FUNDAMENTALS OF APPLiey — 52 St SeRIEQOUGRON tatistical tool for tests of significan., sans, In a situation wk of the difference between two a acpiizrelis ness tor cc re samples to consider at a time — . me population, i.c., they have fe Ee spe plots each of wheat and yield of that all the samples a re applied to fou = example, five fertilizers are app nding out whether the ever of gtBeat : Ve may be interested in finding nose =e Sai Gane ear lonther wands, whether tle Samples hay. tl the yields is significantly differe ion, The answer to this problem is provided jy the | analy. F euienes of Variance’ was introduced by Prof. R.A. Fisher in 199) Pei ats analysis of agronomical data. Variation is inherent in naq@220* ta Sct acciuanhe causes, and (ii) Chance causes. £8 The variation due to assignable causes can be detected and measured hereas variation due to chance causes is beyond the control of human hand and cannot be eS st ly. % oe efinition. According to Prof. R.A. Fisher, Analysis of Variance ( ANOVA) i “Separation of variance ascribable to one group of causes from the variance ex. ribabe group.” By this technique the total variation in the sample data is expressed as the gy. thet non-negative components where each of these components is a measure of the y, ‘e some specific independent source or factor or cause . The ANOVA consists in the esting of the amount of variation due to each of the independent factors (causes) Separately comparing these estimates due to assignable factors (causes) with the estimate due Factor (causes), the latter being known as experimental error or simply error, 5-11. Assumptions for ANOVA Test. ANOVA test is based on the test Statistics Pag Variance Ratio). For the validity of the F-test in ANOVA, the following assumptions are made : (©) The observations are independent, (#) Parent population from which observations are taken is normal, and {40 Various treatment and environmental effects are additive in nature. In the following sequences we will discuss the analysi of variance for : (a) One-way classification, and (b) Two-way classi fication, Before that we state below (without proof) Cochran's theorem, which is of fundamen importance in analysis of variance. Cochran’s Theorem. Lot Xi, Xz, ..., X, denote a random sample from normal Population N(0, o°). Let the sum of the squares of these values be written in the form O14 Qe +... + Qy ist oan Tare Q,is a quadratic form in X;, XX. with n es of freedom) hi f= Xp, .., Xy, with rank (degrees of fre 7-1 aa Then the random variables @;, @,. Q are mutually independent and Q,/ 0° is f with r; degrees of freedom fand only if Sr, i be 0 chang{OVA technique enables us to compare several pop: Oot of savings in terms of time and money 1 sing two populations means at a time. comparitt ariier, the origin of the ANOVA technique lies in agricultural ¢ en is loaded with such terms as treatments, blocs, plots ete I SOnee a OnE ‘rsatile that it finds applications in almost all types of design of experims vi go versatt education, psychology, business ete. er eration food that ANOVA technique is not designed to test the equ eanees, Rather, its objective is to test the equality of several population mare vig of several independent sample means. a ity Sysists_of classifYing and cross classifying statistical results (data) and testing if the WA consi’ cations differ significantly from each other. Thus, it enables us to determine ified Cfeation is important in effecting the results sc to testing the homogeneity of several sample means, the ANOVA technique is now ion Tin testing the linearity of the fitted regression line or the significance of the pplicthe discussion on these topics is, however, beyond the scope of this book, WAY CLASSIFICATION pose that V observations yy, ( ea , 2, ....n;) of a random variable ‘ on some basis, into h lasses of sizes my, no, ..., ng respectively, bv 5 n) 23 “ Table 5-1. TABLE 5-1: ONE-WAY CLASSIFIED DATA ‘Sample Observations Total Mean EAU yz oe Prin} Ty vy Yar Yaa “ Yon, T2 Fe ya yan Yin, yi as vhe Yan, _ Te variation in the observation y, can be split into the following two components : variation between the classes or the variation due to different bases of , commonly known as treatments. variation within the classes, i.e., the inherent variation of the random variable ations of a class. type of variation is due to assignable causes which can be detected and human endeavour and the second type of variation is due to chance causes nd the control of human hand. object of analysis of variance technique is to examine if there is significant een the class means in view of the inherent variability within the separate ar, let us eonsider the effect of k different rations on the yield in milk of N me breed and stock) divided into k classes of sizes 71, "2, Mk respectively, the sources of variation are :__ ) Effect of the ration (treatment) : ¢;;i=1,2,..., k. Gi) Brror (@) produced by numerous causes of such magnitude that 1 and identified with the knowledge that we have and qo ey | ____ variation of random nature obeying Gaussian (Normal) lay. oe cree bi ical Model. In this case the linear mathematical fs Jy Hoi + Y= H+ (Hi -W + Ey . Fie a;+ 6) j where @=1,2,.. ; joa, 9 a ‘Hy 18 the yield from the jth cow, (J) = 1,2, +m) fed on the ith ation ¢ Model wi ws jis the general mean effect given by : mye k peg n= mn/N is the fixed effect due to the ith ration, ie. if there were no treatm, "y eee oS yield of each cow will by j, i, ‘the effect of the ith ration given by: -w 212... af Bren nerones or decreas the snd by an amenity yg we get at k Lng; = Ln -w) =D mai - HDi =N. pp. N=0 opts ; iv) ¢j is the error effect due to chance. NOTE. Fixed Effect Model vs. Random Effect Model Fixed Effect Model. Suppose the k-levels of the factor (treatmey are the only levels of interest and all these are included in the experiment by the hae or auto large number af eases, the Mlases (treatments) in the model ee specifically chosen by the experimenter. In such a case o's are fixed eonstane (unkans and the model (5-1) is known as the fixed effect model. In this model we would likey estimate G;’s and test some hypothesis about o,’s. In the fixed effect model, the concusi, about the test of hypothesis regarding the parameters i's will apply only to the treatneyy (factor levels) considered in the experiment. These conclusions, to the other remaining treatments (factor levels) experiment. Random Effect Model, Suppo: under consideration) and we w: ts) Under erste however, cannot be extende| which were not considered in the se we have a large number of classes (levels of fave ant to test, through an experiment, if all these class fees only a random sample of classes (factor levels) in the experiment and after studying a analysing the sample data, we want to draw conclusions which would be valid for alle classes (factor levels), whether included in the experiment of not. In such a situation te - Parameters a's in the model (5-1) will not be fixed constants but will be random sample i consequently the model is termed as random effect model. Jn the random effect mode if the null hypothesis of the homogeneity of cass tein! tet ee he to teat the significance af the diference between ods i effects, we cannot. apply the ¢-test boeause all dimstesahnecte ene-notineolel experiment, and Lea estima given by zVARIANCE YSIS_OF aan ase eNce MANOVA Po Fixed prices ofiinterest, then thy cant YT Bise ramets * level ny u v de} do, $ =a eve belo 4 Gay 2 ow Xt, shere 04'S are fixed (unknown) “onstants and agg Isto W © Symbo} 2) 2d). ol onan in Model (5.9), Ng Pingg 63 " Mtn ig () All the observations i’S) are independent Qty (ii) Different effects ATE Additive ™ nature Yon {,o. Ui) ey are ind. 1 (0, 0%), ie, 0) =0 an Under the third @ssumption, the mode) (3) W=0y; andj Ram bg, BR ok YS ot Statistical Analysig of Modey (5.3) PO any C : 54) Ul Hypothesis, Ww. ant to tose 4 ual wee of different ration. a . en tin mean Ao: = ye which from (5:25) reduces to 2° 35'e? 23D yy qaye nee Sttstion fr estimating wand te ty coal 3,772 2 Oy-n-a) = 0 .. 222 Gynec. ees oe fe wy +) . ww (55) ( b We get S na;=0 une 6] TU MAY nag = i : i (560) ), we got iis SE avin: 21-53 ee ae os 7 aFUNDAMENTALS OF app, ‘Uep Sy ‘substituting in (6:9), the model becomes * © See 9) 7) | eee tue oth a es treatment) error component) introduce the error compon, ff Squares. We in the deviation within the class which is due tot & nde both sides and summing over {and jy. 4 1 We got Poms ‘equal. This is tothe left, squaring (yg 2 Fe + Fe FP et s=1 PSY Gy 7.84 De HF. P42 1E (7, THY, a : } =0, since the algebraic sum of the deviations of the ra¢; ONS frog, Fr ws, “Ti SY (yy 7.2 = VE (yu -Fi + Zi (Ve-7. P ad #! : be Sp? = LE (yy -..? is known as total sum of squares (ss id ); ‘gy2 = SE (oy ~Jv 41s called within sum of squares or error sum of squares vi a sp). gf = 3a, (F.—F.%, is called SS. due totreatments (sgn +S. Total S.S. = S.S, Then, i ; i. Degrees of Freedom for Various S.8. Sr?, the total SS. whieh is computed hua 1 quantities of the form (yy - J. will carry (N ~ 1) degrees of freedom (d.f.), one 4 f being ig ae because of the linear constraint © & (9g = Fn) =0. ielj=l h Similarly the treatment sum of squares nj (¥;,~y, will have (k ~1)df sw izt 0 and the error S.S. S_2= DE (yi -¥;,.)° will have (N - k) df set J based on N quantities which are subject to k linear constraints : ny E (yy -F.) =055=1,2, ... yal Hence, we see that the d.f. for various S.S. are also additive, since N-1=(N-k)+(k-1) k57 sque o sam she corre eo res (M.8.8.) (Definition). ‘Phe sup sponding variance or the mean suum qy eeares tivided : SST. ” Of squares (My ¢ 2 PY its degre MSs) Theres 87, (Say) , is the M.s.g due to treay Sp", (say), is the Mg wwe) A sis the M.8.5, due to error 4 rious 8.8. For obtaining g an sons of Vari ning appropriate ty: gations or its equivalent hypothesis /1, : Plate test statistic for t, due to each of the independent factors, ©’ ~ %= 0, we need re nett rie SS 6 Pon 0 linear model (5:3), summing ov Bat ‘ er j and dividing ean st the : ; dividing by n, we get dis») = = (n ptm + ey} = Gees i 57 gusspover ‘and and dividing by 0 = Sanfomaee snail [vurEna+2 Ze) = F=H+E. [On using (5.20) ion of Treatment Sum of Squares (S.S.7.) k LX ni (¥--F.¥ 1 k 2 } =BS nj [@+aj+%.)-(u+8)] Using 63) and (59) in SEY nj(o; + €,-€.°= EE n; [oP + (%,-E.)" + 20, (8, -F.)] =LE[n,o,2+n;(€,-€.)° +2 {no F (e.- =} 0,4 =Yno2+BL nie, -€.) (EC e.-€ gj are iid. NO, 62) = ino?+ E[Ln;¢,?-Ne.2] =Lnja?+ DnB (e.)-NE CE Stetyare iid. (0, 0,2, B (Z;,) = 0 and E.)] = 0 Var (€.) = B(,2)—(H(¢,))" = E( 4,2) > BC E;)= Var (@.) =o." Since 4 " "eat neianee of the mean of a random sample of size n from population with Simi Su ity we have EB (€.2)=Var(.) = o2iN ting in ), we get Ss) g X na?+¥ n, (62m) -N (o2IN) =Lniai?+ Y o2—o,2= 3 nj 042+ 0.2 oe = Zao? + (k—1) a2 89)tty )=oe +p X no2 ==... =0%=0, MSs. due to reatmen, ts, | Expectation of Error Sum of Squares 68.z). Rat em BSSE) =E[E Loy -3,.° ef -E[ZE Wt a tesa ee, yr] ry a\ \ A sing =ELL (jz. = BI Beat ces [3 (Se, mes T ye iw SPY Gta) nea : ; an 7 ia ni (62/ nj) = = “a a E(S.S.E.) = (Nk) o2 2 FGSEin_y.,, i, the error mean sum of squares always gives an unbiased estimate of of o2 Thus we have a SST. alsa 7 (se Wo ae 02, under Hy, otherwise E(s?) > E(s,?) Hence, the test statistic for Hy is provided by the variance ratio pe a If Ho is true then F should take the value 1, otherwise it should be rater than unity order to find out if an observed value of F is Significantly greater than unity meine obtain the sampling distribution of the statistic F defined in (5-12). 2 Under Ho, by Cochran’s theorem, ae 2 and Se are independently distributed as}. Ge variates with (k — 1) and (N-k) df respectively. 7m: 2 tlle 5 le ‘ eal : lee ‘wb “oe follows Snedecor’s F (central) distribution with (k —1, N—) df A : plated at Thus, if an observed value of F (6-12) is greater than the tari tas F for (k- 1, N ~k) df. at significance, (usually 5% © rejected at that level, Ay. The above 5.2, kno analysis of presented in the TableTABLE, FOR. ONE-WAY. CLASSIFIED. Data lean Sum of Square torent Fe hough S,* and S,? add up toS.2 ay ——— above that although S/ ey add up toS,? and the fan squares (ie, variances), viz. 6? and 8? do not adds cree ond ite own aS ‘ANOVA, is in fact the analysis of the Sum of Squares, iP to total M. sh kno! ithmetic for calculation of v; ‘i thud! , the arithm a arious sum of squa tical Prot eiowing simplified formulae ‘quares is reduced to a using BS Gay J = DD 5 fee ig pave their ™ We sé N otal ; - RL > eine ma ; total of all the observations and N =n, ++... +7, -gthe eran isthe er ISS pif tes, the sum of the squares of all the observations is. sion Di» nown as Raw Sum of eres ihe expression (@?/N) is ealled the Correction Factor (C.F. as Raw Sum of 590s BSS. ‘Total S.S. = R.S.8.— C.F. (5-13a) ss. =D ED (yj-¥? =D [Dvj-F-" Error mie ahi ] A (2»,) 2B) Bxt- 7 [jet d HE Lyf 2 (7,21 n) (6-136) vviere T,, is the total yield from the units receiving the ith treatment. Traiment $8. = 7.8.8. — S.S,E, = i wea m}o N =X (7,2 /n)-Cr. (5-14) culate total sum of squares and treatment sum of squares by using within or error sum of squares is obtained by subtraction. ter than the MSS due to any factor of classification (s2) then s that the dispersion of sample means (¥, , Yo...» Ye) ber than expected ordinarily from a set of k-samples from the tween the classes is taken as insignificant and no further test there is no need to caleulate F and Hy of equality of means ®) should be calculated only if s; > s/. e variance ratio F which are significant at the level of (*). Double star (**) will represent highly significant valu : ing of e is independent of change of origin, the shifting o ince we ilso reduces arithmetic to a great extent. Finally, sinceiance ratio which is independent of change of gee ier eirganiainand scale facilitates caleulatiee. foe _ It may be noted that the change of origin will not change the value. ‘ change of scale will change the values of the sum of the seuan, Howey Pree) remains unaffected by the change of origin as wel as chan oe th __&: Critical Difference (C.D.). If the treatments show significa” interested to find out which pair(s) of treatments differ ‘ant nificantly ct Student’ ¢ for different pairs of treatment means, we calesien the Lo, F 2 this the ale also, thy 8 ing “at the given level of significance. This least significant difference is Tenon ifiane ED). Masthead ¥ The C.D. between any two treatment means, Say Yj, andy), at leve of sign; Significan, ~ ©D.(5,-F,)=[ The eritical value of at level of significance « and emoray | -4 (0/2), \MSS.B, (t + a ~+(@2), gy | (2 (M.S.S.E. stands for ‘Mean ‘Ss# provides an unbiased estimate of the error variance 6,2, In particular, if n;=nV i= 1,2,...,h, ie, C.D. (Fj, 94) = ty 4 (0/2). SpV 2/0 Here f,(c) is the right-tailed critical value of t for v d f. at level of si PLt> tla) =a Ifthe difference |, ~J, | between any tw: ienificance a 5 thy o treatment means is greate M Said to be significant, otherwise itis not significant Greater than CD (gp Zaeieet of Change of Origin and Scale. if we transform the original data (y variable U by the transformation Wy=Qy-AVh => F,eA¢hi, and 5 =Atha, Error mean square for the original data = Nd, = It [Error mean square for the transformed data) Hence, the t-statistic for testing the significance of the difference betw ¥) is given by : n any two means (L weal) ‘and it follows Student's ¢ distribution with error df ‘Thus, from the point of view of applying Student's est for testing the significance ofthe dis between any two treatment means, iti ah 1\" 3 slatJALYSIS UP VARIANCE AN! — the assumption of normality ig NOt critic, The A ctendard deviation, Fear can be a pegarding If more than two Populations are in afe side’ and one MAY compa, re population pope js on th ris on exror the equality of severa Population can el on. If the hypothesis tributio . ar} of equality ig auality is rejected, a valig test must be q afequality is rej e cel type of data ere might be certain ty oa cans tie observations ane — fons is constant. Howeve; Spaenteiteations ia beeen e differes Cochrs for Which the a, Suitably trangh * the meth, Scope of n, W.G. and Cox, GM. obse ve em as in that ease, their ee See tig is nesien only for testing of he > paatura and their estimated yg estimators ‘Ariances are 522, Variance of the Estima ites, Yy Ht +e: Ge he least square estimates oft and ¢ the least s i Are given by [See (5:5) and (6: r 1 Wee 65) and (54)a 512 FUNDAMENTALS gp A G all covariance terms vanish, since ey/s are uncorrelated bine 1 2,2 , BR By ey Ble eet. +e, J, at Product terms y, i 8 Vanish A [Var (ein) + Var (32) +... + Var (e,, 1 ee oe + = aN ("8% \=% Ly Var AAV ana) Substituting in ("*), we get Ve 2 je a a er reas ears nu Remark. If we assume that each level of the factor is used an equal number ‘ oftimes, i, 2.9% 9% o2(2_ 1) fy n, =n, (say) for alli = 1,2, ..., , then N= Hence, from (5:16) and (5-16a), we get n ie a ay 1 Var (i) = fe War (&,)=02 (I — |= nk ) oe 523 ANOVA for Random Effect Model for One-way Classified p, effect model (5:1), o's are supposed to be constants. In the random air suppose that there is only one source of variation, say, Factor A. w es vs», Ay have been selected at random from a large population of level The linear model (for Random Effects) is given by : yy =HHa+ey; G=1,2,...85) J=1,2,..,n) i ata. In som effect mod 08e F levels 1 et s. yA 4 (Here for simplicity we assume that all the classes are of equal size, gay (2 factor of A is repeated an equal number n of times.) + SAY MaKe, eat woe where itis General mean effect = 7. ly iz. a; is Additional effect due to the ith level of A, ( constant but is a random variable. 2s ---y), Which is nes and &, is the error effect due to chance. We assume that iid ia a, ~ N(O,o,2),i=1,2,...,k; and & ~ N(0,62);(6=1,2,..4 and as and ¢,’s are mutually independent. Var (yy) = Var (1 + 4; + €) = Var (a; + €3) ge of oii [-« Variance is independent of change oft! = Var (a;)+Var(e;) —[." a,, €, are independent] es = Oy? + 0, el 6a (60 9,2 and o,? are known as the components of the variance of ¥ and them known as the variance-component model. aeBe) = and Var are not statistically independent, sg aret er J Howe (x) (Sig) oy? =8 [by- BOD) by 26, }=2 [oye yw) Cov Oy» Fu [asters ey] Weing (5.17), = Ela) + Bla) E (¢,) + Ela) Flay) + Re, B ey) Wsing (5-18), ar (a,)=0,2 (sing (5-18), 19%) ton coefficient p between y,’s is given by + lg Ee correla ot . é rey tu: 8 (5:19) ana (5:19q)) Var (yy) Var (9) Sa +9, normally distributed and p #0, 5,’s are not independent, jsare since =n @,=0,(n, =n) is not valid in this case, since a, effect model is not uv in fixed sed in state a random variable. a '8 are random variables and We cannot ropriate hypothesis for testing the equ: ‘The appr ‘ality ofall the class means in this case would be : Hy: 0,? = 0 against H, :6,2>0, in the case of the fixed effect model, the total varia Asin -»- (520) tion ean be partitioned as follows ‘ Ae ola Fe-F.P+ WD ly, 8, 8 «(6.21) DAGyey? =u Ss H-5 te % 20 in rs ; > TSS. =SS.A.+SS.E, 6210) “P nxe Toalss.crss) = ZL (9-7? rm ‘ Ba Aner DA Ge-7.P (6216) : The (ap) ning of the corresponding degrees of frelon (fis ein 3 *) is given in Table 5.3 Since the h-1 tira Sum of squares and the corresponding SSA Cetra,” freedom are additive in nature, by ee theorem, we Conclude that the various sum lependent. TABLE 5:3 Sum of squares i544 FUNDAMENTALS op APPL D stm ‘Summing (5:22) over i from 1 to . and dividing by k, we get we, y efi, may + dn 4B iss wil where a =% 2a %) s ‘Ng Fe =(Qj- ap) + (Ge - E. ay i L We have SSA 22 (H-F-P=0¥ [G4 B)-Cgaz P Bray “g - rem 2a Let us take : U; =a; +. J “yy ees es z “6 U =; Dim, D (ait B= ayrk. 8) i ie 6 U;,-U =(a;+%,)-(ag+ €..? Ba) eee 24), we ket SSA. =n bs (U,-0¥ > BSSA)= ail Oy] og Since ~ NO, 0,2, = 1,2, ... a] ead ty" NO,oDVi,j > and as and @,’s are independent, ge a sicher MG v(o, 6.2 + ee t Lu k ie seeeeaca #[> U,- BF b)(o2+8) | 6G,’ i=1 * (a+) [<1 X~ 8 then BX) =r] | Substituting in (5:26), we get EAS.S.A.) =riGhealy (o + a => &£ ee =no,2+02 a EMSA)=o2+no2 0 Hence under Hy: 6,? = 0,°"BQLS.A.) 262 Sir WM ll ee 8 ide MSA, provides an unbiased estimate of 6 under Hy 2 Yo awe asi pe Sioa anes5-15, my he Pe oes 2 (e) / Pes [- B@?)=Vare Seton) ee 2 or EMSE)=02 (5, y pin Dae ay 56S E sp always provides an uunbitad oe us jased estimate of o2 529 G2, (5-29), We conclude that under Hp : o,? = : rid independent estimates of the Ge 0, both MSA (44) and a same variance 2, ba and Me (5-30) we fail to analysis can be expressed in the ANOVA Tabl NOVA Table 5:4. WAY CLASSIFIED DATA, Ww) ASSIF ATA-RANDOM EFFEC erento a pet Bfeans,5)Sie, giVariance Ratio, FOR ONE- ses (525) Tnportant Remarks: J "above diseussion, We OST, that the computational procedure for finding the vat cam of squares ing ANOVA table for the statistical Analysis of cee ae el is exactly similar hffect Model. However the conclusions ‘drawn from the ie sruhese models are ai fect model, the conclusions drain 47 valid if levels of Factor A ‘fect model the conclusions draw’ the given set of treatine (treatments mare val nts use’ we.are interested in estimating the variance components 0,” vides an estimate of 0. MS.E. always Pr (63)516 FUNDAMENTALS oF APPLE, Sy iy the analysis A[ie-(2e/5s) where ki For proof of (5-38), see § 525, below 3. If we reject Hy: then we do not proceed with 4. Sometimes, in random effect model, the ANOVA table provides a ne, eee ae eran (as) 0c (6882). Tn soho imation one EBEkAtive eee ea eiee gt es apsdeccs thatthe true value of the variance comme fa that sampling variation ed to the negative estimate S24, ANOVA for Random Effect Model for One-way Classi, fea ‘Unequal Classes. Here the linear model is Ms) yy = N+ a+ Ey fi é where k a | i= General mean effect = > nit dao n,) £ int in is the additional effect due to ith level of A 38 the error effect due to chance. As before: ) 1 4G, ~ NO, 02),i=1, 2-15 Een, iid fy = N00); i= 1,2,.0h; ‘The partitioning of the various sums of squat is as follows [See : Fix ed Effect iy nd a k & » y= L md Fi--¥ Pedy Oy-7,F = T.S.S,=S.S.A. + SSE. k where SSA. =Sy2= ¥ nj(¥;.-7..) and S.S.E. ra ‘Summing (534) w.rt. j from 1 to n; and dividing by n,, we get Summing (5-34) w.r-t. i and j and dividing by N= > n,, we get ‘ ¥ 1 = i. inidcaree Einai+ = + 00+ -. wher ae $ ag = Y nja,/ ¥ nj Substituting from (5-38) and (5-39) in (5-37), we get547 ni( PN ~ a9) Ep] 8 and ¢.» 8 ate ing pend ent] Ne, | 2 POR Biea same , t in avenaa l ha ies: Ts =Lno2-n. Lp aga a aA [ato shecovarance terms are zero since a/s are independent he : 1 =.) Dn- 5. Yn? Ba2)cNor 2 A N.veat Beene Os, ons ( i.) 1 oe if dg = |y- ved nal oz cat 64 Als . & i AY n,22-Ne. >| = njE;2)-NE(z.2) rz lier is. 63 A 62 o2 = x (n, se), N= Roto? =(k-)o2 i549) Substituting ftom (5-41) and (5-42) in (5-40), we get (539 Da E(S.S.A.) = (v= 62 +(k-lo2 (543) on PSSA IR-) =hoPto? = BMSA)=10,4+02 ee (608 where ‘ 2 nf a «(6388 ey Hence, jased estimate o2 inder Ho: 0,2 = 0, E(M.S.A.)= 6,2 = M.S.A. provides an unbiased ‘der Hy, Sa ee thFUNDAMENTALS OF APPLE. 8) 5.18 “és Ly =k {SE Ow - | [2x0 | A aie (SSE) #[3z vy an wy Using (5. en 6 36) ang A =(N-h)9?, (n= Xn) ba in : [Proceed as in Fixed Effect Model, equation (5.49) o = BSSEIN-b) =02 3 EMSE)=00 "Pig 3° tee M.S.E. always provides an unbiased estimate of 6,2. - Equation (545) = 6,2=MS.E. ; Prom, (6:44) on usitig 6-46), we get b 7 62- MSA -MS.p NOP +62=MSA, => 16,2 + M.S.E. =M.S.A, oF Ou" = zee. ‘a & 4 £ a/t ng |v-(2 n/n) =a [ n-(3n2/X n)| ¥ = t fel is iz (5.49, | fen a stated in (5:33a). le 5-1. The following Table shows the lives (in hours) of four batches Of ele -—________ life of Bulbs in Ho eee Life of Bulbs ‘1,600 1,610 1,650 1,680 | Sou 1580 1,640 1,640 ‘1,700 = 1460 © +1,550 +=—*1,600 1,620 i 4,510 1,520 -1,530__1,570 pet Perform an analysis of variance of these data and show that a signifi \% ject their homogeneity. | wit Solution. Let y,; be the jth observation in the ith batch ; i = 1, 2, 3, 4 | Set up the hypotheses : ot es ee ueteeries : Hy: th = ta = Ha = Ht, ie., the four batches of electri lang eae Hypothesis: H,: At least two means are different. sign ifting the origin to 1,640 and then dividing b , : data is expressed in the Table 5:5. B by 10, (cf. Remark 5, § 25:1), the given tob ee TABLE 5.5: LIFE OF LAMPS (IN HOURS) typ fe Decoded Data : u,, = (y;, ~ T T exp No. eh Uy = (Yj - 1,640)/10 T= Sey | Dae re ESS ae ewer | | |S 2 1-6 0 ‘40 =| 28 398 | a 3 a ee | Ne 718 -9 -4 9 = 11 193 | | eres 18) 2100 a ancy ye -3 | 853 \A Ae ~43 515. is Total’ | @=DYu,=-7 Ehu! as ry oy{eRVARIANGE 3 ow RSS, = “M sd jon Factor (CF) - @ correct =e Total S.s, 54 26 9 =RSS._ op petwoen (batches) sum of squares — y ft (23% 1 | Gar eet Vii Biiddysg, 3| Ty be the : NA tt oy shen => Then); ro I Ig ny Al s? =— Y yy2- 52 oe vik Iso Piganpt > ht =i (62452 a which gives the S.S. of observations for the ith strain of typhoid 4 i ie! nit TABLE 5.6 : CALCULATIONS FOR VARIOUS SUM OF SQUARES Ma | Strains of typhoid A B Cc cae a ORS A.) BSR AG LA_ orl n 10 6 8 ves o | Ji 10-9 13-5 15 12 1s | in | Ti=ny; 109 81 92 123-2 "1 Hi if yi M881 18225 13225 125-44 997,16 Ree | 32 16180 35:52 10'50. 3192 13.95 | by nj (s? + 932) 2,806-1 1,20662 1142 1,730-96 1,259, G? (482-2)? : OR = 77 =p = 581292 RSS. = LY yi? =¥ nj(52 +92) = 8237-73 } ij i TSS. = RSS. - CF. = 8,237-73 — 5,812.99 = 2.4948) r S.S. due to strains of typhoid oP digas OF, | (109? | (81)? (92)? | (193. 2 Bg tg ha ee OBL ge = 1,188:1 + 1,093:5 + 1,058 + 1,379.84 + 1,185-8 5812-9 Error S.S. = 7.5.5, ~$,8. due to strain = 2,424.81 - 92.3 249 TABLE 5:7 : ANOVA TABLE Mean § T T Sources of Variation df. | Sum of Squares | Between Strains of Typhoid | 4 92-32 Error 35 2,332-49 Total 39 2,424.81ariaNnce S21 Enel, we gor 4.and 35.af = 5-735. Since calcul al Fos ‘not significant at 5% level of scniniee value of F is less th, and the mull hypothesis Sn a ] valves pec Ott. ai ald Mad s eg A | Muckine | Meio 1 a purchased three new| | An gia Maria e we makes and | +31 * oof alter as 2 a | pines gotermine minether one Of) ae | de “the others in| Observations| 36 2 Pte | SB ho 28 | | output. Five- | % a res are observed | fists 3 2 25 0 Prom ack machine and the (ee ee 2 pen in gable 5°8. Use analysis of variance technique and determine ope OE significa” 58 ont in their mean speeds, Use a= 5% phil Be Bo i ines ng abou date vier shifting the origin to 30. How are the results in Part (a) a aoe ” Here, the factor of variation is machines (Aj, Ao, A3). We set up the ution oe ees is | Ho: oa = Ho= He ig 1 Hy: Atleast two oft a= 0-05 ie., all the machines are equally effective che means are not equal. k=3; N=nmytntng=15 it is given © m= 72>" TABLE 5:9: CALCULATIONS FOR VARIOUS S.S. ‘Sample Observations (i) [i eae" Total 36 38 31 T,=160 | T,?=25,600 42 | 35 Ty, =185 | 72? = 34,225 28 74=135 _ | Ty?= 18,225 > 1,2 = 78,050 Total | G = Ly = 480 ied (ie eee! bas» YD yi? = 25? + 302 +... #25? + 28°= 15,810 Zi Raw S.S. (R.S.S.) = 160 + 185 + 135 = 480 4g0)?_ 230400 _ ae Bie 215860 5,810 - 15,360 450| Treatment (Machines) 250 200 I Error i (Total 0 | + I ue of F for (2, 12) d-f. and = 0-05 level of > The critical (tabulated) value of 0 ; Fi Since the computed value of test statistic is greater than the critical value Ste i . ree een ve reject Ho at 5% level of significance aa conclude with 95% conf “ah noo ESR Eas ee em Nias sia Ris ‘Since Ho of equality of means 16 rejected, we proceed further to find out wh “thy the vntnent means differ significantly. For this, we find out Critical Din t'*h pi, roe re significant Difference (L-S-D.) between any yo means to be gi a S.B. of the difference between any two tr ae F. andy, is even, Dy Sa 1 aire r vat q FF) pl z(t +2) = (bet) - oo tt S.B. (Fey) soe (b+ 5) sys +5) = ee Vi. - 9, 1° LSD. = typ (0/2) SB. Gi ~3)) th = typ (0-025) x 2:58 = 2-18 x 2:58 = 5:62 COMPARISON OF TREATMENT MEANS TABLE 5:11: | Treatment (Machine) | Mean hourly output Difference __| Comparison with Ls) ~~ uPl mein = 59) | gaol AL Fang <5:62) | Not, iene | esierion 5/85 yp amis 2° Ay je, = oT < 5-62 (Not Signi “| dassific 5 ‘enifcant) || gerva .c¥ie5 pistons? | As Fu =e 227 [1 ¥o-Fa.|=10) > 562 (Signiticans | observa a a alt _— t Since the difference | ¥2:—Js. | is significant, we conclude that the machines Ay and, E differ significantly as regards their hourly output. The remaining differences arent | fy significant. il ‘Conclusions. (i) Treatments (machines) Ay, Az and Ay are not equally effective fat th (ii) Ay and Ag differ significantly. Remaining differences are not significant. eck, TABLE 5:12: CALCULATIONS FOR S.S—DECODED DATA pie “a oe T (®) We will now srachines Sample Observations wart ‘2 compute the various (uy = yy — 30) a com of squares anal ay [=o] of 6] @ [a [t= wolne | age complete the = 35) T)2=128 table after shifting|. 43 er | tage [TH ~ the origin to 30. As -6| 0|-2|-5 | -2 Let ujj=94 - 30 Total | G= +5 Correction Factor (Cp) _ G 30! Total S S.(T.8.g) Treatments (machines) 8.8. (8,2 Remark. If we had changed for the transform ned data. Howevo, changed the scale also, thes r, the varia n the various sum of sqi mee ra Mares would have also ‘atio and the concluss ‘on will not be affected, nS are classified into piasses), say A,, Ay, Ay + and into h cate B, 'h According to some criterion B, having kh combinations 4, Bdizto | Re; A; often called cells This scheme of classification according to two For example, the yield of milk may be affected by differences in seen 48 Well as the differences in variety, i.e., breed and stock of the eas ieee * 7 cows are divided into h different groups or classes fipriinata ue Stock, each group containing # cows and then let us canes thee fect ions given at random to cows in each group) on the yield of milk. Yield of milk from the cow of th breed eee A 3 = [ or stock, fed on the ration i ie, rations f to the breed i eee ce olen act Ne i treatments (rations) and the s following kx h two- iN ands oe ates Per expressed as variable ec ld Mast of cows) is iG ieties (breed a s (rations) is ariet i 's (ration rl a item ented al ng the columns an “resented along the rows of the table.5.24 Treatments tions) Column Means = * Nga wipe i Wi ; \/e | poo SN ri is ~ h | 5-3-1. ANOVA for Fixed Effect Model. If we assume Rae “i both the factors the levels used are the only ones of interest, soe Above ai parametric model, is used > then the f° Factor A ; Treatments (Rations) ef tor B ; Variety (Breed and Stock) of cow Fac In the above illustration, the fixed effect: model is : 5? yyehitey > By) = Hy GHZ k5 F=1,2, where y,, are independent N (jj, 04%) and € are i id. N(0,02) ¥ ; ui is further split into the following parts : dius Rx Say ; e (i) The general mean effect w given by : H= > i/N. eo io (ii) The effect 0%, (i = 1, 2, --»» #) due to the ith ration gi 7 & given by: =, ) 1s A 6 a Wie = x ys @2,2)..,m) a) th ke Obviously salto (iit) Th j ia (a i) The effect B,, (j = 1, 2, h) due to the jth variety (breed of cow) given b 73 soni cx a Bp = yb, where py =~ D yy G=1,2,.,h...680 é i Obviously, ns si x =o (Glad (iv) The i ; : ies Raa effect yy when the ith level of first factor (rations) and th ew! reed of cow) occur simultaneously and is given by : : vial Sees : y=OVi=1,2,..,k and Ly=0 Vi=12-+ysis OF VARIANCE 5-25 we have qhus Hy HEA ie I) + (phy = W) 4 (Hy ~ Hy = Hy +H) he model (5:16) becomes, Yy SHG +B HH) + ey jg tho error effect due to chance and k x w= iat consequently t id yore h k 2 ZB Lwe0vs ; Lrja0vi j=l only one observation in each cell, the observation corresponding to the ith there is Aeration and jth level of breed of cow is only one, i.e., 9. But we cannot estimate by one , 1 fno, Hence, in this ¢ase (one observation per cell), the interaction effect 1, = O-and the a 5-18) reduces to Dy = H+ Oy + By + ey (5-54) 1 (6-54), where a, B, and jt are fixed (unknown) constants and, ¢, and x, are ‘the model (5°5 E bles is known as the fixed effect model for two-way classified data with one dom varia mt pvation per cell. eeatistical Analysis of the Fixed Effect Model (5-54). Let us write: ‘ 5, Mean yield of the th treatment ration) = py age Les Q, 5) = Mean yield of the jth variety Loypeg o fewer -- (G-55a) i ih 1 core st. Ss aes 7 ‘The overall mean = 75 EE yy BUH % z3, (5-556) Least Square Estimates of Parameters. The least square estimates of the “ters i, , and p, are obtained on minimizing the error sum of squares : fhe a Be Si Yd ef LT Oy-1=o5- BY je aah - 23 (yj-H- 0% -B,) i =22 (yj--95-B;) Since Da; = 0 = f,, we get from the above equations : Aer REY — “7. 5 B=_ D9y-h=Is-7 (5:56) ia 1 A ; “ike Ey 18-72 ay- he Thus the linear model (5-19) becomes Mb i ae ar WAI + (Fi. =F.) + Gj-F.) + (yj -Fv Fj +) .. (5-56a) “ror term ¢, being so chosen that both sides are equal.——— ys - vn of Squares, Transposini J. t0 the left ¢ w Sto A and jfrom 2 to h, we get Side ” parti -7.) + (95-7) summing both ytd + (Vi j ° 7 paw? = gp + LL Yond: P+ DDG. Pe i SS (yy de PIP ca fs 5 -Fjt¥-422E (FF J aC Lae 4-7.) mee! 7 Fy 5 is vi : 2a LOH? yogererdit )=d [ (5-7 Ly s5, Now par ye 216 ane ) ov )-L.G5-F.}]=0, sh deviations bea set of observations about their mean ig hia 4 since anes easily seen that other pet tarmg,elao yattish 2 i Similarly it ¢ o e k y x a : jah UII Pek Uy jy) + Xv Fin h Lov “J ihe 4 si? =SP+ sp+5e or _7)2is the Total S.S. ee . = LE Ov yy A Ss? SG _7.,)2is S.S. due to treatments, Gy =e G57.) athe SS: due to varieties, B j eal Se = EE (vy y+) is the error or residual 8.8. set up the null hypotheses that the treatment In other words, the null hypotheses for treatments anf, Null Hypotheses. w 4 and vari varieties are homogeneous are respectively : Hog? ae = Hoe = SMe = BG Blog ? Ma = Be = = He = si or and (5-146), their equivalents 7 Hy y= 09= «= 04 = 05 Hoy : Bi = Bo= ... = By = 0 Alternative Hypotheses H,, ; At least two of the j.'s are different ; Hy, : At least two of the sm different. or their equivalents : 5 Hy, ; At least one of the o,’s is not zero ; H), : At least one of Bs is not zero Degrees of Freedom for Various 8.8. The total S.S., Sp” being computed from! quantities (y,; - J.) which are subject to one linear constraint LY (y; - J..) = 0 wile ij (N-1) df Similarly, $2 will have (k -1) d.f,, since (¥, —J-.) = 0 and Sy? will have(t-! 1 b-nee Vil fy since SF. J.) = 0 and sg will carry (N 1) —(k ~ 1)—(h -1)=(h (N= hk).VANCE of VAR 5.27 partitioning of 7. in as follows \e sous Deth-Dth= Ds moyy ( nat the df. are additive a ies th with imp? atistic. 10 order to obtain appropriate test statistic to test the ste wed the expectations of the various mean §.8 : hypotheses H, a due to each of independent factors te onal of one-way classified data, we get s2 Mean S.S. due to treatments = k of [74 (say) ‘ os Sy Mean S.S. due to varieties = eG = 5y/, (say) SZ GD LD = 8 Gay) ing (5-54) over j from 1 to h and dividing by h, we get prror Mean S.S. i (5-60) Ly yng lin hav + E B+ Bey] > Fuanrase (2 £p=0) | similarly summing (5-54) over i from 1 to h and dividing by & and using ¥ a, = 0, we get . Fj = Ut Bye --- (5-60a) Summing (5-54) over i andj both and dividing by hk and using (5-53a), we shall get J. syste. Expectations of Various Sum of Squares: k wisp =elh ¥ ,-7.8]=2 [AXisq+e.-n-e] inh i [From (5-60) and (5-606) =a[a¥ las (e.- =hLo?+he[L &,. P| a) + 2hD a EE, -%..) AYot+hY (%.-€.)2+2hD a G-z =hYo?+hE (DG2—Ke.2) — [ E(G.-F.)=0ase, areiid MO,02)] AL of + h(E BCG h ECE] =h Doe [Y (9)=4.$¢] ko? ..9, h ae +N aes 4 (ME 298) ; y ae nce Pa), ~. 61)FUNDAMEN TA 5.28 a simitarty we can prove that psy ak Spee DOP a pis = sa at). 2 rey ssxpeetation oferror um of square I (Ss) [EX Gu Prady th ") sp wubstituting trom (5:54) and (660) to (5:606) and simplifying, wy i nse) B/D Ee Be By? BS | a5 [EE ey Gite GAPE FA 268, 268+ 2 e484 95 eT [Sdq2+h Der +hD Ej? + hhe2- 2h G,2 I F ; 2 ~ Oh DEP + 2hke “52M ay ~ SAAR AL = EE Bley" 4nd BG.2)+8 3, EG?) + hk ( ~ 2h & 2 X PG,2) - D~ aS = oF +h 5 (,! Mm +k (6,%/k) + hk (o,2/hk) ~ 2h 2 (2). se SFr he hae ee A) = hh o,2+ ho + ho,? + 0,2 2ho,2— 2ho,? VM Lid =(hk-k-h+ Do? ay E(Sz9) = (h- (k= D em =o8, te, Besiees Thus, we see that Error M.8.S. always provides unbiased estimate of ow a s.2 are unbiased estimates of 9.7 under the null hypotheses H,, and H,, as ete ni respectively. en in Bsa By Since 2os2 One we get ; ql Els?) = Hop under H,,, otherwise, E(s,2) > E(sg2) te and E(sy*) = E (sy), under H,y, otherwise, E(sy*) > E(sy2) SU Since various S.S. as well as their respective d.f. are additive and since under the ni hypotheses H,, and H,,, each of sy’, 8,2, sy? and s,* provides an unbiased estimate ofa under the assumption of normality of parent population, we get by Cochran's theorem S2 Se Se are mutually independent x2-variates with (k — 1), (h — 1) and (h— 1) (k- 1) Af. respective Hence under h,, and H,y respectively, we get5.29 5 © go Conforms to 7, a-»,g as Wiha) pln hypotheses of the homogeneity iVed.f and at certain e meted or accepted. ‘very elegantly presented in the ANOVA Tabip 5.14 . ANOVA TABLE FOR TWO-WAY CLassiF} (ONE OBSERVATION PER CELL) c Mean Sum of -] | Squares "| ED DATA af [@-)xlo3, — merical computation of various sums of squares, the following formulae are used 2 RSS.-CF. =3Zy-£ [aa For : am of squares (TSS.) or Sp= i) otal i vA tment sum of squares (SPiey 2 TACK ‘Trea “a Yarietos sum of squares (Sy*) = ae TP -CF. B onSin i Ter Ty ms ond OF = GIN, sam sfequares is then caleulated by subtracting the total of treatment sum of squares end Dror sum Xjuares from the total sum of squares, ie, Sy'=S,'~S*—S\ ie is denote the factor of variation represented along the k-rows by A’ and the factor of sslosropresented along the f-columns by B. The ANOVA table for two-way classification with one ‘erat per cell is as follows: TABLE 5-14 : ANOVA TABLE FOR TWO-WAY CLASSIFIED DATA as] es. af MSS. | Variance Ratio (F) | ae BeRSA 5 {Ryne SSA| (k-1) 4. Figk—),(k- 0h D1) h-¥ MSSB._ pi _ 1), (k-1)(h-DI ? SSB. | Fy=yaag -M0-D, y | e-1)(h—y)| SSE. _ 2 a= Daa MSS | hk -1 oous for the rows (ins of means differ signiticantiy™ ae 5.) whiny Bor i 7 Factor A) means h is given this. ~ critical value of ¢for error df. and ) The *arignificance «for two-tailed test) * ta 4 sans); (4/4, 2,--+ Al, then the ith | 5 cD (row means) anes c Figl Saiy, otherwise not in ¥ tilt Ges fr any te column (FAC B means o/2 3B.) a exp. comp moans) =f 24022) e 3); @ #7 = 1, 2, «.., A), th ora be |e -Fy | > OD. column mean 2, tr Se said to di ignificantly, otherwis ve sae igs oe f Mee. Wartegce of the Ketimated, Parame or We have obtaj Rey fe i Ut 16-50) ofthe parameters in the model (5-54) oe, 8 estimates I : s 60 : its eene +b=54-5 4 synere Fi, Fy and Fare given by lef. equations 5-60) to (5.695) By ie Fi apt opt B 7 Yy=B +B Ey es 4 ft ant and A AeA od Var @) -£[h-e)] =E[7.-E9..)] =E(5--w C B(Y.)= 1 (From (9) = 02k) £E(€..2) = Var ( 2 Var (G) =6[4-E@)]" 2E[Gi--F-)- EG: -F.] HB [(a,+ 8-8.) Blas +B. -F | a on 2 iid 2k [(a;+%.-8.)- 4] =B(§,.-8 few [a2+8.2-23,2.] = Var ( @;.) + Var (€..) ~2E (). €..)OF VARIANCE 5-31 ting in (** ty = 1. I Var (a aie «(i ia) 9? a, (kal var @) = (3?) 02 oe we can prove that Var (B) — | for Two-way Classified Data with one r (Factors) of variation, say, A and B, A being at »p and B being at q levels B,,j = 1, 2, ..,q ; which have boon eulected at from a large population of levels of A and B respectively. Jy denote the observation under ith level of A and jth level of B. Then the linear ical model (for Random Effects) in this case will be - ‘ anaes ca --- G70) aa twa where a; ~ NO,0,9;; *"N(0, 0,2) } tue 18, cy ~ NO, 0952); ey ~ N (0, 6,2) - GID and a;s, 8s, ¢y's and ¢,'s are mutually uncorrelated, and iis the general mean effect, a, is the additional effect due to ith level of A, 6; is the additional effect due to jth level of B, £, is the random error effect, and cy is the interaction effect. Remarks 1. As against the fixed effect model in which ais, b's and ¢;’8 are fixed constants, here 4s and ¢,’s are random variables following normal distribution as gives in (6-71) 2. From (5-70), on using (5-71), we get. Ey) =u and Var (yj) = 6,7 + 2 + 0,52 + 0,2 V iand j (6-72) 4. In the above random effect model (5-70), we need not assume that the interaction effects are all we cannot test for or estimate 0,2. Null Hypothesis. The appropriate hypothesis for testing the equality of all the class due to different levels of A and B are: ® Hoa: 0,2 = 0 against Hy, :0,2>0 } ti) Hog: 03% =O against Hyp : 042 >0 [plitting of the Total Variation and Degrees of Freedom. As in the fixed effect ‘enn’ ‘tal variation in the model (5-70) is partitioned as follows (in the usual ++ (5-73) Zoe 2 2 2-78 = SF [yates ed Ge-7.)+Gy-F1° im fi, 00-7 2 Flow Fe-Fu+7. + (Fn H+ (Fy -F-o-vag-) (S.8.E.) orresponding sum of a Square 8 a g of dest + q-)* (SSB) re additive, the c since the deste independent (CoeBraD® ‘Theorem. Expectation of Various gum of Squares: Summing (5-70) over j fi aividing by 4 We BE J from 1 4, ji eptajt bot Oo & 4 ay gamming (6-70) over irom 119 and dividing by p, we get (5 peat bjt eyt by 58g) if Here, by =D 6 and xd , ae (ay Summing (6-70) over i and, a | and dividing by pq, we get ay Wehave =U +p + b+ ©. +E. ‘ 2 SSA. =9 2 (¥,- (68 Let us take ms 5 U; = a40,.4%, Then Ther Then $9.4 Using 583) bee sl Sint Sinsstocnisetanne ttt wks Pa 1 gatas AMIR, sineg = ji Ont?” * pg Pot Similarly, we shall get TT BCG.) = py (0ab?. P) = pg * Fab tuting these values in (5-88), we get 1 Eley. - Fy +. Peon [142 ‘Subs +pqZ.2 pL E+ 2g 2.24 295.2 > 2p. 2 ‘Taking expectation of both sides, we get 3/33. @ageee zy] 7 = DY Bey?) -¢ LE G;2)—p> Bl @.?) + pq E(Z.2) ej i 4 Saban le 9 2 Var (0)—p 2 Var @,)) + pq. Var(é.) [Bey 2 2 @ 2 i = ZY (2)- (ss) = Oa? Gap?) poo, tia 77. a2 q py (% +Pq = ) [ey '*n@oe zl = PY Gap? — P Oo5?~ 90a5? + Oa5? = (Pq —P - 9 + 1) Gas? =—-D@-Dox?aniance = a ye amet, WE CAN proy, an Od e that Se i o2 (1-2 ya > 3” ao} A Pa l Sa si and & are uncorr ° GP car i related, %o2) per é: e, : precy Be7 BIE) ey ~ os 5-89), (5-90) a = won fo a "E59 in G.g7) ot 53 [out (1 We Bet t Piss BT 5 = ~P~ q*5q)+ dy] j 56 ; Pa) * 23 loa (ya | 2008? PI-I—P+V + 62(p4_ \ cE ee pry Mall = Ou" +02 or ROMS py rea = TOO tee, Sa paaitusine im (5-85), we get sige MSA. = (62+ 042) +9.5 . Out?) + 4.8.8 = MSE 4 gh ” hy _MSA.-MS at ) * eo q ginilarl¥> from (5°86), we get ~ MS.B. =MS.E.+p62. iP P na (594) is Hy: 0,= 02>. = voder nal ate a - sae % > Gu'=0,both MSA, ma Msp . si yulatiol . OT ios estimates 01 pop’ m variance. The test. statistic for teting fast ne ~ Fy-i.g-na@-» jis ejected at ‘of level of significance if P,> F = D,=Vq-9 0s herwise, we may accept Hy. Under the null hypothesis : Hy : by = toh M.S.B. and M.S.E. provide independent estimates of the same population variance The test statistic for testing Hg is : M.S.B. Fa =MSE. ~F_y@-v@ a) e accepted, so i herwise Hy may ‘. is re} ‘of level of significance, ke B> Fs ayqKs then He is rejected a ‘ tistical analy i anova Table 538 Rp galistical lysis can be displayed in the following | — |537 sy tale 8 16 gives quality rating of then ve TABLE 5:16 ; 6 7 85 OS =__10 5 70 88 5 : eee 518 gt ene ht 82 ee 4 86 § 0 80 2 9 7 sphether there is any significant difference betu ween raters or vice Stations vs of variation are : (1) Raters and (2) Ser # ations PO i sil oP service station by the ith rater; i =A, B,C,D, EB, j 10] . ; eo! seo ne there is 20 significant difference between the raters “tyere is no significance difference between the service Ley coey Hp FE different, ie. at least two of the raters differ are different, Ze, at least two ofthe service ey Het 624,100 | 5.62500 | "72,400 | 9 105,625 1.22.5 RSS, = LD? = 9" * (96)? + (95)? + * (39? i = 3,20,000 ‘Total $8, = ees — carn 3,29,948 8:20.00 99°FUNDAMEN, FUNDAMENTALS Soy 5:38 3,20,000 = » of squares = 4,009 Row sum of squares = Column sum _ 7.8.8. ~ Row S.S. — Columns S.s, ~ or sum of Squares = SS. = 9.048 _ 4 5:18: ANOVA TABLE Oe 3 MSS. | Mo Sources of Variation — MSs ay a Eat VR a | 756-67 - ) Between Columns | Tes a (Service Stations) | | | peey 1 ao0f 6=1=4 10000 808 * Bags Between Rows | | peels . (Raters) | i 7 Ted 13 , 5 ~~ Error [2,73 36 | 7606 | a1.) 1 [ 1=49 5 “ed Foo for (9,36) df. is 215 and since the calculated value og value, it is significant and 17 6 of F eater than the tabulated value, it lope i alee yd be ore nt we conclude that there is significant difference ot reiected at 8 ‘Again tabulated Foos for (4, 96) df, is 263 and since the calcu “4 it Jess than it, it is not significant and we fail to reject Hox at 5% level of sig a colo” Jess than Jade that there is no significant difference between the raters, TABLE 5.19 7 ‘ ve different methods of | az Example 55. Three different met tho analysis. My My, Mg are used to determine in M, parts per million the amount of a certain 7 aa at i nstituent in the sample. Each method is used 5 a by five analysts, and the results are given iN) Anglyst —” 7 ; i Table 5-19. 73 fon * Ss (a) Do these results indicate a significant é * a $ variation either between the methods or : = 68 between the analysts ? " 74 a 69 (®) Decode the data by changing the origin and scale as follows : uy = 10 (y, -6 s Perform the ANOVA for the transformed data. How do the results compare with hae obtained in Part (a) for the original data # (c) If there is significant difference between the levels of any of the factor(s) of variate differ significant 5 then determine the corresponding pairs of sample means which (d) Also obtain an estimate of the error variance 6,* Solution. Here the two factors of variation are, say , A: Analysts, represented along the rows of the Table B: Methods, represented along the columns of the table.5-39 cpp ies, there is no si signific ‘ant difference by tween, the _ there is no significant difference j betwe en the mi vethod: s of pss pneses* iC oth | Us. are different. 1s are different. oof Hae ws ie fH ‘ wos oe vethe i fanalyst and the jth method i= 1,2 k=5,h=3 and N=hx 19% 8 on 7 gh 9 52 Ape © we have 0. CALCULATIONS FOR VARIOUS SS. Method M My 1” VW 216 y 2 67 va 466.56 70 69 21-2 69 72 68 216 —_ am ts 457.96 344 Geol | ST 2= 2.29421 1183-36 | LT? = 3,827-45 5) + (70% +... + (71 Raw-wise S.S. = 155-66 + 151-49 + 149-90 + 155-64 = 765-67 a yy = 107-1 G =Grand Total = 2. g@ aoriy | 11,470-41 _ Correction Factor = yy = “15 15 7164-694 Total S.S. =F: sea ag: ae vo tacor A analyes)® = 22043) _ 764-604 = 164 764-694 = 0-976 T,2-CF 764-694 = 0-043 eet eli eee SSB.=S.S. due to factor B (Methods) = tET 2-CF 764-694-079 as _ 7164-694 = 76549 - ' p.-0976 6 -0088- 0.796 = 0337 a Re cays _ssa-SS5-40 TABLE 5:21 : Al Sources of variation | af @ (2) Factor B (Methods) | A — Error |4x2=8 lor | 14-(4+2)=8 Total N- Tabulated Fos (2, 8) = 19-40. Since the calculated value Fy < 1, it is not significant and there is no significant difference between the analy; we Since the calculated value of Fy = 23-27 is greator th, 15-1=14 significant. Hence the hypothesis Hop of the homogenity of thy °°, “®bulatey . oo level of significance Thus, we conclude that the methods differ grr t?°4® is ul No significance. SeiBeantly aust ty | yi (b) Analysis for De-coded Data : leva path uy = 10 (65) = 10y,, — 65 es TABLE 5-22: CALCULATIONS FOR S.S. DECODED DATA U, os | Methods "Tote | Anaiyee 7 | il 5 | 2 9 7 3 8 5 | 4 | ili 1 |, ana 9 6 | Total.) 47 30 Ti? 2,209 900 Raw S.S. (RSS) = (10? + 5? + 6) +... + (92 + 6? + 42) = 161+ 134 + 105 + 179 + 133 (Row-wise 5-5) = 712 ca Ge 96 = == 614. mea GF. NB = 6144 Total S.S. = R.S.S.~C.F. = 712—614-4 = 97-65-41 1,856 — 614-4 = 618-67 ~ 614.40 = 4.97 8 LATO _ 6 5 — 614-4 = 694-0 — 614-4 = 79.6 —S.S.B. = 97-6 - 4.27 — 796 = 13-73 ae __VR®) (4) (3) +2) 1.07 a 39-80 1-72 4 | = both the origin and scale in the given observ vations, the if.) are different for the decoded data as ee ge have cb » 'g.S.B. and S.S. | * vy oe ss *jata, However, the variance ratios Fs and Fz are same in each k or aetasions about the tests of hypotheses are same as in Part (a). 1 Baan the Oe significant difference between the different levels of Factor B pein Te ented 0 find out evnigh pairs of column (Methods) means differ are i 'L.S.D. between any two column means, we find the C.D. or the original data. sa aes FO erg: ‘ sndaris ty)” aera ,, (0-025) « { = 2-306 x 0-083 = 0-1914 OLUMN (METHOD) MEANS __ TABLE 5:24: COMPARISON OF BLE 6-24: CO ee FetorB (Method) | Column (Method) | Difference between | Compart —_ Means eae | : > cD | | : | >cD" \itted | >CD | 1 __ irs of column (methods) “cs the grate difference difference between all the three pai 2. and finally between the | being between Methods 1 and 3, followed een the Methods 2 and 3. py the Methods 15-42 Remark. Wwe had proc testing the significance of t same conclusion, (See Re ate of the error #8 Provided by the « M.S.S.B. = 0.0171 [ANOVA Table 5:27 fo, Ifwe use the ANOVA Table 5:23 for the decoded data °riginal g ano > Vary) » ( 10-* (MSSE for decoded data) = 10-2 2172 2 ry 5. Analysis of Two-way Classified Data with mOngg “Tre Effect Model. In the preceding section, we have considered sr 85 This does not enable us to obtain an estimate of or make test 2° obser, between two factors A and B, (say) , under study. For this we nest, the ine for some or all of the cells. For simplicity, let us consider f a of A and B classifications. Let 7; denote the kth (k = classification and jth B-classification. ‘Then, the linear by Ation, re th bServations Ae 0s 22 — model tion one ho et for earn Obs Mey +72) Obserye 8h con ”) obseryat; a the i $59, St Visg ig i jo a8 discussed int Jie = Hay + Sup =p + (Hy, H+ By) + Hy Hie ty +) Hig, HUF +B ryt en © Dalaran Py a me 1Y'D, 1) ghion Ty where u is the general mean effect, a i the effect of ith A-clasifcatg, 2” effect of jth B-classification, ie., B, and yy is the interaction effect Pi between ith 47,4" Bis and jth B-classification, i.e., A,B, and ej, is the error effect due n ith A-classina (Sy to chan ficas, Lid. NO, 6,2). Here 0, Bj and, are completely enumerated and fixed marie aten 2 2 2 x a= 8 2 =O forall ; ¥ y= 0 for anti a i=l iz v2 Ba) Let us write . = Sum of all the observations = DEY. yy, i Total of all the observations in the ith A-cl; B-classification = yijz lassification and jth t Veg = Lun 5 Ime = Lie 7 7 - = Total for all the observations in the ith A classification = Jy = Total of all the observations in the jth B classification = D3. yse yy Jaye ‘ ; Yim ‘Wa yr j yxam ji yon Jon ees 9p [gene ve0 yu : : r sgh a Jim “) pea e : fe 2? vin eg aia yia2 Yue * Se : 2 ¥ See BE] gy iam ‘im vt ya P : aig | : ‘ a th | pas 2028 nj Biv, ae eS It pam a+ yee ae wae os we 599) i ETat-s th in 7 suistcal ising the residual sum of squares, viz., are | ost B= DEE Oia-H- 0-8-4? J : the solution of the normal equations : %) 5 Oe PELE Oi — Ho BW) re 225 Oia GiB Wy) ih 233 Qu -b-o1- Bn) 3, 23 Oi -H-a)-B)—%5) lqution (531) on using (5-30a) gives LLLVyie=mpg.u = Mo SE Oy - ua, — 8 — essed in the following Table 5.25 VATIONS FOR TWO-WAY CLA: m OBSERVATIONS PER CELL Nig Naga Yam *2qi Yaga Y2qm Yin Yiga iam Yoa Yonge Youn on SSIFIED Data Total | | I alysis of Model (5-3). The least square estimates for various o ffects, (5-99) (5-100) [ppeez(e+)--]aha m Geen ae fiom equations (5:32) to (5.39) ep and % : ual yiudes are ea! : and summing both sides over where fi sn rrransposine svi SEY (yy -F where ee mq AY y gs, due to factor B =P LG AB ue to the interaction 4 Fin Yop tI) and 5.8. due to Error = DLS Oia Nall Hypothesis. As in At Alyy by = Bas = = Hp = B Hog ‘pa = He =Uy =H Hogg oy = 0 Vi andy ie, A and B are independent of each other. Alternative Hypotheses Hig : At least two of the y1;.’s are different, 1. Hyg: At least two of the 1,'s are different, i.e., at least one of the P Hip :A and B are not independent 53, the null hypotheses at least one of the 0,5NCE ARIA! of jreedom for Vay, Biietaum of san ious g fi 58 Sun, iat Biss -Dx(q~1ay wt en f3 ionine FLY. is a8 fol “" sp pg - = P-1) + (q Dag Day soctat the df. are additiy, q~1) the Yethare te fas Tate usual nog = Tato tg . 509 ‘ave 0 sy St the ioe due to factor 4. S4' Of g n SiS. due to factor B Meat eS? are D> 8x2 tations of Various Sur, F Squares Bape gov on using (5:98) Tweirg, Summing (59g ng Ver j an Jor =N +0, 43, "a and din, similarly, we have Jy =N+B+3, and 51039) (SZ) = B [ma ¥ 6 oi Elna (az A 5 B[may lo? +(@,.5 Fr 204% -z.) ist term vanishes, ES) = mq & a2+ mg BLY: ~pe.,] = mq 02+ mq [2 ae 2.)-pe@.)| = mq¥a2+mq| ¥ (%) mad +P 5:08) ES) = E ba (6104ye shall get yet 546 . ; ma g2+(q- Do proceosin ee oa ek pe OF HUSp?) = 0.8 + 14252 ny | / TAI F + ¥ 2 gs Axe =™ BEG F c Bj +E...) =m | put me om? mr int i am oe spe last term is ero, since Cit the last Now LE (8. -& z i ADD Gye bbn +P Ley! + pg e.-2qy 79 2 2 pre j +pqe LL Eg DG? ‘Taking expectations of both sides, we getoF poe ce =PU mn qm ~ B.g~ bm * m 547 D—q+1)Se pq-P-q m = (p Dig 92 we have m in ( nt a5 2 DE, BAD ay pore a ai Sa i Jo, 02 + —_m °-DG-Hu 4 Lee aN: tation of error means, Meal? expect have tee ; at wor k from 1 tom and dividing bY my, we es pe Vik Vij =e e # , q (m ~ 1) 6, B|_s, Waim=1]=92, ie, ing ‘hus, we see that error M.S S. always provides 8S. 848 65 Sas? provides an unbiase 8.844 Sp Sar om 102) to (1030) reane tively jeain( 5.107 sti biased estimate o¢¢ SStimate of 62 under A While each Hy an d Hy, SS. as well as their df are Ge various 5 BONN etpeitaly coche Mes Itusle ofc? and #5? provides smityo parent popul additive and since ‘under th th aN S.8., 62, 42,5 Provides an unbiased an unbiased estimate of ¢ Under the assumption of ation, by Cochran's theorem, we i ave Seo. Se a2 ~%e~vig ca Pee mie say et and SE wemutually independent. Hine, under Hy,, the statistic Thier Hp, the statistic F wdunder Ha, the statistic:JAMENTALS tabi eh the and Ppajoct ray retain fo8 a aes Sin is =p—-D@-D ee pam =D ial OETA. = 1h S8.-CR » oer ¢ fuowin iy by 7 ac: ' Cty Sh-Sie dependent, then we ca: Guy InNOt Proceed sd 2; SBH ie - i rand B are not ind ter inreie aly i Han i Hee 4 pat te a test ae 6:26 show the birth-weights of babies born ‘i e analysis the anal 6, The se a if oe oe oferavida, there being three i a according #0 ieee fmt 16526: BIATHWE WEIGHTS (in Ib.) OF BABIES BORN _ Age's group of mother 0 | 30— if 2 / 51, 50, p48 1 C8, 48,60 | 2 | 52,53, 54 | 53, 53,55 _ 5-2, 5-2, 5-0, 5:5 | 541, 3 58, 57,59 | 60,59, os 158 5-8, 5-5, 55 | 59, 54,55 4 60, 60,59 | 62,65, 60 | 6-0, 6-1, 60 | 6-0, 5-8, 55 | 58,5655 Sand over | 60, 60,60 | 60, 61,63 | 5:9, 60,58 | 59, 6-0, 55 | 5°5,60,62 Test whether the age of mother and order of gravida significantly affect the birth-weigtt5.49 gunor, and () The order of gravida do.not ign iignificantly GALCULATIONS FOR VARIOUS 8.8 other Row Ze gvour of As | 36 and over | Totals (7 7 | 15-0 : 5,655.04 | 154 6,198.69 168 lett 7,516.89 69 7,909-21 rca 7,956.64 ; as | (xr | | =a187_| \ Yeas Be ome te St) 88 eared | Oy? = 95,0007 a tole “p27, are the totals ofthree observations in each ell in the mm =3,p=5,q=5,N=mpq=75 + (6-5)? + (6-0)? + (6-2)? = 2,351:19 + (58 + = 2,337-46 = 2,361-19 — 2,887-46 = 13 73 A ; zr? 35,225.47 GBigeter ga ABST 46 = 1020 3507777 _ 9 97. 6 spmother’ age = se “ma — 2,337-46 = 1-06 Qe (14-9) + (15-4)? + LL? at ae ae + (17-4)? + (17-7)? = 7049-33 S.S. due to order of §.8. due to mother’s age - 50 Po | | ss ae 56 gravida 55 3 52 | oc 33 _ 9 37-46 | — 10-90 - 1-06 = 0:36 =1-41 §S. due to error = 13-73 — 10-90 - 1-06 - 0-36_ TABLE 5:28: ANOVA TABLE 5-50 Source of at, Juana M.S.S Varia varton lay a TLORTO| ne D | es patil ‘ 10:90 2725 | : jor of are¥i cs 4 1.06 0-265 | Mother's a&° ‘ , | 16 0:36 0-023 | ete! nd Ly G | 141 0-0282, pesidual ror pts s Pe | Total | ee —— ve table, clude that observed F above table, we CO” at ob lor intel asi ‘or homogeneity of classes is rejected both for ( action ig the nYjida, Hence, (2) 28 size of the mothers, and (b) the orden nother és ee weight of babies ‘significantly. a of gent Ste 1 peg un Classified Data with Random Effect Model. Consider two factors of variations, s, ym Obsen gi=1, ,p)and B being at q levels (Bj; = 1, 2, .. eA and B, tion, e jst I & HP) ot eandom from large number of levels A 2 Suppo ting Which ane’ that are s itt th ) any “ty d i e (i, > Jt cal have been select are > 9). Let ux denote t mathematical model ( Jie =H a + bj +i + Fuk where a (21 (0, 009) 5 4 NO, oF ey = NCO, Oas?) 0k 4 NO, 02) vs and £,1's are mutually uncorrelated a's, bjs, ey’8 and &/ he Ath observation (& = 1, 2, for Random Effects) will be : G=1,2, ds and and iris the general mean effect, «is the additional effect due to th level of A 4, isthe addtional effect due to jth level of B cj is the interaction effect g; is the random error effect The appropriate hypotheses for testing the equality of means for diffe rent levels of A ang and 6? = 0 y Bare: Hog 10, = a= Hog 3 b= or => can Hoan : Os? =0 The corresponding alternative hypotheses are : Hyg :0,2>0, Hig:o;2>0 and Hyap : O44” > 0.HE (Fy. 5 (AxB) =m EE Ss IBS) SE. = XEE (yaa5, 2 S.S.E. ay esponding partition, corr he BED aia TSS.) (8.3.4); INE Of the " pam=1 =(p | ) is ag follow, +1 lef, SiGe D spat a Ss ) SS axay Sse) chran's Teeee™ Bre, Additive, the “Donn Mm of g “Sug sat Bates of Theorem) Wee ap See ts, ‘arious Sum F Squares « as a = ee asian of Ve : . et 5. ORE 5 “ Py Box’ and k ang dividing by ae Summing (5:10 : 511g) (S1Me) (Glnad) (From 12) nq EB |S, ey Cin + Ei.) —(ag4 2.4% is ((a;+@,. +5, (5-115) (5-116) he +B48 Ui =a; 46; - Let us take es a . > _ .. Pit ima5:52 uj eat + Ej.) — (ay + ©. +E...) jn (6-115), we Bet eae [2-0] 110), from (5116), we get Substituting in ( E(S.S:A) Using assumption in 1 1 yin (0. O42 +4 Sab? + ng o2) 2, , as 2 a ae |e > F i=1| 04? +9 Se tng % ‘ ES (u;-D"| = -v[ot oe wy ald Wy; a] ts 7 24] Substituting in (5-117), we get 24, Sab® 0? H(S.S-A.) =ma.@-1) [044+ att 22) | A Jo o2-+m oui+ma 94% or E(M.S.A.) = 0,2 Fm oy? vt mag. Similarly, we can prove that wy Gar” | O* | i = “ 24 Sab E(S.S.B.) = mp (q vos ‘ier si 52 +m Gai? +mp os? => E(M.S.B.)=0,2 + mg, Further ) +6 )+(&.- =me [EE eee y+.) ] ome [EE %, i the expectation of the product term is zero because ¢y's and €,,’s are uncorrelated Proceeding exactly as in the case of Random effect model for two-way classified data wt one obsérvation per cell, we shall get Heth 2 y+.) ]=@a-p-a+ Dout=-» q- Vou?5-53 y| =(pq-p-4+)) m = P- 14 %, q-0) _yiq-Do2= -DG-v [ -1) [o2 +m g _ 1) oat? +P 24m Ob = E(MS(AxB)] =02 +m 72 + m Gus? E Oot (5-121) [EEE (ore) ]-min-nee of Squares for Fixed Effect Model for two-way cl vay classified (6-122) ies (MSE) =62 i" 3 Bl pa aways provides an unbiased estimate of a] yea sque a ae Gat? 79 - Oa ‘ Ths ees een Oe ‘nd from (6-121) and 12 te pol ie provid ‘upiased and ‘ndependent estimate of ‘the same population get OB) PO S$ Hogs is eiven bY 1) (5-123) fe unbiased and both provid Hence the test "7 ‘variance (0. + ™ oa (6-124) nificance, otherwise we ™4Y Y 1007 = . Under Hop (042 =» MS. gs. (AxB) poth provide unbiased and | Gi ates of the 68° + opulation 24 moat) Hence, the test statistic for testing Hos is given by* ib ie oP MSE or (5128) Fe MS. (AxB) Fyaenet ato level of significanc® otherwise we ma E> Bagg 040 (a), we reject Hos ‘eoept it.5.54 displayed in the ANOVA 7, jstical analys able 6, a ANOVA TABLE FOR RANDOM EFFECT M TABLE 5:29 WITH_m-OBSERVATIONS eS & a BEAVA Sources of | df | SS. Variation | tnt SSA. The above sta ea PURE q-1 SSB. Interaction | (p-1)(q-1) S.S. (Ax B) effect. | Error SSE. Total TSS. The following points are worth mentioning : 1. If Haggis rejected, i. e, if the factors A are B are indepe: analysis. If Hag is not reject lependent, then we ¢, ted, only then we test for H, or H, n 2, From (5-122), we get EIMS.E.)=62 > 62=MS.E. From (5121), weget: EIMS.(AxB)=0,7+moy? = 62 MS. = MSE. +m6qs2 =MS.(AXB) or 6,52 = From (5:118), we get HIMSA) =62+moa2+mgo,2 => (824m (624m S42) + mg a2 :. ot = Ms M.S. (AxB)+mq6,? = MSA [Using (5-127), 9a)’ = mq 6,2 =MS.A.-M.S(AxB) = ara Similarly, from (5-199), we get Sy bd informatio E(MS.B.) =62+mo,2+ mp oy? => mp o,?=MS.B.-Ms her SBMS. (AxR Sal gz = MSB=MS.(A xB) sm) 11.0 3. In may be noted th: a Snod. oa fe e not that ile i : used for calculating each i — pied sie Jl the sume error variance SE: | ari cactati emcee eae IMMA eter falas mean ny i, MSE. for computing the in ti fel we use two the test statistics for both Ho, and Hog. eA BAG ME sabe om ant 13.¥