Gauge R&R: A GUM/Metrological/ Bayesian Perspective Dave LeBlond MBSW-38 May 19, 2015 1 Acknowledgements Thanks to: • Stan Altan (J&J) • Bill Porter (PPP LLC) • Yan Shen (J&J) • Jyh-Ming Shoung (J&J) for organizing this session, inviting me, inspiring discussions, and providing the 3 Gauge R&R examples. 2 Outline • Measurement Uncertainty (MU) from a Bayesian perspective • Computational considerations • Examples 1. Dissolution measurement 2. Particle size measurement 3. Bioassay measurement • Conclusions • Bibliography 3 The GUM perspective ISO/GUM compliant practice Current practice Reportable result = m ±U Q: What is U? A: “A parameter, associated with the result of a measurement, that characterizes the dispersion of the values that could reasonably be attributed to the measurand” (i.e., The part of the result after the ±). 4 MU requires a complete probability model Reportable result = m ±U • Let m be an analytical measurement that estimates m • Let m be the unknown measurand quantity • fixed, but uncertain • MU refers to the uncertainty in m (not m) • ISO: “…values that could reasonably be attributed to the measurand.” • (Don’t forget to put the u in the mu) • The posterior, p(m|m), expresses current knowledge about m • To get p(m|m), we need prior knowledge and Bayes’ rule • Inference about m requires a Bayesian approach 5 U has a probability interpretation Reportable result = m ±U • Bayesian interpretation: posterior distribution of m, conditional on m • Assume symmetrical, location-scale family with scale independent of m m U p(m|m) p m | m d m e.g., 0.95 m U m-U m m+U m 6 MU needs 2 spaces & 2 experiments Space Observed Parameter Experiment p(q) Gauge R&R p(s) p(y|q,s) p’(s) = p(s|y) ∫p(y|q,s)p(q)p(s)dq p(m) Routine Measurement p’(s) p(m|m,s) p(m|m) ∫p(m|m,s)p(m)p’(s)ds likelihood prior posterior y = vector of observed Gauge R&R results q = vector of true levels s = true analytical imprecision m = true but uncertain measurand quantity value m = analytical measurement that estimates m 7 MU needs 2 spaces & 2 experiments Space Observed Parameter Experiment p(q) Gauge R&R p(s) p(y|q,s) p’(s) = p(s|y) ∫p(y|q,s)p(q)p(s)dq p(m) Routine Measurement p’(s) p(m|m,s) p(m|m) ∫p(m|m,s)p(m)p’(s)ds likelihood prior posterior Assume • m is unbiased for m • normal likelihoods • diffuse priors for q and m: p(q) = p(m) = 1 Key outputs • Gauge R&R: p’(s) and U • Routine Measurement: p(m|m) … or at least m ± U 8 Getting ∫p(m|m,s)p(m)p’(s)ds by MC Integration likelihood MCMC draws Diffuse prior l(m|m,s) location-scale m arbitrary s* p(m) = 1 l(m|m=0,s) Plug in each s* p(m*|m=0) l(m|m=0, s*) Obtain 1 MC draw for each s* m* Estimate percentiles P2.5 0 P97.5 m* Estimate U U = ½(P97.5 – P2.5) 9 Sometimes buried in the SAS Log… NOTE: Convergence criteria met. NOTE: Estimated G matrix is not positive definite. NOTE: Asymptotic variance matrix of covariance parameter estimates has been found to be singular and a generalized inverse was used. Covariance parameters with zero variance do not contribute to degrees of freedom computed by DDFM=KENWARDROGER. Covariance Parameter Estimates Cov Parm Day(Site) ExpRu*HPLC(Site*Day) ExpRun(Day) ExpRun(Day) ExpRun(Day) ExpRun(Day) Residual Residual Residual Residual Group Estimate Standard Error Site Site Site Site Site Site Site Site 0.2612 0.9645 6.3519 0.8339 10.3038 0 5.7744 3.0847 8.6020 6.7212 0.5356 0.5322 3.5080 0.9441 5.2917 . 1.1481 0.6446 1.6357 1.2652 G I L T G I L T Lower 0.03723 0.4129 2.7176 0.2028 4.6221 . 4.0502 2.1285 6.1194 4.7965 Upper 503733 4.2521 28.0595 80.8289 39.6342 . 8.8978 4.8713 12.9815 10.0968 MIXED arbitrarily sets a negative variance estimate to zero, effectively performing model reduction and ignoring uncertainty in the estimate. The Bayesian prior would disallow a negative variance so that an interval estimate of the variance is available. 10 Implementing MCMC • An MCMC chain is a correlated multivariate sample from the multivariate posterior distribution of all parameters in the model, given the data and prior assumptions. • Independent “noninformative” (wide uniform or normal with huge variance) used for all parameters. • Square root of variance components as parameters • Run 3 MCMC chains from different (random) starting points • Only retain every 40th iteration as a draw (to reduce autocorrelation) • Discard first 3,000 draws from each chain (“burn-in”) • Save 10,000 draws from each chain (30,000 total) • Checked for convergence of the 3 chains 11 Comparison of approaches Consideration MIXED BUGS Estimation Method REML, ML, MIVQUE0, Type1, Type2, Type3 MCMC, method chosen by BUGS Denominator df Method BW, CON, KR, RES, SAT, DDF=list No worries Multiple Comparison Adjustment Method BON, SCHEFFE, SIDAK, SIMULATE, T No worries Effect coefficients E1, E2, E3 No worries CONVF, CONVG, CONVH, DFBW, Computational options EMPIRICAL, NOBOUND, No worries RIDGE=, SCORING=, NOPROFILE Asymptotic normality? yes No approximations Iteration convergence Warnings provided Manual Speed Generally very fast Can be slow (hours) Syntax mimics… Mixed model algebra Data generation Non-negative bounding of variances Arbitrarily set to zero Handled through prior 12 1. Data i indexes Site (S, 4 unique sites) j indexes Day within Site (SD, 24 unique Days) k indexes ExpRun within Site (E, 48 unique ExpRuns) l indexes HPLC*ExpRun combinations within Site (HE, 96 unique combinations) m indexes Batch (B, 3 unique batches) obs 1 2 3 … 70 71 72 73 74 75 … 142 143 144 145 146 147 … 214 215 216 217 218 219 … 286 287 288 i 2 2 2 j k l m y 7 20 25 2 81.00 7 20 25 3 80.62 7 20 25 1 82.46 2 12 23 48 3 83.36 2 12 23 48 2 85.23 2 12 23 48 1 85.01 1 1 8 1 2 90.39 1 1 8 1 3 95.50 1 1 8 1 1 92.40 1 6 11 24 1 6 11 24 1 6 11 24 3 13 32 49 3 13 32 49 3 13 32 49 3 2 1 2 3 1 81.01 86.82 91.73 87.12 88.72 86.20 3 3 3 4 4 4 3 2 1 2 3 1 76.74 81.62 82.42 93.51 91.59 92.07 18 18 18 19 19 19 35 35 35 44 44 44 72 72 72 73 73 73 4 24 47 96 3 85.47 4 24 47 96 2 90.00 4 24 47 96 1 89.44 13 1. Statistical model Diss30 Site Batch y obs q Si obs Bmobs “fixed”(but uncertain) effects obs 1...288 i obs 1...4 m obs 1...3 Set to zero restrictions Day(Site) ExpRun(Day) HPLC*ExpRun(Site*Day) SD j obs Ek obs HEl obs obs measurement uncertainty effects SD j obs ~ N 0,sigma.SD 2 , j obs 1...24 Ek obs ~ N 0,sigma.Ei2[ obs ] ,k obs 1...48 HEl obs ~ N 0,sigma.HE 2 ,l obs 1...96 obs ~ N 0,sigmai2obs S4 B3 0 14 model{ 1. BUGS model # Priors theta~dnorm(0,0.000001) # Likelihood for(obs in 1:n.obs){ mu[obs] <- theta + S[i[obs]] + B[m[obs]] + SD[j[obs]] + E[k[obs]]+ HE[l[obs]] y[obs] ~ dnorm(mu[obs],tau[i[obs]]) } for(jj in 1:24){ SD[jj] ~ dnorm(0,tau.SD) } for(ll in 1:96){ HE[ll] ~ dnorm(0,tau.HE) } for(kk in 1:12){ E[kk] ~ dnorm(0,tau.E[1]) } for(kk in 13:24){ E[kk] ~ dnorm(0,tau.E[2]) } for(kk in 25:36){ E[kk] ~ dnorm(0,tau.E[3]) } for(kk in 37:48){ E[kk] ~ dnorm(0,tau.E[4]) } for(site in 1:3){ S[site]~dnorm(0,0.000001) } S[4]<- 0 for(batch in 1:2){ B[batch]~dnorm(0,0.000001) } B[3]<- 0 for(site in 1:4){ sigma[site] ~ dunif(0,100) tau[site] <- pow(sigma[site],-2) sigma.E[site] ~ dunif(0,100) tau.E[site] <- pow(sigma.E[site],-2) } sigma.SD~dunif(0,100) tau.SD<-pow(sigma.SD,-2) sigma.HE~dunif(0,100) tau.HE<-pow(sigma.HE,-2) } 15 1. Joint posterior marginals theta S[1] S[2] S[3] B[1] B[2] sigma.SD sigma.E[1] sigma.E[2] sigma.E[3] sigma.E[4] sigma.HE sigma[1] sigma[2] sigma[3] sigma[4] mean 87.86865 -1.85345 -4.80681 -6.78676 1.67055 -0.57506 0.63573 2.81409 0.99404 3.65876 0.56074 0.95419 2.47125 1.82693 3.00323 2.65304 sd 0.56354 1.10904 0.75670 1.32984 0.33997 0.34345 0.42448 0.91456 0.57189 1.09118 0.44593 0.31807 0.25124 0.19887 0.29334 0.25404 2.5% 86.75000 -4.06602 -6.31602 -9.42800 1.00500 -1.24800 0.02972 1.39297 0.08034 2.02900 0.02289 0.21159 2.03500 1.48300 2.49100 2.20997 25% 87.50000 -2.57400 -5.29025 -7.63400 1.44500 -0.80692 0.30180 2.19200 0.58310 2.90500 0.22100 0.76710 2.29400 1.68600 2.79800 2.47400 50% 87.87000 -1.85200 -4.80400 -6.78500 1.66700 -0.57430 0.58095 2.68500 0.94945 3.48000 0.46070 0.97680 2.45200 1.81000 2.98200 2.63400 75% 88.23000 -1.14400 -4.32000 -5.93800 1.90100 -0.34580 0.89732 3.28100 1.32700 4.22000 0.79032 1.17100 2.62800 1.95200 3.18300 2.80900 97.5% 88.99000 0.33300 -3.29998 -4.18898 2.34400 0.10550 1.59600 4.99000 2.25900 6.24702 1.66800 1.52400 3.01800 2.26100 3.64602 3.20302 Rhat 1.00100 1.00098 1.00101 1.00102 1.00098 1.00100 1.00151 1.00095 1.00095 1.00097 1.00109 1.01856 1.00114 1.00102 1.00095 1.00106 n.eff 30000 30000 30000 29000 30000 30000 3500 30000 30000 30000 14000 1100 11000 30000 30000 19000 16 1. Marginal posterior distribution of sigma by site Posterior for sigma 1 L 2 3 4 5 T 15 10 Percent of Total 5 0 G I 15 10 5 0 1 2 3 4 5 sigma 30,000 mcmc draws 17 1. Marginal posterior distribution of sigma.E by site Posterior for sigma.E 0 L 5 10 T 30 20 Percent of Total 10 0 G I 30 20 10 0 0 5 10 sigma.E 30,000 mcmc draws 18 1. Marginal Posterior distributions of sigma.SD and sigma.HE Posterior for sigma.SD Posterior for sigma.HE 6 Percent of Total Percent of Total 6 4 2 0 4 2 0 0 1 2 sigma.SD 30,000 mcmc draws 3 0.0 0.5 1.0 1.5 2.0 2.5 sigma.HE 30,000 mcmc draws 19 0.2 0.5 0.6 0.7 0. 9 0.4 0.5 0.3 sigma.E4 1.0 0.1 1.5 1. Bivariate posterior kernal density 4 0. 0.0 1 0.0 0.8 0.5 1.0 1.5 sigma.SD 30,000 draws 20 1. Variance Components Estimates (SAS, Bayesian) Variance Components Group Day(Site) DissRun*HPLC(Site*Day) Site G Site I DissRun(Day) Site L Site T Site G Site I Residual Site L Site T Estimate (REML, median) 0.26 0.34 0.96 0.95 6.35 7.21 0.83 0.90 10.30 12.11 0.00 0.21 5.77 6.01 3.08 3.28 8.60 8.89 6.72 6.94 Lower95% Upper95% 0.04 0.0009 0.41 0.04 2.72 1.94 0.20 0.006 4.62 4.12 0.00 0.0005 4.05 4.14 2.13 2.20 6.12 6.21 4.80 4.88 5.04E+05 2.55 4.25 2.32 28.06 24.90 80.83 5.10 39.63 39.02 0.00 2.78 8.90 9.11 4.87 5.11 12.98 13.29 10.10 10.26 21 1. Error Budget (SAS, Bayesian) Site G I L T Source % of Total Residual Day(Site) DissRun*HPLC(Site*Day) DissRun(Day) Total Residual Day(Site) DissRun*HPLC(Site*Day) DissRun(Day) Total Residual Day(Site) DissRun*HPLC(Site*Day) DissRun(Day) Total Residual Day(Site) DissRun*HPLC(Site*Day) DissRun(Day) Total 43 2 7 48 100 60 5 19 16 100 43 1 5 51 100 85 3 12 0 100 Posterior quantiles of % of Total Median(2.5th-97.5th quantile) 40(18-65) 2(0-18) 6(0-17) 49(18-78) 56(30-81) 6(0-33) 16(1-39) 16(0-53) 39(17-64) 1(0-11) 4(0-11) 54(26-80) 79(56-94) 4(0-24) 10(1-25) 2(0-25) 22 1. Posterior distribution of Error m m Error Error SD E HE [1] SD ~ N 0,sigma.SD 2 E ~ N 0,sigma.Ei2 HE ~ N 0,sigma.HE 2 [2] sigmai2 ~ N 0, 6 To obtain a sample from the posterior distribution of Error: 1. Start with 30,000 MCMC draws of the 4 sigmas (a 4-vector for a given site i) 2. For each draw vector, a. simulate a random sample of the 4 error contributors using [2] b. Plug these into [1] 3. The result is 30,000 MCMC draws from the posterior distribution of Error 23 1. Expanded uncertainties (U) for each site • 95% credible interval of posterior distribution of Error: 0 ± U • Results should be reported as m ± U Posterior Distribution of Error at 4 Sites -10 Site : 3 0 Site : 4 U = 3.63 (SAS: 3.06) Percent of Total U = 8.38 (SAS: 7.20) Site : 1 25 20 15 10 5 0 0 25 20 15 10 5 0 Site : 2 U = 6.66 (SAS: 5.84) -10 10 U = 3.72 (SAS: 3.21) 10 Error 30,000 mcmc draws 24 1. Adjusting for site bias G 84 86 T 82 I L 80 95%CI of Site Mean 88 Site Means for Batch C 1.0 1.5 2.0 2.5 3.0 3.5 4.0 Site (based on 30,000 mcmc draws) • ISO/GUM philosophy: Do everything in your power to adjust for bias • First must decide what “truth” is • Is one site the “reference” site? • Take the average across sites as the “truth”? • Determine the bias appropriate for each site. • Adjust reported values from each site by subtracting the site-bias 25 2. Statistical model Dv50a Site Batch Day(Site) Day*Batch(Site) y obs q Si obs B j obs SDk obs SBDl obs obs “fixed”(but uncertain) effects measurement uncertainty effects obs 1...252 i obs 1...7 j obs 1...6 Set to zero restrictions SDk obs ~ N 0,sigma.SD 2 ,k obs 1...21 SBDl obs ~ N 0,sigma.SBD 2 ,l obs 1...84 obs ~ N 0,sigma2 S7 B6 0 26 2. Joint posterior marginals mean theta 3.55274 S[1] -0.10220 S[2] -0.00934 S[3] 0.10382 S[4] -0.01691 S[5] 0.00924 S[6] -0.01911 B[1] -0.27059 B[2] 0.35566 B[3] -0.24313 B[4] -0.04624 B[5] 0.04943 sigma.SD 0.02372 sigma.SBD 0.04174 sigma 0.24960 sd 0.06061 0.06714 0.06736 0.06718 0.06750 0.06726 0.06704 0.05847 0.05898 0.05759 0.05878 0.05874 0.01882 0.02713 0.01202 2.5% 3.43300 -0.23440 -0.14100 -0.02814 -0.14980 -0.12190 -0.15110 -0.38580 0.24000 -0.35480 -0.16140 -0.06520 0.00089 0.00200 0.22710 25% 3.51300 -0.14660 -0.05363 0.05949 -0.06217 -0.03564 -0.06396 -0.30962 0.31590 -0.28230 -0.08572 0.00971 0.00914 0.01944 0.24130 50% 3.55300 -0.10230 -0.00958 0.10385 -0.01745 0.00929 -0.01881 -0.27090 0.35510 -0.24310 -0.04664 0.04954 0.01956 0.03869 0.24920 75% 97.5% 3.59300 3.67200 -0.05748 0.03012 0.03472 0.12530 0.14870 0.23650 0.02786 0.11630 0.05394 0.14190 0.02593 0.11090 -0.23100 -0.15500 0.39510 0.47280 -0.20450 -0.12980 -0.00697 0.06928 0.08884 0.16480 0.03373 0.07047 0.06058 0.09957 0.25750 0.27440 Rhat 1.00097 1.00096 1.00104 1.00097 1.00108 1.00110 1.00118 1.00100 1.00101 1.00095 1.00104 1.00096 1.00164 1.00308 1.00098 n.eff 30000 30000 21000 30000 15000 13000 8500 30000 30000 30000 23000 30000 2900 2600 30000 27 2. Variance Component Estimates (SAS, Bayesian) Variance Components Estimate Lower95% Upper95% % of Total Day(SITE_APP) 0 0.0004 ? 8E-7 ? 0.005 0 1(0-7) BATCH*Day (SITE_APP) 0.00047 0.0015 0.00006 4E-6 3.61E+109 0.01 1 2(0-14) Residual 0.06 0.06 0.05 0.05 0.08 0.08 99 96(83-100) Total 0.06 0.07 0.05 0.05 3.61E+109 0.08 100 100 Bayesian point estimate is the posterior median Bayesian % of Total is posterior median(central 95% credible interval) 28 2. Posterior for sigma.SD 3000 0 1000 Frequency 500 1000 0 0.05 0.10 0.15 sigma.SBD 30,000 mcmc draws 0.00 0.05 0.10 0.15 sigma.SD 30,000 mcmc draws 500 1000 2000 Posterior for sigma 0 0.00 Frequency Frequency 2000 Posterior for sigma.SBD 0.20 0.22 0.24 0.26 sigma 30,000 mcmc draws 0.28 0.30 29 8 6 0.0 7 2 0.0 8 0.0 0.11 0.1 0.0 0. 06 4 0. 05 0 Day*Batch variance (% of Total) 10 2. Bivariate posterior kernal density of 2 error budget components 0.5 1.0 0.01 0.02 4 0.03 1.5 2.0 2.5 Day to Day variance (% of Total) 30,000 draws 3.0 30 2. Posterior distribution of Error m m Error Error SD SBD [1] SD ~ N 0,sigma.SD 2 SBD ~ N 0,sigma.SBD 2 [2] ~ N 0,sigma2 To obtain a sample from the posterior distribution of Error: 1. Start with 30,000 MCMC draws of the 3 sigmas (a 3-vector) 2. For each draw vector, a. simulate a random sample of the Error contributors using [2] b. Plug these into [1] 3. The result is 30,000 MCMC draws from the posterior distribution of Error 31 2. Expanded uncertainty (U) • 95% credible interval of posterior distribution of Error: 0 ± U • Future results should be reported as m ± U Posterior Distribution of Error Percent of Total 6 U = 0.5031 (SAS: 0.5027) 4 2 0 -1.0 -0.5 0.0 0.5 1.0 Error 30,000 mcmc draws 32 3. Statistical model Log_assay Analyst Conc y obs q Ai obs C j obs “fixed”(but uncertain) effects Run(Analyst) Plate(Run*Analyst) Rk obs Pl obs obs measurement uncertainty effects obs 1...60 i obs 1...2 j obs 1...6 Set to zero restrictions Rk obs ~ N 0,sigma.R 2 ,k obs 1...10 Pl obs ~ N 0,sigma.P 2 ,l obs 1...20 obs ~ N 0,sigma2 A2 C6 0 33 3. Joint posterior marginals mean sd 2.5% 25% 50% 75% 97.5% theta 4.62917 0.03696 4.55700 4.60600 4.62900 4.65200 4.70300 A[1] 0.02415 0.04812 -0.07098 -0.00528 0.02386 0.05341 0.12050 C[1] -0.11067 0.02275 -0.15490 -0.12580 -0.11090 -0.09574 -0.06523 C[2] -0.03164 0.02257 -0.07554 -0.04675 -0.03179 -0.01675 0.01311 C[3] 0.02867 0.02249 -0.01549 0.01371 0.02860 0.04377 0.07298 C[4] 0.02143 0.02304 -0.02285 0.00603 0.02111 0.03644 0.06818 C[5] 0.00480 0.02271 -0.03912 -0.01040 0.00459 0.01977 0.05001 sigma.R 0.04537 0.03047 0.00243 0.02309 0.04154 0.06171 0.11680 sigma.P 0.06873 0.01772 0.03955 0.05644 0.06687 0.07890 0.10870 sigma 0.04523 0.00574 0.03568 0.04117 0.04466 0.04867 0.05811 deviance -203.08263 10.09257 -219.90000 -210.30000 -204.20000 -197.00000 -180.70000 Rhat 1.00103 1.00098 1.00100 1.00100 1.00099 1.00099 1.00101 1.00105 1.00099 1.00100 1.00100 n.eff 36000 60000 60000 60000 60000 60000 52000 25000 60000 60000 60000 34 3. Variance Components Estimates (in log scale) (SAS, Bayesian) Variance Components Run(Analyst) Plate(Analyst*Run) Residual Total Estimate 0.0015 0.0017 0.0039 0.0045 0.0019 0.0020 0.0073 0.0090 Lower95% Upper95% 0.0003 0.000006 0.0017 0.0016 0.0012 0.0013 0.0032 0.0051 1.4679 0.0136 0.0159 0.0118 0.0032 0.0034 1.4871 0.0219 % of Total 21 21(0-71) 53 53(15-83) 26 22(8-45) 100 100 Bayesian point estimate is the posterior median Bayesian % of Total is posterior median(2.5th-97.5 percentiles) 35 3. Posterior distribution of Error m m Error Error R P sigma.R 2 R ~ N 0, 3 sigma.P 2 SBD ~ N 0, 6 [1] [2] sigma2 ~ N 0, 6 To obtain a sample from the posterior distribution of Error: 1. Start with 60,000 MCMC draws of the 3 sigmas (a 3-vector) 2. For each draw vector, a. simulate a random sample of the 3 error contributors using [2] b. Plug these into [1] 3. The result is 60,000 MCMC draws from the posterior distribution of Error 36 3. Expanded uncertainty Percent of Total -U = -0.0924 +U = +0.0924 10 m m U 5 0 -0.1 0.0 0.1 Error Translate to the Relative Potency scale e m e m U e m e U ,e U Percent of Total e-U = 0.912 e+U = 1.097 10 em em – em-U em+U - em 80 7.1 7.7 102.5 9.0 9.9 120 10.6 11.6 SAS 5 0 0.9 1.0 exp(Error) 60,000 mcmc draws 1.1 1.2 7.8 37 Conclusions Metrological Approach Pros • Scientifically sound • Model based (forces analytical introspection) • ISO compliant Cons • Learning curve for CMC, regulators • Metrological approach still evolving (slowly) • How to deal with site and/or instrument biases? • How to deal with transformed scales of measurement? Bayesian Version of Metrological Approach Pros • Permits direct probability statements (risk management) • BUGS syntax mimics data generation mechanism • GUM revision moving toward Bayesian perspective Cons • Steep learning curve for CMC, regulators • Unfamiliar software tools (BUGS, R, STAN, JAGS,…) • MCMC requires care, maybe long computing times 38 Bibliography 1. Gelman A, et al (2014) Bayesian data analysis, 3rd edn, CRC Press 2. Burdick R, et al (2005) Design and analysis of gauge R&R studies, SIAM 3. Willink R (2013) Measurement uncertainty and probability, Cambridge University Press [gives a frequentist perspective] 4. Working Group 1 of the Joint Committee for Guides in Metrology (JCGM/WG 1), JCGM 100:2008, GUM 1995 with minor corrections, Evaluation of measurement data — Guide to the expression of uncertainty in measurement. 5. Working Group 2 of the Joint Committee for Guides in Metrology (JCGM/WG 2), JCGM 200:2012, VIM, 3rd edition, 2008 version with minor corrections, International vocabulary of metrology – Basic and general concepts and ssociated terms. 6. Eurachem working group (Editors: S L R Ellison , A Williams) Eurachem/CITAC Guide CG 4 (2012), Quantifying Uncertainty in Analytical Measurement, 3rd edn. 7. Hubert et al (2004, 2007) Harmonization strategies for the validation of quantitative analytical procedures: A SFSTP proposal Part 1, J Pharm Biomed Anal 36, 579-586. Part 2, J Pharm Biomed Anal 45: 70-81, Part 3, J Pharm Biomed Anal 45: 82-96. 8. Feinberg et al (2004) New advances in method validation and measurement uncertainty aimed at improving the quality of chemical data, Anal Bioanal Chem 380:502-514. 9. Howson C, and Urbach P, Scientific reasoning: the Bayesian approach 3rd edn. , Open Court, Chicago, IL [argues that scientific inference requires a Bayesian perspective] 39 “… To endure uncertainty is difficult, but so are most of the other great virtues”. - Bertrand Russell, 1950 - Thank you for your endurance!! P.S. Don’t forget to put the u in the mu 40