Bill Afantenou, second year Statistics graduate student at Queensland University of
Technology conducted an experiment to fnd out how, by varying whether he ordered thic or thin crust, whether !oe was ordered with the pi""a and whether garlic bread was ordered with the pi""a, a#ect the time it too for a pi""a to be delivered to the front door of his house$ Being a poor graduate student and limitation of time, he decided to have only two replicates, %ust to get a reasonable estimate of the variance$ &e also tried to repeat the experiment in as nearly as possible identical conditions to reduce 'noise($ &e ordered the pi""a from the same shop, namely )omino*s +i""a$ To be consistent he ordered a Supreme pi""a each time at approximately the same time of day$ The response was measured from the time he closed the telephone to the time the pi""a was delivered to the front door of his house$ Bill wrote each of the eight treatments on a piece of paper twice, put them all into a hat, mixed them up, and too them out one at a time to allocate the order in which each treatment was done$ The three ,ualitative independent variables !rust -Thin./, Thic.01, !oe -2o./, 3es.01, and Bread -4arlic bread5 2o./, 3es.01 and the response variable )elivery time, are in minutes Analy"e the data and summari"e your fndings6 / Applied Statistics Assignment 7 888 SOLUTION Identifcation of Variables To start the analysis, the frst step is to come up with the listing of independent and dependent variables $The following table shows the variables involved in the given problem5 9ariable 2ame Type of 9ariable 2ature of 9ariable :evels of the 9ariable !rust 8ndependent Qualitative Thic !rust Thin !rust !oe ;rdered <ith +i""a 2ot ;rdered with +i""a 4arlic Bread ;rdered <ith +i""a 2ot ;rdered <ith +i""a )elivery time )ependent Quantitative 2$A$ There are total = possible combinations from the above three factors with two repetitions for each treatment$ 0 Applied Statistics Assignment 7 888 Visualization of Data Scatter +lot Box +lot Obserations fro! Bo" #lot and Scatter Plot$ ;rder with thic crust and coe without garlic bread taes longer than all other treatments The treatment with thin crust and coe along with garlic bread taes the least delivery time All remaining treatment di#er very slightly from one another Model for %nal&sis 2ext, we shall analy"e whether the mentioned and other remaining treatments di#er signifcantly and how the each factor a#ects delivery time$ Since the design of the experiment corresponds to completely randomi"ed design, the analysis can be started by taing the model as below5 Y ijkl = + i + j + k + () ij + () ki + () jk +() ijk + ikj ; where 3i% . ;bserved value, > . ;verall mean, ?i . @ain e#ect of crust A% . @ain e#ect of coe k = @ain e#ect of bread ()ij = B#ect of interaction between crust and coe ()ki = B#ect of interaction between bread and coe ()jk = B#ect of interaction between coe and bread ()ijk = B#ect of interaction between crust, coe and bread ijkl . Cesiduals, which is i$i$d$ D2-/,E F 1 %nal&sis of Variance '%NOV%($ F Applied Statistics Assignment 7 888 Girst it is to be checed whether the source of variation in the data set is due to Treatment or due to chance alone5 Analysis of 9ariance Table Source )egree ;f Greedom Sum S,uares @ean S,uare G 7 9alue +7 9alue Treatment H I=$JK= L$JJ00 I$IHIK /$/FM=L Cesiduals = 0F$M 0$MLFM Inference$ Since the p value is only /$/FM=L , it is not due to chance that we observe a variation in the data set $ Thus treatment has an e#ect on the delivery time of the pi""a$ 2ext the source of variation in the dataset for di#erent factors are analy"ed using the model Y ijkl = + i + j + k + () ij + () ki + () jk + () ijk + ikj and the following results were obtained5 Analysis of 9ariance Table Source )egree ;f Greedom Sum S,uares @ean S,uare G 7 9alue +7 9alue )rust 0 0=$/LFM 0=$/LFM 00$ML *+**,- . !oe 0 /$MLFM /$MLFM /$KL /$MLM00 Bread 0 0=$/LFM 0=$/LFM 00$ML *+**,- . )rust$)o/e 0 0/$MLFM 0/$MLFM L$HL *+*-0. 1 !rust5Bread 0 /$/LFM /$/LFM /$/I /$=ILIH !oe5Bread 0 0$MLFM 0$MLFM 0$// /$KILMJ !rust5!oe5Bread 0 /$/LFM /$/LFM /$/I /$=ILIH Cesidual = 0F$M/// 0$MLFM Inference$ @ain e#ect of !rust is signifcant @ain e#ect of Bread is signifcant @ain e#ect of coe is insignifcant however two way interaction of !rust with !oe is signifcant Two way interactions involving bread are also insignifcant K Applied Statistics Assignment 7 888 Three way interaction is insignifcant Thus, the model assumed has to be refned so that the factors with insignifcant e#ects can be neglected$ Thus the refned model becomes5 Y ijkl = + i + j + k + () ij
+ ikj ; Note$ %lt2ou32 t2e !ain e4ect of co/e is insi3nifcant but due to 5ierarc2& Princi#le 6e cannot eli!inate it fro! t2e !odel+ %nal&sis of Variance '%NOV%( for Refned Model $ The source of variation in the dataset analy"ed using the refned model is as follows5 Analysis of 9ariance Table Source )egree ;f Greedom Sum S,uares @ean S,uare G 7 9alue +7 9alue )rust 0 0=$/LFM 0=$/LFM 0I$//II *+**-17- !oe 0 /$MLFM /$MLFM /$IKL0 /$MFFM== Bread 0 0=$/LFM 0=$/LFM 0I$//II *+**-17- )rust$)o/e 0 0/$MLFM 0/$MLFM =$0=JI *+*078., Cesidual 00 0I$0=HM 0$F=J= Inference$ @ain e#ect of !rust is signifcant -as before1 @ain e#ect of Bread is signifcant -as before1 )learl&9 t2ere is a t6o 6a& interaction bet6een )rust and )o/e and an& conclusion on t2e e4ect of co/e : bread on delier& ti!e can onl& be studied b& anal&zin3 t2eir interaction #lots Validation of Underl&in3 %ssu!#tions I Applied Statistics Assignment 7 888 The following are the assumptions 5 2ormality of the residuals &omoscedasticity of the residuals with respect to each factor Residuals Vs ;itted alue and Nor!alized < = < Plot for Residuals$ ;or!al Test for Nor!alit&$ @ethod + N 9alue Shapiro7<il normality test /$KL/HL0K Anderson7)arling normality test /$IH0FIL= !ramer7von @ises normality test /$M/HJLK/ :illiefors -Oolmogorov7Smirnov1 normality test /$L/MKJ0L Shapiro7Grancia normality test /$MM//IH0 Inference$ M Applied Statistics Assignment 7 888 The residuals 9s ftted values are randomly distributed without any pattern, thus proving model is indeed correct All the tests as well as the Q7Q plot of standardi"ed residuals shows that they are indeed normally distributed Test ;or 5o!oscedasticit&$ &omoscedasticity of residuals is checed using Bartlett*s test$ The below are the results of the same on each factor5 ;or )rust$ BartlettPs O7s,uared . /$=F/J, df . 0, p7value . /$KLIJ ;or )o/e$ BartlettPs O7s,uared . /$FFHL, df . 0, p7value . /$LKKK ;or Bread$ BartlettPs O7s,uared . /$F/L0, df . 0, p7value . /$LIJJ Thus it can be inferred from a high p value that its by chance only that we are seeing a variation in deviation in the dataset and our null hypothesis of homoscedasticity of the residuals cannot be re%ected$ Multi#le )o!#arison $ 2ow, the e#ect of di#erent levels of the factors on delivery time are studied by multiple comparison using Tuey*s &S) method for the complete model$ Tu/e& !ulti#le co!#arisons of !eans ,7> fa!il&?6ise confdence leel ;it$ ao'Delier& Ti!e @ )rust A )o/e B Bread( Sl no Comparison Diferenc e Lower Upper P-adj 0 T2ic/ )rust = T2in )rust 1+017 *+CD70, *C -+-D8C*, *+**-17-7 F <ith !oe N <ithout !oe 7/$KHM 7 0$LFI=/J /$=HI=/JF /$MFFM==K - Eit2 Farlic Bread = Eit2out Farlic Bread ?1+017 ? -+-D8C* , ? *+CD70,* C *+**-17-7 I Thic !rust <ithout !oe N Thin !rust <ithout !oe /$M/ 7 0$J0L=00 J F$J0L=00= = /$JFFLH/I M Thin !rust <ith !oe N Thin !rust <ithout !oe 7F$// 7 I$I0L=00 J /$I0L=00= = /$00LLKJI L Thic !rust <ith !oe N Thin !rust 0$HM 7 /$LLL=00 I$0LL=00= = /$0==HJL/ L Applied Statistics Assignment 7 888 <ithout !oe J D T2in )rust Eit2 )o/e = T2ic/ )rust Eit2out )o/e ?1+7* ? 8+,0.C0 0, ? *+*C-0CC 01 *+*81*D7- = Thic !rust <ith !oe N Thic !rust <ithout !oe 0$FM 7 0$0LL=00 J K$LLL=00= = /$IKJLJFK , T2ic/ )rust Eit2 )o/e = T2in )rust Eit2 )o/e -+D7 0+---0C C0 .+0..C00 CC *+**-187- Inference$ The following combinations have signifcant e#ect on the delivery time of the pi""a T2in crust ta/es less delier& ti!e t2an t2ic/ crust 'een 6it2 co/e( T2in crust 6it2 co/e ta/es less ti!e t2ic/ crust 6it2out co/e Eit2 3arlic bread delier& ti!e is !ore t2an 6it2out 3arlic bread 'No interaction 6it2 co/e and crust( Interaction Plots$ Interaction #lot bet6een )rust and )o/e$ Grom the A2;9A table we can clearly see that the two factors N coe and crust interact, hence we plot their interaction$ H Applied Statistics Assignment 7 888 )onclusions$ o Statistical eidence su33est t2at t2in crust #izzas 62en ordered 6it2 co/e do 2ae a si3nifcantl& lo6er delier& ti!e co!#ared to t2ic/ crust ones ordered 6it2 or 6it2out co/e 'refer Si No D G , of Tu/e&5SD co!#arison table aboe( E4ect of Farlic Bread$ As mentioned before, garlic bread has statistically signifcant e#ect on delivery time$ But to understand its e#ects, lets plot K way interaction diagram5 = Applied Statistics Assignment 7 888 &ere, we can interpret the Tuey*s &S) result of e#ect of garlic bread by comparing means when it is ordered as against when it is not ordered N both are statistically di#erent$ The same can also be observed by a simple box plot5 )onclusions$ o If 3arlic bread is ordered t2en t2e delier& ti!e is reduced+ ;inal Su!!ar&: Grom Bill Afantenou*s experimental data, statistically, we can have following inferences5 The delivery time gets reduced on ordering garlic bread as compared to the case when it is not ordered$ The delivery time gets reduced on ordering thin crust pi""a as compared to thic crust one$ The e#ect of coe can*t be stated independently because of its anti N synergistic interaction with the type of crust$ Between the eight treatment groups, the order comprising of t2in crust 6it2 co/e and 3arlic bread taes signifcantly lesser time then t2ic/ crust 6it2 co/e and no 3arlic bread, the exact intuition as observed from the box plot initially$ J Applied Statistics Assignment 7 888