360 likes | 479 Views
Sequential Experimental Designs For Sensitivity Experiments NIST GLM Conference April 18-20, 2002. Joseph G. Voelkel Center for Quality and Applied Statistics College of Engineering Rochester Institute of Technology. Sensitivity Experiments. ASTM method D 1709–91
E N D
Sequential Experimental Designs For Sensitivity ExperimentsNIST GLM ConferenceApril 18-20, 2002 Joseph G. Voelkel Center for Quality and Applied Statistics College of Engineering Rochester Institute of Technology
Sensitivity Experiments • ASTM method D 1709–91 • Impact resistance of plastic film by free-falling dart method 2
Engineer Specify a probability of failure g – 0.50, 0.10, … Find dart weight x=d such that Prob(F; d)=g Statistician Find a strategy for selecting weights {xi} so that d is estimated as precisely as possible Objectives Darts are dropped one at a time. Weight of ith dart may depend on results obtained up to date 3
Data Collection Possibilities Non-sequential • Specify n and all the {xi} before any Pass-Fail data {Yi} are obtained. • Find dose d of drug at which 5% of mice develop tumors Group-sequential • Example: two-stage. • Specify n1 and the {x1i}. Obtain data {Y1i} • Use this info to specify n2 and the {x2i}. Obtain data {Y2i} • Same mice example, but with more time. (Fully) Sequential • Use all prior knowledge: x1Y1 x2Y2 x3Y3 x4Y4 • Dart-weight example. One machine, one run at a time. 4
Objective: Example Estimate weight d at which 10% of the samples fail • So, try to set the {xi} to minimize Model and Objectives 5
(Fully) Sequential experiments Estimating a d corresponding to a given g, e.g. 0.10. The real problem. g = 0.50? g = 0.001? Our Interest 6
Up-Down Method. Dixon and Mood (1948) Only for g=0.50 Robbins-Munro (1951) wanted {xi} to converge to d. Like Up-Down, but with decreasing increments g far from 0.50 convergence is too slow A Quick Tour of Some Past Work 7
A Quick Tour of Some Past Work • Wu’s (1985) Sequential-Solving Method • Similar in spirit to the R-M procedure • Collect some initial data to get estimates of a and b • Better than R-M, much better than Up-and-Down • Performance depends somewhat heavily on initial runs • Asymptotically optimal, in a certain sense 8
Some Non-Sequential Bayesian Results • Tsutakawa (1980) • How to create design for estimation of d for a given g. • Certain priors on d and b • Some approximations • Assumed constant number of runs made at equally spaced settings. • Chaloner and Larntz (1989) • Includes how to create design for estimation of d for a given g • Some reasonable approximations used • Not restricted to constant number of runs or equally spaced settings. 9
This Talk. Bayesian Sequential Design • A way to specify priors • Measures of what we are learning about a, b, and d—AII and Information • Specifying the next setting, with some insights • Some examples and comparisons • Rethinking the priors 12
Specifying Priors • Consider the related tolerance-distribution problem • The r.v. Xi represents the (unobservable) speed at which the ith sample of film would have failed. Say from a location-scale family (e.g., logistic, normal, …) 13
Specifying Priors • Two-parameter distribution • Could specify priors on • (a, b) • (a, d) • (b, d) • For simplicity, want to assume independence so only need to specify marginals of each parameter (a, b) 14
Specifying Priors • Instead of (a,b) … • Consider g=0.10 example • Consider • a • distance from d to a = 2.2b • Easier for engineer to understand (a, a-d) 15
Specifying Priors • Ask engineer for • Best guess and 95% range for a • 5.0 ± 3.0 • Best guess and 95% range for a–d distance • 6.6 / 2.0 • Translate a–d=2.2b into b terms: 3.0 / 2.0 • Translate into normal, independent, priors on a and ln(b) • We used a discrete set of 1515=225 values as prior distribution of (a,b) (a, b) 16
Specifying Priors • More natural for engineer to think about priors on a and d. We let engineer do this as follows. • We created 27 combinations of prior distributions: • a best guess—10 • a uncertainty (95% limits)— ± 2, 4, 6. • a-d best guess— 1, 3, 5 • a-d uncertainty (95% limits)— / 2, 4, 6. • We graphed these in terms of (a,d) (a, d) 17
AII Measure 20
Simple Example • Objective: find the dcorresponding to g=0.10 21
Simple Example • Finding the AII for various x settings 23
Setting increment = 1 First Simulation • (a,b)=(8,1.82). Makes d=4.0 25
d Example with a More Diffuse Prior • a =10 ± 4, a-d =5 / 6 • Simulation againdone with a =8,a-d =4 26
Information on a, b, d = a-2.2b • A serious problem—all the information on d was obtained through b • The simulation trusted the relative tight prior on a … • Another problem: more objective methods of estimation, such as MLE, will likely not work well • Are there other ways to specify priors that might be better? Two methods… 29
Equal-Contribution Priors • For d=a-2.2b, restrict original prior so that Var0(a)=Varo(2.2b) • Results of another simulation • Problem: fails for case d=a: Var0(a)=Varo(0b)? 30
Relative Priors • Consider the tolerance-distribution problem • The r.v. Xi represents the (unobservable) speed at which the ith sample of film would have failed. Say from a location-scale family (e.g., logistic, normal, …) 31
Relative Priors • We observe only the ( x, Yx ) ’s • If we could observe the X ’s, the problem would be a simple one-sample problem of finding the 100g percentile of a distribution. • Assume the distribution of the X ’s has a finite fourth moment. 32
After m runs, observing X1, X2, …, Xm, we have • Using delta-method to find Var(s) and m-1m • So, to a good approximation Relative Priors 33
Now assume tolerance distribution is symmetric and its shape is know, e.g. logistic. Then • So, with k1 and k2 known, Relative Priors • So, in this sense it is defensible to specify only the prior precision with which a is know, and base the prior precision of b upon it. 34
Summary • AII as a useful measure of value of making next run at x. Combination of shift in posterior mean & probability that a failure will occur at x • Informal comparison to non-Bayesian methodsBayesian x-strategy is more subtle • Danger of simply using any prior, and recommended way to set priors 36