Chapter 24 Planning Accelerated Destructive Degradation Tests William Q. Meeker and Luis A. Escobar Iowa State University and Louisiana State University Copyright 2002-2003 W. Q. Meeker and L. A. Escobar. Complements to the authors’ text Statistical Methods for Reliability Data, John Wiley & Sons Inc., 1998. 24 - 1 Work done jointly with Danny L. Kugler and Laura L. Kramer (Imaging & Printing Group, Hewlett-Packard) December 14, 2015 8h 10min Reasons for Conducting an Accelerated Test Accelerated destructive degradation tests (ADDTs) are used for different purposes. These include: • To identify failure modes and other weaknesses in product design. • Improve reliability. • Assess the durability of materials and components. 24 - 3 • Monitor and audit a production process to identify changes in design or process that might have a seriously negative effect on product reliability. A Class of Linear Degradation Models • These models are of the form yijk = µij + ǫijk ′ x )τ + ǫ = β0 + β1 exp(β 2 j i ijk where yijk , τi, and xj may be transformations of the measured degradation, ti, and the accelerating variable(s) AccVarj , respectively. • The model is linear in the sense that for specified AccVar vector xj , the degradation is linear in τi. i=2 ◮ For a scalar x , β ′ x = β x . 2 j j j 2 ′ = (β , β , . . . ), ◮ For multiple AccVar, xj = (xj1, . . . , xjp)′, β 2 2 3 ′ x is a linear combination of the AccVar, i.e., and β 2 j p X βixji = β2xj2 + β3xj3 + · · · + βpxjp. β 2′ xj = 24 - 5 Planning Accelerated Destructive Degradation Tests (ADDT) Chapter 24 Objectives • Outline practical issues in planning ADDTs. • Describe criteria for ADDT planning. • Illustrate how to evaluate the properties of a ADDTs. • Describe methods of constructing and choosing among ADDT plans. 24 - 2 • Present guidelines for developing practical ADDT plans with good statistical properties. Motivation/Example Reliability assessment of an adhesive bond (AdhesiveBondB) • Need: Estimate of the B05 and the proportion failing after 3 years (156 weeks) and 5 years (260 weeks) operating at room temperature of 25◦C. 24 - 4 • Destructive measurements of strength in units of pounds. • Constraints ◮ 88 test units. ◮ 16 weeks for testing. • Test at 25◦C expected to yield little useful data. Degradation Model for the AdhesiveBondB ADDT yijk = β0 + β1 exp(β2xj )τi + ǫijk • The degradation model for the planning and data analysis is where yijk = log(Newtonsijk ) p √ ti = Weeksi τi = 11605 xj = − ◦ Cj + 273.15 (ǫijk /σ) ∼ Φnor (z). • Thus transformed time is square–root of Weeks and transformed AccVar is Arrhenius-transformed temperature. 24 - 6 Required Planning Information The ADDT planning requires some information which includes: • Purpose of the study. • Planning values for the parameters β that characterize degradation model for Dij . • A planning value for the parameter σ. • A plausible distribution to model the variability ǫ. • An specification of the critical degradation level Df . 24 - 7 • The range of AccVar conditions available for experimentation. • The maximum length of the experiment. Alternative Method for Specifying the Planning Model Parameters • Specify planning values for β0, β2, σ, say β0✷ = 4.471 Weeks Aged 2 4 6 12 16 8 6 4 18 Totals 8 31 24 25 88 Assumed Planning Information for the AdhesiveBondB Experiment The objective is to find a test plan to estimate with good precision the B05 and the proportion failing at use conditions (25◦C) after 3 and 5 years of operation. • A Normal failure-time distribution linear degradation model with constant regression parameters (β0, β1, β2) and same σ at each level of temperature. • Planning values for the parameters θ = (β0, β1, β2, σ)′ are β0✷ = 4.471, β1✷ = −864064160, β2✷ = 0.6364, σ ✷ = 0.1580 • Critical degradation level is specified as: Df = 40 pounds. 24 - 8 Result: The planning information defines the degradation curves at all levels of temperature. It also defines the failure time distribution at a given temperature. 0 70 DegreesC 60 DegreesC 50 DegreesC 10 20 30 40 50 9 6 60 4 6 24 - 12 24 - 10 Degradation Paths from Planning Information D(τ, x, β ✷) = exp[β0✷ + β1✷ exp(β2✷x)τ ] √ Weeks, x is Arrhenius-Transformed Temp Square Root–Log Axes τ = 100 50 20 10 Weeks 6 6 Original AdhesiveBondB ADDT Plan 70 6 6 60 16 7 12 Weeks 6 8 4 8 8 2 8 25 0 50 Newtons β2✷ = 0.6364 σ ✷ = 0.1580 • Specify the degradation rate (slope of the line), υ ✷ of µ(τ, x, β ) for a given temperature, say υ ✷ = −0.1025 at 50◦C. Because υ ✷ = β1✷ exp(β2✷x), with x = −11605/(50◦C + 273.15) = −35.912, then β1✷ = −0.1025 exp(0.6364 × 35.912) = −863499883 24 - 9 • The difference in the values for β1✷ obtained from the two procedures is due to rounding in the specification of υ ✷ and β2✷. Original Test Plan for the ADDT AdhesiveBondB Data 0 0 0 6 6 7 6 0 13 8 6 6 20 8 6 9 23 Test plan: 8 units with no ageing were measured at the start of the experiment. A total of 80 additional units were aged and measured according to the following temperatures and time schedule. Temp ◦C 8 8 — 50 60 70 Totals 24 - 11 DegreesC Comments/Questions on the Original AdhesiveBondB ADDT Plan 24 - 13 • The 8 observations at ti = 0 were not aged, i.e., there were never put into an oven. • Too much extrapolation in temperature? • Reasonable to extrapolate accurately to 25◦C? Some Practical Guidelines for One AccVar ADDT Plans • Statistically optimum plans are a research topic. They are not practical plans but provide insight into the problem. • Use three or four levels of AccVar and spread them out as much as possible, subject to model adequacy. • Limit high level of the accelerating variable to a maximum reasonable condition. • Reduce lowest level of the accelerating variable (to minimize extrapolation)-subject to seeing some action. • Measure test units at three or four different times for each level of the AccVar. 24 - 15 • Choose initial sample size using large-sample approximations. • Use simulation to compare alternative plans. Evaluating Test Plans • Large–sample approximations ◮ Fast. ◮ Provides some general insights. ◮ Provides simple sample size rules. • Simulation ◮ Reflects actual variability (no approximations). ◮ Provides visualization of sampling variability. ◮ Requires up-front computational effort. • Use large–sample approximations to get an initial indication of sample size needs and to suggest potential designs (e.g., use of optimization techniques). Use simulation to study the properties of particular designs. 24 - 17 Want a Plan That 24 - 14 • Meets practical constraints and is intuitively appealing. • Is robust to deviations from assumed inputs. • Has reasonably good statistical properties. Evaluating Test Plan Properties Suppose inferences are needed on a function g(θ ) (one-toone and all the first derivatives with respect to the elements of θ exist, and are continuous). • Properties depend on test plan, model and (unknown) parameter values. Need planning values. θ Σb " ∂g(θ ) . ∂θ # b) • Large–sample approximate standard error of g(θ ∂θ v # u" u ∂g(θ ) ′ b =t Ase[g(θ)] where Σb is the inverse of the Fisher information matrix Iθ . θ Weeks 100 150 200 Accelerated destructive degradation test simulation based on originalplan.ADDTplan AdhesiveBondB.Normal.ADDTpv 0.05 quantile of degradation versus Weeks at 25. DegreesC Resp:Log,Time:Square root,x:Arrhenius, Dist:Normal 50 24 - 18 300 24 - 16 For the θ = (β1, β 2, β3, σ)′ parameterization, the computation of Σb can be numerically unstable. Transform to stable θ parameters. 72 68 64 60 56 54 52 50 48 46 44 0 Simulation of Original Test Plan Degradation 0.05 Quantile Versus Time • Monte Carlo simulation is a powerful tool. Newtons 80 Weeks 100 120 140 160 180 Accelerated destructive degradation test simulation based on originalplan.ADDTplan AdhesiveBondB.Normal.ADDTpv Fraction failing versus Weeks for 40 Newtons at 25. DegreesC Resp:Log,Time:Square root,x:Arrhenius, Dist:Normal Lognormal Probability Plot 60 220 3000 4000 3000 5000 0.60 beta2 0.70 Focus quantity 1: 1588 beta2: 0.6509 24 - 19 0.80 b b • In repeated samples V is random and if V > log(b tr log(b tp ) p) V✷ b then the ratio R = t˜p/tp will be greater than RT . log(tp ) e r e 24 - 23 t˜p/tp will be greater than RT with a prob- ability approximately equal to 0.5. • The ratio R = • In repeated samples approximately 100(1 − α)% of the intervals will contain tp. z2 V✷ (1−α/2) log(b tp ) n= [log(RT )]2 provides confidence intervals for tp having the following characteristics: 24 - 21 Mean of the simulation estimates 0.50 260 Simulation of Original Test Plan FT (t) Estimate at 25◦C with Df = 40 Newtons .01 .005 .002 .001 .0005 .0001 .00005 .00001 .000003 .000001 .0000003 .0000001 .00000003 .00000001 40 2000 Simulation of Original Test Plan Joint Distribution of tb.05 and βb2 Estimates 1000 Accelerated destructive degradation test simulation based on originalplan.ADDTplan AdhesiveBondB.Normal.ADDTpv Failure time 0.05 quantile vs DegreesC for failure definition 40 Newtons Resp:Log,Time:Square root,x:Arrhenius , Dist:Normal 0.80 0.75 0.70 0.65 0.60 0.55 0.50 0 1000 t_0.05 at 25 DegreesC for a failure def of 40 Newtons 0 Statistical Properties of a Test Plan With a Given Sample Size n t_0.05 at 25 DegreesC for a failure def of 40 Newtons 15 5 0 20 30 DegreesC 40 50 60 70 Accelerated destructive degradation test simulation based on originalplan.ADDTplan AdhesiveBondB.Normal.ADDTpv Failure time 0.05 quantile vs DegreesC for failure definition 40 Newtons Resp:Log,Time:Square root,x:Arrhenius , Dist:Normal Simulation of Original Test Plan Failure-Time 0.05 Quantile Estimates Versus Temperature 10000 1000 100 10 1 # ! r 24 - 20 b ˜ p) = log(tbp)± √1 z log(tp), log(t V = log(tbp)±log(R log(b tp ) n (1−α/2) e [tbp/R, tbpR] p . e 80% 24 - 24 3.0 24 - 22 , R with RT , and log(b tp ) p r s √ b = t˜p/tp. R = exp (1/ n)z(1−α/2) V log(b t ) p log(b tp ) 2 z(1−α/2) V✷ [log(RT )]2 Needed sample size giving approximately a 50% chance of having a confidence interval factor for the 0.05 quantile that is less than R use condition= 25. DegreesC and a failure definition= 40 originalplan.ADDTplan AdhesiveBondB.Normal.ADDTpv Sample Size to Estimate the AdhesiveBondB t0.05 Quantile at 25◦C when Df = 40 Newtons solve for n to get 5000 10000 2000 500 1000 200 2.5 99% 2.0 Confidence Interval Precision Factor R 1.5 50 1.0 95% 90% 10 20 100 n= b d /tb2 with V✷ • Replace V = nVar p b log(b t ) t where Exponentiation yields a confidence interval for tp " • Approximate 100(1 − α)% confidence interval for log(tp): ADDT Sample Size Determination for a Quantile tp Weeks Sample Size 8 12 4 0 Fraction Failing Test plans with a sample size of beta2 A Two-Variables Linear Degradation Model With Temperature and Relative Humidity as AccVar = β0 + β1 exp(β2xj2 + β3xj3)τi + ǫijk ′ x )τ + ǫ yijk = β0 + β1 exp(β 2 j i ijk • AdhesiveBondD model where yijk = log(Newtonsijk ) p √ ti = Weeksi τi = (ǫijk /σ) ∼ Φnor (z). 24 - 25 • The AccVar transformations are Arrhenius for Temp (Temperature) and logit for RH (Relative Humidity), i.e., 11605 xj2 = − ◦ Cj + 273.15 ! RHj xj3 = log . 1 − RHj Original Test Plan for the AdhesiveBondD ADDT Data Temp ◦C 6 6 6 6 6 6 36 0 6 6 6 6 6 6 36 1 Weeks Aged 2 4 6 8 12 16 6 6 6 6 6 6 36 6 6 6 6 6 6 36 6 6 6 6 6 6 36 6 6 6 6 6 6 36 6 6 6 6 6 6 36 6 6 6 6 6 6 36 48 48 48 48 48 48 288 Totals • A 8 × 3 × 2 full factorial arrangement in Weeks, Temp, and RH, with 6 units at each factor combination. RH % 20 50 60 70 80 50 60 70 Totals 24 - 26 • Might want to re-allocate units assigned to the corners of the experimental region, i.e., highest (lowest) levels of (Weeks, RH, Temp). Also, might want to decrease the number of pens allocated to each factor level combination to reduce size of the experiment. Choosing Experimental Variable Definition to Minimize Interaction Effects • Care should be used in defining experimental variables. Some Practical Guidelines for Two or More AccVar ADDT Plans • Moderate increases in two accelerating variables may be safer than using a large amount of a single accelerating variable. • Guidance on variable definition and possible transformation of the response and the experimental models should, as much as possible, be taken from mechanistic models. 100 • • • • • Volts 300 • • 500 • 24 - 30 24 - 28 • Knowledge from mechanistic models is also useful for planning experiments. • Models without statistical interactions simplify modeling, interpretation, explanation, and experimental design. • Proper choice can reduce the occurrence or importance of statistical interactions. • There may be interest in assessing the effect of nonaccelerating variables, i.e., factor Weeks in the AdhesiveBondD example. • There may be interest in assessing joint effects of two or more accelerating variables, i.e., (RH, Temp) in the AdhesiveBondD example. • Traditional factorial experiments are good as starting plans. • When possible, a fractional factorial could provide economy in testing. 24 - 27 Size (mm) • Choose initial sample size using large-sample approximations. • Use simulation to compare alternative plans. • • • 200 • Comparison of Experimental Layout with Volts/mm Versus Size and Volts Versus Size • • • 150 Size (mm) Examples of Choosing Experimental Variable Definition to Minimize Interaction Effects • For accelerated testing of dielectrics, use size and volts stress (e.g., mm and volts/mm instead of mm and volts). • 100 • • 3.0 Voltage Stress (Volts/mm) 50 2.5 1.5 1.0 • For light exposure, use aperture and total light energy (not aperture and exposure time). • For accelerated humidity testing with a corrosion mechanism, use RH and temperature (not vapor pressure and temperature). 24 - 29 3.0 2.0 2.5 2.0 1.5 1.0 Size (mm) • 50 100 Volts • 200 • • • • • • • • 300 • • 200 • • Voltage Stress (Volts/mm) 100 Technical Details 300 • • 3.0 • 2.5 1.0 1.5 2.0 3.0 • The asymptotic standard error for gb and log(gb) are 1 q Vgb Ase(gb) = √ n q 1 Vgb 1 q Vlog(gb) = √ Ase[log(gb)] = √ . n n g • Easy to choose n to control Ase. 24 - 31 24 - 35 To compute these variance factors one uses planning values for θ (denoted by θ ✷) as discussed later. • The variance factors Vgb = nAvar(gb) and Vlog(gb) = nAvar[log(gb)] may depend on the actual value of θ but they do not depend on n. Given a specified model and parameter values (but without need to specify sample size), one can compute scaled asymptotic variances. Asymptotic Approximate Standard Errors for a Function of the Parameters g(θ ) 24 - 33 The following slides give technical details used in SPLIDA to implement the methodology. 400 Comparison of Experimental Layout with Volts versus Size and Volts/mm versus Size • Size (mm) 2.5 2.0 1.5 1.0 Areas for Future Research in ADDT Planning • Optimum plans (to provide insight and general rules). • Optimized compromise plans (using practical constraints). • Plans that allow the use of prior information on activation energy (i.e., plans for Bayesian analysis). 24 - 32 • Plans allowing for censoring and coarse data (accounting for measurement limitations). ADDT Model Asymptotic Variances # θ " X ∂ 2Lij (θ ) ∂ 2L(θ ) πij E − =n ∂ θ∂ θ′ ∂ θ∂ θ′ ij # Under certain regularity conditions the following results hold asymptotically (large sample) θ " b ∼ • θ ˙ MVN(θ , Σb ), where Σb = Iθ−1, and Iθ = E − πij Iij #′ θ Σb " # ∂g(θ ) . ∂θ Avar(gb). 24 - 34 where πij is the proportion of observations allocated to (ti, xj ). " ∂g(θ ) ∂θ b) ∼ • For a scalar gb = g(θ ˙ NOR[g(θ ), Avar(gb)], where Avar(gb) = !2 • When g(θ ) is positive for all θ , then b )] ∼ log[g(θ ˙ NOR{log[g(θ )], Avar[log(gb)]}, where Avar[log(gb)] = 1 g ADDT Model Fisher Information Matrix • The Fisher information matrix, I , is θ " " # # X ∂ 2Lij ∂ 2L πij E − =n ∂ θ∂ θ′ ∂ θ∂ θ′ ij ij X Iθ = E − = n where L = log[L(θ )], Lij is the contribution of a single observation at ξ ij = (ti, xj ) to the log-likelihood, and πij is the proportion of observations made at ξ ij . E is the expectation operator, the index on the summation operation runs over the distinct time, and stress level combinations ξ ij . Iij is the contribution of one observation at ξ ij to Iθ . 24 - 36 Contributions to the ADDT Model Fisher Information Matrix • The contribution Iij to the Fisher information matrix is ′ f11(ξ ij )uij uij f12(ξ ij )uij 1 Iij = 2 σ symmetric f22(ξ ij ) = 1 exp(β ′ xj )τi = 2 ′ x )τ β1 xj exp(β 2 j i . where and f11(ξ ij ), f12(ξ ij ), f22(ξ ij ) are the LSINF elements for the distribution Φ at the experimental conditions ξ ij and uij is the vector of partial derivatives of the degradation path with respect to the β parameters uij A Reparameterization of the Linear ADDT Model for Numerical Stability • The model is as before ′ x )τ + σǫ yij = β0 + β1 exp(β 2 j i ij exp xj − x̄ γ 2 τi − τ̄ + σǫij ′ • Suppose that x̄ is the centroid of the stress variables [i.e., x̄ = (x̄1, . . . , x̄k )′] and τ̄ is an average transformed time. Then the model can be reparameterized as yij = γ0 + γ1 where γ0 is the intercept for the average stress line (i.e., degradation line for x̄) at τ̄ ; γ1 is the slope of the average stress line; and γ 2 = β 2 are the regression coefficients corresponding to the x variables. • It can be shown that γ1 is the geometric mean of the slopes ′ x ), for the values of x stress variables in the β1 exp(β 2 j j data set. 24 - 38 • The vector ϕ = (γ0, γ1, γ 2, σ)′ denotes the stable parameters. ∂µij ∂β0 ∂µij ∂β1 ∂µij ∂ β2 24 - 37 ∂µij ∂γ0 ∂µij ∂γ1 ∂µij ∂γ2 2 24 - 40 • For decreasing degradation ht(tp) = ν+ςΦ−1(p), for p > Φ − tp 2 b Avar[ht(tp)]. ∂ht (tbp) Avar(tbp). ∂ tbp ✷ Avar[ht(tbp)] = ht (tp✷) ∂h−1(z) t Avar(tbp) = ∂z 24 - 42 • This allows to write Avar[ht(tbp)] as a function of Σb or as θ a function of Σϕ b. Thus Then ignoring the probability spike, we get Asymptotic Variance of ht(tbp) • For test planning, the Fisher matrix is evaluated at the planning values θ ✷. = ξ ij and v ij is the vector of partial derivatives of the degradation path with respect to the γ = (γ0, γ1, γ 2) parameters 1 h i . exp γ ′ (x − x̄) τ − τ̄ v ij j i 2 = h i ′ (x − x̄) τ γ1 (xj − x̄) exp γ 2 j i where and f11(ξ ij ), f12(ξ ij ), f22(ξ ij ) are the LSINF elements for the distribution Φ at the experimental conditions f11(ξ ij )v ij v ′ f12(ξ ij )v ij 1 ij Iij = 2 σ symmetric f22(ξ ij ) • The contribution Iij to the Fisher information matrix is ADDT Fisher Matrix (Transformed Time Scale) • For test planning the Fisher matrix is evaluated at planning values θ ✷. 24 - 39 Relationship Between the Stable Parameters ϕ and the Original Parameters θ • It can be shown that ′ x̄)τ̄ γ0 = β0 + β1 exp(β 2 ′ x̄) γ1 = β1 exp(β 2 γ 2 = β 2. • Solving for the βs β2 = γ 2 β1 = γ1 exp(−x̄′γ 2) β0 = γ0 − γ1τ̄ . Alternative Expression for the Variance-Covariance Matrix • Let Σγb be the variance-covariance matrix in the stable parameterization and Σb the corresponding matrix in the origθ inal parameterization. • Then Σb = AΣγb A′, where θ ∂β ∂β ∂β 0 0 0 0 ∂γ ∂γ ∂γ 0 1 2 ∂β1 ∂β1 ∂β1 0 ∂γ A = 0 ∂γ1 ∂ γ 2 ∂ β2 ∂ β2 ∂ β2 0 ∂γ ∂γ ∂γ 0 1 2 0 0 0′ 1 1 −τ̄ 0′ 0 ′ ′ ′ 0 exp[−γ 2x̄] −γ1x̄ exp[−γ 2x̄] 0 = 0 0 I 0 0 0 0′ 1 24 - 41 0 is a vector of zeros and I is an identity matrix with the same row dimension as x. ν . ς and " cθ = # ∂ht(tp) . ∂θ Asymptotic Variance of ht(tbp) in Funtion of the θ and ϕ Parameters. !′ !2 cθ′ Σθb cθ ∂ht(tp) ∂ht(tp) = A′ = A ′ cθ . ∂θ ∂θ • Define ∂ht(tp) cϕ = ∂ϕ It can be shown that cϕ = ∂θ ∂ϕ ∂z ∂ht−1(z) ∂z !2 24 - 43 24 - 45 t, transformation, = 2ht(tp✷). √ ′ Σ c . cϕ b ϕ ϕ !2 ∂ht−1(z) ′ cθ′ AΣϕ b A cθ ∂ht−1(z) ∂z • Then direct computations yield Avar(tbp) = = = Derivatives ht (tp✷) • For the square root of time, ht(t) = ∂z 2 ∂ht−1(z) = = 2z ∂z ∂z • The elements of cθ are: ∂ht(tp) 1 = − ′ x) ∂β0 β1 exp(β 2 ht(tp) ∂ht(tp) = − ∂β1 β1 ∂ht(tp) = − x ht(tp) ∂β2 ∂ht(tp) Φ−1(p) = − ′ x) ∂σ β1 exp(β 2 Linearization of the Degradation Model h i τi − τ̄c • A first order approximation about ϕ✷ gives n i o ∂µij ∂µij + γ2 − γ2✷ + σǫijk zijk = γ0 + γ1 − γ1✷ ∂γ ∂γ ✷ 2 ✷1 + σǫijk + γ2 − γ2✷ v2ij = γ0 + γ1 − γ1✷ v1ij where h zijk = yijk − exp γ2✷ xj − x̄c ✷ v1ij = exp γ2✷(xj − x̄c) τi − τ̄c h i ✷ = γ1✷ (xj − x̄c) exp γ2✷(xj − x̄c) τi v2ij f22(ξ ij ) 24 - 47 • The asymptotic variance for this model is −1 ′ f11(ξ ij )v ij v ij f (σ ✷)2 X 12 (ξ ij )v ij πij n symmetric Σϕ b = ij i−1 (σ ✷)2 h , F ϕ ✷ (ξ ) = n ✷ , v ✷ )′ and ϕ✷ = (1, γ ✷ , γ ✷ , γ ✷ , σ ✷ )′ . where v ij = (1, v1ij 0 1 2 2ij ∂ht(z) ∂z !2 n ′ Σ c cϕ Avar(tbp) = b ϕ ϕ σ2 ′ F (ξ )−1 c = cϕ ϕ ϕ f11(ξ )v v ′ f12(ξ ij )v ij X ij ij ij . πij symmetric f22(ξ ij ) ij Standardized Asymptotic Variance Formula n σ2 • It can be shown that where F ϕ (ξ ) = 1 ′ (x − x̄ )] γ1 exp[γ 2 c 24 - 44 • The standardized information matrix Fϕ(ξ ) depends only on the test plan (ξ ij , πij ), ϕ, and the derivatives v ij . Derivatives-Continued = − = − (x − x̄c)ht(tp) = 1 = − ′ (x − x̄ )] γ1 exp[γ 2 c " # 1 τ̄c ′ (x − x̄ )] − ht (tp ) γ1 exp[γ 2 c • The elements of cϕ are: ∂ht(tp) ∂γ0 ∂ht(tp) ∂γ1 ∂ht(tp) ∂γ2 ∂ht(tp) ∂σ where (x̄c, τ̄c) is a centroid of the data. 24 - 46 • The elements of cθ and cϕ were obtained directly. A check for the expressions is cϕ = A ′ cθ Optimal Test Plans • Statistically optimum plans are a current research topic. They are not practical plans but provide insight into the problem. i−1 c ϕ✷ . • Statistical optimum plans can be obtained numerically under the following criteria: h ′ argξ,π min Avar(tbp; ξ , π ) = argξ,π min cϕ ✷ F ϕ ✷ (ξ ) Notice that the derivatives, c and the information matrix, F , are evaluated at ϕ✷. • The General Equivalence Theorem can be used to check the optimality of the test plan obtained from the numerical algorithm. 24 - 48 The General Equivalence Theorem to Check Optimal Test Plans • Defining ξ v as a plan with all units at v and where c and Fϕ are evaluated at ϕ✷. 24 - 49 dc(ξ , v ) = cϕ ′[Fϕ(ξ )]−1Fϕ(ξ v )[Fϕ(ξ )]−1 cϕ−cϕ ′ [Fϕ(ξ )]−1 cϕ • The test plan (ξ copt, πcopt) minimizes Avar(tbp; ξ , π ) iff v sup dc(ξ copt, x) = 0. Standardization of Explanatory Variables • Simple scaling of the explanatory variables and time gives s )′ w ]v + σǫ yijk = β0s + β1s exp[(β 2 j i ijk where wj are the scaled explanatory variables and vi is the scaled time variable (all explanatory variables scaled in [0, 1]). • When xj = xj is a scalar, then xj − xL , 0 ≤ wj ≤ 1. wj = x − xL U τi vi = , 0 ≤ vi ≤ 1. τH β0s = β0 β1s = τL exp[β2xL] β1 β2s = (xH − xL)β2 24 - 51 xL, xH are the smallest and largest possible value of the explanatory variable x and τH is the largest possible time in the experiment. General Guidelines • Reduce lowest level of the accelerating variable (to minimize extrapolation)-subject to seeing some action. • Measure test units at three or four different times for each level of the AccVar. • Choose initial sample size using large-sample approximations. • Use simulation to compare alternative plans. 24 - 50