Stat664 About Expected Mean Squares The discussions below are for Balanced Multi-Way Mixed-Effects Models. Unrestricted Models Any random effect (including any interaction that involves fixed effects) has its elements sampled from normal population with zero mean. For instance, assuming that βj ’s are random, even if αi ’s 2 ). are fixed, the random variables (αβ)ij ’s are i.i.d. N (0, σAB The general formula given on page F.9 in classnotes Variance Components Model is for unrestricted models (click the link above and then click on § 4.1 on the navigation tabs on the left). That is, no side conditions were imposed on summing over the fixed-effect margins of random effects. The two examples on page F.10, assuming unrestricted model, follow through this formula. Three-Way Unrestricted Mixed-Effects Models Fixed: A; Random: B, C. u A B C AB AC BC ABC Error Vu {B,C,BC} {A,C,AC} {A,B,AB} {C} {B} {A} — — E[M Su ] 2 2 2 ] + nσABC + nbσAC σ 2 + nbckA2 + [ncσAB 2 2 2 2 2 σ + nacσB + [ncσAB + naσBC + nσABC ] 2 2 2 σ 2 + nabσC2 + [nbσAC + naσBC + nσABC ] 2 2 2 σ + ncσAB + [nσABC ] 2 2 ] + [nσABC σ 2 + nbσAC 2 2 2 σ + naσBC + [nσABC ] 2 2 σ + nσABC σ2 Fixed: A, B; Random: C. u A B C AB AC BC ABC Error Vu {C,BC} {C,AC} {A,B,AB} {C} {B} {A} — — E[M Su ] 2 2 σ 2 + nbckA2 + [nbσAC + nσABC ] 2 2 2 2 σ + nackB + [naσBC + nσABC ] 2 2 2 2 2 σ + nabσC + [nbσAC + naσBC + nσABC ] 2 2 2 σ + nckAB + [nσABC ] 2 2 σ 2 + nbσAC + [nσABC ] 2 2 2 σ + naσBC + [nσABC ] 2 σ 2 + nσABC σ2 1 Restricted Models A restricted model has side conditions imposed on the fixed-effect margins. For the example above, the random variables (αβ)ij ’s are no longer independent, though still are normally distributed. Additional side conditions are imposed on these random variables. One key restriction, among others (about covariance structure), is that a X (αβ)ij = 0, for j = 1, · · · , b. i See the discussion on the two-way restricted model on page F.11 of the abovementioned classnotes link. Expected Mean Squares for Restricted Models The general formula on page F.9 still holds except for the replacement of Vu by Vur : Vur = {v : v is a pure random effect not involving u} Note that multi-way random effects model is unaffected by this definition. Three-Way Restricted Mixed-Effects Models Fixed: A; Random: B, C. u A B C AB AC BC ABC Error Vur {B,C,BC} {C} {B} {C} {B} — — — E[M Su ] 2 2 2 ] + nσABC + nbσAC σ 2 + nbckA2 + [ncσAB 2 2 2 σ + nacσB + [naσBC ] 2 σ 2 + nabσC2 + [naσBC ] 2 2 2 σ + ncσAB + [nσABC ] 2 2 ] + [nσABC σ 2 + nbσAC 2 2 σ + naσBC 2 σ 2 + nσABC σ2 Fixed: A, B; Random: C. u A B C AB AC BC ABC Error Vur {C} {C} — {C} — — — — E[M Su ] 2 σ 2 + nbckA2 + [nbσAC ] 2 2 2 σ + nackB + [naσBC ] 2 2 σ + nabσC 2 2 σ 2 + nckAB + [nσABC ] 2 2 σ + nbσAC 2 σ 2 + naσBC 2 2 σ + nσABC σ2 2 MINITAB By default, MINITAB assumes unrestricted models in both GLM and AOV (Balanced ANOVA) unless otherwise specified in Options sub-dialog box. The example follows through Multi-Way ANOVA Model using bottler example. The original data were obtained from fixed-effects experiment. Its use here is for illustration only. For convenience, the verbal names of the factors were changed using more convenient names A, B, and C. For each case of unrestricted/restricted × (A fixed, B&C random)/(A&B fixed, C random), only the table of expected mean squares were retained (note that MINITAB produced ANOVA table using appropriate error terms, and using Satterthwaite pseudo F when appropriate). Fixed: A; Random: B, C 1 2 3 4 5 6 7 8 Source A B C A*B A*C B*C A*B*C Error 1 2 3 4 5 6 7 8 Source A B C A*B A*C B*C A*B*C Error Variance component 3.52083 1.77083 0.52083 -0.06250 0.08333 -0.08333 0.70833 Variance component 3.69444 1.75000 0.52083 -0.06250 0.05556 -0.08333 0.70833 Error term * * * 7 7 7 8 Expected Mean Square for Each (using unrestricted model) (8) + 2 (7) + 4 (5) + 4 (4) + (8) + 2 (7) + 6 (6) + 4 (4) + (8) + 2 (7) + 6 (6) + 4 (5) + (8) + 2 (7) + 4 (4) (8) + 2 (7) + 4 (5) (8) + 2 (7) + 6 (6) (8) + 2 (7) (8) Term Error term * 6 6 7 7 8 8 Expected Mean Square for Each Term (using restricted model) (8) + 2 (7) + 4 (5) + 4 (4) + 8 Q[1] (8) + 6 (6) + 12 (2) (8) + 6 (6) + 12 (3) (8) + 2 (7) + 4 (4) (8) + 2 (7) + 4 (5) (8) + 6 (6) (8) + 2 (7) (8) Q[1] 12 (2) 12 (3) Fixed: A, B; Random: C 1 2 3 Source A B C Variance component 1.77083 Error term 5 6 * Expected Mean Square for Each Term (using unrestricted model) (8) + 2 (7) + 4 (5) + Q[1,4] (8) + 2 (7) + 6 (6) + Q[2,4] (8) + 2 (7) + 6 (6) + 4 (5) + 12 (3) 3 4 5 6 7 8 1 2 3 4 5 6 7 8 A*B A*C B*C A*B*C Error Source A B C A*B A*C B*C A*B*C Error -0.06250 0.08333 -0.08333 0.70833 Variance component 1.77778 -0.10417 0.05556 -0.08333 0.70833 7 7 7 8 Error term 5 6 8 7 8 8 8 (8) (8) (8) (8) (8) + + + + 2 2 2 2 (7) + Q[4] (7) + 4 (5) (7) + 6 (6) (7) Expected Mean Square for Each Term (using restricted model) (8) + 4 (5) + 8 Q[1] (8) + 6 (6) + 12 Q[2] (8) + 12 (3) (8) + 2 (7) + 4 Q[4] (8) + 4 (5) (8) + 6 (6) (8) + 2 (7) (8) SAS SAS basically uses unrestricted models in PROCs GLM, MIXED, and VARCOMP. However, users still have the flexibility to declare appropriate error term in a TEST statement. Note that SAS sort of advocates unrestricted model form (see “Sum-To-Zero Assumption” section in the help documentation of the RANDOM statement in PROC GLM). Assume that it’s desired to use unrestricted model. When using RANDOM statement with TEST option in PROC GLM, SAS does all the tests using appropriate error (pseudo error) terms and produces the correct output. See Multi-Way Mixed-Effect Model for example. For restricted model, use E = error-term in the TEST statement in PROC GLM. The error-term must be an effect (one only). 4