Vanderbilt University Medical Center SRC Presentation Vincent Kokouvi

Vanderbilt University Medical Center SRC Presentation Vincent Kokouvi Agboto Assistant Professor/Director of Biostatistics, Meharry Medical College Assistant Professor of Biostatistics, Vanderbilt University Medical Center

Introduction to Experimental Designs in Biological and Clinical Settings.

1. Overview Introduction 2. Examples of Classical Designs 3. Optimal Experimental Design 4. Other Designs Issues 5. Conclusion

1. Introduction l Experiment: Investigation in which investigator applies some treatments to experimental units and then observes the effects of treatments on the experimental units through measurement of response (s).

1. Introduction Treatment: Set of conditions applied to experimental units in an experiment. l Experimental Unit: Physical entity to which a treatment is randomly assigned and independently applied. l

1. Introduction Response variable: Characteristic of an experimental unit that is measured after treatment and analyzed to assess the effects of treatments on experimental units. l Observational Unit: Unit on which a response variable is measured. l

1. Introduction l Experimental design procedure: l Basic idea: Appropriate selection of values of control variables. Three Fundamental of Experimental Design Concepts: Randomization, Blocking, Replication. (R. A. Fisher) l Decision before data collection.

1. Introduction l Important stages of an Experimental Research: Background of the experiment; Choice of factors; Reduction of error; Choice of model; Design criterion and Size of the design; Choice of an experimental design; Conduct of the experiment and Analysis of the data

1. Introduction Classical (Standard) Designs l Optimal Experimental Design: Only alternative when the standard designs do not provide us with adequate answers l

2. Examples of Classical Designs Example 1: Soils Moisture and gene Expression in maize seedlings. l Example 2: Drug and Feed Consumption on Gene Expression in rats. l Example 3: Treatments on Gene Expression in dairy cattle. l

Example 1 l l l Experiment: Effect of three soil moisture levels on gene expression in maize seedlings. Total of 36 seedlings were grown in 12 pots with 3 seedlings per pot. Three soil moisture levels (low, medium, high) randomly assigned to the 12 pots. After three weeks, RNA extracted from the above ground tissues of each seedling. Each of the 36 RNA samples was hybridized to a microarray slide to measure gene expression.

Example 1 (continued) l l l Treatment: The three moisture levels Experimental Unit: Moisture levels randomly assigned to the pots Pots: experimental units. A pot consisting of 3 seedlings is one experimental unit. Observational units: Gene expression was measured for each seedling Seedlings: Observational units. Response variable: Each probe on the microarray slide provide one response variable. This is the Standard Experimental Design (CRD).

Example 2 l l Experiment: Gauge the effects of a drug and feed consumption on gene expression in rats. A total of 40 rats were housed in individual cages. Half of them calorie-restricted diet (R); Another half Provided with access to feeders that were full so calories intake unrestricted (U). Within each diet group, four doses of an experimental drug (1, 2, 3, 4) rats with 5 rats per dose within each diet group.

Example 2 (continued) l At the conclusion of the study, gene expression was measured for each rat using microarrays.

Example 2 (continued) l l l Treatment (factors): Diet and Drug. Factor Diet (R, U); Factor Drug (1, 2, 3, 4) Each combination of diet and drug: Treatment (R 1, R 2, R 3, R 4, U 1, U 2, U 3, U 4). Each rat: Experimental unit/Observational unit. Response variable: Each probe on the microarray slide. This is a full factorial treatment design. It was used because all possible combination of diet and drug were considered.

Example 3 l l l Experiment: Study the effects of 5 treatments (A, B, C, D, E) on gene expression in dairy cattle. A total of 25 Gene. Chips and a total of 25 cows, located on 5 farms with 5 cows on each farm are available for the experiment. Which of the following designs is better from a statistical standpoint?

Example 3 (Continued) l l Design 1: To reduce variability within treatment groups, randomly assign the 5 treatments to the 5 farms so all 5 cows on any one farm receive the same treatment. Measure gene expression using one Gene. Chip for each cow. Design 2: Randomly assign the 5 treatments to the 5 farms within each farm so that all 5 treatments are represented on each farm. Measure gene expression using one Gene. Chip for each cow.

Example 3 (continued) l Design 1 Farm 1: B B B Farm 2: D D D Farm 3: A A A Farm 4: E E E Farm 5: C C C Design 2 Farm 1: A B E D C Farm 2: E D A C B Farm 3: C D E A B Farm 4: A B E C D Farm 5: C A D B E

Example 3 (continued) l l Observation Units: Cows in both designs. Experimental Units: Farms in Design 1 and Cows in Design 2: a randomized complete block design (RCBD) with a group of 5 cows on a farm serving as a block of experimental units. Design 1 has no replication because only 1 experimental unit for each treatment. Design 2 has 5 replications per treatment.

Design 3 (continued) l l l Design 2 is by far the better design. We can compare treatments directly among cows that share the same environment. With Design 1, it is impossible to separate difference in expression due to treatment effects from differences in expression due to farm effects.

3. Optimal Experimental Design 3. 1. Motivation Example 3. 2. Comments on Orthogonal Designs. 3. 3. Some Examples of Non-Orthogonal Designs 3. 4. Optimal Designs

3. 1. Motivating Example Suppose that the yield is linearly related to temperature whose range is [50, 150]: Y= a + b X l If we want conduct experiments at two points, which of the following will we choose: Design 1 at 50 and 150? Design 2 at 70 and 130? Design 3 at 90 and 110? l

3. 1. Motivating Example What is the optimal design in this case? l Better design among the three designs mentioned l

3. 1. Motivating Example l It is the design 1 because it gives the smallest confidence region for the parameters (D-optimality) and also give the smallest maximum variance for the predicted responses (G-optimality)

3. 2. Comments on orthogonal Designs l Pros (Many desirable properties) - Easy to calculate - Easy to interpret - Maximum Precision (in some sense) - Tabled designs widely available

3. 2. Comments on Orthogonal Designs l Cons: Not applicable if - Irregular design space - Mixture experiments - Sample size not power of 2 - Mixed qual and quant factors - Fixed covariates - Nonlinear models

2. 3. Some Examples of Non-Orthogonal Designs 16 -run design with 8 two-level factors with main effects and 6 interactions: BC, CH, BH, DE, EF, DF l 12 -run mixed level design with one 3 level factor and 9 two-level factors l

2. 4. Optimal Designs l l l Optimal Experimental Design (OED): Standard alternative when classical designs not applicable. Choice of a particular experimental design: Depends on the experimenter’s design criterion (optimization problem). OED: Reduce costs of experimentation by allowing statistical models to be estimated with fewer experimental runs; Evaluated using statistical criteria.

3. 4. Optimal Designs Ynxp ~ N (X + , 2 I), Xnxp: design matrix, : unknown px 1 parameter vector and 2: known l y(xi) = f’(xi) + i l X=[f(x 1), …, f(xn)]’ l

3. 4. Optimal Designs Design : Probability measure over a compact region with (xi) = i l places weight (xi) on xi l Problem: n (xi) is not necessary an integer l

3. 4. Optimal Designs l Approximate design: = x 1 x 2… xn 1 2… n with (dx) =1 and 0 i 1 l Exact design: n (xi) must be an integer

3. 4. Optimal Designs n. M( )=X’X= m(x) (dx)= f(x) f’(x) (dx) = i f(xi)’ : Information matrix of l Optimality crietria: * = arg max (M( )) l

3. 5. Some Useful Criteria D-Optimality: max |X’X|: l A-Optimality: min{trace (X’X)-1} l G-Optimality: min{max d(x)} where d(x) = f’(x)(X’X)-1 f(x) l V-Optimality: min{average d(x)} l

3. 5. Some Useful Criteria l l D and A-Optimality: Estimation based criteria. G and V-Optimality: Prediction based criteria.

3. 6. Algorithms for Optimal Designs l l Development of efficient computing methods and high power computer systems Great interest in algorithmic approaches. In general: Difficult to find exact designs analytically. Finding exact designs Solving a large nonlinear mixed integer programming problem. In practice: Find designs close to the best design locally optimal introduction of exact design algorithms.

3. 6. Algorithms for Optimal Designs l Typical Exact Design Algorithm steps: - Choose an initial feasible solution design - Modify solution slightly, by exchanging a point in the design for a point in the design space .

3. 6. Algorithms for Optimal Designs Fedorov algorithm (Fedorov, 1969). l Modified Fedorov algorithm(Johnson and Nachtsheim, 1983). l K-L exchange algorithm (Donev and Atkinson, 1988). l Coordinate exchange algorithm (Meyer and Nachtsheim, 1995). l Columnwise-Pairwise (CP) algorithm (Wu and Li, 1999). l

3. 7. Software for the Computation of Optimal Designs l l l SAS JMP Matlab R C++

4. Other Designs Issues l l Supersaturated Designs Bayesian Designs Model Robust Designs Model Discrimination Designs

5. Conclusion l l l All problems are different Statistical knowledge will help improve the design. Get involved with the statistician (biostatistician) early in the process. Collaborate closely with people who know the background of the study. Even the most sophisticated statistical analysis could save do much to save a study based on a “bad design”.

References l l l Agboto V. , 2006. Bayesian approaches to model robust and model discrimination designs. Unpublished Ph. D. dissertation, School of Statistics, University of Minnesota. Agboto V, Nachtsheim C, Li W. Screening designs for model discrimination. Journal of Statistical Planning and Inference, 140: 3, 766 -780, 2010. Atkinson, A. C & Donev, A. N. (1992): “Optimal Experimental Designs”. Oxford Statistical Sciences Series: 8, 1 -328. Chaloner, K. (1984). “Bayesian experimental design: A review”. Statistical Science 10, 273 -304. Cook, R. D. & Nachtsheim, C. J. (1982). “A comparison of algorithms for constructing exact D-opitmal designs”. Technometrics 22, 315 -324. Li, W. & Wu, C. F. J. (1997). “Columwise-pairwise algorithms with applications to the construction of supersaturated designs”. Technometrics 39, 171 -179.