB, The model is Y = mean + A + B + AB + ABC + Error (so A and B appear in ABC by convention). Source A B AB ABC Error Expected Mean Square 2 + rcσ 2 + rσ 2 σ2 + rbcσA AB ABC 2 + rcσ 2 + rσ 2 σ2 + racσB AB ABC 2 + rσ 2 σ2 + rcσAB ABC 2 σ2 + rσABC 2 σ 2 The appropriate denominator to test for an A effect is M SAB; for a B effect it is M SABC; for an AB effect it is also M SABC, and for an ABC effect it is M SError. • E.g. # 3: A fixed, B and C random, all crossed, so the model is Y = mean + A + B + AB + C + AC + BC + ABC + Error Source A B AB C AC BC ABC Error Expected Mean Square 2 + rcσ 2 + rbσ 2 + rσ 2 σ2 + rbcσA AB AC ABC 2 2 + rcσ 2 + raσ 2 + rσ 2 σ + racσB AB BC ABC 2 + rσ 2 σ2 + rcσAB ABC 2 + raσ 2 + rσ 2 σ2 + rabσC BC ABC 2 + rσ 2 σ2 + rbσAC ABC 2 + rσ 2 σ2 + raσBC ABC 2 σ2 + rσABC σ2 This seems to be a perfectly sensible design, and yet there is no appropriate denominator to test for any of the main effects. In particular you can’t test for the fixed effect A. This is weird. Minitab will (these days) produce a “synthetic F-test” in such situations; but such synthetic tests are approximations and should be interpreted with even more than the usual amount of caution. • E.g. # 4: (a weird one; I’m not completely sure that I’ve got this right or that it even makes sense . . . . . . ) A fixed, B random, C random and nested within A, A crossed with B and B crossed with C. The model is Y = mean + A + B + AB + AC + ABC + Error Note that since C is nested within A we represent C as AC, and then B crossed with C is represented as B × AC which becomes ABC . (E.g. A is different measuring devices; B is technicians; C is specimens to be measured; each technician measures each specimen r times and each technician uses all devices; however the specimens have to be prepared in a particular way for each device and so a specimen to be measured by device number 1 cannot be measured by device number 2, etc.) (O.K., so it’s a bit far-fetched, but I’ll bet you that weirder designs have happened in real life!) Source A B AB AC ABC Error Expected Mean Square 2 + rcσ 2 + rbσ 2 + rσ 2 σ2 + rbcσA AB AC ABC 2 2 + rcσ 2 + rσ 2 σ + racσB AB ABC 2 + rσ 2 σ2 + rcσAB ABC 2 2 + rσ 2 σ + rbσAC ABC 2 σ2 + rσABC 2 σ 3 There is no appropriate denominator to test for an A effect, which is probably what you’re really interested in — but that’s not too surprizing since it’s such a weird design. • Just for the sake of completeness, let’s run through the foregoing examples under the restricted model. E.g. # 1 (restricted model): Source A B AB Error Expected Mean Square 2 + rσ 2 σ2 + rbσA AB 2 σ2 + raσB 2 σ2 + rσAB σ2 Under the restricted model we see that the right denominator to test for the A effect is the mean square for AB, and the right denominator to test for the B and AB effects is the mean square for Error. E.g. # 2 (restricted model): Source A B AB ABC Error Expected Mean Square 2 + rcσ 2 + rσ 2 σ2 + rbcσA AB ABC 2 + rσ 2 2 σ + racσB ABC 2 + rσ 2 σ2 + rcσAB ABC 2 2 σ + rσABC σ2 Under the restricted model the appropriate denominator to test for an A effect is M SAB; for a B effect it is M SABC; for an AB effect it is also M SABC, and for an ABC effect it is M SError. E.g. # 3 (restricted model): Source A B AB C AC BC ABC Error Expected Mean Square 2 + rcσ 2 + rbσ 2 + rσ 2 σ2 + rbcσA AB AC ABC 2 2 + raσ 2 σ + racσB BC 2 + rσ 2 σ2 + rcσAB ABC 2 2 + raσ 2 σ + rabσC BC 2 + rσ 2 σ2 + rbσAC ABC 2 σ2 + raσBC 2 σ2 + rσABC 2 σ Under the restricted model there is no appropriate denominator to test for an A effect (but B and C are O.K.) 4 E.g. # 4 (restricted model): Source A B AB AC ABC Error Expected Mean Square 2 + rcσ 2 + rbσ 2 + rσ 2 σ2 + rbcσA AB AC ABC 2 2 + rσ 2 σ + racσB ABC 2 + rσ 2 σ2 + rcσAB ABC 2 2 + rσ 2 σ + rbσAC ABC 2 σ2 + rσABC 2 σ Again under the restricted model there is no appropriate denominator to test for an A effect. • We will now display Minitab’s version of all this. Minitab doesn’t work in the abstract so it had to be given some data. The data used to generate these analyses were white noise, so don’t expect to see anything sensible in the actual tests. (Things should always turn out to be “non-significant”.) Specific numbers had to be chosen for the numbers of levels of each factor, and for the number of replicates in each cell. The numbers used were a = 3, b = 5, c = 4, and r = 3. Note 1: Minitab writes the EMS terms corresponding to fixed effects as “Q[...]” and those corresponding to random effects (variances) simply as “(...)” where the ... bits are the numbers corresponding to terms in the model. Also the coefficients of the “Q[...]” terms are omitted under the unrestricted model (although they are put in for the restricted model). Minitab (misleadingly, I think) indicates a sort of dependence on interactions between fixed effects under the unrestricted model. E.g. Minitab’s MS for A in the model Y = mean + A + B + AB + C + AC + BC + ABC + Error where A and B are both fixed effects and C is random, a, b, c and r as above, is (8) + 3(7) + 15(5) + Q[1,4] where [1 refers to A, 4] refers to AB, ((5) refers to AC, (7) refers to ABC, and (8) refers to Error. The rules tell us that the MS for A is 2 2 2 σ2 + rbcσA + rbσAC + rσABC with no allusion to AB. (There would be an AB term if B were random.) There’s always got to be something weird going on to confuse you. Note 2: The following Minitab output was produced with an “older” version of Minitab. These days, where it says “ * No exact F-test can be calculated.”, Minitab would actually calculate the approximated or “synthetic” F-test. . . . . . . (continued over page) 5 First the unrestricted model: MTB > # E.g. 1 (unrestricted model): MTB > ANOVA ’Y’ = A B A*B; SUBC> Random B; SUBC> EMS. Source A B A*B Error Total DF 2 4 8 165 179 Source Variance Error Expected Mean Square component term (using unrestricted model) 3 (4) + 12(3) + Q[1] -0.00394 3 (4) + 12(3) + 36(2) -0.00944 4 (4) + 12(3) 0.92079 (4) 1 2 3 4 A B A*B Error SS 1.9874 2.6626 6.4598 151.9308 163.0406 MS 0.9937 0.6656 0.8075 0.9208 F 1.23 0.82 0.88 P 0.342 0.545 0.537 MTB > # E.g. 2 (unrestricted model): MTB > ANOVA ’Y’ = A B A*B C(A B); SUBC> Random B C; SUBC> EMS. Source A B A*B C(A B) Error Total DF 2 4 8 45 120 179 Source Variance Error Expected Mean Square component term (using unrestricted model) 3 (5) + 3(4) + 12(3) + Q[1] -0.00394 3 (5) + 3(4) + 12(3) + 36(2) -0.01734 4 (5) + 3(4) + 12(3) 0.04341 5 (5) + 3(4) 0.88528 (5) 1 2 3 4 5 A B A*B C(A B) Error SS 1.9874 2.6626 6.4598 45.6975 106.2333 163.0406 MS 0.9937 0.6656 0.8075 1.0155 0.8853 F 1.23 0.82 0.80 1.15 P 0.342 0.545 0.610 0.275 . . . . . . (continued over page) 6 MTB > # E.g. 3 (unrestricted model): MTB > ANOVA ’Y’ = A B A*B C A*C B*C A*B*C; SUBC> Random B C; SUBC> EMS. Source A B A*B C A*C B*C A*B*C Error Total DF 2 4 8 3 6 12 24 120 179 Source Variance Error Expected Mean Square component term (using unrestricted model) * (8) + 3(7) + 15(5) + 12(3) + Q[1] -0.02670 * (8) + 3(7) + 9(6) + 12(3) + 36(2) 0.00729 7 (8) + 3(7) + 12(3) -0.01580 * (8) + 3(7) + 9(6) + 15(5) + 45(4) 0.02329 7 (8) + 3(7) + 15(5) 0.09102 7 (8) + 3(7) + 9(6) -0.05510 8 (8) + 3(7) 0.88528 (8) 1 2 3 4 5 6 7 8 A B A*B C A*C B*C A*B*C Error SS 1.9874 2.6626 6.4598 3.5322 6.4157 18.4703 17.2793 106.2333 163.0406 MS 0.9937 0.6656 0.8075 1.1774 1.0693 1.5392 0.7200 0.8853 F * * 1.12 * 1.49 2.14 0.81 P 0.384 0.225 0.055 0.714 * No exact F-test can be calculated. MTB > # E.g. 4 (unrestricted model): MTB > ANOVA ’Y’ = A B A*B C(A) B*C(A) ; SUBC> Random B C; SUBC> EMS. Source A B A*B C(A) B*C(A) Error Total DF 2 4 8 9 36 120 179 Source Variance Error Expected Mean Square component term (using unrestricted model) * (6) + 3(5) + 15(4) + 12(3) + Q[1] -0.00394 3 (6) + 3(5) + 12(3) + 36(2) -0.01546 5 (6) + 3(5) + 12(3) 0.00748 5 (6) + 3(5) + 15(4) 0.03592 6 (6) + 3(5) 0.88528 (6) 1 2 3 4 5 6 A B A*B C(A) B*C(A) Error SS 1.9874 2.6626 6.4598 9.9478 35.7497 106.2333 163.0406 MS 0.9937 0.6656 0.8075 1.1053 0.9930 0.8853 * No exact F-test can be calculated. 7 F * 0.82 0.81 1.11 1.12 P 0.545 0.596 0.379 0.316 And now the restricted model: MTB > # E.g. 1 (restricted model): MTB > ANOVA ’Y’ = A B A*B; SUBC> Random B; SUBC> Restrict; SUBC> EMS. Source A B A*B Error Total DF 2 4 8 165 179 Source Variance Error Expected Mean Square component term (using restricted model) 3 (4) + 12(3) + 60Q[1] -0.00709 4 (4) + 36(2) -0.00944 4 (4) + 12(3) 0.92079 (4) 1 2 3 4 A B A*B Error SS 1.9874 2.6626 6.4598 151.9308 163.0406 MS 0.9937 0.6656 0.8075 0.9208 F 1.23 0.72 0.88 P 0.342 0.577 0.537 MTB > # E.g. 2 (restricted model): MTB > ANOVA ’Y’ = A B A*B C(A B); SUBC> Random B C; SUBC> Restrict; SUBC> EMS. Source A B A*B C(A B) Error Total DF 2 4 8 45 120 179 Source Variance Error Expected Mean Square component term (using restricted model) 3 (5) + 3(4) + 12(3) + 60Q[1] -0.00972 4 (5) + 3(4) + 36(2) -0.01734 4 (5) + 3(4) + 12(3) 0.04341 5 (5) + 3(4) 0.88528 (5) 1 2 3 4 5 A B A*B C(A B) Error SS 1.9874 2.6626 6.4598 45.6975 106.2333 163.0406 MS 0.9937 0.6656 0.8075 1.0155 0.8853 F 1.23 0.66 0.80 1.15 P 0.342 0.626 0.610 0.275 . . . . . . (continued over page) 8 MTB > # E.g 3 (restricted model): MTB > ANOVA ’Y’ = A B A*B C A*C B*C A*B*C; SUBC> Random B C; SUBC> Restrict; SUBC> EMS. Source A B A*B C A*C B*C A*B*C Error Total DF 2 4 8 3 6 12 24 120 179 Source Variance Error Expected Mean Square component term (using restricted model) * (8) + 3(7) + 15(5) + 12(3) + 60Q[1] -0.02427 6 (8) + 9(6) + 36(2) 0.00729 7 (8) + 3(7) + 12(3) -0.00804 6 (8) + 9(6) + 45(4) 0.02329 7 (8) + 3(7) + 15(5) 0.07266 8 (8) + 9(6) -0.05510 8 (8) + 3(7) 0.88528 (8) 1 2 3 4 5 6 7 8 A B A*B C A*C B*C A*B*C Error SS 1.9874 2.6626 6.4598 3.5322 6.4157 18.4703 17.2793 106.2333 163.0406 MS 0.9937 0.6656 0.8075 1.1774 1.0693 1.5392 0.7200 0.8853 F * 0.43 1.12 0.76 1.49 1.74 0.81 P 0.783 0.384 0.535 0.225 0.067 0.714 * No exact F-test can be calculated. MTB > # E.g. 4 (restricted model): MTB > ANOVA ’Y’ = A B A*B C(A) B*C(A); SUBC> Random B C; SUBC> Restrict; SUBC> EMS. Source A B A*B C(A) B*C(A) Error Total DF 2 4 8 9 36 120 179 SS 1.9874 2.6626 6.4598 9.9478 35.7497 106.2333 163.0406 MS 0.9937 0.6656 0.8075 1.1053 0.9930 0.8853 F * 0.67 0.81 1.11 1.12 P 0.617 0.596 0.379 0.316 . . . . . . (continued over page) 9 Source 1 2 3 4 5 6 A B A*B C(A) B*C(A) Error Variance Error Expected Mean Square component term (using restricted model) * (6) + 3(5) + 15(4) + 12(3) + 60Q[1] -0.00909 5 (6) + 3(5) + 36(2) -0.01546 5 (6) + 3(5) + 12(3) 0.00748 5 (6) + 3(5) + 15(4) 0.03592 6 (6) + 3(5) 0.88528 (6) * No exact F-test can be calculated. • It is actually possible to convert from the restricted to the unrestricted model by making appropriate changes of variable, and to show that the two are really the same thing!!! So what’s going on when the restricted model tells you to form your F-statistic one way, and the unrestricted model tells you to form it another way? The long and the short of it, it seems to me, is: You are simply testing different hypotheses. For instance, in E.g. 1, the two F statistics for testing for a “B” effect both look like 2 = 0. However, if you go through the change of variable they are testing H0 : racσB 2 under the restricted model is precisely equal to jazz, you see that the value of racσB 2 2 that of racσB + rcσAB under the unrestricted model. Thus testing for a “B” effect using the restricted model (i.e. using M SE as the 2 + rcσ 2 denominator of the F statistic) really amounts to testing H 0 : racσB AB = 0 2 2 which can only be true if both σB and σAB are 0. I.e. you are really testing that both the “B” effect and its interaction with A (as they would be interpreted under the unrestricted model) are null. This does not seem to me to be what one would want to do. I strongly recommend sticking with the unrestricted model. 10