Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecturer: Dr. Michael Stuart, Department of Statistics email: mstuart@tcd.ie Lectures: Tuesday, Thursday, Laboratory: Thursday, March 12th, Tuesday, March 31st, Postgraduate Certificate in Statistics Design and Analysis of Experiments 6.00 - 8.00pm 6.00 - 8.00pm 6.00 - 8.00pm Lecture 1.1 1 © 2015 Michael Stuart Design and Analysis of Experiments Course Outline • The need for experiments – experimental and observational studies – cause and effect – control Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 2 © 2015 Michael Stuart Design and Analysis of Experiments Course Outline • Basic design principles for experiments – Control – Blocking (pairing) – Randomization – Replication – Factorial structure Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 3 © 2015 Michael Stuart Design and Analysis of Experiments Course Outline • Standard designs – Completely randomized designs – Randomized blocks – Two-level factors – Split units Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 4 © 2015 Michael Stuart Design and Analysis of Experiments Course Outline • Analysis of experimental data – Exploratory data analysis – Effect estimation and significance testing – Analysis of variance – Statistical models, fixed and random effects – Model validation, diagnostics – Software laboratories Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 5 © 2015 Michael Stuart Design and Analysis of Experiments References Mullins, E., Statistics for the Quality Control Chemistry Laboratory, Royal Society of Chemistry, 2003, particularly Chapters 4-5, 7-8. (EM) Available as an electronic resource Montgomery, D.C., Design and analysis of experiments, 8th ed., Wiley, 2013. (DCM) Dean, Angela and Voss, Daniel, Design and analysis of experiments, Springer, 1999. (DV) Available as an electronic resource Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 6 © 2015 Michael Stuart Design and Analysis of Experiments Further reading Box, G.E.P, Hunter, J.S. and Hunter, W.G., Statistics for Experimenters, 2nd. ed., Wiley, 2005. (BHH) Daniel, C., Applications of Statistics to Industrial Experimentation, Wiley, 1976. (CD) Mead, R., Gilmour, SG and Mead, A, Statistical Principles for the Design of Experiments: Applications to Real Experiments, Cambridge University Press, 2012. (MGM) Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 7 © 2015 Michael Stuart Design and Analysis of Experiments Lecture notes and supplements Module Web Page Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 8 © 2015 Michael Stuart Assessment • 3-hour written examination – 3 questions. Questions 1 and 2 carry 30 marks each, Question 3 carries 40 marks. – Appendix gives tables of critical values of the t-distribution and selected critical values of the F distribution. – Non-programmable calculators are permitted for this examination Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 9 © 2015 Michael Stuart Assessment Examination dates: Monday 27 April to Friday 22 May 2014 inclusive Examination Timetables will be available in March "The onus lies on each student to establish the dates, times and venues of examinations by consulting the relevant timetable on the College website. No timetable or reminder will be sent to individual students by any office." Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 10 © 2015 Michael Stuart Course assessments • Module assessment, as for Base Module • End of Lecture, Minute Tests – How much did you get out of today's class? – How did you find the pace of today's class? – What single point caused you the most difficulty? – What single change by the lecturer would have most improved this class? Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 11 © 2015 Michael Stuart Lecture 1.1 1. Introduction to Course 2. What is an experiment? 3. Case study: Industrial process improvement − three design principles 4. Case study: Clinical trial − how randomization works 5. Multifactor Designs 6. Other application areas 7. Experimental vs Observational Studies 8. Strategies for Experimentation Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 12 © 2015 Michael Stuart Part 2 What is an experiment? Try something, to see what happens Try something different, to see the difference in what happens Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 13 © 2015 Michael Stuart Experiment as demonstration Pendulum – length L – period T T 2 L g L g 4 2 T 2 Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 14 © 2015 Michael Stuart Newton's colour demonstration Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 15 © 2015 Michael Stuart Newton's colour demonstration Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 16 © 2015 Michael Stuart Newton's colour demonstration Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 17 © 2015 Michael Stuart Thought experiments • Aristotle (4th century BC): – speed of falling objects is proportional to weight • Galileo (17th century AD): – not true! Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 18 © 2015 Michael Stuart Comparative experiments • Galileo's pendulum experiments • A comparative experiment is a programme of actions undertaken to study the effects of making changes to a process or system. • “To find out what happens when you change something, it is necessary to change it”. (BHH, p. 404) Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 19 © 2015 Michael Stuart Control, a key feature of comparative experiments • Complete control – the counterfactual argument • Practical control of study environment – chance variation if no change introduced – comparing results of change to no change involves a test of statistical significance Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 20 © 2015 Michael Stuart Lecture 1.1 1. Introduction to Course 2. What is an experiment? 3. Case study: Industrial process improvement − three design principles 4. Case study: Clinical trial − how randomization works 5. Multifactor Designs 6. Other application areas 7. Experimental vs Observational Studies 8. Strategies for Experimentation Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 21 © 2015 Michael Stuart Part 3 Case study on process improvement • Comparison of standard (old) process and new process for manufacture of electronic components • Key criterion: – number of defective components Ref: EM Notes, Ch 4, Example 1, pp. 3-6 Hahn.xls Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 22 © 2015 Michael Stuart Experimental design • 50 components sampled per day, • 6 days per week, • 8 weeks, • Systematic layout, as follows Week Number 1 2 3 4 5 6 7 8 Monday Old New Old New Old New Old New Tuesday New Old New Old New Old New Old Wednesday Old New Old New Old New Old New Thursday New Old New Old New Old New Old Friday Old New Old New Old New Old New Saturday New Old New Old New Old New Old Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 23 © 2015 Michael Stuart Sampling plan 50 components sampled per day Measurement: X = number of defectives in sample of 50 Why 50? Why not 1? 100? the whole lot? For fair comparison, let p = X/n SE(p) = (1 ) n Ref: EM Notes Ch 3 p 2 Measurement precision Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 24 © 2015 Michael Stuart Results Numbers of defectives per daily sample of 50 for 48 days (8 weeks) Day Defectives 1 0 2 0 3 6 4 3 5 3 6 3 7 4 8 1 9 0 10 2 11 0 12 0 Day Defectives 13 1 14 0 15 3 16 1 17 0 18 2 19 0 20 1 21 2 22 0 23 1 24 3 Postgraduate Certificate in Statistics Design and Analysis of Experiments Day Defectives 25 0 26 0 27 0 28 2 29 0 30 0 31 1 32 1 33 0 34 0 35 0 36 2 Day Defectives 37 2 38 0 39 0 40 0 41 0 42 0 43 1 44 0 45 2 46 0 47 0 48 0 Lecture 1.1 25 © 2015 Michael Stuart Comparison of two processes over eight weeks Numbers of Defectives in Samples of 50 Units Day Old New Difference pair Process Process (New – Old) 1 0 0 0 2 6 3 –3 3 3 3 0 4 1 4 +3 5 2 0 –2 6 0 0 0 7 1 0 –1 8 3 1 –2 9 0 2 +2 10 1 0 –1 11 0 2 +2 12 3 1 –2 Postgraduate Certificate in Statistics Design and Analysis of Experiments Numbers of Defectives in Samples of 50 Units Day Old New Difference pair Process Process (New – Old) 13 0 0 0 14 0 2 +2 15 0 0 0 16 1 1 0 17 0 0 0 18 2 0 –2 19 2 0 –2 20 0 0 0 21 0 0 0 22 0 1 +1 23 0 2 +2 24 0 0 0 Lecture 1.1 26 © 2015 Michael Stuart Comparison of two processes over eight weeks Numbers of Defectives Summary Old Process New Process Difference (New – Old) 25 22 –3 2.08 1.83 –0.25 Total 8 week averages per cent Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 27 © 2015 Michael Stuart Differences in numbers defective, with control limits 8 6 4 2 Difference 0 -2 -4 -6 -8 4 8 12 16 20 24 Day Pair No statistical significance! Ref: EM Notes Ch 1 § 1.7 Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 28 © 2015 Michael Stuart Calculating the control limits Numbers of Defectives in Samples of 50 Units Day Old New Difference pair Process Process (New – Old) 1 0 0 0 2 6 3 –3 3 3 3 0 4 1 4 +3 5 2 0 –2 6 0 0 0 7 1 0 –1 8 3 1 –2 9 0 2 +2 10 1 0 –1 11 0 2 +2 12 3 1 –2 SD(Differences) = 1.57 Numbers of Defectives in Samples of 50 Units Day Old New Difference pair Process Process (New – Old) 13 0 0 0 14 0 2 +2 15 0 0 0 16 1 1 0 17 0 0 0 18 2 0 –2 19 2 0 –2 20 0 0 0 21 0 0 0 22 0 1 +1 23 0 2 +2 24 0 0 0 Control limits: 0 3xSD = 4.7 Ref: hahn.xls Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 29 © 2015 Michael Stuart Formal significance test D0 Z SE(D ) D SD / n From Summary table, sum of differences = – 3 From control limit calculation, SD = 1.57 Z 3 / 24 1.57 / 24 0.39 -4 -3 -2 -1 0 1 2 3 4 not statistically significant Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 30 © 2015 Michael Stuart Alternative design (proposed by engineers) Week Number 1 2 3 4 5 6 7 8 Monday Old Old Old Old New New New New Tuesday Old Old Old Old New New New New Wednesday Old Old Old Old New New New New Thursday Old Old Old Old New New New New Friday Old Old Old Old New New New New Saturday Old Old Old Old New New New New Assume this design was used; check for no effect Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 31 © 2015 Michael Stuart Defect rates, per cent, with differences, for the first and second four week periods First Period Second Period Difference Both Processes 3.0 0.9 2.1 Old Process 3.3 0.8 2.5 New Process 2.7 1.0 1.7 Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 32 © 2015 Michael Stuart Testing statistical significance P̂ P̂ P̂P̂11 1P̂ P̂P̂22 2 Z P̂ Z P̂ (100 1P̂ ) 2 P̂ (100 P̂ ) Z 1 (100 1 P̂ 2 (100 2 P̂ ) P̂ ) P̂ P̂ P̂ ( 100 P̂ ) 1 1 2 2 1 1 2 2 P̂1 (100 P̂1 ) P̂2 (100 P̂2 ) n1 n2 n n n n1 n 22 1 Ref: EM Notes Ch 3 p 11 2 .000.0.99.9 333.3.00 .0 0 .9 3397 97 9 99 99 .1 99 .1..11 3 97 9700.0 0.99..9 99 1200 1200 1200 1200 1200 1200 1200 2 2...111 2 00..56 56 33..75 75 -4 -3 -2 -1 0 1 2 3 4 highly statistically significant! Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 33 © 2015 Michael Stuart Classwork 1.1.1 Assess the statistical significance of the difference in defect rates, %, between the first period and second period for the old process. Homework 1.1.1 Assess the statistical significance of the difference in defect rates, %, between the first period and second period for the new process. Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 34 © 2015 Michael Stuart How can this be? Numbers defective in time order 6 5 4 Defectives 3 2 1 0 6 12 18 24 30 36 42 48 Day Long term downward trend, systematic bias Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 35 © 2015 Michael Stuart How to avoid systematic bias • Make comparisons under homogeneous experimental conditions • 1 Systematic arrangement, as implemented: avoids known biases • 2 Random allocation: within each day pair, allocate old and new processes at random avoids known and unknown biases Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 36 © 2015 Michael Stuart Random vs Systematic allocation Week Number 1 2 3 4 5 6 7 8 Monday Old New Old New Old New Old New Tuesday New Old New Old New Old New Old Wednesday Old New Old New Old New Old New Thursday New Old New Old New Old New Old Friday Old New Old New Old New Old New Saturday New Old New Old New Old New Old Suppose there is an additional "other factor", unknown to the experimenter with settings Up, Down, settings alternate every day, including Sunday Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 37 © 2015 Michael Stuart Random vs Systematic allocation Week 1 Week 2 Experimental Factor Other Factor Experimental Factor Other Factor Monday Old Up New Down Tuesday New Down Old Up Wednesday Old Up New Down Thursday New Down Old Up Friday Old Up New Down Saturday New Down Old Up Sunday Up Old and Up always coincide, New and Down always coincide. Factors are "confounded" Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 38 © 2015 Michael Stuart Random vs Systematic allocation Random allocation minimises chances that experimental factor settings pattern coincides with other factor settings pattern. Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 39 © 2015 Michael Stuart Two design principles • Blocking (or local control) – identify homogeneous blocks of experimental units – assess effects of experimental change within homogeneous blocks – average effects across blocks • Randomization – allocate experimental settings to units at random Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 40 © 2015 Michael Stuart Another design principle • Replication – 24 comparisons • Why 24 • Why not 1? 50? 100? SD( D ) / n – power calculation Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 41 © 2015 Michael Stuart Lecture 1.1 1. Introduction to Course 2. What is an experiment? 3. Case study: Industrial process improvement − three design principles 4. Case study: Clinical trial − how randomization works 5. Multifactor Designs 6. Other application areas 7. Experimental vs Observational Studies 8. Strategies for Experimentation Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 42 © 2015 Michael Stuart Part 4 Clinical trial of heart disease treatments • 596 patients suffering from heart disease • to be treated by drugs or by surgery • each patient assigned at random to one treatment – 310 (52%) assigned to Drugs – 286 (48%) assigned to Surgery • Was the randomization successful? Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 43 © 2015 Michael Stuart Was the randomization fair? Z P̂ 50 P̂ (100 P̂) / n 52 50 52 48 / 596 2 2.05 0.98 Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 44 © 2015 Michael Stuart Balance with respect to Covariates Drugs per cent Surgery per cent Limitation in ordinary activity 94 95 History of heart attack 59 64 Heart attack indicated by electrocardiogram 36 41 Duration of chest pain >25 months 50 52 History of high blood pressure 30 28 History of congestive heart failure 8.4 5.2 History of stroke 3.2 2.1 History of diabetes 13 12 Enlarged heart 10 12 High serum cholesterol 32 21 Covariate Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 45 © 2015 Michael Stuart Balance with respect to Covariates Drugs Surgery per cent per cent Limitation in ordinary activity 94 95 -0.5 Histroy of heart attack 59 64 -1.3 Heart attack indicated by electrocardiogram 36 41 -1.3 Duration of chest pain >25 months 50 52 -0.5 History of high blood pressure 30 28 0.5 History of congestive heart failure 8.4 5.2 1.6 History of stroke 3.2 2.1 0.8 History of diabetes 13 12 0.4 Enlarged heart 10 12 -0.8 High serum cholesterol 32 21 3.1 Covariate Postgraduate Certificate in Statistics Design and Analysis of Experiments Z(Diff-0) Lecture 1.1 46 © 2015 Michael Stuart How randomization works • Balance with respect to – known covariates AND − unknown covariates (not achieved by systematic assignment) • Minimize experimenter bias Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 47 © 2015 Michael Stuart 1. Class count 2. Random number Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 48 © 2015 Michael Stuart Lecture 1.1 1. Introduction to Course 2. What is an experiment? 3. Case study: Industrial process improvement − three design principles 4. Case study: Clinical trial − how randomization works 5. Multifactor Designs 6. Other application areas 7. Experimental vs Observational Studies 8. Strategies for Experimentation Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 49 © 2015 Michael Stuart Part 5 Multi-factor Designs • Traditional versus statistical design – efficiency – interaction Ref: EM §5.2 Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 50 © 2015 Michael Stuart Multi-factor designs are efficient Illustration: • Yield of a chemical manufacturing process affected by – operating pressure, – operating temperature • Choose between – Low and High pressure – Low and High temperature • Resources available for 12 experimental runs Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 51 © 2015 Michael Stuart Traditional “one-at-a-time” design, Y5 Y6 Y7 Y8 High Pressure Low (best) Y1 Y2 Y3 Y4 Low Postgraduate Certificate in Statistics Design and Analysis of Experiments Temperature High Y9 Y10 Y11 Y12 Lecture 1.1 52 © 2015 Michael Stuart Fisher’s two-factor design Y10 Y11 Y12 Y7 Y8 Y9 High Pressure Low Y1 Y2 Y3 Low Postgraduate Certificate in Statistics Design and Analysis of Experiments Temperature High Y4 Y5 Y6 Lecture 1.1 53 © 2015 Michael Stuart Calculation of effect estimates Pressure main effect, traditional design: (Y5+Y6+Y7+Y8)/4 – (Y1+Y2+Y3+Y4)/4 SE: 2 4 Pressure main effect, Fisher design (Y7+Y8+Y9+Y10+Y11+Y12)/6 – (Y1+Y2+Y3+Y4+Y5+Y6)/6 SE: 2 6 Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 54 © 2015 Michael Stuart Multi-factor designs find best operating conditions 75 best 60 High Pressure Low best 65 Low Postgraduate Certificate in Statistics Design and Analysis of Experiments Temperature High 70 best Lecture 1.1 55 © 2015 Michael Stuart Multi-factor designs reveal interaction Classwork 1.1.2: Calculate Pressure effect at Low Temperature and at High Temperature; calculate the difference Calculate Temperature effect at Low Pressure and at High Pressure; calculate the difference Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 56 © 2015 Michael Stuart Multi-factor designs reveal interaction 75 60 Pressure effect Low T: 60 – 65 = –5 High T: 75 – 70 = +5 Diff: 5 – (–5) = 10 High Pressure Temperature effect Low 65 Low Temperature Postgraduate Certificate in Statistics Design and Analysis of Experiments High 70 Low P: 70 – 65 = 5 High P: 75 – 60 = 15 Diff: 15 – 5 = 10 Lecture 1.1 57 © 2015 Michael Stuart Interaction defined Factors interact when the effect of changing one factor depends on the level of the other. Interaction displayed Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 58 © 2015 Michael Stuart Lecture 1.1 1. Introduction to Course 2. What is an experiment? 3. Case study: Industrial process improvement − three design principles 4. Case study: Clinical trial − how randomization works 5. Multifactor Designs 6. Other application areas 7. Experimental vs Observational Studies 8. Strategies for Experimentation Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 59 © 2015 Michael Stuart Part 6 Other application areas • Agriculture • Genetics • Biological Sciences • Physical Sciences • Engineering • Psychology • Social Sciences? Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 60 © 2015 Michael Stuart Lecture 1.1 1. Introduction to Course 2. What is an experiment? 3. Case study: Industrial process improvement − three design principles 4. Case study: Clinical trial − how randomization works 5. Multifactor Designs 6. Other application areas 7. Experimental vs Observational Studies 8. Strategies for Experimentation Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 61 © 2015 Michael Stuart Part 7 Experimental vs Observational Studies Example: Process improvement study, old or new process Observational study: new process is run, old process inventory is sampled, product from old and new processes compared Experiment: process is changed from day to day, under controlled conditions • Current control vs historical control Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 62 © 2015 Michael Stuart Example: Clinical trial, drugs or surgery Observational study: check patient records, compare drug and surgery Experiment: assign patients at random, compare drug and placebo Retrospective vs Prospective Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 63 © 2015 Michael Stuart Number of Deaths from Cancer Lurking Variables Number of Churches • Lurking variable = Population size • Covariance Analysis ? or try number of deaths per thousand Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 64 © 2015 Michael Stuart Lurking Variables 80 Population ('000) 75 70 65 60 55 50 100 150 200 Number of Storks Postgraduate Certificate in Statistics Design and Analysis of Experiments 250 300 Ref: BHH Ch 1 p 8 Lecture 1.1 65 © 2015 Michael Stuart Experiment vs Observation • control of input factors; • no control of input factors (happenstance); • control of environment; • environment may vary; • blocking to control known non-experimental factors; • matching to control nonexperimental factors; • randomization to minimse the effects of unknown non-experimental factors • randomization impossible; "lurking" variables possible Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 66 © 2015 Michael Stuart Cause and effect • Fisher's randomized controlled experiment, – the "gold standard" • Rubin's matching via propensity scoring • Pearl's Structural Causal Model • etc. Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 67 © 2015 Michael Stuart Caustic comments ... large segments of the statistical research community find it hard to appreciate and benefit from the many results that causal analysis has produced in the past two decades. Pearl (2009) Statistics Surveys Vol. 3 96–146 I appreciate the opportunity to expand on the essential point of Shrier’s and Pearl’s letters, because I think that it has fostered, and continues to foster, bad practical advice, which is based on an unprincipled and confused theoretical perspective. Rubin (2009) Statist. Med., 28:1415–1424 Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 68 © 2015 Michael Stuart Fisher on smoking and lung cancer "The evidence linking cigarette smoking with lung cancer, standing by itself, is inconclusive, as it is apparently impossible to carry out properly controlled experiments with human material. Observations not fulfilling the requirements of decisive experimentation might be suggestive, not conclusive, and may be afforded a confidence which is more than their due. Association is not causation." RA Fisher, quoted in "Cigarette-cancer links disputed", New York Times, Dec. 29, 1957 Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 69 © 2015 Michael Stuart Regression analysis and cause and effect "The justification sometimes advanced that a multiple regression analysis on observational data can be relied upon if there is an adequate theoretical background is utterly specious and disregards the unlimited capability of the human intellect for producing plausible explanations by the carload lot". K.A. Brownlee, 1965 Big Data Analytics Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 70 © 2015 Michael Stuart Lecture 1.1 1. Introduction to Course 2. What is an experiment? 3. Case study: Industrial process improvement − three design principles 4. Case study: Clinical trial − how randomization works 5. Multifactor Designs 6. Other application areas 7. Experimental vs Observational Studies 8. Strategies for Experimentation Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 71 © 2015 Michael Stuart Part 8 Strategies for Experimentation Box on strategy: When you see the credits roll at the end of a successful movie you realize there are many more things that must be attended to in addition to choosing a good script. Similarly in running a successful experiment there are many more things that must be attended to in addition to choosing a good experimental design. (BHH, End notes) Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 72 © 2015 Michael Stuart Robinson's outline Ref: GKR p.6, see also p.7 Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 73 © 2015 Michael Stuart Lecture 1.1 1. Introduction to Course 2. What is an experiment? 3. Case study: Industrial process improvement − three design principles 4. Case study: Clinical trial − how randomization works 5. Multifactor Designs 6. Other application areas 7. Experimental vs Observational Studies 8. Strategies for Experimentation Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 74 © 2015 Michael Stuart Minute test – How much did you get out of today's class? – How did you find the pace of today's class? – What single point caused you the most difficulty? – What single change by the lecturer would have most improved this class? Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 75 © 2015 Michael Stuart Reading Lecture 1.1 EM Sections 4.3, 4.5.1, 4.5.3, 4.6, 5.2 Supplementary reading: DCM Chapter 1, Section 2.5 DV Chapter 1, Sections 2.2, 2.3 Next lecture: EM Notes, Chapter 4 Postgraduate Certificate in Statistics Design and Analysis of Experiments Lecture 1.1 76 © 2015 Michael Stuart