AMSADesign of Comparative Experiments ‘This book should bo on ths shelf of oveny practising statistician who dosigns experi ments. ‘Good design considers units and treatments fint, and thet dllocsies (reaiments 10 Units. Itdoes not choose (rom a menu of named desiges. Tus approach requires a ‘polation fur units that does not depend on the treatineats applied, Most stucture-oa the setaf observational units, oran the sel of treatments, can be defined by factors. This ttook develops acoherent framework for thiaking ehcut fastors ard theirzetationships, including the wse of Hasse diagrams. Thess ere wed to ouside structure, edloulate cegrees of freedcm and allocate treatment subspaces tn appropriate stata, seal an ‘one-term course the author has taught since 1989, tne book is ideal for advanced tndergraduate and, beginning graduate courses, Examples, exeicties ancl discussion questions are drawai from a wide range of real applications: from drug development, to agricaltare, to manufacturing. A, A. Bati-EY ban been Professor of Statistics at Queen Mary, University of London since 1994. She is a fellow of the Institute of Mathemaltcal Statistics and a past Provident of he Intarnational Biometric Society, ritish Region. This hook refleets hee extensive experience Wuching design of experiincats and advising on its application, Her book Association Scenes was published by Cambridge University Press in 2004,‘CAMBRIDGE SERLES IN STATISTICAL AND PROSARILISTIC MATHEMATICS storie Board R. Gill (Department of Mathematics, Utrecht Univessiy) A.D. Ripky (Department of Statistics, Usivonity ef Oxon) '5.Ross (Departament of Uidusiiel end Systzins Engineering, Universty of Southern Cafifonsia) 1B.W. Sitwerman Gt, Pete's College, Oxford) M. Stein (Department of Statistos, Unigersity of Chisago} ‘This series of high-qualiy uppcr-division textbooks end expasiiory menugrapis covers afl aspeeis of tochestc applicable mathematics. Tre tapi ange fram pure and appied satities (0 probability theory operations rexeasel, optizain, and rutbennaticet programming, The hacks contain clear presentations of new developments in the Feld and alsa of the sete of tear incastical methods. While emphasizing rigorous vesimert of theoretical net, the tocks also contain apalisaions and ciscussions of new techuiques rade possibleby advances {n computational peel, Mecady published! octsinap Methods end Thete Applicuton, by A. C, Davison acd D. V-Hinkiy Markov Chains; by 1.Noccs Asyuiptote Statstice,by A. WE sa dor Vaart I. Mavefer Methoddsfor Pine Series Anatyss,by Donald 6. Percivaland Andrew. Welden Fayesian Metiods, by Thomas Leanard od loti, J. Ha Ennpirieat Processes in iiiaction, by Sara van de Csr Nuvacricul Methods of Statvtcs by Solu E, fonalan A Uses's Guida 0 Measure Theoretic Probatitiy, by David Polkird 9. The Lstimarion nd Tracking of Frequency, by W., Quien ond G3. Flannan 10. Data Anebysisand Graphics esivg &, ry John Madpoald and Tah Biv 11, Stausticat Medets, by A.C. Davison 12. Semiparcimente Regression, by D. Ruppert, M.P: Wand, RJ, Caroll 19, Braves in Probability by Loie Chaurnont end Mans Yor |, Staisical analysts of Siochaste Processes in Tune, by }, K. Lindhey 13. Measure Theory ae Filtering, by Lathidar Aegoun and Robect EU 16. Essentate of Sestie! inferanor, by Ga, Yoong and R.L, Saith (7, Perteas of Disrtsion Theory, bf Thoms A, Sever 18, Statstival Mechaniss of Dizondered Syste, by Anon Bevier 20. Ronda Grape Dynartess ny Rick Wacrst 21, Newark, by Perec Wale 22. Sadétenoint Approximations with Applications, by Rensld W. Ftier BR. Applied Aepuptotes, by AR, Deszzala, A.C Davison and W. Reid 24. Random Heworks for Cometwrication, by Massiaia Frarecscheti and Ronaké MecsterDesign of Comparative Experiments R. A. Bailey Queen Mary, University of London w CAMBRIDGE UMIVERSITY PRESS.CAMONIDOE UNIVERSITY races ‘Canbbiidge, Now York, Melboume, Madi, Cape Tow, Singapore, ‘So Prulo, Delhi, Dubzi, Foley, Mexico City ‘Cambridge Univesiny Prcer “The Edisboigh Building, Cambridge co Sau, UK Prise in the United Stars of Americn by Cambiidge University Press, New York snmnicambridge.org Toformauion on dis tle: wow cambetdgesorgly P0528 65057 RA. Biiky 2008 “this pabicaion isin copyright. Subject ro scasuory exception ang 10 he pidtisions of rlevanc collective icnsing agreement, ‘ne reproducton of any part ray teke place withort the wcten permission of Cambie University Pree ‘Hea published 2008 A catalogue record for thie pbleaton epee from the Betis Libvory ISIN 978-o-yar 6506-7 Hardback ISBN 978-0-521-68597-9 Paperback Cambridge Unireuity Pres hat no responshitty fr she persistence of scouregy oF URLs for external or third-party internet websites efron ‘dis publican, anil doesnot gearantez that any content on such webshes I, ‘or will emai, zccurite or appropriate, Lafermaion regurcing pricés.travel timetables. and other Factual informetion given in shis work iecoreect 28 the tne offs printing but Cembridge Universicy ress doce nor gutraince the accuney af sch infornacion theater S a 10007020638@a ans FERPUSTAKARS SULTAN AncuL sinu UNIVERSIN PUTRA sasLAvsia Contents Pretace paged 1 Forward look 1 LE Stages’ « saaically designed experimen 1 LS.L” Consolation, 1 132° Seatities! design 2 1.13 Data collection 2 Uta Daa seruty 3 115 ialgais ‘ LLG Tnterpretati 5 1.2 The ideal and the reality 5 1.2.1 Purpose of the experiment 3 122 Repliation 5 123 Local eonirl 6 124 Constraints 6 125 Choke 1 13 Anexainple 7 1A Defining terms ' Ls Linear meet 14 US Summary Is ‘Questions for discussion 16 2 Unstructured experiments 19 21. Complely zandomizcd designs 9 2.2 Why and how ( randomize 20 23 “Thc treatment subspace 21 24° Orthogonal projection 23 25 Linger model a 26 Estimation a 2.7 * Compasison with matrix notation 2% 28 Sumsorsquares 26 WB 02.11 Allowing forthe overall mega 2.12. Hypothesis testing 2.13 Sullicient plication for power 2.14 Amone general model Questicns for diseuesion ‘Simmple treatment structare 3. Replication of control treatments 3.2, Comparing new treatments in the presence of a contol 33. Other treatment groupings ‘Questions for discussion. Blocking 4. Typosof biock 4.11 Natucal discrete divisions 4.12 Continuous gradients 4.13 Choice of blocking for tial management 4.1.4 How and when to black 42... Orthogonal block designs 43° Construction and randomization 44 Models for block designé 45, Analysis when blocks have fixed effects 46 Analysis when blooks have random effects 4.7 Why use blocks? 48 Loss of pawer with blocking Questions for discussion Factoriat treatment strxeture Sal Trertmont factors and theie subspeces 52 Interaction 33 Principles of expectation mosols 54 Decomposing the treatment subspiee 35 Analysis 56 Taree trcatmentsfogtors 5.7 Factorial experiments 58 Construction and randomization of ficiorial designs 5.9 Factorial tenuments plus control ‘Questions for discussion Rov-cohnnn designs 6.1 Double Hosking 62 Latin squares, 63. Conshntetion and randomization G4 Ontogonal subspaces 6.5 Fixed row and column effects: model and analysis Coments 30 3 35 38 au“ 3 BR 44 "7 2 8 3 3 35 0 56 51 By 59 6 7 B 0 " 8 6 7 ia ca 80 2 9 98 9 99 15 105 195 108 to 0 Cont!lenis 0 38 35 8 4l a 43 44 a7 m2 5 53 35 3 56 7 59 59 61 or 8 50 n 8 75 7 4 "7 0 2 7 oe 99 9 105, 105, 06 rr 10. 110 Comente 68 Random row end coluuaneteets: mmodet and analysis, Questions for discussion. 7 Experiments on people and anitnats 7.1 Introduction 72 Historical comrols 4 Crnss-over rials 7.4 Maiched pairs, mawched threes, and so on. 5 Completely randomized designs 7.6 Body par as oxperimental units 7.7 Sequential allocation to an unknown number of paticnts 7.8 Safeguards against bias 79. ubical issues 7.10 Analysis by imeation to eat Questions for discussion 8 Sivall unitsinsidelarge aults 3.1 Experimental units bigger than abservarional wile 3.11 The context 8.1.2 Construction and randomization 8.13 Madct and strata 8.14 Analysis 8.15 Hypottesis esting 8.16 Decreasing variance 8.2 ‘Treatmont (actors in difforent strate 8.3 Split-plot designs + 831 Blocking che large unite 83.2 Conisiruciion and randomzation. 533 Model and strata 8.34 — Analysis 5 Evaluation 8A The split-plat princi Questions for discussion plo 9. More about Latin squares S.1 Usesat Latin squares SLL One treatment faeter ina square 9.12 More general row-colume designs 9.13 Two teaument factors ina block design 9.14 Three tasotinont factors in an unblocked design 9.2 Gracco-Latin squares 93 Uset of Grasca-Latin squares 9.3.1 Supestmpased design in a square 83.2 ‘Wo treatment fiotors in a square 9.33 Three Weatment factors ina block design 2 16 7 17 aR 1g us 120 120 121 m 124 196 19 131 131 131 132 132 132 135 137 138 146 a6 lay lag a9 (52 152 156 187 157 187 158 159 161 162 166 166 166 (669.34 Four treatment factors in an unblocked design Questions fordiveussion 10 ‘The caloutes of factors tol 102 103 104 los 106 107 108 103. 10.10 10.11 10.12 Introduetion Relations on factors 10.2.1 Factors and their classes 10.22 Altasing 10.23 One factor finer thin arather 10.24 Twospecial factors ‘Operations on factors 103.1 The infimum of two factors 10.3.2 The suprenwm of wu factors 1033. Uniform taetors Hesse diagrams, Suilspaces defined by factors 105.1 One cubspace por factor 10.5.2 Filted values and cnide sums of squares 10.5.3. Relations between subspaces Gnhogonal factors 106.1. Definition of erthogonatity 16.6.2 Projection mattices commute 1016.3 Proportional meeting 1064 How replication caw eftec! orthogonality 186.5 Achein of factors Orthogonal decomposition 107.1 A second subspace for each factor 10.7.2 Mifects and sums.f squares Calculatious on the Hasse diagram 19.3.1 Degrees of freedom 108.2 Sumsof squares Ontiogens! weninent stsuctures 10.9.1 Conditions en treatment Feetors (0.9.2 Collections of expectation mndels Orthogonal plet structures 10,101. Conditions on plot factors 10,102. Varidnee and covariance 19,203 Maveix formutstion 19.104 Sires Randomization rthogonat desigas 10.121 Desirable properties (0.122. General definition 19,123 Locating teatraent subspaces 10.124 Analysis of variance Comtears 167 7 169 10 160 1@ 70 m im 17 im im 15 us 78 178 178 18 ie im 19 180 181 181 182 12 126 185 185 (a7 189 189 196 193 193 194 195 136 196 7 or 198 198 200 oa! w| | : ' i | ni i i ' i i \ i i 8ier let 169 69 69 169 170 (7 m Mm ma 7 175 Ws 178 178 178 178 178 178 119 180 rT I81 182 182 4 85 85 87 89 89 SRRAIIRRGEE Contents 10.13 Fusther examples Questions for discussion LL Tneomplete-block designs Vi Tntvoduction 18.2 Balance 113 Latico designe 114 Raxdomization 115 Analysis of Balneed incomplete block designs U6 1L7. Analysis of lattice designs 18 Optimality 119 Supplemented balance 11.10 Row column designs with incomplete colons Questions for ciszussion designs in incomplete blacks (2.1 Confounding 12.2 Decomposing interactions 12.3 Constcucting designs with speaified confounding, (2.4 Confounding mote than ove characier 12.5 Prcudofaciors for mixed numbers of levels (2.6 Analysis of single-teplicate designs 12.7 Sovoral replicates Questions for discussion 13 iractional Fuetorisl desiyos (3.1 Fractional replicates 13.2. Choice of defining conuasis 1a Weight 134 Resolution 13.5 Anelysia of feastional repliestor Questions for discussion 14 Baclewarl ok V4.1 Randomization 41.4 Radom say 14.1.2 Ratdoo permailions ofthe plots (4. Random choice of plan 414 Rendomizing westicat labels 14,1.5 Radomizing instaxces of cach treaument 1.156 Raxdom allocation topos 14.7 Restieted randomization 1422 Factors such ste, sex, age and treed M3 Weling a protocel 202 215 ne 29 a9 21 mB 226 229 70 233 24 23s 28 m1 241 242 aS 249 351 253 387 288 259 259 260 262 265 266 zm an mm m mm mm mm ms m5 28 79 282x Comtents i 143.1 What i the purpose of te experiment? 282 143.2 Whol ore the treatments? 282 1433 Methods 285 ! 1434 What sce the expesincatel units? 23 1435 What ete the observational nnits? 284 — 1436 What measures ave 10 be recorded? 283 143.7 Whats the design? 283, Pre 1438 hustifiation forthe design 28 1439 Randomization used 285 143.10 Pla 284 | 143.11 Propored stastcal enalysis 284 144 The eight siagus 285 ‘This te 145. Astory 286 Be Questions for discussion 200 aie Exercises 291 satis 1 Sources of examples, questions aud exercises M3 snot roe! Further veading 9 ie References 321 chan i Index 227 desig how t tore tH : neus! : inal Ea ana spo that cbser! wit daar! one the ty fi incl toe i FO) Tha) aval! romeHens 282 282 283 283 283 | 283 ‘284 284 284 284 284 285 286 290 291 a3 a9 321 war Preface This textbook on the design of experiments is intended for scents in their final year of a BSc in Mathematics or Staistics in the British system or for at MSc far students with a ifferent background. It is based on Iecuures that I have given in the University of London tnd elzowhere since: 1989, I would like it to become the book on design which every working suulstictan has on fis or her shelves, Tassame a basic background in statistics: estimation, variance, hypolhesis testing, linear models. also assuine the necessery finear algebra cn which these rest, including onthogonal projections und eigenspates of symmetric matrices. However, peopke’s exposure to these, lopies varies, asdous the notation they ws, sa summarize what is needed at various point in Chapter 2, Skim that chapter to see if you ned to brush up your knowledge of the background, My philosophy is that you should rot choose an experimental design frerm alist of named ‘designs. Rather, you should think about all asgecte of ee current experiment, ad then decive how te put them together appropriately. Think about the observational units, and what struc- ture they have before treatments are applied. Think about the number end nature of te treai- iments, Only then should you beain to chink about the design in the sense of which treriment is allocated to which exporimental unit. “Todo this requires a notation for observational units thatdoes not depend on the treatments applied. ‘The cost ia Kule moze wotation; the gain is s lot nose clarity. Weiting Yos for the response on the fut anit with treatment 2 goes with a mindset that ignores randorntzadon, that manages the experiment by treatment, and thal docs noses the need for blindness, I tabel observational units by lower-case Greck Tellers. thus F (on) is Oe rexpinse an ebservetional unit eo and Tc) is the trestment on that unit This notation merely minnies goud practice in ‘data recording, witch has a row for cnch obsorvatinaal unit: three.of the columos willbe tho ‘ove which names the unis, he on which shows the reniments apptiec, and the one showing the responses. In this took, sondomization, blindness and management by plot structore are recurring theme Mast siructore on the set of observational units, or on te set of teatsnents, can be defined by factors. Lave developed a mrthod for thinsing about factors ad their relatiouships, inchidltig the use of Hasse diagrams, which corers all othogoral designs. The metho! uses tho inflonre FAG of two factors (which almost everybody olso, excopt Thur [113], waites as FG) and the dual concept, the suprennum F VG, which almost nobody else (again apart from ‘Tjug secs dhe need for, until degeocsof freedom mysteriously jo weong. Bveryone that [have (aught this method to nas reacted enthusiastically and adopted il, However, you need ta have some idea of simple structure befere you can appreciate the generality ofthis approach, which xixii Prefrce ts therefore delayed unitl Chapter 10. “The Hasse diagrams, and the insistence on raning observational wnits, are (wo features ‘of this buok Dat do nol appear in most other books on the design of experiments, The third difference, which is relatively ntinor is my notation for moéels in Factorial desizns, Expres: uch a5 2-4 ay + By + (0B)iy are compact ways of paramotrizing sevoral models st ones, but they do-encourage the fiting of inappropriate models (wina: Nelder has called the neglect of marginatity in (83). [take the view, explained in Chepter 5, that when wo analyse data we first chaose which model tot and then estimace the paremeters ofthat model; wede not need toknow how to paranetrizeany of the models that wedid not fil, Alsoin Chapter 51 spell out n) arecoften implicit, btthsir neglect can lead to contradictions. The thied is Orthogonality: not everyone will agree with this (see Mead (77), for:example}, but Ibcliove thot we should aim for orthogonality between treatment factors wherever possible ‘Another relatively minor difference in my approach is that my snslysisanfwaciance tables atways include the grand micaa. This is parly to make all the celcufations casicr, especially when using the Hasse diagrem, A more important reason is to demosirate thal filing a larger model aftr a amalfer eng (such as a complete two-way table aficr an additive model) isin principle no different from fitting teezlment effec aftr removing the grand mean. Unlike some topics in mathematies, Design of Expetinients can set cut its stall early. Thus ‘Chapter 1 introduces most ofthe issues, aed points the way forward to whore inthe book thay are covered in more detail, Read this chapler to see if this book is for you. ‘Chepter2 covers the simplest caperitaeit there is no structure on either the abservatioual units oF the treaunen's. This gives an opportunity to discuss randomization, replication and analysis of variance, widhont entra contplications, a8 well a (o revise prerequisite knowledge, Steucturc on the ebscrvational units is developed ie Chapiers ¢ (simple blocking), 6 (row column designs) and 8 (observational unite smaller thon expesimeental units), Structure on the tweetments ie developed in, parallel, in two indspordent chapters. Chapter 5 deals with Jactoriel reaiments (crossed factors) while Chapter 3 covers conirol treatments and otter ‘ways of meunively spliding mp the treatments (nested factors). Chapier 3 can be omitted ina short course, but there are some aress of application where Ciapter 3 Is more relevent then Chapter 5. The ‘nrixed’ case of factorial treatments phis a control is covered in some detail in Chapters 1, 5 and 10; this occurs surprisingly often in practice, and is frequently misunderstood. Choptar 8 deals with tho situation when one er more of the tretinent fastors mutt he applied to something larger than observational units, This topie is often misunderstood in Iréetice, 48 a glance at too many journals of experimental science shows. Every working stintician should be awate o€ the danger af fase replicaion. Chepters 7 and 9 are somewhat light relief from the main development, and could be omiticd without making later chapters inaccessible. Chaptcr 7 applics the ideas so far to expesiruents on people; iL also describes issucs peculiar 0 suchtexperiments, ‘The reader who Js conoerned exclusively with such experiments is adviaed to eoatinus with one of the more speciatized texts, such as those recemnmended in the Furter Reading, Chapter 9 takes single combinatorial objeet—the Latin square—and uses it in several ways (o design different types of oxporiment, This demonitrates (hat ere 1s no such thing es a ‘Latin-aquave desig, 07 Pret at appreé ehapie a, of: in Arye! mere Fi of the shoul Rach! vex nny ef disou ser share work we Ake, T shot Pattu! the of desig ba statis eri ‘owe nate wen b aa! these F word alsoface bird tect out heit vith hes ally ‘ger sin sal and ige. vith ted ant, be 1 vho ele, or, Preface perhaps, that the phrase has nrany interpretations, ‘Chaptor 10 ie my favourite, Itpuls cl the preceding material together into asingle general approach. Because i is 50 general, the proofs are more abstract than those In the earlie, hapiors, and you may want to omit them at the first reading, ‘Chepters 11-15 introduce shuce snore advanced topics that a statistic of: incomplete-block desiges, confounded factorial designs, and fractional factorial designs Anyene who needs to use those techniques frequently should probably fallow this with come more advanced reading on these topics: some suggestions are made in Further Reading, Hinally, Chapter [dig zoran of Chapter { in the light of what has been covered inthe rest af the book, Contronied with en expeciinent ro design, how should we think about it aud what shoud we do? noedsto be aware Bach chapter is followed by questions for discussion. Because 46 many aspects of designing an experiment have no single “ight answer, Thave used thise discussion quostions with amy students rather than requiring writen homework, Each student is required te lead the dissuosion atleast once, Apt fron the initial dificuky of persuading stadants that his ie not a tervifying ordeal, this tothnlque nas worked remarkebly swell, Qher stedents join in they share ideas and offer nelpfal criticism. At the end, Lcamment on both the presenting stuctst's ‘work and the gexeral discussion, commending what is good, correcting any outright mistakes, and pointing out any important featur thet they have all missed. Every year 3 new set of stuedents finds new nwanecs in these questions Some instnictors may want to supplemica: the discussion questions with written honie- works, The Bxeroites at the end are provided for They are les closely linked to the individual chapters tan the questions for discussion. Acknowledgements - 1 should like to thank the following people. .K, Waller taught mo the basics et the Ait Pollution Rescarch Unit: careful éata recondiig, meticulous verification at every stage, and the exeitemon! of extrseting pallens from data, H.D. Patienion taeght me so much show designing experimenis while T was working as a postclactoral resecrcher under his guidance, D.J. Finney was bold cnough to fet me teach the couse on Design of Experiments in the Statistics MSc al the University of Edinburgh within twelve months of my meeting the mat- tial. J.A. Nelder appointed me as a statistician at Rothamsteé Experincental Station even ‘though I had no formal statisticat qualieation, DJ, A. Preece introduced tne to such practical vmatters az data sniffing. D.R, Cox hia heen gently nagging mc tu write this book for over lavenly years. ‘Thanks, too, to all the scientists whose interesting experiments T have worked on and te all those statisticians whe continue to bring me theic imeresting design problems, Many of these are named af the back ofthe book. Finally, thanks to all those student, friends ad colleagues, in the widest sorse of the ‘word, who have road dinfs of part of the material and mace helpful suggestions, Of course, all opivions sud any eorasining errors are my ov R.A. Bailey December 20071.1] The) 1. Thel exp, end s Bat wl 00 stan, hog thet, stati ort) sei nex elu this und, fox, waisibpenitnn Chapter 1 Forward look 1.1. Stages ina statistically designed experiment ‘There are several stages in designing an experiment and carrying it out. 14.1 Consultation The scientist, or ether investigator, comes tothe ats cexperiinent, Sometimes at: appointment is made; sometimes the approsch is by vekephane or comail with the expectation of an instant answer. A fortunate statistician wil already have 8 good working relationship wilh tie seicntit. In some eases the scioatist and tatilicfan wil ‘both view their joint work a a collaboration, Ideally the consultation happens in plenty of time before the experiment, The statistician will Mave fo ask questlons to find out about the experiment, amd the answers may not be inedia‘ely available. Then the statistician needs time to think, and to comnpare differct possible designs. In coniplicaied cases the statistician may nead to consult other statisticians tore specialized in soine aspectof design, Unfortunately, the stetistician is sometimes consulted only the day before the experiment starts, What shoutd you do shen IFilis obviows that thesciemlist has cantacted you jut that he ean write "Yee" on a form in response to the question ‘Have you consulted a statistician?” Inznbe is not wort spending time on, More commonly th: scleatist genuinely has no idea dee stntistical design takes tithe. In that case, ask enough questions te find ont the main features ofthe experiment, and give a simple desiga that secins to answer the purpose, Dapress on the scientist tha this design may not be the best possile, and that you ean co beter if given more fotie, “Try to find out moe ebout this sort of expsriment so that you are batter prepared the next timc thal this person, or ons of her eolieagues, cares to you. Usually the scientist deus not come with statistically precise re elucidate this information: ay careful questioning, About 90% of the sudstician’s input st this stage is esking questicns. These havc to be phrased i terms that a won-staistiian ean understand, Equally, you must rot be shy about wsking the scientist to explain echnical teams trom his etd if they seem zelovant, IF the scientist docs have a precoxccived ides of a ‘design’, it moy bo chosen from an ally short list, based on luck of knowledge of wist is availeble, “Too many beaks and taask sdvice on the design af the2 Chapter |. Forward jock, ‘courses give a ist of tree or Cour designs and manage te suggest that there are no alters, Your job may be to persuade the scientist that a better design is available, even iit did not figers in the textbook from which she Tearnt statistics Example Li (Ledybirds) A famous company (wiich 1 shall not name) fad designed an ‘experiment fo compare a mow pesticide which they haa devotoped, a standard pesticide, and “no weainent’, They wanted to convinee the regulatory authozity (the Ministry of Agricul ture, Fisheries and Hooch) that their newe pesticide was effective but did not harm Indybirds. Tinvestigeted the data from the expesimeat, and noticed that chey Med divided 2 fed ito three areas, applied one pesticide (or noting) to each area, and made measurements on throt samples from each area, Tasked the poople who had designed it what the design was, They said that it was completely randomized (Sce Chapter 2). [said thai T could see that it was nt completely randomized, beeguse all the earmples for exch pisticide came from the narra acct of fi itd, They replied that it ast be completely randomized because thare were no blocks Goe Chapter 4) and iL was not a Latin square (see Chapier 6). In defence. of theie argument (hey quoted « respectable textbook which gives only these three signs, 14.2 Statletical design Mot ofthis iovk is about staistcal desiga, The oaly purpose menilening Itherv iste show how it fis into the process of experimentation. 143 Osta co¥ection Tn collaboration with the scientist, design a form for cullecting the data. ‘This should either be on squared pares, with squares large enough to write on conveniently, or use the modern ‘electronic equivalent, a spreadsheet or a hand-held data-logger, There should be a row fer ‘tok: observational unl (ee Section 1.4) and « column for each variable thatig (o be sevorded, is better if these variables are decided before the experiment is started, but always leave space to itclde extra information whose relevanes is rot known unti Iter Bmphasive io the scleniist deat all relevant dain should be recoded 28 son as possible, ‘Taoy should never be copied inty 4 *nester” format; human beings almest always make errors swihea copying data, Nor should they be invented lates Asxample 12 (Calf feeding) Ina calf-feeding trial each cal was weighed several times, once at birth end theresfier on the nearest Tuesday lo certain anniversaries, such as the nearest ‘Twenday to its cightwack birthday, The data included all these dates, which proved to be imurtally inconsistent: some vere not Tuesdays and some were the wrong length of ie apart ‘When [queried this Twas told that only the biuhdave was zefiable: all he otberdates had heen ‘writen down atthe end of the experiment by a temporary wo:ker who was deing her best to follow the ‘nearest Tuesday! rule after theovent. This labour was utterly pointless I the dates rad veen recorded when she calves were weighed licy would have provided evidence of how closely the ‘nearest ‘Tesday' rule had been followed deducing thedates aftr the event could more accurately and stiaply have been donc by the eomppter 3s part ofthe data analysis. Sometimes a scientist wants totako the data from his feld notebooks and eeorganize them into a more logical order for the statistician’s benefit, Discourege this practice. Nol asly does was ay ‘one of was g! agron! (sides! Only! right he dic fell in A junior surpri that ttJoot not . c in fs Ie they not woke rent ern tor ed. hls, 0 es ald em 1. Stages ina statistically designed experiment 3 Pia Pas o 6 Average Average BE Fig. 14. Data sheets with intermediate calculations id Example 4.3 inroducs copying errors; reordering the dala loses valuable information such es which plots ‘were next to each other or what wae Ihe time sequence in which messrénients were made: see Example 1.5. For similar reasons, encourage the scientist to present you witt the raw data, without ineking intecmeciate catculetions. The data will be going info a computer in any éxse, 30 intermediate calculations do net produce any savings and may well produce errors. The only Ttenoft besught by intormodiato calculntions is © rough check thal certs mumbers are the eorreel order of magnitude, Example 1.3 (Leafstelpe) In en experiment or leafstrips disease in barley, nav mossurement ‘was apparenily the percentage of disease on eacit plat. A preliminery graph of the dats showed fone outlier far away from she west of the data, [asked to see the data For the outlying plot, and ivon a collection of pieces of paper like those shown in Figure 1.1. Ilranspired that te aginnonnist had taken a random sample of tea quadrats in each plot, had inspecied 100 tillers leshoots) in each quadat (o vec how many were infected, and averaged the ten mumbers. ‘Only the average was recorded in the ‘ollicial” data. For the outlying plot the agronomist rightly thought that he did not need 2 calculator to add wine zeros to one nonzero aumber, but he did forges to divide the total by 10, Once: had corrected the average value for this plo, it fal ca Tine with the est of the data ‘Also Iry to persuade the solentist that data cellection is too important to be delegated 19 Junior staff, especially temporary ones. An experiment cannol be bette: thai its date, but a surprising number of goad scientists will put much elfort inta their science while believing, thal (he data ean take care of themselves. Unless they really feel part of the team, junior or temporary staff siruply do not have the same motivation to record the data carofully, oven if they eve conscientious. See alsa Example 1.2. 11.4 Data scrutiny ‘After the experiment is done, the dati sheets or data files should be sent to the statistician for analysis, Lack over (hese as coon as possible for obvicus a.omalies outlicrs or evidence: Of bad practice, Can that number really be a calf's birthweight? Experienced statisticians4 Chapter 1. Poowartt Jook become remarkably good at “data suiffing'-—leoking over a sheet of figures and fieding the ‘one or two anomalies. Thee is how the crrors in Exzmple 1.2 were found, Simple tables and graphs can also show up earurs: in Hrurmple 1.3 he outlior was revealed by ¢ graph of yield in tonnes per hectare egcinst aercentage of infected tillers. ine the final digits in the dats, Tf the number of significant figures changes at one polut, this may indicate a change in the person recording the dats or the machine being used. Occasionally it indicates a change such as from weighing in pounds to weighing in kilograms ing by 2.205. Any such change is likely 10 coincide with a change ln conditions which is more serious than the appearance of the data. These ebccks are easier ta conduct on paper data than on electronic dsta, because mos: spreadshoots ive na facility for disinguish- ing between 29 and 29.09, Exanpple 14 (Ktwi fruit) Alan agrioulmral research station in New Zecland, an inst called peneirectr no used to masts de hardness of ‘wrong with the data, He looted again at the data akeot, and noticed (ut wo difforent banduriings had been uscd, He re-analysed the den, lelimg the'data in one handwriting be an unknown constant snultiplo of those in the other. ‘The titted valu of the ‘constant was 2.2, indicating tai one person fad recorded fa pounds, te ether in kilograms. ‘Query dubious data while it ix lll (esh in the scientist's memory, Thal way théte is 1 chance that either the data cun be corrected or other explonstory information recorded. Exwuple 15 Rain at latest) In an experiment whose response was the yicld of went on ‘each plot, the numbers recorded on the Last 12 plots out of a total of 72 wore noticeably lower (han the others. T asked if thers was any reason for this, and was told that it hnd started to sain during the barvest, with the ain staring when the harvester was about 12 pls irom the end, We wore therefoo able to include an extra variable ‘rain’, whose values wore 60 zeros followed by 1, 2, ...4 12. [ncluding ‘rain’ a8 a covariate in the analysis removed a lot of ‘otherwise unexplained veriation Example 16 (Eueatypts) In n forestey progeny tal in Asia, differea families of cucalypts ‘wore grown in fivectree plots, After 36 moiths, a ferestey worker measured the diantcter of ‘each treeat breast height. Inthe preliminary cxamioation of hedata, the statistician calestatod tho sakiance ofthe five responses in each plo, and found iat every plot had exactly the same variance! Challenged on this, the forestry werker admitied hat he had sncasured every tee in the frst ple, bul thereafter measured just ce Vin each plo. For ees 2-S he had acted the consant¢ 10 the meesuremen's ftom plot J, where ¢ was the difference betwseun the diameter at bresot height of ee 1 in this plet end the diameter at brcoat hcight of erce 1 in plot 1. {In this caso, the statistician’ preliminary seretiny showed thatthe data were largely boaus. 1.1.5. Analysis ‘This means calculations with the data, [¢shoutd be planned at the design stag, bosause you ennnot decide if a design is gocd until you know Inv the dasa will be analysod. Als, this planting enables the experimenter to be sure thal se icollocting the retovant deta, IT neces sary the analysis tay be modified in the lightof wrferesesn cieumstances: see Lampe 15. way san el Se md Chapt dad 4 m4 118 The stand isthe) unde i 12 Here thee, 1.24 Why! publit ist statis at This how sin ive ae the ¢ of ae} Pek okt 75 12h Wis: © assur *the and ield one ol. ms on ish ment inary kere that othe me. E, =.“ = visa se youl 0, this fe 15. UNIVER AUTRR HALAS UL coli va, clita be Ey 5 For a simple design the statistician should, In principle, be able to anelysc the data by hand, with a caloulalor. In practice, itis mre sensible to use a relate statistical computing package, A good package shavld ask the user to distinguish plot structure from Weatrrent structure, a in Section {42 should be able to Kandle all he structures given in Section 1.4; ‘and it should automatically calculate the correct variance ral for experiments Fike those in ‘Chapters 8 and 10. J: is a good ide to do the planed analysis on chummy data before the weak data arive, to avoid any unsceeseery delay. ‘Many ether statistics hooks are concerned aimost exclusively with analyst we cover paly enough of i to help with the process of designing experiments. - Ir this book 41.1.8 Interpretation ‘The dota analysis will produce such things as avalysis-of-varinnee tables, lists of means and standacd emrors, P-values and so on. None of these may micaa veiy mach to the ssientit, (t is the stetitician’s job to intenpeet the results of the analysis in lems which the scleaist cab Understand, and which are pertinent to his atiginil question, 1.2. The ideal and the reality Hcro Idinauss« fow of the tensions berweun shat the stalstcia thik is desirable and what Ihe experimenter wants. 1.24 Purpose of the experiment “Why is the expotiment being done? If the answer is ‘to use un empty greenhouse! oF ‘to publish another paper’, do not put much staistical effort into it, A more legitimate answer is “to find out about me differcnces fenween so-and-so’, but even this fs too vague for, the staiaticsan to te really helptal, Tecally, the aims of the exporinont should be phrased in terms of specific questions. The aim may te to estimate something: for example, ‘Flow muck better is Drug A than Drug 67° “Thia quoation necds refining: how much of exch drug? how administered? to whom? and ow will ‘etter’ be mensurcd? For estimation questions we should sim co obtain urbiesed estimators with love variance. ‘On the oxher hand, the aim may be to teal a hypothesis, for examapl that there is nocffeet- ivedifference between orgaricard inorganic scurces of nitrogen frilizes. Again the question deeds refining: how much fertilizer? applied tc what erep? in whal sonts of circumstances? is the effect on the yield or the taste or the colour? For hypothesis testing we want high powor ‘of detecting differences that ave big encasgh lo mater inthe selencé imvelved 1.2.2 Replication “This is the word forthe munnber of tines that ench (reste is tested, "The vell-kaown formula for the variance of the mean of sr numbers is a? my on the sumption that he numbers area random sample fiom z population with variance”, Incteas- ing the repLcotion usualy decreases the variacs, because i irereaven the valu: of ‘Onthe other ond, increased replication may false the variance. Typially, a targer mumioe:6 Chapter 1. Forward look. of experimental unis are mere voriable than 4 small number, so nsenensing the repliestion may increase the velue of 6?. Sometimes this incresse outweighs (he inczesse in 1. Incrouned replication asually vaises power. ‘This is because I: uscally raises ihe nunber of residual degrees of freedom, and certain important Families of distribution (such ss} have ‘immer fails when they have more dogrees of freedom, “The one thing thet is almost cerain about increased replication is that it increases costs, which the experimenter usually wants to keop down, 4.23 Local control ‘This meane divicing the set of experimental units nto blocks of alike amis: soo Chapter 4, isso cated Bfocking. Tf it is done wel, blocking luwers the variance, by renoving some soitees of veriei ity from teetuent contrasts. If cact block Is representative rather than homogencous then blocking tas the opposite effet locking cam incronso the variance i eri, ectuse blocking ales always decreases dhe variance, it usually mises power, However, it decreases dhe number of residuel degrees of freedom, s0 ican reduce power jFnuiibers re sail: see Brample 4.15 Blocking incivases the complexity ofthe desiga. i cua tis fc only increases the com= DPlexity of the analysis aud imezpretation but gives more scope far mistakes in procedure d= ing ths exposirt, it forocs the design w be nonorthogonal: see Chap- 1.24 Constraints ‘The most ofvious constraint fs coxt, Frenybody will be pleased ifthe same cesults can be achieved for less money. If you can design a sinaltsr, cheaper experiment than (ke scientist Proposes, this is fine fit prodoces good estimatore. On the othorhand, i: may be ‘enpossible to draw clear concluslons from an experiment that tou smal, so then the entre cast is wasted. Partof your duty isto watn when you believe that the whole experiment will he wasted. The availability ofthe test materials ty provide constraint. For example, i testing new varieties of wheat there rnay be fimiled quantities of seed af some orall of the new varictes Asailabiity of the exporimontal units provides « dflcrent sort of canstraim, There nay ‘be competition with cther experimenters ta use land or Bench spuce. If results ere needed by a cermin desdlina then time limits the number of exporimonta units, In a clinical el it Is unethical to use far foo many patients becuse this unnecessarily increnses the nunther of alienis who de not get the best treatmeat. On the other and, it is also unethical to ue so few peticats that uo clea: conclusions can be draw, for then al. the patients have been used in vain. Similar remarks anply to experiments on anivals in which the animals have to be At shere are nctural “blocks” oF divisions among the experimental units these may force ‘constraints on the way that the experiment ean be cartied out. Hor example, it may be impos sible ( have all vaccinations administered by the same nurse. ‘There are often uther consteaimts imposed by the management of the experiment. For Avex, Era of ye meas the pe ™ ilook may Jaber thaw josts, 4 it then, hap- lever, fs are pom dur 13. An example 3 example, temporary apple-piekers like to work with their friends: iC may be unrealistic 10 cexpost them each to pick from separate rows oF tees, 1.2.5 Choice Given all the constrains, ther are sill wo fundamentally immportvot choices that have to be ‘made and where the statfalician cat provide wavice. Which troutments are to be tested? The scientist usually has aclear idea, but questions ean Hil be helpful, Why élethe decide on these particular quantities? Why these combinations sul not cthers? Stould he consider chenging (wi factors ata ime? (see Chepter 5), Does the inclusion of Fess intresting treaiments (such es tho boss's favourite) mean that the replication fer aif wentments will be oo Tow? ‘There isa sleung belief i sciemife circles that all'new treatraeats should be cafipared ‘with “ao treatment’, whick is often called contvet, You should alvays ask if a conical is necdcd. Scientific orthedoxy says yes, but there are experiments where @ control ean be harmful. Af there ts already an eifective therapy for a discase then i is unethical to run an experiment comparing a nox therapy (0, ‘do nothing’; in this case the treatments should be the how therapy and te ove curren in vse. In a tril of several pesticides in one tied, if there isa ‘do eothing’ restrtent on some plots then lie pest may multiply on those plots and then ‘spread to the othors. A ‘du nothing’ treatments also not useful iF this would never be used in is already known thet th ‘do nothing” treatment has a very different effect ‘om all te other tréarmerts, ‘The the experiment may do aothing snore then confirm this, Bxamples 32 and 6.3. In such cases. i is better to omit the “do nothing” reaiment so iat more resourecs can be devoted to finding out whether these is any difference between the cofher treatments, “Which experimental unite should be weed? For example, is it heller (0 use portions of representative farmers’ elds or a well-controlled experimcatl farnt? Tae later is better if the effect ta be detected is Hkely 10 be smell, or if one of the trealments is sufficiently Lunknosen that I anight have disastrous ccoromic or ewsironmiental eoncoquences. The former i hotter for a lage cnnlizmatory experiment, before recommiensing varetles or weatinens for use on a wide scale. Similarly, is it beter to use 36 heifers Frown the same her o: 36 bought at the snarker specifiealy for this experiment? Univershy studemts are a comeniont source of oxpexinental cits for psycholngisis, but how far can cesuits valid for such studamts ‘bo extrapotated te the general population? 1.3. An example Example 1.7 (Rye-grass) An experiment wax cinducted to compare thee diferent cular ‘of 1ye-grass in combination with Four quantities of nitengen fertile, ‘Wo tesponses were measured: one was the total welghtof dry mater harvested from each plot, and the other was the percontage of water-solunte earbohydrate in the crop. “The three cullivars of tye-prass were called Croppor, Milic and Melba, Tie four amourts8 ‘Chapter 1. Forwand ack 0 |} ico 2s wo] oa ° ist 90 |] eo o || 10. 40 0 of 10 240 o |} 20 240 6 s0 | 20 |] 60 “T T TT T T Cropper Melia Melle Mab Cropper Melle Fig. 12. Layout of te rye-grass experiment In Example 1.7 of Feciilizer were 0 kwiha, 80 ke/ha, 160 kg/ha and 240 kp/ha. ‘The oxperinenial rca consisted of two fet, cach divided into three sips of land. Bach atrp covsisted of four plots. Cultivate vers sown oa whole strips begauie itis not practicable to sew them in smal arees uniess sowing is done by han. tn coatasy, it is posoctly feasible to apply ferlizers to smaller areas of find, such asthe plots, The layout forthe experiment i shown in Figur 1.2. Notice ths puter Bach anova of nitrogen is applied 1o ove plot per strip, and exci cultiver is applied to one strip per field. This pattern is the combinotorial devign, Notice ths lack of pttere. Thee is no systematic eer i the allocation of culivars 10 ‘Strips in cach field, nor any systematic order in the allocation of amounts of nitrogen to plots Ineach strip, This lek of patiom isthe rwadernicaton, 1.4 Defining terms Definition An eaperimental seit is the smaltes| unit to which a treatment can be applied, Definition “A tronsorent is the entire description of what ean he epptied to an expociraental anit: Although the previous two defiitions appear to be circular, they work welt enough in practice. Definition Au ofservivionat unt is he sralest unit un which a response wilt be measured, Example 1.6 revisited (Eucalypts) The experiments! units were the plots, The observationat Units should have bon the tress. Example 1.8 (Wheat varieties) The experiment compures different varieties of wheal growe inplots in afield, Here the experimental units ate the plots and the reatiments are the varieties, ‘We cannot tell what the visservational wit is without more Lafurmation. Pratably a plot Is the observational unit, bal it ovight be an individu plant. I: night even be sho whole field ta. Exar! amor ‘obse: Exaj feedleak leach sal 1810 12. leach jas zatall hin wed. ass | sma goons ‘SIERO paw TE evinwle every day v ¥ ig. 1.3. Powe Ceetmerts in Example 1-10 Example {.7 revisited (Rye-geass) Hore the teestments ar. the combinations of oulivars with amounis of fertilizer, so here are twelve treatments, ‘The experimental unit is the plot. “The obge-vational unit is probably the plot but might be a plant or a stip Example 1.2 revisited (Calf feeding) Here the treatments were differeat compositions of feed for calves. “The calves were avi fed individually. “They were housed in pene, with ion calves per pen. Hach pen was allocated to a certain type of feed, Batches of this type of feed Were put into the pen; calves were free to eal as much of this as they liked, Calves were weighed individually. ‘The experimental units were the pens but the observational units were the calves. Example 19 (Aathna) Several puicnis take pact in a experimentto compare drags intended to alleviate the-symptoms of chronic asta, For exch patient, the drogs are changed exc ‘month, From fime to time each patient comes into ike clinic, wher the peak floweratein their tongs ts measured. Here he treatmentaare the drugs. An experimentel unit isa patient oath combination, so #€ 20 paticnis are ssed for G months thea there ase 180 experimental units. The observational utto the efiai: we da not know how tis celates to the patient-monis without farther informtion, Eanmple 1.14 (Mental arithmetic) After calculators became widespread, theze was concert that children in primory achools woreno longer becoming proficient in mental aitimetie. One suggested remedy was wholo-ctass xcsstoas, where the teacher would call out a question sect as $477 and ebikdron would put up (aeir ands to offer to give the corect amewer An alternative suggestion was to do this in small grcups of about four children, to encourage those whe were shy of responding in dont of the whole class. Another question was; is il beiter to have these sossions far one hour once a week or for 10-12 minutes every éay? ‘The treatments are the four combinations of group size and timing shown in Figure 1.3, Bach treatment can be applied only to a whole clase, so the experimental units aro classes, However, (o measure (he effectiveness of the Irentmients, esch child must eke an tndivideal lest of mental arithmetic after seme wel Line. Thus the observational units are the children, Example 1.14 (Detergents) in a consumer experimen, ton hoosewives test new detergents, Bach housewife tests ene detergent per washlosd for each of four washloads. She assesses ihe cleanliness of each Washload on a given S-polut scale, ‘Here he 40 washloads are te experimental units and tho observations units: the decrgents are the treatments Example 112 (Tomatoes) Different suuieties ef tometo are grown in pots, with ditlseent composts and different amounts cf water. Esch plant is sepported on a vertical stick until itis 1.8 meiees high, thew all further now growth is wound around a herizontel rail, Groups1000702068 Chapter 1. Forward look of five adjevent plants arc wound around the sume rail, When tie tometoes are tipe they are ‘harvested and the weight of saleable tomatoes per rail is recorded, Now the treatment is the variety-cornpost-water combination, The pots are the expeti ‘montal units but the radsare fe cbservational unit ‘These examples show that there are four possible rolaionchips batween experimental sand observational units, @ The experimental units and the observational units arc the same. This isthe most usual situation, Ie occurs in Exorapte 1.115 in Examples 1.7 aod (2 if there is one messure- ‘meng per plog; in Example 1.9 if thers is one measurement of peak flow rate in lungs rer pationt per month, {ii) Bach experimental unit consisis-of several observational units. This is ustally forced by prctical considerations, as in Examptes 1.2 and 1.10. Examples 1.7 and 1.8 are of this (pe if the observational unit ie.a plant, So is Example 1.9 if the observational unit is & pitiemt-woek, This situction is fine so long as the data are ai Chapter 8. Bach observational unit conssta of several expec ‘Exaniple1.9 if each pied had theic drugs changed monly but their peak flow rele rneasured only every thee oaths, Tt wiuld also ccour in Essiagtee 1.7 and 1.8 if the cbservatenel unit were the strip or etd respectively. (a these cases the mensuTements canmot be linked lo individual treatments so there is nu point in conducting such an ‘exporter Example 1.12 also appearnto bo of this form. Because the experiment would bo useless iT aitterem: pots in whe same group (allocnte to the same rat fad eifferenttreaumens, jn elfectivs the group of pou that i the expetimenial unit not the individual po. Jn fact, (here are sons unusual experiments whore the response on the observational luni oan be coxideced to be tho stm of the (uvkaown) responses on the experimental Units conteinad within i, However, these are beyon¢ the scope of this book. iv) Experimental units and observational units have a partial overlap, but neither is con- tained in the other. This caso is even sillier than the preceding unc, is useful to wrlle down the experimental units and the observationel units in the experi= mental protocol. This should draw attention to eases (i) and fiv) before ili toc lute fo change the protocol Delinition In cases (i) and {ii} an observational unit will often be called a phot for br ‘This usage is justified by the lage antcunt of early work on experimental design that Look place in agricuiural research, However, it can be a litte disconcerting ifthe plot is aotwally 8 porson er halla leaf. It is ¢ useful shorthand in this hook, hut is not recommended for your conversations with sciontiss Notation in this took, general piots are denoted by lower-case Gnock letters, such a5 , B. 44,@. The whole set of plois is denoted hy 1, and the number of plots by NY xa fall sy wean xan, sould ments ‘Thes did ng Exan| tes these’, Vigor, of tal there ising applic Exar rot appl cconkrtnok awe pti sual lungs 20n- veri wnge ook ally 4.4, Defining terms u spray mti-season no sni-teason spray spray late | ootate spy | spray la | no late sorey spray eudy an ard Z no eaely spay vi ¥ Hig. 14. Factorial restment comsbinctions in Example LAS Example 1.13 Pallets) ‘Thro feeds for pullets are to be compued for their effect on the weight of their epgs. Ten pulleis are solected forthe exeriment, are isolated and are indi vidual fed, with Bye gullts recciving each ofthe twe feeds, Aftor tho Zeeding vegimo.bns deen in place for one month, Cheeggs lei thexvalterby'each pallet are individually Weighed, “The individual feeding impicx tha the pallets ave the experimental anit, but wht, are ihe abservational units? Hf the egy are the observational unity then have two diticiltes wwe do not know the number of observational units In advance aad ths nurs will vary from one pullet ta snotaes. Both of these difficulties can be overcome hy decleting that only ‘he Gtst so many eggs laid (or, more practically, collected) will be weighed. On the other nnd, ifthe feeds effect the nuraber of ogg laid as well as thoir weight fie it might be mere seasible lo measure the total weight of eggs iid by eact poll; in this case the pullets are the ‘bservetional nits. Example 1.44 (Slmpte funglelde) In « fungicide wriel the treatments are doses of fungloide: full spray, half spay snd ‘na trestmeat’. ‘The experimenter might sey that there are twa wcamments and a coutrol; in our vocabulary there axe thece tremens. Example £.45 (Fungielde factorial} In another fungicide trial on winter wheat the fungicide could Le sprayed carly, mid-season or lats, or at any combination of those times. The tec ‘ments consisted of all combinations of *spray' snd “no-sprey” at ench date. Sce Figure (.4. Thus there were eight troxtments; the experimenter Cold mie that there were seven, booatse he di not consider ‘never spray” tu be a trealment, Example (16 (Fungieide factaciat plas control) {n an important variant of Exaraple 1.14, the spray is applied only once, tut this can be early, midesvason or tas, ‘Thus the treetmenls se combinations of ansount of fungicide with time of application. How many treatments are there? Ic fs quite common to see the weaiments fu dis exanaple Tall ou scheitalicnly as in Figure L Sta), which suggests ihat rere are nine treatments, being all combinations of smut of fungicide with Give of application. Ihave sean data from such experiments analysed as if thore were nire ircatments. Howeves, i here is no fangielde hen it does tot make sease to dbainguish beoween line of application: tho ine of application should be regarded ax ‘xo! applicable This ves the sovén tealneems showa tn Figure 13), Example L.A7 (Oilseed rape) An caporiment on methods of controlling the disease sole. rola ln ollseed z2pe compared four new chemicals, coded A, B, C.D, with both ‘no ireat- arent’ and the current (expensive) standard chernical X. Each of the new chemicals could be applica either early or Inte; the standard X was applied at both dimes. Thus theze were (vo control reatinwents, and the teeatmeats had the structure shown in Figure 16.2 Chapter 1. Forward took wid nid ects | sean | cauy | seni | ite |, full spay wv Fal speey. t. 7 tuttsray [J J itsony oy TOI woop LY v sway [J ¥ (a Tangri destipon (1 Arpepinedeeition Fig. 1.5, Two descriptions of de reaumeaisin Example 116 Notatlon fa this took, general treatments are denoted by lower-case Latin letters, auch as 1p The whole set of wenunéns is denoted by T,and the aumber of uestmients by f, “The experimental protocol needs to contsin 9 pscise description ofeach treatment, This :nuxt include complete dems, uct as ingredients, proprietary name (If anyl, quantity, ime of application, variety, species, ete. Then give each treatment a simple cede like A, B, C oF 1, 2,4. for rfsronce Inter Detinttion” Treauuenr structere means meaningful says of dividing up T. ‘Beamples of treatment structure include: ‘unstructured ‘This means that there is no structure tothe. treatments at all, several new treatments plus contri! This further in Chapeer 3. the structure in Reurmple 1.14, Tt is examined all combinations of ty factars See Examaple 1.10 and Chapler 5, all combinattons of twa factors, plus control ‘See Example 1.16 and Section 5.9. all combinattons of thrze factors Soe Examples 1.12 and 1.15 and Chapters 5, (2 and 13, {mereasing doses of a quantitatlve factar This is tot caveréd explicilly in this book, apart from the discustion of ixample 5.10, because the relevant ideas ean be Found in most books on regression. Hactots sucli as ihe mount of fertilizer a Fxaraple 1,7 are indeed ‘Quantitative, bu! not really on a comtinuous scale, hevauxe the Farmer will use a whole number of haga of fertilizer Definition Péot structive means meavingtul ways of di tle treatments, ing up the set Qol piovs, ignoring Examples of plot stracture inelude: Unstructured There is uo structure to the observational units at al ‘experimental unlts containing obsorvational units This is the siructire Isdiscussed im Chapter 8, be i Bean, expat rortelook ich a This time’ "orl, ained apart most eed whole oring 1.4, Defining terms, 18 none’ kOe Fig. 6, Treatment staccune in Braniphe 1.17 blacks ‘This means local contrl: dividing the set of experimental units info homogetieous hocks. In Example 1.11 te holisewivés should he inated as blocks, See Chapter 4. ‘blacks envtaining subblacks containing plats This is the structure in Example 1.7. It is vlvo discussed in Chapter &, nits This is mentioned in blocks centalulng experlnental units contalalag obscrvati Chapter 8. two different sorts of blacks, nvitier containing the ather This plot stuiciure oecurs in Example 1.8, where he two sorts of block are patents and anonths, Sec Chapter 6. All these striclures (except for quan Clinptet 10. In principle, any type-of treatment siructure can occur with auy type of plot structure, Tat ic why it ia neither possible nor sensi to give a shot list of useful designs, ‘reatutents) are desciibed in a ueified way in Definition ‘The design is the allosation of treatments (o plots. Although we'speak of sllocsting treatments fo plots, nathematicelly the design isa fone- tion T from @ to, Thus plot 6 is allocated trontmei T(o). ‘The Funstion has to be this ‘way round, because each plot ean receive caly one treatment. {¢ may seer strenge ta apply a function T to an actual object such as arat or ¢ plot of land, but this is indeed what ihe design ‘Weusually choose F ta satisfy certain combinatorial propertics. The designhes theoretical plots (pertteps numbered 1, ..., 8) and coded treatmcais, Definition The ple er layont is the design tanslated info aciucl plas, Some randomization is usually invelved in this sranslation process. ‘The, actual plots must be fabelled or drawn in such a way thet the person applying the treatments ean Wentify the plots uniquely. For example, in a field trial the North arraw will ‘usually be shown on the plan. Example 117 revisited (Oilseed rape) The Nenih errow was omitted from tke plan for this experiment, The porson applying chetniects atthe later dat he'd the plan upside downclative (othe person applying them cary. ‘The eect of this is shown in Figure 1.7, where there are ome quite unintended treatments.14 Chapter 1. Korward look Pot x 9 wo " 1” 13 4 sls 6 n 8 6 mH beret B ° © x BD 4 Li aeiaxoet lexan basi me lg. 1.7, Rosut of hoiging te plan upsiée down a the later date iu Bxample 1.17 ‘The trcatercnts io the plan usually remain in coded fon, Pactly thi is fur brevity. Pandy. itis t0 prevent unconscious biases from the people applying the treatmen's or cecording 1he. dam: see Chapter 7, . 1.6 Linear model ‘The response on plot ais a random variable Yo whose observed value aficr the experiment 1 yoy Thus we have a data vector y which isa reatizaion of the rantom vector ¥, ‘We assume that we have measured on 4 suitable stale so that Ya Zoro ay ‘Where Trio) is a constant, depending on the treatment (0) applied to plot «, and Zs is 8 ranean veriable, depending or ca. Thor Zo can bo thought of ws the contribution of le ploto (© the response, while ts the contvibotion of treatment f. fe cannot assume the additive model (1.1} then w sannot extropolate the conclusions of the experiment to any cther plots. Im suet a case, ix it worth doing the expesiment? Even if we can assume made! (1.1). it gives us more unknowns than measurement, owe have to suger something about the values of tho Zo, ‘The probatility space is (aken tobe the set of occasions and uncontrolled candiliens under hich the experiment might bo certied out, The non-epeatabitity of the experiment gives Zy Usrandomness, 1£a and ate woditferent plots then Zq and Zp are clferent random variables 16 The 6 steuctu using ten of pio ofHV look actly 2 the nee Ze bles 16, Suawnery 1s cn ine same probability space, so they have a Join distitation, including a cerreltion, which may be ero ‘What is the joint distibution of ths (Ze)yca? Hore are soine commen assumptions: shuple texthook made! the Zy sve independeat ideniically distributed normal random varie ables with moan € and varinnce a? fixedeftects model the Zy ar ndepsaen¢sorma’ random vices each with vavieaee o, and the mean dH» of Zo depends on fe position af w within Q; randomcffects model the Zy have idenien! distibutions, and the covariance cov(Zay25) depends on how @ and Bate related i te plot sructare: randowzatlon model the Zu hive identiesl distributions, and cov(Zq, 7%) depends an the method of eandcinization Using our assumptions about the Za, the analysis ofthe data sould give ‘© mninimunn variance unbiased estimators of the treatment parameters t; and linear com- Dinations thereof, such as tj - t ‘= estimates of the variances of Ihose estimaters, + inferences about presence 6r absence of effects. 4.6 Summary “Tho top hall of Higue 1.8 sunnmarizes this chapter. ‘The plot structure and the teatment -deuglure must both he ‘aiken into secount is chosing the design. Randamizing the design ives the plan, ‘The mettad of rardemization ts usuily dictated by te plot structure, tut occasionally we use the method of-randomizalion te define the plot structure. ‘The bottom hal ofthe Gyere ves a proviow ofthe rest of the book, The plot stucture celen #8 the null analysis, as we shal] sce in Chapter 1: this has no treatments and no daia, Pauing this together with the design gives fhe skeleton inalysa, whit haa reatenens bu sill roduia, ‘The skeleton nralysis enables us to make educated guesses about variance and power using prefininary guesses about the size of ¢2. Ifthe variance is too big orthe power tue low thea westiould go tack and changéethe design (possibly changlag ie plo tucture by adding plols or the treatment structure by removing treatments) and begin the cycls egain.16 Chapter |. Forward took, plot steacinre ‘eaten sxcur {combinations} dosien (uinrce feta} centr eon ee ay Cer rote “ [no data) 18. Overviewand preview Questions for discussion Li A professional ayple-grower has weitlen to you, making an appoiitmest to discuss en ‘experiment which he is proposing to conduct in the coming growing season. Part of reads: There is a new type of chemical spray available, which is supposed co make the apple flowers more autractive to bees, Sixce bees are essential for seting the 1 want co jnvestigale thos chemigals. Two manufacturers ere selling the thew sprays, under the Irate names Buzzi! and Attractahes, T propose dividing my archacd into ice parts, [ shall Spray Auractehee onto, ‘ou part, and Huz!¥ onto the second park, The ihind part wit! be managed in the normal, way, with none of these new sprays. I shall then see which part of the orchard does best, ‘Mae notes oi whar you should discuss «ch him—and why!—at your meeting. 1.2 Scverat studies have suggested thot drinking red wine gives some proveetion against earl disease, but it is not known whether the effect is eaused by the aicchel or by some exher ingrodiont of red wine. ‘To investigate this, medical scieuiisis ewolled 40 volunteers inco a trial lasting 28 days. Fo: the first 14 days, half the volunteers drank wo glasses ef red wine per day, while the other holf hed two stendard drinks of gin. Forth remaining Mt days the, ‘rinks were reversca: (hose whe had been drinking red wine changed to gin, while those wha had been drinking gin changed to red wine, On days 14 and 22, tho scientists took a bloc sample from cach volunteer and measured the amount of ailammary substasce in the blood, Identify the experimental units and observational unite, How meny are there of wach? ‘What is she plot structure? ‘What afe the Weatments? Wher is the treatment structure? Ouest in Fool in so inca release carbo! cacy) w news]took ‘etter nea wher uo a wine sthe who lood ood. sch? Hisababites wisitabenins es yok adtituboiei, essing ied ty Questions for discussion n Table Lt. Data for Question 1.3 Hours after drug administration 2 3s 6 8 B Persen Day Dry a a p ¢ © A a ¢ D 4 o 8 8 4 5 a © . A » A c © # 1.3 ‘Twelve people too part in a study to compara four formulations of lithisii carbene, Formulations 4, # and € were pills, white formulation D bad the titiam carbonate dissolved in solution, Pills A and 2 were sivilar, bul with different excipients (tis is ihe name For the inactive substawe that holds We active ehenvical). Pill C contained something to delay de rofease of the lithium carbonate, The formulations contained different quat carbenoto, as follow in ruilligrams por dose, 4 B CD 300 250 450° 400 On one day, exch person was given a dose of onc formulation, Blood samples were taken (610, ioe, five, six, nine and iwvelve hours later. The quantiy of lihiom in the sample, afer sealing tou common dose, was recorded. One week tater, the procedure wes repeated, with ‘each person receiving a different formulation fra the presious one. Table 1,1 shone the daa ‘What are tho exporimental units? Wht aro tho obsorvat meats? How many are there of exch? ‘Comment on the dava recording.8 Chapter 1, Forward look, 1.4 ‘Two iypes of conarete mlx ae 1 be compared. Ingredient _‘Type(@)___Tywe() cement 220 kghes 390 gZn® wale 172 kgm? — 210 kg/m? aggregate 868 kg/m? 1839 kg/m? Each of five oporators mixcs two brtctes, one of erch type, then exsts threo cylindrical sannples from each batch, ‘The samples are let to harder for 21 days. Aller 7 days, each eperator randomly chooses ove eylinder of exch type and ascertains the bresking iond of wach by using a standard corercwstesting meckine, After (4 days, ane of the remaining (wo cylinders of cach type is chosen al random, and treasured in the sme way. The final eglindor of each type is measured after 21 days, ‘What are the treatments? How many ate ther ortify ths experimental units and ebsorvet 22 Wha is the treatment steuctore? al unite, What is the plot structure’? 1.5 Read the paper "The importance of experimental design in proteomic mass spoetremetry experiments: some cautionary tnles” written by Jianhua Hu, Kevin R. Coonibus, Joffrey S. Mortis and Keith A, Baggerly, which was published inthe journal Brieiags it Fevtetionat Genomies and Protecmics in 2005 (Valome 3, pages 322-331). How fer do you think dheir ecommendadéus apply to experiments in other areas of scfence? ‘Trestra canon Tod ar aya s iy ¢ Exam, Sand, Supporlocke ley rey mat heir wb Sr th atari ope a iy yi atte neni Chapter 2 Unstructured experiments 24° Completely randomized designs rthere is no resson to group the plots inia blocks then we say that 0 is unstructured. Suppose that weatment is applied to ¢; plows, in other words that dis replicated: times, Then ‘Trestmeute should be alloceted to she plots randomized. ‘Taconstruct and randomize the design, proceed as follows. G) Number the plots 1. 2... spply wentment 2 10 plots r+ 1, ..-n bry, and ii) Apply reatment 1a plots 1... 50.00, fo obtain a systematic design. Gi) Choose a random pesnneation of {1,2,. i} and apply it to be design, Example 2 (ictitious) Suppose thal here are three treatments coded A, # and C with 74 = Sand 7 =ro—4. Then there are 13 plots. The systematic design is as follows. plat 1234S 6789 ORD name AA AAABBGEBEC CC C ‘Suppose thatthe random permutation is 1234 56789 On 21D (Sisuntas?s 93 ia} ‘where we ere using the usual two-line way of displaying a permuution, which here indicates, that 1 ond 6 cre interchanged, 2 dacs not move, 3 is moved to 8, and so on. In applying this permutation 1 the design we move the teatment A on plot | to plot 6, Jeave the teaser A ‘on plot 2 will on plot 2, move the treatment A on plot 3 to plot 8, and so on. This gives the Following plan, ou 2B plat ras BACCABA 1234.56 teument BA CC BA 920 Chapter2. Unstructured experiments 2.2. Why and how to randomize Why do we randouize? tts 1a avold systematte bias for example, doing all the ts oon teoatruent Bin Maret; on treaiment A, in January thea all the Wests aclection bins for cxanpple, choosing the most leaky pauiems far vie treamicnt thal you are lying to prove is hest; accidental bias for example, using the fist rats that the anjonal handler tnkea out of ths cage. for one weatment and the Test vais for the ater; tieating by the experimenter, Cheating is nol akviys badly intestioned. Por exzmpie, an experimenter may decids to give the cxtra ilk rations to those. schootchitdren who are most undernourished or shee may choose to pul a paticn: ia a ial ifske thinks thot the pation: will pnricularly bereft From the naw teectni@at. This gor! of cheating is for the beneit of the (small) number of people ta the ‘exporimont bul, by biasing the results, may be to the disadvantage of the (large) number of people who could benefit in future from a trment which has been demonstrated, without ‘bas, to be superior, As enother cashnple, she may bé seereily trying toremove bias by trying, {o balance auurbers over some nuisance factor without troubling the statistician, but this too can produce false results unless this “belancing’ is taken indo account in the analysts, Yet ‘again, she may be tying to make life a litle easier forthe technician hy telling him todo all of one treatment first. ‘Thus doing at objective eandemizztion ard presenting dhe experimenter withthe plan hes, the added henefi that it any forceer to ill you something which she had previously thought vaanecessary, such as We cannot Jo it chit wey because ...° or ‘that will put aft replicates a ‘reatment in the shady part ofthe fie How do we choose a random permutation? The precess must be ubjective, sa hat you have no chance to cheat either, Simply writing down the nambers i, ....N in an apparently haphazard onfesis rot good enough. (One exceltent way to ranfomnize isto shulltea pack of cards. A normal pack of 52 playing cerds shoiild he thoroughly randornized alter soveniffe shuttles. 1s Example 2.1 one con desl ‘ou the shutfled pack, woting ke umber orca Jack = 11, Le.) hut not the sult an. gnoring repeated aunibers, Avothor good nicthed is to ask » computer to prualuce a (pscwdo-Jrandian auder of the umabers 1... Even a paim-tep ean co this. The permutation in Excmple 2.1 corresponds tothe eandom ordes 628 Hts 7493 0 Two offer methods of randomizing use sequences of random numbers, which can bo gen comted by computers oF calculators or Found in books of tables. Both methods will be ila tated here for a candom permutation af the numbers 1,2... 13. My calculator piociuces revo aunvbcrs unifornuly distributed between O and L. We need 16 different numbers, so record the random nunnbers to two decimal places, These are shown 2.3. TH 7 0 2 a i a7 on 1¢| Fig. 22) 1 ~) in me the oni leave «| ue ligt! Coating final Th, All ‘ourube! a fow| tross0} mae! 23 Deft: Th nambe is, nero sual Dean each tmers aying 1 deal ocing f the ands gen ius need iowa 2.3, The treatment subspace a 072 034 9.465 096 0.01 0.23 0.30 057 017 04 0.14 1.00 B54 029 GOs OR ORE Bw 5s 9 1 5 w 7 7 3 4 8 5 ia x 5 xk 1 3 OX K RX XX 9 RX 5 959 051 0.22 0.30 005 0.24 0.78 063 0.35 0.13 O12 0.84 096 O67 i uo 2 0 5 4 Bw 3 I BB Mw 7 Ku 2 R R 4 ¥ RX BRX KF Pig, 2.1. Using a sequence of randem nuimbe.s 10 generate a ruidora permutiton of £3 numbers 0.71 0.46 031 0.71 0.78 G11 9.11 O41 091 O24 0.99 020 0.53 058 O21 003 woe 3 xk l 2x 62 4 1B 3 7 9 KF ie. 22. Second erste! of wsing 8 Sequcnce of cardum mmbers (9 garenale a randeni permutation of 13 numbers Figure 2.1. To urn these ino the numbers 5-15 ina simple way, keeping the uniform distribution, multiply cach eatdam number by 160 and subtract multiples of 20¢0 leave a number ia the range 1-20, Crore oul rumbsrs bigger than 12: thogo are madked Xin the figure. Remove any number thal nas already occurred: (vese are marked K in the figure, Coatinue to produce random numbers until 42 different numbers have boon Fisted, Thea (he final number must be the aissing one, 80 H eau be written down without further ado, ‘Thus the sequence of random numbers giver in Figure 2.1 gives us the random order the top row Bsiswyxene24 78 Although dhis process is simple, it usually requires the genexation of for more ransom sumbers than the number of plots being permuted. In the second method we generate only fow more random numbers than, the nimbor of ptots, because only the exacl repents are crossed out, Taen place a under the smallest number, 2 uider the second smallest, and so ‘on, zoplncing each random wiiaber by its eank. This process us the random order Ws 512624 I7TL 2.3. The treatment subspace Definition The function P:0—+ ‘Fi called the reament facion. “There is an N-dimensional vector space assoctaied with 2, IL consis's of W-tuples of veal rumbers, with each place in the N-tuple associnted with ane plot. Formally this vector space. SE, bucie willbe called ¥ for ihe reiwaluder ofthis book. Because data sheets usually have one row forcach observational unit, with each variable as.a column, yeetors in this book are usually written as coluran vectors, Definition ‘The rreanmen subspace of ¥ consists of those vectors in ¥ which are.constant on cach treatmenta Chapter 2. Uncructared experiments Table 21. Some vectors in Excanple 2.2 a7 Tieed Dus | some restos | crenogorat | UEP? ued veour weer | intr | besisfryr | SRO ale nos fm uw uel @ 4 ee) Notation Since the treatment fectoris 7, the treatment subspace will he denoied Vr. Defi Avector vin V isa treatment vector iv & Vos ili a treatment contrast if v © Vr and Tuer Ya = 9. Example 2.1 revisited (Fictltious) ‘The teft side of Table 2.1 shows the set 2, the teeaunent factor 7, a typieal vector v in V and the data vector y. Beyond these are sors veotars itt the eeatment subspace Vp. The vectors are shown as columns. The veetor v3 is @ treatment contrast. Notation Hor cach weatment i let be the vector whose value an plot is equal to {i itT(o)=i © otherwise. ‘Table 2.1 shows the special vectors tl, thy and Uc. clear hat every vector in Vr is a unique lincar combination of the vestors tty «4 0. In Toble 2:1, 4) =Ula+ SMe, Vo = (1/S]uy and vy = (1/34 — (1/4) Recall the sealar predict: for vetoes v ani w in V, the scolac product v: w of v and w is Uefined by vray dogs boob rte = vO, where Y denotes transposo, In parliculan, ¥-¥ = Tuyen vf, which is sometimes called the su of squares of ¥ and sornetinies the squared length of v: itis otso written as ||yf. 24, 0 wet it tour. Propedl Propod basi fol Cart Al ouplot the fata, torn in Tabs 24 Ontos to ran varia wi) iaiments evr rent ment sunt Bibs store yu 2.4, Osthogonal projection B ‘We say that the vector v is orfogonat to the vector w (wien VL w) if vw = 0. Tn Table 2.1, ugema = Sand up-uy siete = 4. Also, Uy tty = 0, 80 ta is exthogoral top, Propositfon 2d Fer each treatment fw uj Proposition 2.2 ffi j thet uj 4.uj aud 50 the set of vectors {uy 1 T} fs en ovthogonal esis for Vr. cr Corollary 23 if there ove trecanenis hen dim(Vr) =, Also in the tzeattuent sutspaue isthe voolor tof unkeown treetment parametors It value ‘on plot avis equat to Trg). Under the linecr model assumed ia Seation 25, B(Yo}= Tr(a 50 {he fied valve on plot o is equal to tho estimated value 474 of fro). Thus we have avector ‘tof fitted vaiues, which is also in the treatment subspace, Taese {wo vostors ave also shown in Table 2.1. 2.4 Orthogonal projection Orihogondlity is important in statistics, purly because orthogonal vectors often ecrrespond 1p random variables with zero carlatie. Many procedures in estimation and-analysie.of sariance are ncthing-more than the desorapesition of the dats vector into ortaogonalpicces. Definition [fW isa subspace of V then the orthogonal cenplement of W isthe set {ve V: vis orthogonal to w for ail win W}. Notation The orthogonal complement of W is written W, which is pronounced ‘W perp’ ‘Theorem 24 Let W be a subspace af V. Then the jollowing hold. (9 WH isaibo a subspace af V. Ga) Qty ew, (80) dien(¥r) = dim ¥ — ain, i) V ie the itera direot sum W@W; tis nears ha given any vector Vv 1a ¥ there is unique vector xin W end a unique vector 2 ie WS such that Y= X+2, We oa x the orthogonal projection af ¥ omto W. aud write x= Fy (v), See Figure 2.3, Onehogonad projection isa linear transformation, so Ry is effectively an ¥ XM mates, 50 we often write Pyy(x) as By% #) Py) (vi) Fora fixed vector vin V aad vector Ww iit W, Exc (¥ey — Wa)? = |v — Wi]. As we varies over W, this ston of squares of differences is wininstced when w ayoxsyoPyy, (0) yy) an odhogondl Bra for W tor Gane ger (Se) By4 Chapier2. Unstractured experments Fig. 23. The vector xis he anthogsal projection ofthe vector vonto ths subspace W 2.5 Linear model . Forunstructored plots we assume that Youtz, where & Vy, E(Z) =0, VarlZ») = of forall «in 9, and cov(Z,%4) = 0 for different plots ‘and B. In otter words, E(Y) = x, which is an unknoven vector In Vp, and Cov(¥} = 04, ‘where Lis the Nx identity matrix Under shese assumptions, standard car model sheory gives te following results ‘Teorera 2.5 dissuane that HCY) = and that Cov(¥) = "I, Let W be a ativiensional subspace of ¥. Phen * 0 BRYY) = PelBO) (2) (Ry ¥?)= [wt + ot. ryt ‘Theorem 2.6 Annu hat E(Y) —-€€V; and thes Cov(¥)= 0%, Let x and ube any vectors tne. Then (8) she best (hed f, anim variance] finear iabived estimator of the scalar x:Aivx-¥: (ii) the variance of the estimator x-¥ is \x|*0?; (i) the covariance of x-Y and 1-Y (x z}o. 2.6 Estimation Proposition 2.1 shows that, ifthe uy a the vectors defined in Seetion 2.3, thon uy-my = 1% Moreover, if v i any other vector in ¥ then wj-¥ is equal to the sum of the values of vom (hose plots with treatment, In particatar vg t= te. Write SUM forthe sum of the vatues 26 at¥9 reatnt Abo, si Then’ whore thegd TH w¥ estime, Si ‘Theor Pxenl dies | They | euctid T suns! estaiments plots =a, ‘ors pie ikea Ste cn sbi q 2.6, Essimation 6 of Y on tho plots with treatment é, and sumyoy forthe muon ofthe values oy en the pls, with freatiment é Then we¥=SUMn, and Also, write the means SUM —i/ty and sonrynz/rras MEAN; aud mean. respectively Similarly, let ug be tae all-1 vector; thet is, up = Yi. Ur. For every woetor vin ¥, write v= Delt. om ‘Thenug-7 = Eyen Yo 7 for alvin ¥, Tn particular yt nN, g-¥-SUM—NY and = sum == NF, ‘where SUM and sum ave the grand totals Tiga Yo and Dagan Tespectively and P and ¥ ane the grand meaus SUM/N and sum/A respectively. . ‘To estimate the treatmeet pararicter 1), put x= (I/rzuy. Then x-2 = x- uj/ry = 7% and SUMrai/ti=MEANp=i. Therefore Theorem 2.6() stows that the best Tincarunbiosed, estimator of tis MEAN; sc, with conesponding esliniate 4; equal te meattr- ‘Similerly, to estimate & Tinear combination such as Ft. Ati pula = Bt oro fu(aes) Bae dufryu. Now and a(dac¥) Meat ‘ 0 Listy is the best lincar unbiased catimate of EA. Tv panicle, put @ = Pt 1rG/M, which is the linesr combination of), .... % whick tas Ig = aN. Then x = (1/N}ua, 80 - SUM Bysree anil (his iste hest linear unbiased estinator of. Now we luk at Theorem 2.5 with W = Vp. Since t€ Vp, we have By ‘Theorem 2.5G) confinss thal this i an unbiased esfinatyrol z Example 2.2 (Milk prodaction) Au experiment to compare the effects of” theve different «iets on milk production usedl 32 Holstein dairy cows at siutar poims in cheir lactation cyctes. ‘They were fed the diets For thrac weeks, During the third week, the average daily alk pro- duction was recerded for each cow, These data, in pounds per day, are shown in Tattle 22. Thos 14 = re 8. Furtheanore, sumraa = 6600, 80 4 = 60.00; sumyag = 7234, 30 168.4, so 2c = $8.55. Therefore the best estimate of the difference: A between covson dicts A and 8 is 4.35.6 Chapter 2, Unstiuciued experiments, Table 2.2, Date for Bample 2.2: mith yield per cow iu pounds per day LietA | 607 59.7 619 GIB G26 G25 579 59.9 596 STI 572 Diet| 55.6 529 527 5h) OLS 380 OOS M3 392 3E2 S05 S60 349 viet | 628 558 560 62.3 61.1 549 605 55.9 2.7 Comparison with matrix notation Some authors use maices (o express linear models, Let X be the A xt matrix defined by Ref wh where uy, .!., ty are repardel as cohiran vectors. Thos, in Heample 2.1, X consists of the Shvee columns of Fable 2.1 headed ws, up and uc. Let Bi be ths vector ft1,ty...y%]. Note that fis wot the same as ¢, because B has feolries while © has Thea * = XB, and so the nr Foxulation of the expectation ancl is 2(¥) — KB, whove least squares solution it « “1 f-(x'x} vy. Qn Yow, wr-ur ar no yu ap-my a wun Ut: wt oo and so . 4 ° 3 ° (x74) - : ° o * Morcores, ok] PSU ee wy || SUNT yep EP ue¥ SUMz— Hlence Equation (2.1) gives t= mean for soa fs Which agrees with Section 2.6. 2.8 Sums of squares Definition Let 1¥ be any subspace of V. The sven af sgtares for W means either [fir ¥||” ar Wewell?. The degree: of freedom for iV is nother nae for dim W. The nrean square for W is sunt of squanes for W ‘degrees of freedom for W* 28, Say for othe expecta Fira the sum Now becaus Si sol, associ sofiments 519 ff? or ido nc pani wnlvaneateamat iat 28, Sums of squares 2 for ether sense of sua of squares. The expected mean square for W, writen EMS(W), is the ‘expectation of the menn square in the random sease: the is Ei\Py¥ 7 exw) — ERE, First ive apply these ideas with 1¥ = Vr. Sinee 1..SUM, ¥ Gr) ES) whenever if j.40 this sure of squares Now, uot ‘The quantity Y(sunt}. fri} is called the crude sim of squares for treatments, whieh may be abbreviated to C88 treatments). ‘The nursber of degrees of freedom fer Vy Is simply the dimensicn of Vr, whic is equal tor. ‘The ancan square for Vp is equal Theos 2.5¢i) shows that EA(|Pog YIP) = Pevyell +664 = [ef +40? ~ Yo a to, because ¢ = Zi, yup. Hence the expected mean squire for Vp is equal to Drv jt +02 Secondly we apply the ideas with W = ¥4!. By Thearsm 2.4), s- Sotee dala vector vector of filled values Bey = Pay = residual veotor, 0 [[Pypyl/* is equal to the sum of the squares of the resishuats. For this reason. all the quantities associated with 1 are named ‘residugl". (The word ‘er? is sometimes use, but this can be centsing to nou-statistcians, who tend to interpret is “mistake')28 Chapter 2. Unstructured experiments Now yiethe sun ofthe erthogotal vecion Fy, ¥and iy 0 Pyagonsss Theor shows that a Re 2 eae Yxb= WP = [Baal + [Post wea Ta “The quantity Yea is jul the foral sum of squares, x0 the sum of squares for residual i qual to the diference Serwoen the total sum of squares aml the crude sume squares for treatments; indeed, tis is usually the easiest way toerlentate iL "The munnher of degrocs of freedom forvesidul, which may be writen as df(residual} is equal othe dimension of W, which is Wf, by Theorem 2.4Gii). Hence the mean square for residual is equal to sum of squares for residual Net “This will be denoted MS(residual) “We know that Piyt = # because 6 Vp. Thus Theorein 256i) shows that (Py YIP) = (¥—1)6? and hence FMS (residual) = 6%. (22) Example 2:2 revisited (Mlk prochtetion) Here he erude sum of squares for Weatmeats is 660.F 773.7 | 468.42 ta ta ‘The toral sump OF squares is 60.7? +--+ 55.9 ~ 1O7515.62. Therefore the residuum of qos is 107515,6200 — 107279.2477 = 250.3123. The number of Uegiees of ficcdom for tesidvil is 32-3 — 29, so the rosiduat mean square is 236.9723/29 = 8.1508, = wi219.2477. 2.9 Variance ~ ‘Scetiog 2.6 showed thatthe bes linear unbiased estimator of the Vicar cominaiion Lavy is Xe¥, where x= L(hi/noy. By Theorem 2.6(8), the variance of his estimator is equal to IsbPat. Now “(} EV. Bn ‘Two cases ore particulsely impartant, ‘Yo estimate the eatmeat parameter % far a fixed Vand iy ~+0 for j 4. Thee the varineeis 0/1, To cotimate the sample .Ay= Land he @. Thus6 ‘Chaprer2. Unstructured! experiments Table 25. vattes ofa sucit that PrIX 20] | <001 al = Felt < =a] = 0025 and Psfx > a] = Pix PHO< X < 2a] =05 Pr > 2o}. Table 2.5 shows tha! Prix > 2a] is neatigisly small so p is unacceptably W8/ VAT is bigger thane then a 6/ WaT < —2a, Therefore Ns PAX 2a] 0, and v0 p = PAX < a3) fo) - Pel < ma ~ B/ Vil] = Pek a. The solid curve shows he probability density function of a ¢ random vasable on d dgrecy of freedom. ‘This defines the interval [2,4] used in the hypothesis test. The dashed curve shows the probabitiy density function ot Af VOR. With probobility 0.9, A/ VUE > 8/ A0T —b, and we want this lower Timi to be outside the interval (—e,a). Hencea + < 6/AT. Replicing Py its expectation 0? gives (aroyv<(5/o% eo Conkider the ingredients in Fauation (2.6). We assume that [8] is # known quality, the size of the smaflest difference that we want to detect, Th vaviaice ¢? is assumed unknaven, bur previous experiments or similar msterial may give « rough estimate for its value. In more covrslicaued expesinonta (ave Sections 214 ond 446and Chapters 8 and 10), we shall ned to replave 6? by the appropricte stratum vatiance. Ii many experiments the variance v0? of tho estimator of 8 is given by » = I/ry + 3/2, but Chapter UI shows thet « more complicaied formuln is needed in non-orthogonal designs. The values « and & deacad purly ou the awmbe: ‘offdegrees of freedar, which depends on the desig. They also depend an some ole choices: «con ihe signidicance level of the Lest & on the upper limit of acceptability for p. If we have even a rauzh ides ofthe sizeof fd /o, asd have set ihe value of a by choosing a siniticance level for the toss, then Equation (2.6) gives an inequality to be satisfied by » and b, There ace two ways in which this car be used Jn soine areas, such as agricullural research, itis pica ta propose the nuarber of trent- ‘ments and thelr replications first, according to resources available, ‘This gives a value for ¥ from which Equation (2.6) gives an upper bound for b, froma which tables give a value Far p. I¢this value is accsptably low the experinnent procoods. 1 not, a modified experimont is pro- peséd with a smaller value of v, usually by increasing resources or omviting loss imtaresting teaunenis. In other reas, such a5 cliical trials, i is more common to set both the significance level and the powsr (which is equal to Ip) in advance, Then, assuming equal replication 1, Bquation (2.6) is used to update vehucs of ratd d alvecnaely enti convergence is achieved2 Chapter 2. Unstructured experiments Example 2.4 Calculation of replleation) Suppose that thre are to Ireaunonia with egusl replication r and that [O]/G= 3. Then v = 2/r and Equation (6) gives eoFaray. Start with d =», for which the tdistibution ja standard normal and 20 a ~ 1.960 and b= (282, Then FSBO 1 280)2/9.00 2.3 Thon ¥ ~6 and ‘Take 1 to be the smallest value that salisfies this inequality, namely r d=4, Repeat the oyele, Now that d = 4 we have a= 2.776 and b = 1.533. Hence > 22.7764 1.393) /9 5041 5. Then d now value of d gives a’=2.206 end b = 1.297, s0 ¥ > 2(2,206 + 1.397}7/9 3.04, Thue we put r =4, Then d =6, Now a= 2.447 and b= 1.44080 so pat ‘th 1 2247 4. J.A40)2/9 0 3.4. “Thisissatisied by the current velue ofr, and weave already scun that the valve immediately below does not satisfy Bqustion (2.6), so we stop. We conclude ht sight experimental uaits should suffice, Note that power can he increased by including extra treatments (because this ineraaces d} ‘but that thisdocs not aber the variance of tncestinatorof a difference benveca lve Leas, 2.44 Amore general model The chapter concludes with a slighty more genecal xodel than the one ii Section 25. As before, we assume that B()=tey. However, we change the assumption about covariance 9 of ito pot ita zB. {nother words, the comelation between responses on pars of aiforent ples isp, which may stot be zero. Complete randomization justifies this assumption, as we shal shew in Chapter 14 Write = Cov(¥}. Under our nssureptions, . cov(Zon%) = { Cao'l4 psI-D=o'((t-p+e4l, 21 Al Nate soit x and Je 5) a soupiss Txe ‘The this a8 change Ten at of Cw relevant ‘Theores Howevs tweachs tea by ‘Theore, ay ad iy) whovens equal b= 6 and ately ‘units bes d) ens. may ler 14. te ab icici ti ee et ses | 4 2.14, Amore general model where Jig the M27 allel miata, Note that ay 4 8 su if xis any vector in V then every entry of J is equal to n+. In particular, Jup = dup, and Be = #1 Uo, Of course. [lg = wo. Therefore Cay =9"(1 —P+NP}e0y 0 ap isan eigenvector of C with eigeavalue o°(1—p+Np). IWxe V and x L wp then Tx = xand Jx = 0, 20 Cx=o%(1—9)x, and therefore x is an eigenvector of Cov(8} with sigenvalee 6%(1— 9). “The results from Theorem 2.5 onwards have assumed that C isa scalar matrix, Changing, {his essumption makes no difference to expectations of livesr functions of Y, but it docs change the expectation of quidratle functions of ¥, that is, sums of sqeacs, If =o? thon oll formulas for variance or expected mena square. inyolye 07, If x is an eigenvector of with eigenvalue & then C.acts oa x just Uke EL. Thus cavefil ropiteernent of o? by the relevant eigenwalue gives the correct results, There is one possible dificulty in goseralizing ‘Theorem 2.6¢i) to the case when x and x are eigenvectors of C with different cigenvalues. However, Cis a symmeltic matrix, so eigenvectors with different eigenvalues are orthogonal to edeh other, so that x-2— O a this ease. The other places where different eigenvalues right cuir are the generalizations of Theorem 2.7() and Thcorem 2.10Gi): we deal with both of these by restricting the cesulls to-cigeaweazors with the sare eigenvalue. ‘Theorem 211 Suppose that E(Y) = TE Vr and Cov(¥) = C. Then the following told. (0) PW is any subspace of V then B(Py(¥)) = Py fii) IEW consists entirely of eigenvectors of C with eigenvatue E and ifdimW =, then HPAI) = [ECO (ili) Ifx € Vp end x43 an eigenvector of C shen the best lincar unbiased estimator of x5 ie ey. {ivy ffx is x eigenvector of C with eigenvalue & then the varkance ofx¥ is xl) (0) Suppose thot x and + are eigenvectors of G vith elgenvalies & and 9 respectively: If Bn then cova: xn} if 1 then cov(s-Y.z V) = 040 Chapter2. Unstructured experiments “Table 2.0, Anaiysis-of-vaviance tnble for unstraciured plots aud wistercuared treatments under the more geeral node! Steam | Source | ot EMS ¥% ‘aciw | mean [1 wea Lede i Yt “plas | teewments +e Total ” 601) Suppose that xis an eigemvecior of C with eigenratue thal x= Ys(Oy/rilMe that W is a d-istensional subspace consisting of eigenvectors of C with éigenatiee&, and thet W és orthogonal ca Vp. YY has 2 nulrivertave norut distribution then xY-E Wi i) x MS(W} has a \odisirioution.on d degrees of freadons and S5(W}/8, has 0 42-distribution ou d degrees of freedom. (oii) TF nel Wo ate subspinces with dimensions di and d3, both consisting of eigenvectors Af C with aigenvahce 8 orthogonct to each otter with Py t= Uy, t = 9 anal if heey 4 mutivariate siormel distribution then MS(W}/ VAS(W) has et F-distribution on ds and da degreos nf freedom. Definition A strate in sampling). an eigenspace of Cov(¥ ) (note thet this is mut the same.as a stratum ‘The analysis of variance proceeds Just x8 before, except thet we first decompose V into the Uifferent strala. Under the assumptions of this eeetion, Vp is one stratum, rension | ard cigenvaluc 7(I — p++Np), while Ye is the osher straw, wilh dimension N—1 and eigenvalue o°(1 - p). Call. hese cigenvatucs go’ and & respectively. We then obtain the snalyeis-of-verinnes table shown ia‘Tuble 2.6, Now we caleulate the variance ratio only for tcrms with the sane eigenvalue. ‘There is wo way of estmaling fp, and hence no way of saseasing whethce Fis (statistically anlly) different from revo, and no way of estimating the Verlanee of the estimator of aay reatmect parameter %. Howetey, all Irainent contrasts ate in Vso thoi noe ‘combinations and te just as before. Hxperinians in whieh we are interested ouly in trcatrcent contrasts arc walle comparative experiments | acs Ques BEAL the stud Fico) wee] alloca ‘watehe! Der to sud 22 Ai thed Swans rel Like to| pint tank anda He! 30 cor, ‘Ad | 23a] soil a onder | of ill wo peat! 24 Al ould lestatt 190, fe valuefimenis vows pW a id at foe on lactors IY kas Jon dy foturn tothe ion 1 F ana tw the ally nator tinea hwo a a “4 Sica soutinos oiying s paananeMceabnbbachecdine aM cteidathatde, ‘Questions for discussion 4a Questions for discussion 2.1 A psychology course tas the 21 students shown jn Table 2,7. ‘The profestor wants 10 use the suidonts to tet wo now Lypos of gill for kozping poop awake, called Wekey-Wakey and Ziaaaway. He tas only six pills of Wekey-Wakey and five pills of Ziazaway. He plans to use eleven students, Hach student will be shut alone in the observation rcom, swallow their allocated pit, and shen Follow a set programme: of activitics until they fal wsleop. A hiddon wwaicher will record when they swallow the pill and when they ‘al asleep. Design the oxperiment for the professor, to tho oxi of giving him a plan sllucating pile (o tudents, 2.2 A marine engineor is investigating woys of treating the standard metal componenis used inthe construction of underwater structures at sea, such as piers and vil-deiing platforms; Me. ‘wants to protect chem against corresion. A colleague hat developed 1 new sort of paint forthe components. ‘The exginesr would like to see whethor two coais of this paint give better protection than e single cent. So he will paint scme metal components once, some twice, thea inmerss them all is his experimental tank ofsen water. ARorthree months, he will remove alle metal compancnis frm the (nk, sand measure the amount of corrosion on exch. He has a virtually uaiimited supply of the meial eompuncnts. The tank bes wom for apts ‘30 compenents. However the pain’ is nase, and there is cantgh for only 24 coats of pent ‘Advise the engineer how bss: to use his weaoutves in ns experiment, 2.3. technician has to measure the ceidity of four scils. You give him three samples of each soll and ask bin to make the twelve measuremeats fa random order, He says that a random axder will confuse him and that i wil! be beder if he measores the acidity of all three samples ‘of sgil A, then all theze samples of acil 2, and so on, Make avtes on arguments you will use (o persuade him that a random order is better. 24 A physician wants to test a now doug fo compare it widh the current standard drag, He would like to have 90% power for detecting a difference of (5 units {Tue does a hypoihests test st the S% level of significance. He believes that the value 07 62, in the poptlation, is about 109, for measurements in those uults, IF be uses x people for each drug, what isthe smallest vyahie of rthathe should use? ‘Table 2.7, Available students in Question 2.1 Nave Sex Age NaWe Sex Age Name | Sex Age Adin M19 Helen 20nd 20 Belinda «=F 20nd" 0 Peter M8 Carole F198 dames, M20 Quentin MOF David = MoO Po Rukh Bo Baker FO BOR Gmbh FO Fina = B20 Michael M20 Tunis 20 Greory M19 Naomi FI Urals 20a Chapter 2, Gnstractured experiments, @ 3 24 16 TPS Pay sys| ca wl 7] 4| sofa 2 2(7]altlat7l2l4 | _7| 18] 30] 18] 16 [a6] 4 sPifefelifefiq2 otal ifis[se| sfc] 4 Hig 27, Field plan in Question2.5 25 A Gompletely eundomized experiment was Conducted to compare seven teatmerts for their effectiveness in reducing sob disease in polatoss, "The Hild plum is shown in Figure 2.7, ‘The upper figure in each plot denotes she tretinent, coded ~7. Ti lower figure dovates an index of scabbiness of potatoes in that plol: C0 potatoes were randomly sampled from the plot, foreach one the percentage of the surface area infected with seals: wat assosted by uy and recorded, aud Ue average of these 100 percentages was calculated lo give the scabbiness index. @) Give the analysfs-of-variance table-fer these data. (0) Is there any evidence that the mean scabbiness is different according to different treat- ‘ments? Justify your © Baimace ts mean seatbiness produced by each treatment. (@) What is tue standard ercor of the above estimaies? © Wh the standard ercor of the differences between means? Chap Sim 34 ‘Suppose which From B For giv Put se that without which | conty wl lope! rust Sor sume shealt have tr treatmeaty its for 27. tes an ma the yeye treet 4 a “| | 4 Chapter 3 Simple treatment structure Tren 3.1. Replication of control treatments: ‘Suppose that treatment | is aecntro! Westmont and that restments 2, ...,# are new trextment which we wantto compare with the control. Then we want oestimate tj— 6, for F From Equation (2.3), the average viriance of these estimators is equal t0 Litieje ( ! Ee For given valics ofr) and N, Preporiion 2.8, with R = A —n, shows that chis average variance is ritinolzed when ry = 43> =f = (8 n1}/—0). Pur =r and T= so thet weneed to choore r wy minimize g(r). Now, gis dtferer without limit ag 7--> O and as r— 17. Moreover, 1 eo Ware tol roe we eon (0,07) and increases, #O= whic is zero when, and enly when, 2 = (W—79/JF=1. Thus afr) is minimized when, and only when, ¢ = (¥— ef vF= Ts that is 2 (te Un/ViRT = viv Tn. Guy In practice we have te use approximate sofations to Heuafion (3.1) because ali the replications aust be integers. ‘Sometimes there is more than one conical ireatment, and we went (o compare every new: treatment with every control Wootinent. Proposition 28 shows that al te enol eestrents should have the same replication as ese ar, say cy, wht al the new treaxmens should have the sume replication as each other, myn. tTthere are contol treatments and nr new treatments we then have mrt = N, and nged ( winimtzc mt rt a ne a4 Chapter Simple treatment structure subject to this constrains, Put Nowe alr) Then #0 wowP vwhishis sno wheat c= (W—ar}/ yi. Thus th avenge variance of eimatot of erences between control treatments and new treammenis is minimized when Vary = Yair = iN nr) /vin= fain. 2 ‘Example 3,1 (Example £17 continue lilseed rape) In experimehi there were hwo con teen aidlght oo troutment. Equion (K2) gives Y2n, — VB, 9 the replication ofthe cones siould fave been lice tat ofthe new teaunets, hn fetal ireximents were elie! o tw plats each Perhaps the comparisons belwaen pare af tow ‘teametus were deemed as inctng a5 those between new tealeals and contol! There are some experiments with a single control restrient where it is known in advance hat (he control Ucaumrent is inefteclive, In such circumstances, comparisons beiwcen new Irealmemts are more informative thé ny emmparison berween & new treslement and the con- traf, This suggests that the oesr treatments should have highr seplication than the control. ovever, the person wito wants fo include the control trentnteit probably waats lo compare all pew treatments with it, rather than among themmeclves, so will want higher ropliestion for she comtrol. You may have ‘0 compromise on equal replication, 3.2 Comparing new treatments in the presence of a control Suppose that ireatmeot 1 is a enetrslireutment ard chat troataterts 2, ...,# are new treatments. Ruther shan asking ingen for 2, +f W9 could ask the tove questions Giemsu= Gi 871 equal to the average of ty... ‘To-test the autl hyporhests Hy agalast th alternacive hy pothesis Hy oy 5% ure not all egual, we calculate the mean square forthe non-control treatnenis, This isequal 10 SS(now)/(¢—2), where 2 7 a y sum, sson)= ferme _ stm This il such tH We ob) wren] Th +L! Define a) Thusceture fences @2 3 Wwe the ct, all now new atta, pare, bn Zar fons, =), 13.3) 3.2. Compering new treaiments in the presence af a control 45 Elf) — sr BW) evr Mot] , aired (OOS! Bev ro E(t) eqyevs Ba) =0) Eye {0} Fig. 3.1. Four model: tor the expeceation of ¥ when one treated isa coniral ‘This te then compered with the residual mean square. Uf we decide that theie is a constant such that) =¢ for == 2,...4f then it is reasonable to go on and test the null hypothesis agabist the alternative hypothesis Hien #0. We obixin the sum of squares Tor this by pretending that all the new trealments ate a single , 2 aD, Heals? suet 4) ‘These tests can he related to 2 chain of vector subspaces similar tothe one in Figore 2.4. LetC te the ‘conteal” factor cin 2: itis defined by cc)={! 17 2 arherwise, Define the subspace: Ve of V to contist of all those vectors ¥ far which va = vs whenever Ce] =C(B). Then dim Ve =2 and Vac Ve G Vy. Further, deftie We =Venve, which has dimension 1, and We avrnve, which has dimension ¢ —2. The the null hypothesis Hy corresponds to the model EQ) EV. If we bave accepted El then Hg correspond to BO) eM, which is PCB) ‘Thus we obtain the chin of models shown in Figure46 Chapter 3. Simpte treatment siruetuce Table 3.1, First hee cofanes of the analysivoofiverionce table when there is one contral srecconent anne enher structare Sumof squares Degrees oFfeedom 88(rmean) U Source meen ‘onal ‘C83(contral) —C$8(mean} t new teatrients | CSS{reutronts} — CSS{cantcal) 2 resicual by subtraction ‘Towel Lovd ‘Totest Ho, welook atthe size of Py,y—P),.¥, which is equal fo Py, ¥. Since Dey —meancanm +meancaa( =H), we have 8801) = fray] = Weal (sums asin Equation @.2). In words, he Suny df squares for new treatments = crude sum of squares for Heatments rude sum of squares fer commok, Similarly, to test Hf, we lonk al the size of Py. y, whi is equal to Ry.y — Py. We have leet? leas? SS{Wie) = Presll? 1s in Equation (3.4). Ia wor te crude sum of squares for control --sum of squares for tne mean. ssum of squares for control ‘The first three columns cf the analysis.of variance table are shown in Tablo 3 Now ittar $S(conire!)-+8S(cew reatments) = S5(eatments), s¢ we have decomposed te sum of squates for treatments into two mesningful parts, exch of which is used to investi- gate a clevont hypothesis mae ioepl tn The i the 13 Tobie 4 4 collect There! aed the inetd possib!sere ytred have vased vesti- 3, Other mestment groupings 4a Treble 2.2. Detain Example 3.2 Total FungieideA [293 06 387 BO] 1762 rungicites [352 253 36.1 31.5 | 1364 FungicideC ]35.7 39.3 368 28.0 | 1398 PungiciteD [44.6 306 312 97.1 | 1195 No fungicide [20.9 22.7 278 18S] 999 55 ‘Table 3.3. Analysis-of-vertance table for Exampte 3,2 Soures ss fal Ms ve 19560. 5105 | 1] 19502.5125 | 1667.43 conirel | 387.2000) 1] 387.2000 | . 33.00 fungicides | 24.7750) 3] 8.2583 0.70 wsidaat | 175.9825] 15 | atzaza | — "Tuat,|2OIS0.A7O0 | 20 Example 3.2-(Rungielde on potatoes) Four fungicides (AD) werecompzted with each other and with » no-furigicide control for their effect on potatoes. Bach treatment was applica to four plots. Tuble 3.2 shows the yields of potatoes in tomnesihectare. The dita have been rearranged inte {reetmeal arder for easo of showing the treatment totals. ‘The total surm of squares is 2150.47. S95. B90 625.5% + ss(connot) = cw 387.2000; 1262? 136 | 1398 | 1525? ” sas SS(iungleides) = "= Ee SE 2.7150, the residual sum of squares is 2015047 — 625.5820 — 387.2000 — 24.7980 = 175.0825. ‘Table 3.3 gives the analysis of variance, 11 is evident from the Inble that the fungicides differ very much from ihe contral in the effect on the yicid of potatoes, but that there is hile difference between the Four Fungici ‘The relevanl means are Bungicide No fungicide SATS (22.475 /TWIDIR SING, which is 1.915. and the standard error of their difference: 3.3 Other treatment groupings ‘The treatments may be grouped into two or nore types for reasons other than that oné (ype -vousists ef eontrols. The consequences for replication and analysis depend on the reasons fat including the different types in the experiment, The foltowing exempts illustrate some of the peasibitties,4“ CChopier 3, Simple treatment nruotare WKY} depends cai Vatlery Vr dimension 7 1R(%2) is constant | dimension | 4 Fig. 3.2. Four moses for expectation mn Example 33 Example 3.3 (Rubber trees) Seven ver compare ylelds of the variedes. [chappers that the varieties have (we visually differeat feal> types: three of the varieties have round grean lenves Cleaf-iype 1) while the other four have Tong serrated grey leaves (leaftype 2). of rubber tree sce planted in an experiment to Vario [1 6 Cestiype}2 12 1 2.2 F ‘The main purpose of the experiment is to compare the soven varictie, s0 they shoutd be equally replicated. However, itis useful © ask if any differences among dhe varieties can be ‘exphied 28 differences between the two feal-types. Thus the analysis should fit the choin of Imodels shown in Figure 3.2, where ¥7, isthe subspace of Vp cansisting of Veetons which are ‘constant on exch leal-type. Calculations sivallar (0 dose in Section 3.2 give the pavtisl malysis-of-vaiznee table shown in Table 34, where r denotes the commen renficaion, There are five degress of freedom lo investigete the question ‘Do varitien differ within leafype?” and one degree ‘of freedom For the question ‘Do the leaf-types differ?” sample 34 (Drugs at difforant stoges of development) A company which develops and manufactures phanmaceuticals was to compare six wearmneats for a certain disease. The initial (cial will use: hexttay volunteers, simply to measure the amount of ceriain cheniicals released iuto the blood two hours after the treatments arc aduninistered. Thece of the treat- ents are three different deses of'a fonnalation (coded A) that has been under development for some time, The ofher three ere thre differant dsee of « aew formulation (coded 8) that has not been so extensively studied, The main aim af the tral is to compar: the doses of formulation A: the secondary aim is to compare the new formulation with the old one; snd she lowest priority Is given to comparing doses of the uew formulation, ‘The main effort in the experimen! shoutd go into comparing the doses of the ok! formu tion A. Sufficient eplication must be used for these thee treatments, guided by the principles inScctions 2.| Gard 2,13. The gampany decides o use 12 volunteers foreach dose, hus using 36 voluntacra, However, it has walficiznt renourcea to uce 48 valunleers, so the thres doses of Formulation 8 are assigned to four valunteets cach. ‘Compared to the more limited design with three doves of A and only 36 voluntuers, the design wid the extra £2 volunteers Increases the pricision of dhe estlmate of @?. Iealso gives 3. ot the dos Nori re ora, 4 the avel precice! compat Nol ad we t ait) 1 Deracwore meat to tear lr have | sould be ican be shain of nich wwe = ule rees-of degree ps and, fe The fe cane opment 2) that wes of brant ciples i using losos of the Lo ives vend atone toearep oaubdthiebs 28 3.3. Other goatment groupings 6 Table 34, Fist utree columns of the analysis-af-variance table ia Example 3.3 Senve Sum of aque Degrees of feeont vem “ess(meay | teartypes | Css{tesraypes) - C88 nes) u uiotien | CSS(oarstin) CSS(teat ps) s vsdamt | - by abt: ~ Tord Tat ” ‘more residual degrees of freedom, snl hence mor’ power for detecting differences belweea the doses A. In addition, it gives some information abort doves of B. Variances of estimators of some contrasts are as foslows. botseeen two doses of A between a dese of A anda dose of B between the averege effect of A and the avornge effect of B between two duses of B IC the company is correct in is judgement thal ceplication 12 is suiciect for the descs of A, then comparisons aniong, those aze sufficiently precise, ut is the comparison between the average etfect of A ind the average cfect of #, Comparisons between doses of B are less ‘procs, bul may yield visstul informatio if there ar large diferences, which should help the ‘company decide whether to procezd with the development of formulation By Nowit isensibie o split the sum of squares for treatmenssinta farce parts: {0 the sum of syuures for differences botwore formulations, which is ealeulnted Tike the sum of squares for leaf types in Baample 3.3; Gi) tho sum of sqyrates for differences between deses of A, which is calcufated as in Fqua- tion (3.3) under the pretence that all deses of 3 are a single control; thee sum of squares for differences between doses of B, which is calculated as in Pgua- jon (13) under the pretence thst sll doses of A are asingle vonttol. Nefine treatment factors FA aod B as follows. ld formulation New formulation Trentniest yo? 3 a5 6 one von50 Chapter 3. Sinnple reamment structure ‘Table 35, Models jor E(¥y) in Example 34 Coordinate (and parameters) Vector (and subspuce) Elo) eye) mv € Hy te) Cy itroaeney BME wea {i Troha BVH srg P TOYS ey ite) Eye Bie) <2 (x) © {0} 3.3. Relationships among the subspaces in Example 3.4 ‘These cefine vector sparcs Vie, V4 and Vp analogous to Yr au Yo. The relationships between these spnces are chown in Figure 3.3. The expectation medel enmesponding t0 each of these. is given in Table 3.5. Example 3. (Reducing feed for chickens), Au experiments wobecondueted to see ifchick- ns can bo ted a aighily inferior dict in the 16 weeks before slaughnior without affecting thelr final weigt, The chickens are housed in 40 cages of 20 birds each, and foods are applied to ‘whole cages, The ten tesaumon's are shows in Table 3.6, Thus the now-coutel (catmonts eat be grouped into threeditferent methods, each of which has several variants ‘Tae main interest in the experiment is probably in comaparing cach mnvthed with the con- ol teatment, witch is the normal feed. Thus Bquatton. 2.1) suggests that we should have, 1 cages with the contrat and r cages with each method, where ri a Vr, With 40 cages, ‘ve could have r= B and r; = 16 or r= 9 and ry = 13, All dhe comments pettining 10 i each me, replicas Howeve then wel fa prac any bl34. Other treatment groupings a ‘Table3.6, Treanments tu the chicken-feeding experiment in Example 3.5 “Treatment Conwal Reduce protein consi by opie 5% oF volume of roughage of type a 9 exch ruethod should be equally replicated, For the ren rcatments in Table 3.6 this eugpests seplication 3° for all non-contre! treatments 13. for the eonire) reatmient However, if we replace one ‘quantity of protein’ Keatment by mother, ‘roughage’ treatment, then we might choose replication 3. forthe Stine tectments 4. Foe the 2qsantity restmenss 2 forthe 4 roughage westments {5 for ihe control treatment Ta practice, such a range of different replications will minke the desig any blocking (ce Chapter 4), so it may be beter to opt For equal replication, a last Por the ‘not-comro! treatments. Tn this example the sur of squares For treatments should Le spline ve part, wihioh os All be caloulated following the principles piven earlier in this chapter. For generality, suppose {hat there aren; treatments pertaining ro method é ford = 1 2,3: the sum of squares for ihe difference between ihe contre treztent and taerest (1 degree vf freedom); (ii) thesum of squares forthe differences between methods (2 degrevs of freedoun): the sum of squares for te differences betweezu treatments ofthe quantity type (i ~ 1 degrees of frealom) Gu) the sum of squares for the differences hotween trcatinonts of the time type (wi — ogtons of freedom}: (8) fae sun of squares for the differences between treaumene of the roughage tye (ts ~ degrens uf freedom). We shall return to these examples in Chapters § and 10.52 Chapter 3. Simpis treatment structure ‘Questions {or discussion . 3.1 Supyose that there aren contol tximents,oach replicated ry times, and ze new Woat= meats, exch replicated tines. Pod te opuimad rato 71/r ial extent comparisons are of interest except those between conteol tn:atments, 344 Consider the scabbiness experiment in Question 2.5, The seven caded treatments con sisted of one ‘do mothing” control nd six spray (eatmients, 28 shown below. The amounts of sulphur are given in pounds per sore, ‘Tecatment 1 2 3 4 2 6 7 Amount ofsulphur G 30061200300 6001200. Tiening NYA autem sutuny autane spring spring spring Give two plausibte reasons for thé particular choice of untequal replication made in this experiment, 3.3 Re-unalyse the dats from Question 2.5, taking aeconnt of the fic that treatment 1 is & control. 344 Write down farmutas for she various guns of ites in Exampte 3.4, For each sum of squares, also write dows the null hypathesis for which it is apprapriate. 3.8 A clinical trial op asthma compared three doses of a new formulation of a ronchadilator, lnree unrelated doses of a standard formulation of the bronchodileter, and placebo, (0) Define three relevant factors on the treatments. (b) Hence spceify relevant models for the expectation of the response, inthe format given jn “able 35. (@) Shows the roltionships btwesn the corresponding model subspces, in a diagram like the one in Figure 3.3. {@) Hence explain how the treaiment su of squares should be decomposed, and give the smwurer of degrees uf freedom for each part, (©) For cach part of the treatment sum of squares, write down the conespariding question about the trostments. are hal rakes ratueaf teets| ud sven is boca! inthe | whichacture je ext ons ere (5 CON nls of ing inuis lisa som oF wiator, uegiven, an ike sive the. vestion Ee Chapter 4 Blocking SE 4.1 Types of block If the plots are not all reasonably simifar, we should group them together into blocks in such ‘There are three main types of block. 4.4.4, Natural discrete divisions ‘These divisions between the experimental units ore already preseat, IC the experimental units are ness-born animus then liters snake satura blocks, In an experiment on peopte or animals, ihe two sexes make obvious blocks. In testing tags on sows" cus, th ears are the experimental nits and cows ore the blocks. In ch industrial process. a block could be 2 batch of chemica! or ore shed for patt of une process. Example 41 (Insect repollénd) Midges in Scotland are a severe iritant in July and August ‘A researcher amis t0,try ou some insecl repelonts, which ave applied 10 people's skin “Twelve people volunteer for the expesiment, is knowe that people vary widely ln thei= auractive.ess to midges, si the woscarcher usee poeple as blocks, applying different repellents toosch arm. After a fixed period of exposure to midges, the number of midge bites on exch arm of exch volunioer is ressinded, yea units Sometimes there is more shan one typeof natural discrete block. If the exps are halfloaves of tohaseo plants then who'e teaves make one sort of block while the plants make another. In a coxsumer experiment, such as Example 1.11, westers and weeks are bath natural blocks. Tn av experiment in she laboratery,technickans, benches and days may all be biocks. , an experiment is caniod out on plots that hed previously been usod for another export ‘ment then you should consi whether to deom the previous testments to be blocks. This, is because the previous textments may have tet some feslJue that may affect the responses n the new experiment, This type of block is particularly important in experiments on trees, sshich may have to be used for different experiments year afl year, 53a Clapwerd. Blocking | | i! 1! | | HT ld He} | Me} It Hy | Hy WW LW ye a | i Ky Fig. 4.1, Inigaton chunels in abe itce experiment Exannple 42 (reigated rea) Rice is usually grown oi inigated lend, Figuro 4.1 shows 32 plols in a tice paddy tte used for am experiment, Inigation channels branch of ihe ‘mais. ixigation channel, cach one watering a tong stip of plots, These sips, or iigation ‘exoupings, should be considered as blocks, Example 43 (Roud signs) A toad (canspott esearch Ighomtory in Wales wented to inves- igato whether (he proposed introduction of bilingual toad signs would prove distcacting to drivers, They made two sets of road signs, one with Welton top and English mderacath, the ‘other with English on top and Welsh underneath, On exch of severe afiriogns they erected ‘one sei of signs around an off-road test driving track, and invited volunteers Wo doa lest drive, Dating cach drive, a rescarcher satin the car with the drives aad asked a series of questions about fice driving, designed (o evaluate whether the driver's cencentration was lowered when near one of the bilingusl road signs. Sirce ittook sorte lime to erect n sat of signs, it was Kot practicable to ask voluntecrs to drive sound the track twice, once with each sot of signs. ‘When the statlstician came 10 analyse the date, he noticed that all the voluntecrs were cither retied poopie or university otudonts. This in itself was ol surprising, because these are the lasses of people who cen easily be froe in the normal working day. Unforivnztely, the person designing the experiment had net noticed thi, and fad unveitdngly alloceted all of the retired volunteers to the signs with Welsh on (op and alt of the student valunteers to ‘he signs with Engliss on top, There was, therefore, no way of telling whether the recorded Alifferorces in concentration wert dus to an inherent uifference between relied people and university studerts or were caused by the relative pusitions of the lwo languages on the signs. ws wy wo clas 442 | anes, tying ts of plots: bound ai variable donot have as vit: Exany soll st goos or] or hu cone ond encom om a carhol of varie OW Ia 44 | Some & these dil ina doctors The and pla: of the exam aftreat th. lange aoking 4.1, Types of block 55 IF the researehers had thought'about this problem in advance, they should have used the Uo clastes of volunteers as blocks, 412 Continuous gradionts fan experiment is spread ou! ia time er space thea there will probably be continuous under. fying tends bur no natural boundaries. In suck cases the plots can be grouped into blocks Cf plois ivhich are contiguous in time or apace, To some extort the positioning of the black boundaries is atbitrary, Inn experiment on peaple or animals, age, weight and slate of healih are continuous vatiabies which ate often suitable for deterinining blocks. To be in the sare block swe peuple do not have to have exactly the same weight, but weight raaiges can be-chosen so tat blocks have « suitable size. Similarly, severity of disease cin be used to black pationts in a clinical Wiel Exampts 44 (Laboratory meusorement of samples) Consider the techaicina-mensuring soil semples in Question 2.3. His experimental units fellow one another in time. As tine ‘goes on, he may gol inom: accurate, are may get tired. Ovtsde fectors, such as temperature ff Hunmidity, may change. Dividing up the sequerce.of experimental units io three four blocks of consecutive plois should remove these unnecessary sources of variation from the ceonelusions ofthe experi Hexampte 4.§ (Field triat) ‘The plots in an agricultural field trial may cover quite large area, ‘eacompassing changes in fertility, Sometimes itis possible to form natural blocks by rvacking oat a stony area, a shedy area atd so on, More often it is simply ascunted thet plots close 0 ‘each otter are more likely ta respond similarly than plots Far apart, so small compact areas are chosen as blocks, Fo Example 4.2, the distance ftom the main iigstion chennel gives 4.conlinuous source of variability that stoulé a's0 be used for btocking, bur now there is someé freedom to cheure how large a distanee each black should cover. ‘4.1.2 Chotce of blocking for trlal management Soune arpecis of tril management force differences between tle plots: As far 26 possible, these differonces should match (some of tic black boundaries. Inactinical wial, patients may have to be divided into groups te be attended (o by different doctors or nurses, These groups soutd be blacks In alsnoraory experiment, technicians may be thought of es natural blocks if thelr times and places oF work: aro already fixed. Mowevor, if technicians ean be allocated to lasts as part ‘ofthe management of the experiment, then 1tmnay be possible te adjust thelr work se thay, for ‘ecannpte, ihe nomber of samples analysed by one person inone session isequal tothe nurmber of teanents ‘There ace many experiments where ott oF more treatment factors can be applied only 10 large areas: see Example 17. These large areas Toran & sort of Blook,36 Chapter 4. Blocking Faxumple 4.5 vevisted (fete tris) Tn the developed world, most agricultural eperstions aze by vactor, Typically a tector ty driven as far as possible in siraight line before belng turned round. This suggests that blocks in fietd trials should be long thin arens corresponding to & few prsses of the aster. Example 4.6 (Citrus orchards} Shoilaiy,citeus orcherus sre planted wih the toes ina pect- angulct grid. The space between rows is bigger thar the space between colurans, so nat larties ‘can drive along the rows for operations such 2s applying pestizldes oF Larvestig. Therefore, both contiguity and management considerations suggest that rovis should be blocks, 4.1.4 How and when to dleck Upussible, blocks should all have the seme si2z; Gi} blocks should be big enough Co allow each Heatment to occur af Ieast once: in exch block. Natural discrete blocks should alwaysbe used once they kare been recognized. IT possible, ‘choose plots and btosks to satialy i). Example 4.7 (Piglets) If the experimental ualts are pigtets den finers are natural blocks. Litters are vot all ofthe same size, typically being in the range 8-12, depending on the breed, 1 would be sensible 0 use only some fixed nuniter, say nine, of pigleis fron cech liter. Then yeu need ar objective rule for which niglets ta choose from the larger liters, such he heaviest Piglets, Akernatively, if larger blacke are needed, oturi with more eowe than necessary and se oniy those litters large enough to give, say, ten piglets Natural blocks havi an upper limit on their size, so it may be impassible to satisfy Gi). In the coves" art example, blocks have size two mo matter how macy treatmens there ars. JHlccks should aliveys be used for managemea!, Then all (rial operations—sowing, har ‘vesting, interim applications of trestments, measuring-are done block by blook, in care there are interruptions, improvements in technique, replacement cf staff eo. Tas ensures that cy extra variation caused by changing conditions is aready accounted for by the blocking. Man- agement Stecks can usually be chuwsen t satsly botln() and Gi}. Example 4 (Weed control) Feld rials tocompare methods of ccntretling weods are some- times planted by hand. Tf a teal ie too large for one porson to do all the planting, chow it ix recommended that cher cach person should plant one or more wiolebioeks, ar cach person should plant the same number of rows in every plo. ‘To sliminate the effects of a continuous trand, blocks can also be chosen tu skisfy both ) and (i), Usually such blocking is helpful, bul it mey be Letler not to use this sort of block if doing st would make the mumber of residcal degices of freedom very small: ses Brample 4.15, Asnoted in Example 4.5, the requirements of blocking fr trial management my coufict with those of blocking to remove a cowinucus tre, You may have to deide which is nore important for the experiment stand, 42, ont veavel Rxarmpl meat tod the tans grows 3 tunel! than byt is each str replicat 42 ¢ For the havea t The! block. F ‘Then yj for Ya No Exam ‘ecs| showal 4 Note t Defiolt other. | i Theos of tin sil1 eaca assible, blocks. breed. f. Then eaviest ary a e ng, hare se there hat any s. Man- escme- | person fy tet sort-o al conflict is more ies? 4.2, Onhogonal block desig an ‘Wehave alco noted examples sshete more then one sort of block isneeded, This peint will ‘pe developed further in Chapters 6 and &. teample 49 (Mushrooms) Mushreonis are grown ingore in tomels or she In an expert near ty compote difaront act of mashircoh i therefore socms cbxious that we should wee the tunnels as natural dsorete Books, However, fort scans of mustroor ate norlty grown ul different terperates and it vay not be pactcnble to vary the camporeti® within arrnel, Asa tos, stars will grow ot diferent rates, s0 that harvesting by stn, Taher than by tunnels, bacorees inevitable, Thraueh a situation, it may be Beller to scent the tunnels as experimental units, so that ‘each slain can be grovn in ts appropriate eanperaiure. Tis may lead to issues concerning replication and conract date analysis: see Chapter 8. 42 Orthogonal block designs For the rest ofthis charter we suppose that ©. consiats of b blocks of equal size We thas have 8 Mock factor B which is defived by B(@} = the block contzining ©. “The block subspace Vp consists of those vectors ia V which take a vonstant velue on cach ‘lock, For j= 1, 2.0: fel vy be the vector whase entry on plet © is equal to {i it en is in block fi, 0 otherwise: Iyeeusb} isan onthogoral best "Phen vj-vj—2k while vy-w = Oi Fb. Therefore {¥ + For Ypy anda yp Now, uy = D5 y ¥7 € Vrs 10 Yo C Vir Hust ag we defined Wy, we pat Wav {0 € Ya: vis onhogonal to} =Va Wer. Exeraple 410 (More Bitton) Table 4.1 shows & block design for thice weatments two aan zs ovr, The eostvont veloc a ste andthe block vectors and Ya ate serve, Trenument A cecure tive I Oc 1, So.-¥s = 2. Two farther vectors shown 45 Uy —2ily EVs and w= ¥1 —¥2 € Va Moreover, xu wep = 0, 50x Wy and ve Wa. Nole that x-w = 0,s0x LW. Definition A block design ie orthogonal if the sraces Wr aad Wy are orthogonal to each. copher, “Theoremd.t Given a block design fort meatirent in b blocks of site hy fe ng be the rniber ef times that tectment i aces’ tn block j. fort tad j 1b. Then the block design ix ordhogonal ifand only ifs; = vif fort stand wb,8 Chapter 4. Blocking ‘Tuhle4.1, Some vectors in Esaanple 4.10 Pla Block 4 se veces le Proof First note thet ay =0/-). Since Wy is orthogonal 10 Ve, Wr L We If and only if Wp L Vy, which happens if and only Ge) whenever Pyar =O; that is, asi) =O For all j whenever Dany = 0. Wey = 14/6 foc eath then Fi oysy = Eylati}/d, which is ze wheveves Eyes = in true for all j,50 Wr 1 Wp. Conversely, suppose that Wp 1. Wp. Fix i different trom 1, and put ey = Ir. a = Ife and a7 =O Fi ¢ {1,1}. Then ayry =O while it Ofer} ). This Yas, = 2 5 for all j. Since Wy 1 Wy, it follows that Sb ME mt forall 7, This is true for all, including ¢ = |, Soccunting the plotsin block j gives Flin Mae, ny be YY Therefore ty =r1/band hence sy=nyfbfor aN. Definttion A conplete-block design has blocks of size f, wilh esch treatment accurtlag once ineach block. Corollary 4.2 Complere-block destgus are orihogonat, ‘We consider orly onhogoral btock designs forthe remainder of this chapter. 43 Con 43 6 Consirug ao @ Exampll A plats, Plots wi The! Tox Here and tale the 1001 Bray gent: new 0 eck ol evry id The| fer ea Figures 4a Recall tere ml where? the cor,jecking ad only © This Va Fg once 4.3. Consimietion and randomization 38 4.3 Construction and randomization Construct and randomize aa orthogonal block design as follows / and randomize, justas for a Apply treatment f to r/b plots tn Wack By for i= completely candomized design. Ai) Regent fr ench block, singe fesh randomization exch fine independent of the ceding randomizations. nd evaluated by each of eight judges. Feample 411 (Wine tasting) Four wines are tastod a there ane eight blocks and 32 plots. {A plots one tasting by one jndges judges are blocks, lots within each judge arc identified by onder of tasting. ‘The aystemtis dosign is the same for each judge. Judge i wing]! 23 4 e [123 4 ‘to ramlomize this design we need eight independent random petmutasiens of four objects Slere-we vse the meitod described at the end of Section 22, using a seeary of random digits and taking 28 many a are nec foreach successive block. Tac random dite ae shown in Ihe ep row of Figure 4.2 and the rindorsized plan in Figure 4.3. Bxample 4.12 (Example Ith coutirued: Detergents) Suppose that there ae three dete” ena io be tested, bot thal che Tescarch orgaization has only Hnited quttes of the ww ow anes (A and), while it has effectively unlimited quantities of itxstanad detergent (C). Ears housewife participating inthe ial will do four washlords, so the researchers decide (hat every housewife will wash (vo loads with dotergent Cand one with cach of A anh B. ‘The systematic design is intoad [ T 3 Detergent [A BOC for each oF he ten housewives. After randomizalion, the plen could appear Fike the one i Figured, 4.4 Models for block designs Recall that Yo = tray: Zoy WHEte Ze i He effect of plete. There are tre commen models sr now tle blocks alfect Za» nthe frst model, the blocks atfect dhe expectation but act the cow HlZa} = Sto» where Ey in an unknown constant depending on te black B(@) cowl hy covariance stil has ts simplest foram tha is e01(Lou%) {3 nce, THES 4, However, colhervrise.cy ‘Chapser 4, Blocking tie Sudye 3 Tang a 4 “thaing) 1 2 3% “hig ]! 2 3 4 Wine 394 wine |e a2 vo wne [a 1 23 Judged bodes Jualge 6 Ting] 1 2 3 4 “Tsing 1 23% “hate ]1 2-3 4 We [2 4 3 1 wim ft 4 3 2 Wine foe 19 Judge? judges Taine |e 12a ¢ Wine [4 2 5 3 4342 Fig. 43. Randomized plan ia Exarple 41 Th this maou, we sny dat che blocks have fixed epfecis. Inthe second model the blocks mnke ae contritation to the expectation, so that E(Zey) = 0. Howewur, the coverianse between the responses on plots & aid f depends on wither o = By (cane B are different but in the same block, or cand f i indifferent blocks. ‘has oe ife-B {ine itor bal Ba) poo? iF R(ee) # ALB): + Of course, 12 py and (> py. Usually we expect that pr > Oz, hovause plots ia the same block should respond in a more alike manner than plots in differen: blocks. Now we say that the blocks have remclon affects. Let Jy be the M x 7 nate whose (a, B-eney is equal to {i it B(o) = 2¢G) © otherwise, covldanzp (3) Then, i Son are best than 20° lass top end years, than to 45 wheaty that 260 wi & & a BI tht outeeruplies 43. Anelysis when blocks have fixed efters ol Housewife | . Housewite2 Housewife 3 Wasiloa’ | 1234 Washioad TT 2S Wasilogd [12 Dowrge [CoC B A Deere {CoA BC Deeg [A C _ = Howsewied Howwewies Howenie 6 ‘Wage [1 2 3 4 Wahid [1 2 9 4 Wasted [71 2 9 7 Betrgem | 6 GC A Detergom |S A CB — Detergent fA COC B Housewife? HassewifeS Howsewifed ‘Wand [i 23s) Waid tS A Wasled [T23 Detergent [A CC Getergen [CB AC Deere |B CAC Housewife 1G Wahiost [23 Detergent | CoC OB A ig. 44, Rendomized plan im Example 4.12 Forexample, f~2 and & ~ 4 and the first Four plats are i block 1 thea ri t1ooo0 a birttoooe! | rietteaos, |# weflitrooool fut w-lono0o9r1tt vt oooortir) | oogortii vt goaooritl vt "Then, ia the randont-effects modct, Covl¥} = Flt pic’a—1)4-p20°I-Ia) FL Put (eu pals + Pall Some natural discrete classifications with a stiall number of possibilies (such as sor) are beat considered as fixed. For example, 20-year-old human méles might always be heavier ‘han 20-yenr-o1d bimaut females and we might want to find out how nnueh heavier. Most other clzasifivations are just a nusauce and age best thought af 28 random. For example, plots a ste {op end of die field may do better then plots at the bottom ond in weet years and worse in dry yenss, bul, on the whole, plots atthe top endl will tend to perform more sinilarly 10 exch other than to plots at the bottom end, 45 Analysis when blocks have fixed effects When blocks have fixed effecis, the expectation part of the mode! is thal Ua) = rea) + Soe aDa Chapter 4. Blocking EY)= 4G, were TE Ve and §€ Vp. Uquation (2.5) shows thas in € Yo and ‘Tr = tay € Wy. Similarly, © = Eq 4+0q, where by = Ceo [tas 920 1033 | 2978 6 7 8 9 878 955 949| 2782 W026 044 1055| 3075 532 946 1012! 2910 569 wna 14] 3117 w | ser 932 sno | 2032 ‘Tooaonent tel {966.7 9929 10162 | 29758 Table 44, Anatysis-of-vasiance table for Example 4.13 when cords have fred effects _souce_|__s8 ms_|_ve wien | DSVASE| 1 | STE | TOBOD.A7 cone | 8470 200 a0 womens | 12268 sin] 392 retidst | 231.85 tsi6| — ‘Tunal” | 9967-32 | 30 © -Coh ¢ + «(o> s ~(lF 6 40 a dh i ps me 10 0 -2-0 0 2% ona eat (0) Dee afer ubtacting Hoek nam Fig. 46, Bonplots ofthe data in Exsenple 4.13 46, Ang Tile 46 f Pur c W plot ¢ she ane Sua. so that Ix tn Sect ana so! Finy ‘Thy Nobd a For| priate & identicd the offackiug 46, Analysis when blocks have random offeets a ‘Table 4.5. Annbysis-of-variance table for blocks and wistructured treatments when blacks have random affects Stcatum | Source a EMS vR Vo mean [men i ol + Wy blocks | blacks bl WE plow | west rm residual | 6— N--1) oral x 4.6 Analysts when blocks have random effects ‘Pur C = Cov(¥). Then we have CHoll- py) 1PM t Pal] day Wf plot @ is in block f theme O-row of Jy is just ¥j. Hence I x 8 aly vecior In ¥ ten the @ ontry fa Jot ia equal to vj-%. In particular, if x = 49 then ¥j-x =k for all j and so ‘Title = Stig. Since Ip = up and Juy = Huy, we see PII —pi) + Me — #2) 4 Npalun, Cup = +50 thal tg is an cigenvectar uf C with oigorveluo Eo, where : & = 0?|(1 pr) +4091 — 22) + Nal Tox Vy thon x =F, yy; forsome scalars 21,..- ys hence +8 = Kay and 50 Jax =k. In Seetiou 2.44 we saw that x= O iF x € Vg), Hence if € Wa = YarnVGt then Cx=07(1— pr) 4p Pr) and xo xt an eigenvector of C with eigensnlueE), where S1=0"{(1 pi) +(e: — pa) Finally, ifx € V2 C Ve then Jax =0 and JX =0 so Cx = Spx, where &y = 97(1—p1). “This the eigenspaces of C (the strata) are Vo, Wy and Vit, with dimensions 1, 6— {and ¥M—b and eigenvalucs &, & and & respectively. Usually we expect thal & > , because &) =ta-+ko*(p1— m2). For ax orthogonal bask design, Wy © Vp. Hence Theorem 2.11 ohowe that the appro: priate analysis-of-variance table is that shown in Table 4.5, “The arithmetic calculations are ‘emia! to these forthe fxed-ofactsrmdel. Assosstatinent differences just x fore, For the effects of blocks, do a two-sided test using, MS{blacks) MStresidaal)”eB Chapter 4. Blocking ‘Table 46. Analysis-of-rariavce table for Bxanple 4,13 when cords have randen effects Source | ss fur| ms mean | 28517950 | "T | BSI7.R poids soo, mel vwowments | 122.85] 2] 61.22 | mess lig] ts65 | — Toul 2962.32 | 30 | I MS(Liocks) >> MS(tesidusl) thea the choice of locks Was goud: de il similarly next lime, [f MS(blocks) << MS{cesiduat) then sithar_ E) < fa because plots withia a block conipete (for example, i atl plo's in chamber In a'greesnouse share a single system of citvulating liquid nutticnts, so p1 is negative} or 8 <&p and there is 2 better way of blocking ‘or ‘rial management has nol been by bleck (for example, if the plctn we in a rectangle, ‘columns ivewe boca used as blocks but harvesting las been dare: by rows) OF Ihe scientist is fiddling the data, and is.not expecting you 10 notice very iow values-of the variance ratio (6uch fiddiing usually Jeeds to (oo small estinmtes of veriability, because, the data are “100 good’), Example 4.13 revisited (Metal eords) Given that tie metal conds used in the expetiment are just ton out of many coming off the production fing, Wis more Hkely that whey produce random effects than fixed ellacts, This gives the analysis of variance in Table 4.6. ‘This, ‘contains exuctly the same nusrbers es Table 4.¢ {apart from dhe final column, bar the layout ‘nd the interpretatin are subidy lifferent 4.7 Why use blocks? Tf we should use blocks and do noi, what happets? If tae blacks conttibute fixed eltecis then the vector y is slmost certainly nol 2er0. Ifthe feoatmonix are not allocated orthogonally to blocks then Gy will not he athogonal lo Wr, $0 Pijpky will be nonzero, The estimator of tris Py, Y, whose expectation is ty + By, f,. Thus the treatrient extitasiors will be biased. IL is raact likely tha by in also ne orthogonal to Ye so the estimntor of o wil asa be blased. An fae, Theorem 2.5(i1) shows that the expectation ofthis estimator will he sothat the varianco will be overestimated. Example 4,14 (Hay fever) Two new treatments for hay fever, from diflerent manufacturers are compared with a placebo, using 18 voluntzers inthe hay-Fewer season, Por conveni in one week the physickans give treatment A to six volunteers and the placcho to threes ii xq varia ce | feel ! | winch Pacing square! oly lyst 48 | ‘Tae fo; statisti advint of ee, ationlocking, fects iment reduce This layout Ittke Yr. 0 | has ov, hation 48. Lose of power with biocking @ second week they give ueaLment B ta six mre volunteers and the placebo to three more. They mezsure the respiratory fiction of each volunteer fares hours after irealment Unfortunately,» cbaage in the weather has the effect tht efferent pollens are inthe siz the second week, with the result shat respiratory ftetion incressex by 4 units, on average, ‘Thon the expoctation of the catimalor of ty —th is equal 9 “y--Ta +4. IF tho result of this trial is that the physicians recommend trealment 8 rather than treatment A, dhe manutaciurer of treatment A might well ty (9 take the physicians to court, OF course, # much larger trial ‘wontd normally be run hefore such a recomarerdiaion isimad, but the point about the danger of failing to block is alll valid} Moreover, Py by now ns coordiawos zero on the (wo new treniments and ordinates :820n the plies, so the expestation of th resdal mem square ied/15 464 Tf the blocks contribute axdom effects then treatment estimators are unbiased fut thei variances are larger thaw ths need be: on averse, fa Wil be repltoed by = Le —O)Bs N-T . whic i bigger than 2a when Ey > 62 TevE do use biceks in she design ba forget 1 vclud che in ihe snalysi, what happens? Now the Ueatment estimators are unbiased, but in both models our estimates of their sesiances are to High, so we may fil to dotee: genuine treatment differences. For ixed ‘effects, the expectation of the estimator of a # equal to, tl wont? forrandow fects, te expectation ofthe estimator af E is equalto Nb: _ = 1G bd Jgger thas Ea when Ey > Eo Example 4.13 revisited (Metal cords) if tx sualysis ignores blacks then the reside ae square becomes (647.90-+ 281.85)/27 = 34.44, so the variance tatio for treatments becom ly 1.78, which is well below the 95% point of F on 2 and 27 degrece of freedom. The analyst may woll fail to detect treatment differences. 4.8 Loss of power with blocking ‘The fobtoiving example, which is then from a ease whens the meuufacturer tied ls sue. the stisticians for asing blocks. shows the only cireanustances where blocking may be a dis- advantege: there are no natural block hounderios, thers area smell nursbee of residual degeces of fceedom, and the purpose ofthe experiment is (arguably) hypothesis testing rtber than esti- mation.0 Chopler4. Blocking ‘Table 4.7, Analysis of variance for the two devigns in Example 4.15 Seaum | seuee [at] was Strawn | Soure |di{ cas nen [mem [1 OPH mean {1 Ha pats J treatment | 1 | See) ea 4 bee 12 & west | 4 § tas focaunens| 1 | $e) +62 Thal 6 residual | 2 & Tal {0) Compiety radimizal design ©) Complow-bock design example 4.15 (Pasture grass} A new additive is claimed to vastly improve the quality of pasiure grass. Are farmers wasting their money in buying iC? There are two treatments: the new additive, and nothing, lots must be large enengh for several sheep to graze frealy, Hance the repli be lange: replication 3 ix chosen, Should the design be completely randomized or in thes yandomized comptete blocks? = ropes oo opment aie a8 “(a7 Using the model for the coanpteiely randomized design trom Section 2.44, but iting & in place of & (¢0 avoid contirsion with the next moucl), we obtain the analysis of variance in Table 4.760). Now we consider the complete block design, ‘Thereare no natual block boundaries, sothe fancom-efiecis model is appropriate, ad we oblaia the analysis of variance in Table 4.7(b). ‘The completely randomized desiga mixes up the five degrees of fivedem orthogonal to Vp, s0 5628, +g, Hence § > & iy >h. ‘The variance of the estimater of t1 —% is 2, inmecomplae-lac design so the. complete-block design gives smaller variance ani to is better far estimation. Por hypothesis testing, we consider the one-sided alternative that zis bigget than 4), Te (oil al the 596 significence level, we need the 0.95 point of the taitribution, which ie 2.020 | on 2 that > ospectivel detecting ¢ ‘Thos ‘Typical detected now addit Quest Aas ans 0 omental a she expelocking +0 ee lee iy of is: the leanmot three sotho. 0). 10 Ve, fa. To 2.920 (Questions for discussion n on 2 dogroos of freedom und 2.132 on 4 degrees of freedom, To beve Sk power of detecting that ty > 4), we also need the 0.90 points of there distributions, which are 1.886 and 1533 respectively. The erguiment in Section 2.13 shaws tia to have probability al least 0.9 of ‘desccling that > ¢) when doing w one-sided testa the Sth eigoiicance tevel we need acu ranensaeyfS the completely eandomized design This he block design isbater <3" 48063 < 3.665 VE es E> 1.7202 > 2.8. “Typically we have E; ar LSky for suct a ural, so stmaller differences are mow likely ( be detected by the unblocked design. scichtist who is mose interested in proving thal the ew addiive 1s exter (than in accurately estimating how much bots) might exmplain if the experiment is conducted in blocks rather than in a vomple:ely randomized design. Questions for discussion ALLA scientist at honticultual research station is planning an experinveat on cabbages. fe 12 19 compare six different metheds of keeping the cabbages tree from slugs. ‘The expert mental area hs 24 plots. He shows you the sketch ia Figure 4.7. Adviashim on how to design the experiment, .—L & 5 Fig. 4.7, Sketen of expetomertal aren in Question 4.2 Chaper 4 Blocking, 4.2 Redo Question 2.1 under the assumption thatthe professor has ten pills of Wakey-Wakey den of Zizzaway. ‘There is onty exe sbservavion room, 30 only ene pill can be tested per dy. Your plan should show which stucant should tak which pill on which day. “Whit information skovld you give the professor aboot he plan? 4.3 The plan in Pigure 4.8 isthe field layout of an experiment coxducted ix: 1935 at Roth ‘sted Expecimental Staton (an agitcalwural research station founded in 1843), Bach plot had a notice on it shoving the block number and the plot number, These are the top two numbers aiven in each plot in the plan, The purpose of the experiment was 19 compare various Lypes cf furigant, in singte and double doses, for tncr zhitiy 10 conteoleeksorms in the sei] where ‘cats were being grown. A.‘control! weatment (i.e, no aniigane) wes included. In dhe plan, enc plot shosis, i order below the plot nurnher, the level ofa factor called Fomiganl, then the dote, then the type of ebemiel. Tn early March, 400 gm of seil wore ennipled froma cach plo}, and ths number of eetwor cysts ia each sample counted and recorded, ‘The fumiganis ‘wore ploughed into the soil ia mid-March, and the ants were sown one week Inter. They were ‘gota, and horvested in Augu. In October, the plots were sampled egaln inthe same way, and the num of cysts rovorded. The variable dageouaré was calculated 28 logcour = log(number of eolvvarm cysts a Fisivext) log(number of eelsiorm cysts in spring before (restment), ‘where the logarithins ure tn baie ¢. This variable i8 shown al the bettom of each plot in the plan, (@) How maay treaunonis were there? (b) How were the plors groiped into blacks? (©) After sampling soil from plo! 16 of block 12, which plot should the setentist sumipt: ext? (G) Devise a better way of numbering the pois. () Why do you think logarithms wore used to present the data in the Form dngcount? 44 Aproring the factorial structure of ihe restments, calculate the analysis-of-variance tebe and the table of mens far the date fogcount i the eclwoem sxperinvenl Quesioatms wakey lr plan foie luda Imbers ls types where sample Questions for diseussion type of Famnigan fy urnigane alerodiitobenzete “Cymag’ eatbaw udp jlly “Seekay” RoOERN Vg. 48. oie oF funigant a Le 2 to Famigart single dose double dose 1 yout far the expovimant in Question 4.3 rpadtafaupaudo rfalsjfatsits rf2t2}2}fa fa ef2}fufau]uda z{«jlwla|s | az 0.549 |-0.011] 0.487 | 0.599 0.241 | 0.784 www ho +| at | as | 99 rfafrofaetafe 2)2)2 rt] afoot w]n ye a} ad2 rf arda faq fa wile de eyez zt ry oy? 0739 | 2.208 | 0.574 a|[uMj)s|ala|nw ‘| 0.189 } 0.365 | 0.277 | 0.107 | 1.187 | 0740 wlo]w wf a | 2 ui} m of mwfolmlo], | )e nf] fas | iw] 2s [as jaz tae} g | i da 2frt2}2}a27euta2)2ya)enjw tepoe pr | 2] a fe PT bt | page bacon | 3.316 xtaz|s|«e[x [2 |x lw a7 | 0873 | 0803 | o.6a9| eata | o.s2i fois |Loss] av | ow |v - | ela | 6s wy parefar ya fm pmo bof o le fy fa wv] fis | ao] of ofa} a2} i |g do rlatejayif2fa2,rtala2|z a ziwis|wlz]uls|z i. | 1.269 | act | oes | 1.685 | 2.170 | 2325 | os | emia) ov Joy] ie ~ a | a | as wf fu fu fo pa tue fm} yf 3 ye aja2famyma) fa | as) ato |e fa 2farj2faf2)2)s1)2}za)yls apogee ta yr yb |e ) 2 | aise} tage | 127 a] ziuwl|2)s |e) alm gia | 10a | tass') ere | 127 | doe | 1 | 1.826 Ba” Chapter 4, Blocking. 45 A group of ecologias is plonaing an oxperiment to compare 36 species of smal insects ‘or ther effect in decomposing a certsin niixture of leaves. Identical quantities of the leat rmixiute will be pat into 180 glass joes. ash species will be allocated to five of these jars several Insects from that species will be pat inio hase jars. The jars wil then be covered with (canaparent film and feft onthe bench in the Tahoratory for several wks before the atnount of leaf decomposition is measured, Because of the shape of the available hench space, the jats must be arranged in a 5x36 rectangle. One sow of far i next o the window and receives the most sunlight dic fifth row receives ihe leat. The ecotogiss suggest the foliowing differeat meihods for dealing with (his ditference: in am (q Pleve the 180 jars in their positions in the rectangle in a Uaplivzard niauner, tis then likely that no speciss is especially favoured. (6) Start as above. Hach morning, reaerange the jare within each column by moving the Jit nearest co the window fo the position furdvest from the window, and moving all dhe other jars in that column one place nearer (o the window, That way, the differences in sunlight shout he avernged out for each jar. {¢) Treat esch rows ablogk, Put oié jar of each species in each Tow, In each row indepen dently, ehonte a random permutation off, ..., 36 and uss i to randomize the positions of dhe jars in hit rox. The differences in sunlight will be allowed for in.the analysis, (4) Pur all five jars for each snceien le colunta, to ensure that each snecias is ‘exposed ta the full range of sunlight conditions, Choose a single random pessustai of 1, «4 36.and ase itto allocate species ta columns, Advise the ecologists of the edvantages and distavantages of these methods, sa Tr Inthiss of all eon Example! binaioas( ‘The treats awed teen wwe ‘ Nota} inoeking sects Je leat pets bit imount 5x30 hrow tn mts sthen Vg the al the ces in ies is tation —_— Chapter 5 Factorial treatment structure jE 6.1 Treatment factors and their subspaces Inthissectiou we eoasider experiments where the eeatmentsteucture is that treatments consist ofall combinations af levels of (wo treatment factors, Bcample 5.1 (Example 1.7 continued: Rye-grass) Hers the iveive treattients are all com binations of tae levels of the folloiviag two treaiment factors. ___+ Levels ‘Cropper, Melic, Melba 0, 80, 160, 240 kg/ha Feotar Calva (G) Fertitizer (F) “Ine treatments may be labelled 1, |.., 12 avcording to the following tals. F ca ‘e169 240 Gomrji 2 3 # Mele [5 6 7 & Mea [9 10 U2 “Thus ireatoient 6 is the ordeved paie (Melle, 80). In particular, fea) ~ Mette ond F(a) = 80. Tey =6~ el) ons { Notation Uf teestments are alk combinations of levels of treatment factors F and G, wile FAG to syow that each level of Tis a combination of levels of F and G. In Pxampte 5.1, 7 CAF. Notation IF is a reainent factor, writeny for the sumber of levels of F. Tn Exomple $.4, 1g — Fand nye 4. 816 Chapter 5, Factorial treatm n structure Fig. 6:1. Orlkogonal subspiees of th: Leatmiat space wha treatmente ae all combinations ofthe evele ring treaumen faciors Now we defing subspaces Vp, We, Ye and We of V analogous to those defined for weat- ments in Sections 2.3 and 2.41 and for btecks in Section 4.2. Thus we put Ve_ = {vootors in ¥ which are constant on each level of F) , We = verve, Vg = {vectorsin V which are constant on each level of G}, We = van. Then dim = ap, dimivy = ap—), dine = ag and diay, mg ‘Theorem SiL {fevery combination of tevets of factors F ad G occurs on the same number of plots then Wr A. Wes. Proof Sirsilar the proof of Theorem 41> a ‘Theoren 5.1 shows that every combivaton of levels of the factors # and G occurs on the same nuraber ef plots then Vet Ve Veo We ONG 60 orthegonlly, and hence that dim(¥s + Ya) = 1+ (ur ~ 1+ (ag — 0. Ihe treatment factor 7 is defined hy all comb combinations uccur at east once, then dion ¥p ions of levels of F nnd C, and all shexe seg, NOW, Ye +¥a CY 30 pu Weert = V2 (Ve Ve). (52) Then Siro — i Wp — dim (Veta) = mana — (nee 1) (aes ~ 1) = (te Ling — Wp Figure 5.1 may be helpful. Compare this with Bigare 4.5, 52, Inve, Table 51 5.2 We eon Table 5: The me among | rE, pena stow prea Bxamy prea! shown | Cropps wl Kxamprene “cate uber son (a) i ance cs as Epeiat 2 n 52, Inervetion “Tnble 5.1. Sonne atodets for ET) when tremens are all combinations of tevels of F aud G EW) = 3710} Ey) eM full rceatment model Oa) Aegag + Ho40) RY) Vet Mes edditive in F and G Dera BOE vy Fony BM} e¥e Gonly wy) e% snl mad Ea) =0 Be {0} 00 mode Yr Ve+¥o {0} Fig. 52. lelatlonstips among the subspaces fisted as posslife models in Table St 52° Interactlon ‘We continue fo suppose thet treatments consist of aif couibinations of fevels of F and G. ‘Table 5.1 shows some plassible models for the expectation of the respons i this situation. “The models are listed i decreasing order of their number of priametore, The relationships mang the subspaces considered as models are shown in Figere 5.2. IFE(Y) & Ve-+-¥y then the difference between ¥ values for different levels of ¥ does not depend on the evel of G, and vice versn, Tae expected resporses for four combinations ere shovon in Table 5.2. ‘Thus plotting meany:.; agiust far exch level of G gives approximately parallel curves. Js exporiment the response on each plo! wax the the crop, (thé twelve treatment moons are as Hixample S.L revisited (Rye-grass) In percentage of water-soluble earbohydrate shoren in Kigure 5.3 then we hove approximately parallel curves. ‘Thus we can ceport, for example, that the percentage of water-soluble eatochydmate in the grain iy OOS higher for Cropper than i is for Melle, icespective of the amount of fertilizer TE E(Y) is not in Ve +Ve then we say thol there is an interaction between F and G. Tn Example 5.1 there is no interaction6 Chapter 5. Eyctorial ueatment structure Table $2. Hlustration of differences when feotors ure additive Level 1ofG Level 20 | nittorence Level Lote [hata Level2ofe | dota ~pitferenes [XA Mean espense 23 Cropper + a Male « Melba + w + 19 . + wv x Fig. 8.3. Nointeracton between Cultivar and Ferttzee in Exariats 5.1 Exaviiple 5.2 (Herbieldes) If the combination of two herbicides gives mere weed couttal than would be expected by adding the effects af the two separate herbicides, this is called, synergism, On the other hand, if the combination gives lesx eantrol than the sum of the effects, then there is ssid to be autagouists benween the herbicides, These are bath forms of intercction “Tne most extreme forra of interaction occurs when the separate curves actually cross over each other, Example 5.3 (Cow-peus) In an experiment in Sout Arica, the treatments caasisted of five varieties of cow.pea in combination with (ieee methods of cultivation. The moan yields, in Commes/acetare, are shown in Figure 5.4, Here there Is crossover ielezaction, Xt is impostant (o report this interaction, Clearly cultivation method # is bast for vad for the other threo, while cultivation method 2 is worst for va ios C and D ané worst iow C aud 2 and best for the 52, ] a ther ane 0 Sa thse fi of po sugges! irr wa} such a nensen} wd lines 9, the ini caster | axis inquan Kxamy ey of cereal, old vat respostImeture ‘antral called of the fans of 5 ever ot five hls, i orleat worst lor te 52. Interaction 2 ‘Mea response 4 + x + t x aby & . x + ‘ + x 7, A 8 c ’ 2 Vasiey Fig, 54, Ciossover interaction batween Variety and Cultivation method in Example $3 other three. Calivation method 3 is intermediate for every vattety aad so might be the rafex: (one 10 use i a farruer is trying a new variety Hot atnong the five tested here . Some peopte draw lines between successive points wilh ihe sarve symbol in diagrams Dke those in Figures 5:3 and 5.4, The lines ceraily aid the tumen eye (see if (he successions ‘of points are parallel, However, they cen also be misleading. tn Figure 5.3 such a tine would sugges: a value for the percentage of water-soluble carbohydrate in the crop if niteogen fertil- ier were applied et 100 kha, ane the value might ot be toe fa eon the uth, In Figure 5.4 such a line would saggest a variely intermediate between varieties A and which may be unsense If one reatrieat factor is quantitative and the oller is qualitative, then eople who draw ines would plot the levels of the quantitative factor along the x-axis, asin Figure 5.2, 90 that the intermediate values suggested by the lines might have seme meaning, However, itis often oF to read the diagram if the fiotor with the higher number of levels is ploticd along the: “axis, aad the levels of lhe olher ‘acter are shown by symbols, whether oF not either of them inqua ive. Example 5.4 (Madera cereals) Modern cereals have bron bred to procluce kigh yields, but they often need (cree amounts of fertilizer and othe: additives in arder to de-so, For such a Creal, the yield rises rather stoeply in-eesponse to extra fectilizer, By comparison, atraditional old varicly will not be capable of producing the highest yiclds, but it may have a radhec mile response to fertilizer, 30 thet it does better (ban the new variety fnct much fertilizers applied,0 Chapters, Factorial treatment structure Mean response Second chenvoa! 8 Aven + Present 7 + 6 5 * ‘ ah ‘Absent Present Frstebercal Fig. 5.S, Threshold interseion in Exurnplo 5 ‘This is another example of crossover interaction. A farmer with low resources, or with poor |. inight de better to grow the teadtional variety. 3m psychology, a crossaver interaction is called dverdinal. Example 5.5 (Enzyme in blood) Two chemical preparations ware injocted ino mice to see itchey affected che quantity of some enzyme in fheir Hood, ‘Thus there were two treatment factors, F and G, white fevels were atwence and presence of eich of the chemicals respect ly, The mean respunses are shown in Fignre 5.5, tt is clear hat both chenvicals are needed in order to increase the quantity of enzyme, Thus their effects are not addilive, and there is interaction between them, Example 5.5 shows threshold interection: exch treatment facter needs to he prosent at 2 certain level before the other vin tke effect. In more complicated cases the «reatinen factors may have rnore than tw lovels, and it may be that enly one of them acts 2s a thteshld far the other Example 5.4 (Saplings) At a forestry research station, poplar saplivgs were: phanted in an exporimeat to compare trealnients for getting the young trees extablished. Rout iypos of coliae ‘wore pul sround the saplings to prevert them from predators and other damage: these won ‘combined with five different heights ofthe saplings at planting time. fer some months, the height ‘was measured again, Tteumned out that on: (ype of collar excluded so much fight that all the plants died, se the zesponse on those saplings was zcro irrespective of the height at slanting time. This is an extreme form of threshold interaction, Unless the qualitative factor Collar-type is at aleve that permits the plents to grow, no olher treatment fieter ean have any effec, 52. tnt Any thee exampl ingthe yore in thei, thom Figu fine for for tte carci | tevmpere ‘thei faetoe vg Branig indose| reaction logos. resid the san efstanc} eure poor to see meat speat~ codid ere is ital a acters or the collar 3, the that gh at factor any 82. Interaction 81 Mean nising time, in minutes pe ermperane, ine 120 oe . Boe oo ca . 4 0 * o =} “4 50 Mixing temperature, in °C Fig, 5.6, Useful intesaction in Example 3.7 An interaction with the above behsviour ean bo good if it imple that setting one Factor the correct level means tal we-can be slack about controlling the olners, Example $.7 (Tablet manufacture) A pharmaceutical company ean vary several factorsdus- ing the manfacture of medicinal tablets. fn ons parccular process, the staff can vary the sem perature al which the ingcedients are mixed: and, independently, they con vary the temperature the pipe carrying the liquid ingredient into the mixture. They mesnuce how loag ittakes for ‘he mixing process to be complet2-—the shorter the better. Figure 5.6 demonstrates that not only doos the higher pipe tempersture reduce the mixing (ime for each mixing temperature, It also reduces the difference berwan the mixing times for the two mixing temperatures. ‘This suggests that, during production. it is worth being careful to miaitain the higher ple temperature; thes i will not matter foo mach the mixing Aernperature varies. “The oppasite ofa threshold itoraction isa trigger intoraction, Here a lestone (eaiment factor nceds to have the correc! level fo change the response but uhere Is mo furincr gata Trem, Iaving beth a tho correct level Example 5.8 (Catalysts in a chemical reaction) A chemical process in the preducien of an ‘industrial chemical has several stages. At each siage a catalyst may be present to improve the reaction, Figure 5.7 shows the effect of having a ctalyet probunt or absent ot exch of two sages. The response is a measure of the qaality of the cttemical produced, Relativerto the tesidval mean square, the tes responses with at feast one eatelyst present were judged to be le seme, ‘Two further types of interaction are worth mentioning. If there is no crossover then the distances bervieen the eurves iiay increase as the gencral wesponsc increases, Thin ecu be82 (Chapter §. Pretoria tccatnent sirectire ean response ‘Second stage 4 Absent Present + + 2 0 + a 6 sp Absent Present Fiat tage Fig 5.2. Viggerinemotion ia Exemple SR an indication thet we are moaruring on the waong scale, The basi treatment Z adds a constant 7 to the response en each plot whet assume in Rqvation (4.1) that each block adde « fixed ecitaat. assumprion (1.1) is that s applicd. Similarly, we Example 5.9 (Counts of bacteria) IF the purpose of the treatments is to reduc the umber of bacteria in milk then it igmch more Holy thatthe effect of each treatment ix to matiply the initia? naniber of bacteria by a positive constant (which we hope is lest than 1). Thus we ‘would expect Equation (1.1) 0 apply to the logarithms of she counts rather than to the counts themselves. Lat Yo be the sureber of bacteria per wilifirein sample @. Suppose that trezlment factors P and G each act moliplicatively on the covnte, 10 that Yy = Zy% Reguy if only factor F 1s epplizd aud Y = ZX oo) If only factor G is applied. these factors continue to act tmultiplicatively in the presence of each olher. then Ve = Za > Rrra) X ted. with the result ta log (Yq) = log(@e) + 108(2r4a)) +Tog(gta(0}): in other words, the model is additive in & ‘nd Gon the log scale bu nat on the original count scale, Similarly, we may recesure volume but expoct the linear mode (1.1) co apply Co the linear measurement, 50 thal we red to lake cube roots before analysing the data. Suppose that there is anon-linear monotonic fucction f suck that /(Y) sazilies the linear Model (1.1), IF there is no interaction between treatment taciors # and G when we consider the traneforraed data (()), then there wild be infexaction between them when we consider the sntransformed data y. This is rather hard to spat graphically. If f is nearly linear over tke cof the data, zero interaction. ca the transformed scale laoks like zero iteration on the nal scals, HF f is not ncarly linear over dhs range tfc calculating means of the daa on the wrong scate gives a seiiousiy misleading impression Fig. 5, Bxaitrple : rind sew Sano inthe peracre| highest the imp Corea 1s vet cal voraid OFg point “4 ore pee avesucts is hat ly, we umber whipty Ws we counts ‘actors ctor F fo act result einF linear linear asider or the er the: an the aa on 52. Interaction 8 Mean respone Noaddednittogen + 16 ‘Added nitrogen x. a x * x . x 12 . . 0 . a Wok ows TSthonts SMa hous Ohaus Of veich 2546 vetch —5D% vetch 75h vetch 100% veteh Hig. 58. The improverent due to added nitrogen destines as tie proponion of voich increakes. see Example 5.10 Finslly, the distances between the curves may increase as he levels of oneofthe treatment factors increases, There may be a simple explanation for this. Example 5.10 (Vetch and oats) An expesiment on forageorsps compared five seed mixtures in he presence nnd absence of nitrogen ferilzer. Figure 5.8 shows the mean responses ia tous por Bere. The seed erixiures cange from all os, no vetch to np oats all vetch, The yield is ighest For the 50:50 mixture. Added nitrogen improves the Yield of all of the mixtures, but fine improvement dectines a8 the proportion of veich increzses. “This makes perfect sense. ‘Cetoal esops, such as oars, need (o eke nittogeu from te soil Cty can, while legumes, such, a8: wich, detually fix ntragea in the soil Examples 5.55.10 all exhibit imcraction even though there ia simple explanation ofthe: non-additivily ia seme cases. OF course, there is no need to remember the némés for all hese types of internction. For one thing, thoy are not really well defined. What is ¢ crossover Interaction from the point of viow oF one factor may or may net be a crassoxer irteraetion from the point of view of the ofter, Threshold snd tigger interactions are, in fae, the satne thing from different perspectives: in both eases there isnt Feast ene level of one fictor al which the response des fot vary withthe levels of the other factor. Whal is important is go bs exvaro that interaction ean cecur and that it sometimes, but not always, has ¢ relaively simple explenstion.84 (Clupter 5, Fretorialteeament structure Tf there ere no fixed block effects then B() = €, Using Section 4.5, wean put w arguments to those given in u i = Rs a and ty = Tyg = Py, tte The vector ty is called the nals gfer of weatnen factor #. Somes te eas fs 5p, which must sum to zero, are called the effects of F. Similarly, tc; is the main effect of reat- att facon G. Write Trg for the projection of F onto Wea. “This vector teq (or its entries) is called the F-by-G interaction: Thus the P-by-G interaction is 2400 it and only Wc Vy Vow I Wy J. We tea Bquaien (5.1) shows tha tp -+tp + 8 Une projection of ont Ve + Vs, so then Equation (5.2) shows that tee = t—Ty — Te — Te. ‘Of course, what wo incan by “approximately parellel the size of the variance, The estimate of the inlereetion is a messure of the departure ofthe fit in the full model ¥p (rom the fi inde submodel Vie-+ Ye. The sigalticence of this departure is assessed by comparing its size (divided by its degrees of freedsin} with the residual mean square, as we show in Section 53. 5.3 Principles of expectation models This time ta come clean over sn issue that have been fudging until now. Te Chapter 1 T suggested that there was a clear dichowmy between estimation and wsting. That may be true when dhe weauments are unstained, However, we have now met severtl cases where ‘we are imtererted in many different models For the expectation: seo Section 2,11, Chapter 3 and Table 3.1, to these circumstances we umvally do Hypothesis fests (0 select the smallest snodel supporied by the datz, and then estimate the parameters of that model. A cofection of subspaces of ¥ wich ane to serve as expectetion models cannot be arbi- trary, but shold ohey the following principles. Principte 5.1 (lutersection Peluciple) 1£¥j and Vj are both expoeiation models hen i NV; should also be an expectalien model, ‘The Intersection Principle is there 10 avoid ambiguity én mudel-itting. Uf y is in (or is ‘slese to) Yi F1¥2, then our fited model should be ¥; (or z subspace of it) aud it should also. be Vs Cor a subspace of i): i the Intersection Pri model ¥;r¥» Out thercis no way of deciding bervera tne models ¥) and Va. Principle 6:2 (Sum Prineigle) WY; and Vz are both expectation models then ¥y +¥% should also be an expectation mest. ‘Thereare three reasons forthe Sum Principle, First itis a feature of all Finese models that if v ond w sre allowable as vectors of fited veluss then so should ¥-+9¥ by: apply this with vin ¥y and w in Vp, in graphs such as Figure $.3dependson * Seco where ¥ model, ¥ ‘hi, For sof differen, mak i “ne “Tha expt bythe “Te vessial, Prine Wavi Fu Vi+¥y so also ted Sut we aa Pu mi dimiv th Wy the oxy, Thy two avant vty interct,tecture ven ia in te. bt teat Yo. It called Ve, 80 nds on Irube at lpartuce Year, pler ray be where pee le abi. Prineges of expettion models 5 W+¥ werner { YN } orsmanns 4 % mona Nba Ww Fig. 59, Kiting submodels of W +49 Secondly it avoids arcbipuity, jus like the Jmersestion Principle. Suppose thal y = v-+ 6, where v and ware nonzero veetorsin Vi anc ¥; respectively. Unless vi +¥2 Is an expectaion cde}, we shall be Foread to make.an arbiteery choice between V; and V4 “Thirdly, itimplies that all the expectation models onc cemtafned ina single maximal model, For mest of this book the maximal model is Vr. Sa Chapters 9, 12 and 13 we shall (eke a Afferent maximal model when we Gan assume that some interactions ave kero. The maximal litode! isthe starting point for testing hypotheses about mode's. It also defines the residual: nothing in the maximal model is ever pul into residual, even when we Fi a siaaller medel. ‘The Intersection and Sum Principles are Voth awomnatical.y savisied if ihe eollection of expectetion models forms «chain, as in Figures 2.4, 2.1 and 3.2. Moreover, neither is affected by the mumbets of replications of the treatments, “The thitd principle can be affected by tho, auinbers of replications, It is also more contra: versial, Prinelple §.3 (Orthogonatity Principle) [1 ¥) and V2 are both expectation models and if W =i NV) thea YF should be orthogonal co Vy AWW, Pat Wy = Vin We and We = Vyas; Then Wi = WW +0 and vo =W + Wa, and so Vie ¥, =F EW FIV; = VW, JF isorthogonal oW, ther is orthogonal to Ys snd 50 (Ve + Vp] OVE = Wy similarly (44.4 2] Vz = Ik san be shown thatthe converse is lao true: if ethar of these equations holds then Ws is orthogonal to W. Suppose that we have accepted the hypethesis fut E(Y) © W +¥2 and want to sce if wwe cau rena the expectation model oY. Ted thi, we axarin te sizeof the veto: By av,¥ Py. by comparing the toa ay- Py yl] far wit MS(restdul), where dim{¥j-+¥_)—aim¥, = im¥y ~ dim(¥, 1%) = dim — did. Now, Fyayy yy just the prejestion of y onte (Vi + Va) GV¢, which is Pyyy IFW; is orthogonal to Wi, If we accept the liypothesis that (X) ¢ ¥j then we go on to tet whether wo ens, reduce the expectation mode! to W, We do this by examiing the size of P, ¥— Fy, Which is iy, J. ‘The models we aro discussing ace shown in Figure 5.3. Staring at Vi + ¥2, there sre tyro routes dawn to W, and there is no good reason to choose one rater than the ater, The edvantago of orthogonality it dat both routes give the same result, beeause in oth cases ‘we are examining the sizes of Py,y and Py,y. In other words the test for reducing model VV) 10 ¥ is exacly the same 25 the test fr vehicing Veto Wad similaty with i and V2 Imerctanged,a Chapter 5. Factorial teatime structure TE; is not orthogonal to Wa, then thore are nme values of the data vector y which give Contradictory results. Inference can be difficult in these circumstances, OF course, there is unlicely tobea problem if Wy and isnt ‘nearly’ orthogonal. In some practical circumstances ‘we are Forced (o deal with non-aribogonal models, especially when the data are observational, Nonetheless, i this book we iri ourselves to collections of eansectation models that conform tall tree principles ‘The proof of Theorean 4.1 shows that, in general, wheiher or not the subspaces defined by ‘ovo treatment fictors satisfy the Orthogonality Principle depends on the numbers of replica- {ions ofthe conibinations of levels, ‘Thus the Orthogenslity Principle has implications forthe design of experiments, xamuple 5.11 Example 34 continred: Drogsat different stages of development} Figure 3:3 shows that the nly pair of subspaces that we need womy abouts ¥y and Yj. Now, + is in Vs if and only if va is the stme for all eo which receive'a dose of the now for- mulation, while v is in Va if and only if ya is the same for al 6 which revive a dose of the ole formulation, Hence Ya Ws —'Ve. Moreover, Ya + Va © Fr and dim{V4 + Vi) = ddim( Va) + dion(¥p) — dim(Vq V5) == 4-+-4— cim(Yp} = 8 2 = 6 = diom(¥p}t therefore a+ Vg = Vp. Finally, Pv @ YaFIV} then ry =O whenever T(6) is dese of the new Forme tavion, while if w € Ypr1¥¢t theie Wig =O waenever T() 1s dose of the old fonrmutatioa, and so the spaces V4 ¥p snd Yer VV! are orthogonal to each other. ‘Therefore the collection of ‘expectation mindels in Figusc 3.3 satstics the three prineighes. la arguments about intersection and orthogonality apply fo all the models given in (Chapter 3. To ouisty she Sum Principle, we have to explicitly ensure that suis of models ate Included, For example, ihe calcctien af expectation models for Hxampte 3.5 is shown in Figure 8.19, Here Vo is dhe mode! with only two iceatment parameter, ene for the control ‘weatmient and one for the rest, and Vy 15 the model whh four teeatment parameters, one for each method and one for the comtro treatment. The space Vy has six teestment parameters, one for the contol treatment, one for cach treatment ofthe ‘quantity’ (ype, and one for each of the other two methods. The spaces Ye ard Vs are defined simitarty hy the “roughage” ancl ‘time? typas. Thus the model Vg + Ve Foress all treatments of the ‘time type" to have the san parasiccr, but ciherwise allows for afferent trntment parameters, Given a collection of expectation models thst satisfoe the three principles, we test sub models by stacting atthe meximal model and working downwards. At ench stage we test the ext submodel by exanvining the difference behwecn The sums of squates forthe ft in fhe cur- onl model and forthe fit inthe submedel, dividd by the difference between their dimensions. This mean square is always compared to thc original residuel meau square. Tf-ve accept the submodel, we move dawn the diagram wo it and continue front there. Ifa any stage we havc reece all suhmodels immediately telow the current model, we decide thatthe current onc 's the smellest that is supported by the dala. Because of orthogonality, it does not maar in wlat onder we test submodels when there is a chcice. nce we have decided on the smellest model, we estimate ils parameters, which are unt ally shown in a table of moans, along with their siaudard errors of differences, Forwuiately, the calculations needed for estimation and testing ace virtually the same. For estimating the parameters of the expectation madel Yq we need Py,y. I the submodel Vj is 54, Dec a immedie equal 104 table. Ce thedigy cof -vasia thease: 54 q We ca eed the Sun} subsp model Supt ather pl! theses] that all in Cra, combi,stances ation: >aforr ned by eplica fo the ven in roclols sn ant} aw for reters, eeach Pad t sub st the ecure pe the shave one ter i > sue 34, Decomposing the wealment subspace aI Ve ee a My {0h Fig. 5.10, Callzcion of expeciation models in Heample 3.5 immediately below Vj, (hen lesting for V, needs the sum of squares [Py —Ty,y||?» whieh is equal tof) sf? —[Px/2 These sums of squares ere displayed in an analysieof varias. table. Convention diciates the opposite order fer the analysis-of-variance table (o thet used i Ahedingrae of submeels, 30 anode! testing proccods by sterting athe bottom ofthe analyss- __ of-wariance Inble and working upwverds. In the next two sections tas is described in deal for thease where trealmeats consist of all combiuatiGns of two treatment factors. 5.4 Decainposing the treatment subspace ‘We rau lo fhe case where itvatments consist ofall combinations of factors F and G. We ced t0 show thas the calleetion of modals in Table 5.1 satisfies the Satersection Principle, the Sum Principle and the Orhogonality Principle, Figure 3.2 shows that the only pair of subspaces that we noed to check is Vand Vz. The collection of models contaias ihe additive model ¥p-++V¢, $0 he Sum Principle is seisfied. Suppose thal v.€ Vp Vi; and thal «8a plot for which F(et) = G(o) = 1. Let be aay other plot and suppose that G(B) = j. All combinations of ths tevels of F and G occur, 80 there isa plot y for which FG) = | ané G(r) =f. Now, ¥ ¢ Ve s0v¢ = Fy AlSO VE Vi 0 vq vp, Hence vg = vs forall B and so ¥ € Vp This shows thal Ve OVg = Yo, até 0 the Intecsection Principle is satisfied "Taoorem 5.1 howe thut the Onthogonality Priveiple is satisfied if all combinations of levels of F and G have the same replication, ‘That is why we insist, im most of this book, that all combinations of treatment fretors must eecar equally often. (The apparent exceptions in Chapter 3 will be explained in Chapter 10) White ¢ for eambinations,88 Chapter. Factorial reauent steucture Teble 5.3, Quantities arsociated with the V-subypaces symbol ° r @ rao Vsbspnce % a Ye Vine Saasion 1 a ms seg Maver oft vals Py yy Pay Bay coordinate nf mean mast; = meateng —— meanpaL np hu? =< a pees We now fave Ve = VoBWe DMB Wace ‘xthogonally. As in Section 4.5, this decomposition leads ta decompositions ofthe dimension (of ¥p and ofthe vector of fitted values fer the Cull treatment rode, as follows. = % OW @ We O Were Hae— 1) veoter Puy = Ty + Pyy + Py + WisncY dimension name = 1 + fne-1) + @te—1) + (ae Here Py, is tho ft For the ll (eatment mecel, Wy isthe fit forthe aul evedel, Pop i the catiiate of the maiu cileet tp of P, Py, 9 is the estimate of die main ctfoct tg of G, and Phpjg¥ 15 the estimate of the F-by-C interaction tye. ‘We need to calculate various.jquantiies associaced with the W-subspaces. ‘The easiest ‘way te do this is to calculate them by subtraction from the quanttics essocizted with the V- subspaces, which ane shown in Table 5.3. Hore we stito Vp as Vex, because T = FAG, ‘Since each combination occurs on r pls, each Level oF # occurs on ngr plots end these are Avtar plots altogether, From these we calculate the quantities associated with the W-subspaces by suceonsively subtracting W-quanttios from ¥-quaniities, as shown in Table 54 Example 8,12 @rotsin in feed for chickens) Bight newly hatched chicks took par in y feed ing experiment, Four diflerenfeods (4, 8, C and D) were made available (o two chicks each, ‘The protein in feeds A and B was groundnuts, while'the protein in foods C aad D was soya beau. Moreover, feod # and D contained slded lishmeal. ‘Thus the treatments consisted of all combinations of levels of two treatments factors: P (protcin} with tovels g and s; and A Gshancal) with levels + and -. ‘The fourth eoluma of Table 5.5 shows the weights of the chicks (iu grams) at the encl oF six weeks, . fe rican annntndithih WD ewaninnas rae tate? te chara fae tnoOSES OB _O'GOTLIET OLE _ SPOLY _CTEELOSL S'RapeDCE CTRGPOST GRLESESI ‘DBCS | _ BEOLpavenbs Oso Gor os sar sve Ores scisy scaor ori | se [ + 8 OF wr ase sez. Oreay STI9y seer ory | wos | + Fo > ODE sce O0'2rr. ses ore | oie | 7 8 wl wr CON sce OTe orter | fre Fo 8D oe~ GOOLE scre- oy tsr sR og ors = tor OEY sept- Ose 7 8 @ oss OF Os ee Seba OF 10F - 8 oy ose wr ogee ScvO~ 0510 ay wines td «AY dO 25 w wees wkend YI AUSUREE ONeCRW TL TES apcuongy uy avons fo swans pus sivalfa ‘sty Suzo}n2709 °S'S TL piss —(ubss—(oroa)ss—(ovalssd (swan)gs—(o)gsd, {uesu}se—Lulsso (eraulgsp sats ows = paul urau + f=ueau —?=Jueaw—!-O'=surew — veel — f=Oueau. uesu— sueow co yoaye U} seeUTpINCD Phys hy aye sy Ay — ag ag — ey waa iy yess = aye (= 2aX1 =) (Pe int U ogee Dea, om ot n= sor Ove 2 a o penis pa poreuso, sof wos st safe, aray -ssondsqms. yoko Wh porosvesso apnHuENG “°C ACL mucture: a ension wv is 3, an” Chapter $. Factorial treatment structure The next four calusins of Teble $.5 are the fits for four ofthe expoctation mudets: ¥, Vp, Vis amd Ve. All tke coordizales ae obtained as simple averages of entries fromm the ‘weight’ vector, The squared length of exch of these vectors ean be enleulnted chor as the sum of the squares ofall 1 enries oF by using he appropriate formula fora crude sum of squares. For example, sviiy=g = 1669 and sump. = 1863 50 meanpug = 417.25 and nea, Tas CSS{P) = 4x 417.257 4-4 x 465,75? ~ 1669/4 4 18632/4 = 1S6C082.5, ‘The eolurm for the estimate ofthe main elfect of Pis obtained by subirxcting the column for the fit Zor tho null model from tiae-colunan for the fi for P. The squated length of this vector can be calculated either as the sutt of the squares ofall its enutics that is, 8 x 24.25?) or by using the appropriate formula for a eum of equares, tha is, by subteucting the squared Teagthof the ft forthe null model from the sguated length ofthe for P (that i, [564082.5 — 1559378.01. Ths column forthe estimate of the main elfect of M is similar. ‘Tho colina for the fit far PM is obtained by adding the columns for the fit forthe mull ‘rodel, the maineffect of P and tho main effect of M The squared length sequal tothe sur of the squares of all its ontrios (that is, 2 > 397,50? + 2 x 437.00? +2 x 446,00? +2 x 485.507), {tis alto th sur of the squared lengths of three vectors which have been edéed (o oblaia i (chat is, 1559378.0.4-4704.543190.5). ‘The column for the estimaw of the P-by-M teraction is obtained by subtencting the ceolviie for the fit for P+ from the column for the fi for 7. "The squared length may ‘be obtained by. la) calculating 4um of the squares ofall lis entcies (8 x 4.007), or (b)laking (ho eilference between the squared length ofthe fil for T and the squared length of the fit for P41 (1367331.0 — 1367208.0),cr (6) using tke formula forthe sum of equates for interaction shown in Table 5.4 (1357331,0-- 1559478.) — 4704.5 ~ 3120.3). Finnlly, the coluran for the resédual isthe difference bstweon the origina ‘weight’ wootor ad ihe fit for 7. Its quate lengdh is equal to the sum of the squaes of al ils entries; iis also the difference between the lotal sum of squares (Squared fength of the ‘weight vector) and the crude sum of squares for tcatments(equared length oF the ft for 7), 28 described in Section 28. 5.5 Analysis binstions of to Weament factors F and G then we replace the Ineatmients line io the analysis-of-variance table by lines forthe tw matin effects and one for the interaction, Theorem 2.5(i) shows tha, if Cov(¥) —o7F, theu tieexpected incan squares for these Its are these show in Table 5.6. First we use the ratio NS(EAG) Ms(residual) to teat Zor interaction, If we cauinot ansuare that the interaction is 2ero, then scpont that we cannot use the simpler, adcitive made, give a table of the treatment msans and standard errors of their differences, and atop. If tho interaction eva be explained in a sienple way, then do so, Consider wheter is more hepiul to report the effect of F at each level of G, or vice vera : Te we cau assume thatthe inveracton is zero, report this clearly, Then test sepecately for sach main effect. f either main effet is nomzero, give its table of means and the etandard | | { Exam analysi Wi that th Dour | Tor pre, dist Ht, sme precist presey TH who tfAsuctuee: Ya, Me, weight” not the res, Por 465.75. colurin, of This 24.25%) squared 82,5 the null aim of 5.50%}; Dian it jeo the ne Tor ares pat we dard , then vice ly for lndara 52. Analysis, 9 Table 5.6. Treaturn tives in dhe aualysis.of- variance table fora factorial experiment with to trecenent factors: dg = (0 — Wl = 1) Var Sumat | Degrees of Source squares frcedom BMS Au a Use go} MSO F Sst) wot apo re MS(residval} 2 o | so bel vet byt : Itrel™ MS(FAG) Poy-ci | ssgrag) eg Srl eo) as (edunal Table 5.7. Anatysis-of-variance table for Uxanapte 5.12 Source ss. far{..Ms vR Tea Tssos7EO | | 155997800 | TL protin 4745| 1 | 47o4s0| 3537 fcbineal ios) 1] 3ianso| 23.60 frocinatisimeat| “iano 1) 13860) 097 residual svo|4| isezs| — "Polat Tra R600 | 8 " errors of their differences. Example 8.42 veststed (Protein in feed for chickens) The cetculstions in Table 55 give the ‘analyse of variance in Table 5.7. ‘The variance ratio forthe protein-by-fishmeal intetsction is apcroximately 1, so itis clear hat there is ro interaction, «UC there bad been an interaction, it would probably have been ‘hetier to report the effee's of adding fishrweal to each type of protein.) The variance ratios for protein ard for Tohnical are hath grealct than 21.20, which iv the 99% point of t divtibution on { and 4 degrees of Geedom, $0 we can report that added flswncal inereanes ‘welgh by 39.5 gm respective of type of protein and tht replacing groundnuts hy soya bet increases weight by 48.5 gm whether or Hol Ashuneal Is addod. The staedard errors of both of these differences is //132.25% (2/4), which is 8.13. Here is a waming about vocabulary. Many sciemtists say stat (wo factors “intectet! to seem that they ‘act together” inthe sense thal youcan add their paraté main etfocts, This is precisely what stalstiiuns cell "zero intecaction". Thus iti always a good iden to report the presence or absence of interaction by pointing out what this means in 1he particuiar case, “The phrase ‘main etfect’ can also be miisinterprsted, In Example 5.12, x chicken breeder who thinks that grounds are the obvious source of protein tay say that tke ‘sain effect”92 Chapter 5. Ractorial wenement structure ‘Table 5.8, Trearment totals in Excaple 5.10 Nitrogen Percentage of vetch Niuogen Cr ee Yes [1277 16.4 1429 2.80 aay | GH.ad No. Wnes_u186 1277 1203 9.89] 56.29 Seed mine wet | 22.81 26.00 2706 2483 2005] 120.78 CAEN) —Cwithoul Ny |E -228— TO? a0] 20 Tiaeae 2.67 216 1641.12, ost] 0.20 fnmims men | 1.03 0.52 000 0.52 —Los] 0.09 of fishmeal is ft ~ ty, = 493.00~ 401.50 = 31.50. Example 5.14 revisited (Vetch snd eats) Table 5.8 shows the treatment totale, In fast, the ‘oxperiment was conducted in five complete blocks, so CSS(mean) = 120.747/50 = 201.7562 ‘nd CSS(seed) = (22.81? 4-26.00? + 27.06 + 24.8% + 20.087)/10 = 294.8275, finther more, C8S(nittogen) — (64.49? + 56.297)/25 = 293.4010, and the crude auia of squares for reatmems Is (12.777 4----+-9.597)/5 = 296.4640. From these we calculate 88(sead) = 294 8275 ~ 291.7562 = 3.0713, SS(oitregen) = 293, 1010 ~ 291.7862 ~ 1.2448 and the sum of squares for ifteravtion Is 295.4640 — 291.7562 — 3.0713 ~ 1.48 0.2917. Hence the ‘ean square for interection is 0.0729. The complote set oF dat nat shown here) gives the residual moan square wo be 6.0402, AL frst sight, Items that there is no evidence of ine action. However, Figure 5.8 shows thatthe offoo of nitregen decreases ws the percentage of vetch increases, which is exacily whal we would expect, Table 5. has some extra rows, The fist shows the difference between the two levels of nitrogen for cach seed mixture, The second azves the lincar fi, to is eifference, as a polynomial in the percentage of veich: “The third subtracts the mean of the previous row. Since ezch treatment tctal comes fom five plots, this final ow conesponds to valuss of 40.103 on the five plots sth pitogen bit no YA, 103 sin the five plo's with neither nitsogen nor velc, and so on. These values sum to zero fon each love af nitrogen aad on each level ofthe sed mixture, so the voctic they represcat Is pert of the interaction, called the “atrogen-by-linear-vetca’ pal of Ihe interaction, Its sein of ‘squares is 20% (0,103* 401.052") = 0.2663, which accounts for most of the euia of squares for tne interaction. 1K has one degroc of freedom (fer ting the straight fine), s0.s%e obtain (he analysis of sariance in Table 5.9 (using infomation abont block totals from the vomplote ata set), This shows clearly thet the effect of nitrogen changes tinearly with the-perzentege ‘oC vetch in the sce mixture, ad thal there i ao evidence of any further interaction ‘When one of the eaunent factors is quantitative, iis often instructive ts decompose the ineraction in this way. 5.6 Three treatment factors Now suppose that we have threo tact Lg and Ly respectively, Thea se wal factors FG and M whose sets of levels are Ly, leaf te = [Eel and ag = [of We assume thet the 346. Tu treat andt Nal functo| noth oer Defini! Tevel od “el ively. Asin projec ofyor is tho | r vowed "|