Statistical Analysis Professor Lynne Stokes Department of Statistical Science Lecture 6 Solving Normal Equations and Estimating Estimable Model Parameters 1 Regression Models Model y X e Residuals r y Xˆ Least Squares Choose ˆ to Minimize () (y - X)(y - X) Sum of Squared Residuals Solution: Solve the Normal Equations XXˆ Xy 2 Regression Solution ˆ XX 1 Xy Under usual assumptions, the least squares estimator is Unique Unbiased Minimum Variance Consistent Known sampling distribution Universally used 3 Analysis of Completely Randomized Designs Fixed Factor Effects Factor levels specifically chosen Inferences desired only on the factor levels included in the experiment Systematic, repeatable changes in the mean response 4 Flow Rate Experiment Fixed or Random ? Filter A B C D 0.233 0.259 0.183 0.233 Flow Rates 0.197 0.259 0.258 0.343 0.284 0.264 0.328 0.267 0.244 0.305 0.258 0.269 MGH Fig 6.1 5 Flow Rate Experiment 0.35 Filter A B C D 0.30 Average Flow Rate Effects -0.028 0.030 -0.014 0.013 Conclusion ? 0.25 0.20 A B C D 6 Filter Type Statistical Model for Single-Factor, Fixed Effects Experiments Model yij = m + ai + eij Response i = 1, ..., a; j = 1, ..., ri Overall Main Mean Effect (Constant) for Level i Error ai: Effect of Level i = change in the mean response 7 Statistical Model for Single-Factor, Fixed Effects Experiments Cell Means Model yij = mi + eij i = 1, ..., a; j = 1, ..., ri mˆ i yi Effects Model yij = m + ai + eij i = 1, ..., a; j = 1, ..., ri Fixed Effects Models Connection: mi = m + ai ~ y y mˆ y , a i i 8 Solving the Normal Equations Single-Factor, Balanced Experiment yij = m + ai + eij i = 1, ..., a j = 1, ..., r n = ar Matrix Formulation y = X + e y = (y11 y12 ... y1r ... ya1 ya2 ... yar)’ 1r 1r 1 0 r X= r ... ... 1 0 r r 0r 1r ... 0r 0r 0r [(1r 1a ) : I a 1r ] ... 1r m a1 ... aa 9 Solving the Normal Equations Residuals ~ r y X Least Squares ~ Choose to Minimize () (y - X)(y - X) Solution: Solve the Normal Equations ~ XX Xy 10 Solving the Normal Equations Normal Equations ~ X X = X y n r r ... r r 0 ... r 0 r ... ... ... ... ... r 0 0 ... ~ r m y ~ 0 a1 y1 ~ 0 a 2 y 2 ... ... ... ~ r a a y a Check 11 Solving the Normal Equations Normal Equations ~ ra ~ + ... + ra ~ y ~ + ra nm 1 2 a ~ ~ + ra rm y1 1 ~ ~ rm + ra y 2 ... ~ rm Linearly Dependent 2 ~ y + ra a a a + 1 Parameters, a Linearly Independent Equations Infinite Number of Solutions Check 12 Solving the Normal Equations Normal Equations ~ nm ~ ~ + ra rm y y1 1 ~ rm ~ + ra 2 ... ~ rm One Solution a ai 0 i 1 y 2 ~ y + ra a a a a~ i 0 i 1 ~=y m ~ y y a i i ~ y ~a mˆ i m i i13 Solving the Normal Equations Normal Equations ~ ra ~ ... ra ~ y ra 1 2 a ~ ra y1 1 ~ ra y 2 2 ... ~ y ra a a Another Solution ~0 m0m ~ y a i i ~ y ~a mˆ i m i i14 Solving the Normal Equations Normal Equations ~ ra ~ + ... + ra ~ y ~ + ra nm 1 2 a -1 ~ ~ + ra rm y 1 1 ~ rm ... ~ rm ~ + ra 2 y 2 ~ y + ra a -1 ( a 1) Another Solution ~ 0 m ~y a a a ~ y y a i i a i = 1, ... , a - 1 ~ y ~a mˆ i m i i15 Solving the Normal Equations All solutions to the normal equations produce the same estimates of “estimable functions” of the model means Solutions are not estimates Estimable Functions All solutions provide one unique estimator Estimators are unbiased 16 Solving the Normal Equations Two-Factor, Balanced Experiment yijk = mij + eijk = m + ai + j + (a)ij + eijk Matrix Formulation y = X + e i = 1, ..., a j = 1, ..., b k = 1, ..., r n = abr X = [ 1 : XA : XB : XAB ] m , a1 , ... , aa , 1 , ... , b , a11 , ... , aab 17 Solving the Normal Equations Two-Factor, Balanced Experiment yijk = mij + eijk = m + ai + j + (a)ij + eijk i = 1, ..., a j = 1, ..., b k = 1, ..., r Matrix Formulation y = X + e n = abr X = [ 1 : XA : XB : XAB ] m , a1 , ... , aa , 1 , ... , b , a11 , ... , aab Number of Parameters 1 + a + b + ab rank( X ) < 1+a+b+ab 18 Solving the Normal Equations Normal Equations ~ X X = X y n br ... ar ... br br ... 0 ... ... ... ... ... ... 0 0 ... 0 ... ~ y m ~ r a 1 y1 0 ... ... ~ ... 1 y1 r ... ... a y ab ab Check 19 Solving the Normal Equations Matrix 1n XA Linear Dependencies None 1 : Columns of XA Sum to 1n One Solution aa = 0 Eliminates a column From XA a – 1 “degrees of freedom” 20 Solving the Normal Equations Matrix 1n XA XB Linear Dependencies None 1 : Columns Sum of XA to 1n 1 : Columns Sum of XB to 1n One Solution aa = 0 b = 0 Eliminates a column From XB b – 1 “degrees of freedom” 21 Solving the Normal Equations Matrix 1n XA XB XAB Linear Dependencies None 1 : Columns sum to 1n 1 : Columns sum to 1n 1 + (a - 1) + (b - 1) : Sum over all columns = 1n One Solution aa = 0 b = 0 (a)ab = 0 Eliminates a column from XAB 22 Solving the Normal Equations Matrix 1n XA XB XAB Linear Dependencies One Solution None 1 : Columns Sum to 1n aa = 0 1 : Columns Sum to 1n b = 0 1 + (a - 1) + (b - 1) : Sum over all columns = 1n (a)ab = 0 Sums of columns over each i = 1,...,a-1 & each j = 1,...,b-1 (a)ib = 0 equal one of the remaining i=1,...,a-1 columns of XA and XB (a)aj = 0 j=1,...,b-1 (a – 1)(b – 1) “degrees of freedom” 23 Solving the Normal Equations Matrix XA XB XAB Linear Dependencies One Solution 1 : Columns sum to 1n aa = 0 1 : Columns sum to 1n b = 0 1 + (a - 1) + (b - 1) : Sum over all columns = 1n (a)ab = 0 Sums of columns over each i = 1,...,a-1 & each j = 1,...,b-1 (a)ib = 0 equal one of the remaining i=1,...,a-1 columns of XA and XB (a)aj = 0 j=1,...,b-1 Constraints : 1 + 1 + {1 + (a - 1) + (b - 1)} = a + b + 1 Degrees of Freedom : (1 + a + b + ab) - (a + b + 1) = ab = 1 + (a - 1) + (b - 1) + (a - 1)(b - 1) 24 Solving the Normal Equations ~y m ab ~ y m ~ a i ib ~ 0 a a ~ ~ j yaj m ~ b 0 i = 1, . .. , a - 1 j = 1, . . . , b - 1 ~ ~a ~ (a ) ij yij m i j i a ; j b (a ) ij 0 i = a or j = b Check 25 Solving the Normal Equations Another Solution ~y m ~ y y a i i ~ j y j y i = 1, ... , a j = 1, ... , b (a )ij yij yi y j y i, j Check 26 Flow Rate Experiment Fixed or Random ? Filter A B C D 0.233 0.259 0.183 0.233 Flow Rates 0.197 0.259 0.258 0.343 0.284 0.264 0.328 0.267 0.244 0.305 0.258 0.269 MGH Fig 6.1 27 Quantifying Factor Effects Effect Change in average response due to changes in factor levels Factor Level Average 1 2 3 ... k Overall Average y1 y 2 y3 ... y k y Effect of Level t : y t - y 28 Quantifying Factor Effects Effect Change in average response due to changes in factor levels Factor Level Average 1 2 y1 y 2 3 ... k y3 ... y k Effect of changing from Level s to Level t : Overall Average y (y t y ) - (ys - y ) = y t - ys 29 Quantifying Factor Effects Main Effects for Factor A y i y Change in average response due to changes in the levels of Factor A Main Effects for Factor B y j y Change in average response due to changes in the levels of Factor B Interaction Effects for Factors A & B (y ij - y j ) - (y i y ) Effect of Level i of Factor A at Level j of Factor B Effect of Level i of Factor A 30 Quantifying Factor Effects Main Effects for Factor A y i y Change in average response due to changes in the levels of Factor A Main Effects for Factor B y j y Change in average response due to changes in the levels of Factor B Interaction Effects for Factors A & B (y ij - y j ) - (y i y ) y ij - y i - y j y Change in average response due joint changes in Factors A & B in excess of changes in the main effects 31 Two-Level Factors Effect of Level 1: y 1 y Effect of Level 2: y 2 y Common to Use y 2 - y 1 Note: If r1 = r2 , y 1 y = - (y 2 y ) 32 Factors at Two Levels Most common choice for designs involving many factors Many efficient fractional factorial and screening designs available Can use p two-level factors in place of factors p whose number of levels is 2 33 Calculating Two-Level Factor Effects: Pilot Plant Study Main Effect Difference between the average responses at the two levels M(Temp) = Average @ 180o - Average @ 160o = 75.8 - 52.8 = 23.0 M(Conc) = Average @ 40% - Average @ 20% = 61.8 - 66.8 = -5.0 M(Catalyst) = Average @ C2 - Average @ C1 = 65.0 - 63.5 = 1.5 BHH Section 10.3 MGH Section 5.3 34 Calculating Two-Level Factor Effects Two-Factor Interaction Effect Half the difference between the main effects of one factor at each level of the second factor M(Conc @ C2) = Average @ 40%&C2 - Average @ 20%&C2 = 62.5 - 67.5 = -5.0 M(Conc @ C1) = Average @ 40%&C1 - Average @ 20%&C1 = 61.0 - 66.0 = -5.0 I(Conc,Cat) = {M(Conc @ C2) - M(Conc @ C1)} / 2 =0 BHH Section 10.4 MGH Section 5.3 35 Calculating Two-Level Factor Effects Two-Factor Interaction Effect Half the difference between the main effects of one factor at each level of the second factor M(Temp @ C2) = Average @ 180o&C2 - Average @ 160o&C2 = 81.5 - 48.5 = 33.0 M(Temp @ C1) = Average @ 180o&C1 - Average @ 160o&C1 = 70.0 - 57.0 = 13.0 I(Temp,Cat) = {M(Temp @ C2) - M(Temp @ C1)} / 2 = (33.0 - 13.0) / 2 = 10.0 36 Cell Means and Effects Model Estimability Three-Factor Balanced Experiment yijkl = mijk + eijkl i = 1 , ... , a ; j = 1 , ... , b ; k = 1, ... , c ; l = 1 , ... , r mijk = m + ai + j + gk + (a)ij + (ag)ik + (g)jk + (ag)ijk 37 Cell Means Models: Estimable Functions All cell means are estimable m ijk y ijk 38 Cell Means Models: Estimable Functions All cell means are estimable m ijk y ijk All linear combinations of cell means are estimable c ijk m ijk c ijk yijk Does not depend on parameter constraints (includes m i , m ij , etc.) 39 Cell Means Models: Estimable Functions All cell means are estimable m ijk y ijk Some linear combinations of cell means are uninterpretable m1 m 23 Some linear combinations of cell means are essential m1 m 2 40 Cell Means and Effects Models mi m ai g (a )i (ag )i (g ) (ag )i Imposing parameter constraints simplifies the relationships; makes the parameters more interpretable 41 Parameter Equivalence: Effects Representation & Cell Means Model Parameter constraints a i . . . = (a) ij = . . . = i ij (a) ijk 0 ijk Means and mean effects m i m a i m ij m a i j (a ) ij a i m i m (a) ij m ij m i m j m ( m ij m j ) ( m i m ) 42 Contrasts Contrast A Linear Combinatio n of Parameters whose Coefficien ts Sum to Zero k a j j j1 k with a j j1 Contrasts often eliminate nuisance parameters; e.g., m 43 Contrasts Main Effects m i m i a i a i Interactions m ij m i m j m (a) ij m ij m il m kj m kl (a) ij (a) il (a) kj (a) kl 4cr (a) ij2 {(a) ij (a) il (a) kj (a) kl } ij 2 ijkl Show 44