The Posterior Probability of Passing a Compendial Test (Pa) Dave LeBlond, Principle Research Statistician, Abbott [email protected] Linas Mockus, Research Scientist, Purdue University [email protected] May 10, 2012 Bayes 2012, Aachen Outline • Process Qualification (PQ) – Compendial Tests – Test Pass Probability, Pa • ASTM E2709 • USP<905> Dosage Uniformity – Hierarchical process model • Bayesian PQ Approach • Prior Calibration • Number of Batches for PQ? – Operating Characteristics • Application to Sample Data – Cost Estimation • Summary 18th Century Statistics 19th Century Regulations + 20th Century Computing 21st Century cGMPs 2 Posterior Probability of Passing a Compendial Test Dave LeBlond May 10, 2012 Process Qualification Prior Calibration ASTM E2709 Number of Batches? USP905 Sample Data Bayesian Approach Summary What is a Compendial Test? Bad • “Bright line” standard of quality. Good – batch should always pass. • Benchmark for setting batch acceptance criteria. • May use multi-stage sampling. – USP<905> & <701> have 2 stages – USP<711> & <724> have 3 stages Fail • May use complex limits – Indifference zones – Limits on means, individuals, RSDs, counts – Zero tolerance limits 3 Posterior Probability of Passing a Compendial Test Dave LeBlond May 10, 2012 Pass Process Qualification Prior Calibration ASTM E2709 Number of Batches? USP905 Sample Data Bayesian Approach Summary FDA Process Validation Guidance • Establishes 3 validation stages8 – Process Design (QbD) – Process Qualification (PQ) – Continued Process Verification • “ [PQ] criteria … [should] … allow for a science- and risk-based decision about the ability of the process to consistently produce quality products… • … [and] include … statistical metrics defining both intra-batch and inter-batch variability.”8 Ø Place acceptance limits on Pa, the probability that future batches will pass the compendial test. 4 Posterior Probability of Passing a Compendial Test Dave LeBlond May 10, 2012 Process Qualification ASTM E2709 USP905 Bayesian Approach Prior Calibration Number of Batches? Sample Data Summary ASTM E2709 Approach to PQ • Derive g: Pa >= g(θ ) • Identify an acceptance region (AR) in parameter space: g(θ ) >= LB • Identify θ est . • Choose a 100(1-α)% confidence region (CR) method, given a sampling plan and data. • Generate an acceptance table (AT): θ est such that CR is within AR. • Choose a sampling plan so that Prob(θ est is within AT) > some desired value. • Obtain data. If θ est is within AT, there is 100(1-α)% confidence that Pa >= LB on repeated sampling of that batch. 5 Posterior Probability of Passing a Compendial Test Dave LeBlond May 10, 2012 Process Qualification Prior Calibration ASTM E2709 Number of Batches? USP905 Sample Data Bayesian Approach Summary Praise for ASTM E2709… • “This paradigm shift fundamentally changes how our industry should develop in-house specifications.”2 • “[This new] concept of meeting specification … may begin to include estimates of statistical confidence as part of cGMP.”3 • “E2709 was highly effective in identifying nonconforming material.”4 • Available for USP<905> & <711> as CUDAL, a validated SAS program,10 and as an Excel Spreadsheet.11 • Clearly ASTM E2709 is a very positive step… but ... Is anything missing? Can Bayesian tools make further improvement? 6 Posterior Probability of Passing a Compendial Test Dave LeBlond May 10, 2012 Process Qualification Prior Calibration ASTM E2709 Number of Batches? USP905 Sample Data Bayesian Approach Summary ASTM E2709 Limitations • Parameters are fixed (cannot have a distribution). • ∴ Pa is also denied a distribution. • Prediction of failure rates for future lots requires integration over the uncertain Pa … Not allowed. • Confidence region approach – Conservative approximations à Biased predictions. – Confidence regions are not unique. • Qualifies 1 batch at a time. No inference about the process. • What about cost? 7 Posterior Probability of Passing a Compendial Test Dave LeBlond May 10, 2012 Process Qualification Prior Calibration ASTM E2709 Number of Batches? USP905 Sample Data Bayesian Approach Summary USP<905> Uniformity of Dosage Units: Quality Stage Sample Size Requirements 8 Need stage 2 6 1 SD 10 4 Pass 2 10 0 70 100 X 10 130 8 6 2 SD 30 4 Fail 2 20 more 0 Pass X 30 130 (+ individual limits met) 70 100 8 Posterior Probability of Passing a Compendial Test Dave LeBlond May 10, 2012 Process Qualification Prior Calibration ASTM E2709 Number of Batches? USP905 Sample Data Bayesian Approach Summary A Hierarchical Process Model Process mean potency: Process β Batches 1 2 … B Batch i mean potency: iid ( ui ~ N β , σ B2 ) Tablet j(i) observed: Tablets iid ( yij ~ N ui , σ T2 9 Posterior Probability of Passing a Compendial Test Dave LeBlond May 10, 2012 ) [3] Process Qualification Prior Calibration ASTM E2709 Number of Batches? USP905 Sample Data Bayesian Approach Summary Actual Production Data: Joint 95% CI β σB Product Batches LB A B C D E F G H I J K L 3 32 85 49 2 32 10 4 14 16 4 4 100.0 100.8 100.7 100.2 99.7 100.1 98.3 98.7 99.4 98.9 98.9 101.7 est 100.0 101.8 101.6 100.6 100.1 101.2 99.8 100.8 100.7 100.0 99.8 102.6 σT UB est 100.1 102.8 102.5 101.0 100.4 102.2 101.3 102.8 101.9 101.2 100.6 103.6 0.0 2.2 3.5 1.2 0.0 2.4 1.6 1.7 2.0 1.8 0.7 0.7 UB 0.0 3.1 4.2 1.5 9.9 3.3 2.3 7.4 3.4 3.0 3.0 3.4 • Each batch passed USP<905>. • σB > 0 (Pvalue < 0.05) 10 Posterior Probability of Passing a Compendial Test Dave LeBlond May 10, 2012 est 3.1 3.0 3.3 1.5 2.5 2.5 2.1 1.1 2.6 3.5 0.8 1.7 UB 4.3 3.2 3.5 1.6 3.9 2.8 2.2 1.5 3.0 4.0 1.1 2.2 Process Qualification Prior Calibration ASTM E2709 Number of Batches? USP905 Sample Data Bayesian Approach Summary Pa is a deterministic function of model parameters • Use Monte-Carlo simulation to generate lookup table. 95 0.700 0.800 0.900 0.950 σB : 5 0. 0.3 200 0. 00 40 0 0.1 0.2 00 0.3 00 0.05.40 00 00 0 00 0 0.3 00.6.5.040000 000 0.1 σB : 3.5 0.500 0.600 0.700 0.800 00 0.2 00 0.3.400 0 σT 0.200 00 0.1 00 0.2.30000 0 0 .4 0 0 0 00.5.6.0700 0 6 0.800 0.900 0.950 0.990 105 00 0.1 00 0.2 .300 0 0 0 0 0.04.50600 0. σB : 2 100 0.900 4 0.990 0.999 0.950 0.999 2 95 100 0.990 105 95 100 105 β • Based on 10K simulated batches for each of 18K grid points. • Use tri-linear interpolation to obtain Pa for any desired θ. 11 Posterior Probability of Passing a Compendial Test Dave LeBlond May 10, 2012 Process Qualification Prior Calibration ASTM E2709 Number of Batches? USP905 Sample Data Bayesian Approach Summary A Bayesian Approach to PQ • Generate (by simulation) an interpolation table for Pa = g(θ ) • Choose PQ acceptance criteria (AC) – lower bound (LB) for Pa. – require that 100(1-α)% of the Pa posterior mass be >= LB. • Choose a sampling plan based on the simulated Operating Characteristics (OC). • Obtain data. – Obtain a posterior sample of θ. – Obtain a posterior sample of Pa. • If Pa posterior is acceptable, there is at least 100(1-α)% probability that Pa >= LB for the process. • Mean of Pa posterior = expected probability that future batches will pass the compendial test. 12 Posterior Probability of Passing a Compendial Test Dave LeBlond May 10, 2012 Process Qualification Prior Calibration ASTM E2709 Number of Batches? USP905 Sample Data Bayesian Approach Summary Benefits of a Bayesian Approach to Qualification • Direct inference on the parameter of interest (Pa) • Posterior distribution of Pa – Quantitative risk assessment (i.e., ICH Q9) – Production planning (expected cost and throughput) • Leverage prior knowledge (if justified) – “[The qualification report should consider]… the entire compilation of knowledge and information gained from the design stage through the process qualification stage.”8 13 Posterior Probability of Passing a Compendial Test Dave LeBlond May 10, 2012 Process Qualification Prior Calibration ASTM E2709 Number of Batches? USP905 Sample Data Bayesian Approach Summary Proposed Bayesian PQ Acceptance Criterion • “The confidence level selected can be based on risk analysis as it relates to the particular attribute under examination.”8 Ø Propose: median(Pa posterior) ≥ 0.9 (α = 0.5, LB = 0.9) mean would be computationally simpler, more discriminating • Examples of simulated borderline cases… = 105, σB = 2, σT = 4, 10 units/batch, Noninf Priors, 10K MCMC draws Percent of Total 25 20 15 24 batches 30 Percent of Total β Median(Pa)=0.89 10 5 0 25 20 4 batches Median(Pa)=0.92 15 10 5 0 0.2 0.4 0.6 0.8 0.2 Pa 0.4 0.6 Pa 14 Posterior Probability of Passing a Compendial Test Dave LeBlond May 10, 2012 0.8 Process Qualification Prior Calibration ASTM E2709 Number of Batches? USP905 Sample Data Bayesian Approach Summary Prior Calibration Using Simulated Data σT 5 2 5 σB 2 102 107 σB σT β 2 5 2 5 2 5 2 5 2 2 5 5 2 2 5 5 102 102 102 102 107 107 107 107 Pa* Ps* 1.00 0.98 0.97 0.78 1.00 0.87 0.49 0.50 0.00 0.03 0.27 0.44 0.02 0.17 0.72 0.67 * based on 100K simulated batches each. β • Simulated qualification data from 3, 4, 6, 12, or 24 batches • 4 weakly informative priors examined 15 Posterior Probability of Passing a Compendial Test Dave LeBlond May 10, 2012 Process Qualification Prior Calibration ASTM E2709 Number of Batches? USP905 Sample Data Bayesian Approach Summary Weakly Informative, Independent Priors Used None Least More Most N(100,10002) N(100,10002) N(100,152) N(100,102) RIG(.001,.001) Half Cauchy* Half t* Half t* 0-20 Undefined mean infinite mean mean = 70.7, mean = 14.1 df = 0 df = 1 df = 2 df = 2 RIG(0.5,8) RIG(.001,.001) Mean = 4 Undefined mean β 70-130 σΒ σT 0-20 RIG(.001,.001) RIG(.001,.001) Undefined mean Undefined mean df = 0 df = 0 df = 1 16 Posterior Probability of Passing a Compendial Test Dave LeBlond May 10, 2012 df = 0 Process Qualification Prior Calibration ASTM E2709 Number of Batches? USP905 Sample Data Bayesian Approach Summary % Bias in Pa Posterior Mean More Most 5 % Bias in Pa estimate β : 102 σB : 5 σT : 2 None 10 15 20 15 Most 5 β : 107 σB : 5 σT : 2 0 -10 -20 -30 -40 β : 102 σB : 2 σT : 5 0 -10 -20 -30 -40 10 More β : 102 σB : 5 σT : 5 β : 102 σB : 2 σT : 2 5 Least % Bias in Pa estimate Least 20 None 10 15 β : 107 σB : 5 σT : 5 β : 107 σB : 2 σT : 2 β : 107 σB : 2 σT : 5 0 -10 -20 -30 -40 5 10 Number of Batches 15 20 Number of Batches • All priors exhibit conservative (low) mean(Pa) • Non-informative prior is least conservative. 17 Posterior Probability of Passing a Compendial Test Dave LeBlond 20 May 10, 2012 0 -10 -20 -30 -40 Process Qualification Prior Calibration ASTM E2709 Number of Batches? USP905 Sample Data Bayesian Approach Summary % RMSE in Pa Posterior Mean More Most % RMSE in Pa estimate 5 β : 102 σB : 5 σT : 2 10 15 β : 102 σB : 2 σT : 5 15 Most 5 40 30 20 10 0 40 30 20 10 0 10 More 20 β : 102 σB : 5 σT : 5 β : 102 σB : 2 σT : 2 5 Least None % RMSE in Pa estimate Least β : 107 σB : 5 σT : 2 10 15 β : 107 σB : 2 σT : 5 40 30 20 10 0 10 15 20 Number of Batches Number of Batches • In most cases Non-informative prior has lower RMSE. 18 Posterior Probability of Passing a Compendial Test Dave LeBlond 20 β : 107 σB : 5 σT : 5 β : 107 σB : 2 σT : 2 5 20 None May 10, 2012 40 30 20 10 0 Process Qualification Prior Calibration ASTM E2709 Number of Batches? USP905 Sample Data Bayesian Approach Summary 90% Credible Interval Coverage of Pa More Most 90% Interval Coverage for Pa 5 β : 102 σB : 5 σT : 2 Least None 10 15 0.95 0.90 0.85 0.80 0.75 β : 102 σB : 2 σT : 5 0.95 0.90 0.85 0.80 0.75 5 10 15 Most 5 β : 102 σB : 5 σT : 5 β : 102 σB : 2 σT : 2 More 20 20 90% Interval Coverage for Pa Least β : 107 σB : 5 σT : 2 None 10 15 β : 107 σB : 5 σT : 5 0.95 0.90 0.85 0.80 0.75 β : 107 σB : 2 σT : 2 β : 107 σB : 2 σT : 5 0.95 0.90 0.85 0.80 0.75 5 10 15 20 Number of Batches Number of Batches • Non-informative prior has coverage closest to nominal 19 Posterior Probability of Passing a Compendial Test 20 Dave LeBlond May 10, 2012 Process Qualification Prior Calibration ASTM E2709 Number of Batches? USP905 Sample Data Bayesian Approach Summary Summary and Final Choice of Prior • % Bias in Pa Posterior mean – Bias increases from -10% to 0% as #batches grows (3 to 24) – In all cases, Non-informative prior shows least bias • % RMSE in Pa Posterior mean – RMSE asymptotes from +10-30% to zero as #batches grows – Non-informative prior shows least RMSE • Coverage of 90% Credible Interval – Nominal coverage in most cases, some cases only 80% – Non-informative prior coverage closest to nominal Ø Non-informative prior used here… 20 Posterior Probability of Passing a Compendial Test Dave LeBlond May 10, 2012 Process Qualification Prior Calibration ASTM E2709 Number of Batches? USP905 Sample Data Bayesian Approach Summary Establishing a PQ Sampling Plan • Operating Characteristic Curve(OC) • Acceptable Quality Level • Rejectable Quality Level Pa 0.4 0.95 1 Prob(Pass PQ) • Pa = quality metric 0.10 • How many batches RQL (more/less) ? • Requires Monte-Carlo simulation Good Quality Quality 21 Dave LeBlond AQL AQL Poor • Does OC depend on θ ? Posterior Probability of Passing a Compendial Test RQL May 10, 2012 Process Qualification Prior Calibration ASTM E2709 Number of Batches? USP905 Sample Data Bayesian Approach Summary OC of Proposed Approach Number of Batches σT 5 3 8 2 8 4 6 12 24 0.95 4 0.5 2 102 107 (Prob(median(Pa) ≥ 0.9))1 σB 5 112 β • 27 grid points • 1000 data sets per grid point • Vary number of batches • OC independent of θ ? 0.1 0.05 0 0.0 0.2 0.4 Population Pa 22 Posterior Probability of Passing a Compendial Test 0.6 Dave LeBlond May 10, 2012 0.8 1.0 Process Qualification Prior Calibration ASTM E2709 Number of Batches? USP905 Sample Data Bayesian Approach Summary Using AQL and RQL to Set Number of Batches Number of Batches 0.95 3 4 6 12 24 0.10 0.95 (Prob(median(Pa) ≥ 0.9))1 4 0.5 0.1 0.05 0 0.0 0.2 0.4 0.6 0.8 1.0 Population Pa AQL RQL 23 Posterior Probability of Passing a Compendial Test Dave LeBlond May 10, 2012 Process Qualification Prior Calibration ASTM E2709 Number of Batches? USP905 Sample Data Bayesian Approach Summary PQ Pass Rates(α=0.5, LB=0.90) : Bayes vs E2709 Popn.sig.B=5 Popn.sig.B=2 Popn.beta β Popn.beta β 102 102 107 1 0.4 0.2 Pa=0.97 Pa=0.49 Y 0 2 4 6 8 10 12 2 4 6 8 10 12 Number of Batches Number of Batches 0.4 Pa=0.98 0.2 0 1 Pa=0.78 0.8 Pa=0.50 0.6 5 Y 0.6 5 5 0.8 Y Popn.sig.T Y Popn.sig.T 1 0.6 0.4 0.2 0 2 4 6 8 10 12 2 4 6 8 10 12 Number of Batches Number of Batches σΒ = 2 σΒ = 5 24 Posterior Probability of Passing a Compendial Test Popn.sig.T 0 Pa=1.00 Pa=0.87 0.8 2 0.2 Pa=1.00 Popn.sig.T σΤ 0.6 2 2 0.8 5 2 PQ Probability of Passing 1 0.4 107 Dave LeBlond May 10, 2012 Process Qualification Prior Calibration ASTM E2709 Number of Batches? USP905 Sample Data Bayesian Approach Summary 115 Product C Product D Potency 100 10 25 Lot 40 85 85 Potency 100 115 Actual Production Data 10 • Note: Each lot passed USP<905>. 25 Posterior Probability of Passing a Compendial Test Dave LeBlond May 10, 2012 25 Lot 40 Process Qualification Prior Calibration ASTM E2709 Number of Batches? USP905 Sample Data Bayesian Approach Summary Joint Posterior Samples after 3, 6, and 20 batches Product C σB σT Product D β 26 Posterior Probability of Passing a Compendial Test Dave LeBlond May 10, 2012 Process Qualification Prior Calibration ASTM E2709 Number of Batches? USP905 Sample Data Bayesian Approach Summary 0.05, 0.50 & 0.95th Posterior Quantiles for Product D Quantiles of Pa 0 0.5 1 • Cumulative Analysis (lots 3 to 45) • For very good process 3 may be enough. REML Point Estimates β = 100.6 σB = 1.2 σT = 1.5 10 25 40 # of lots 27 Posterior Probability of Passing a Compendial Test Dave LeBlond May 10, 2012 Process Qualification Prior Calibration ASTM E2709 Number of Batches? USP905 Sample Data Bayesian Approach Summary 0.05, 0.50 & 0.95th Posterior Quantiles for Product C Quantiles of Pa 0 0.5 1 • Cumulative Analysis (lots 3 to 45) • For challenging processes >3 required. REML Point Estimates β = 101.6 σB = 3.5 σT = 3.3 10 25 # of lots 28 Posterior Probability of Passing a Compendial Test Dave LeBlond May 10, 2012 40 Process Qualification Prior Calibration ASTM E2709 Number of Batches? USP905 Sample Data Bayesian Approach Summary How Many Batches for a PQ? 2 approaches: • Prospectively: use simulated OC curves, or • Cumulative stability of Pa quantiles. 29 Posterior Probability of Passing a Compendial Test Dave LeBlond May 10, 2012 Process Qualification Prior Calibration ASTM E2709 Number of Batches? USP905 Sample Data Bayesian Approach Summary Cost Estimates • Not a regulatory consideration • Important for Manufacturer – Stage testing increases analytical costs – Failures risk supply of critical drugs – Resource planning • Need Pa to estimate cost • For multi-stage compendial tests, also need Ps, the probability of stage testing. 30 Posterior Probability of Passing a Compendial Test Dave LeBlond May 10, 2012 Process Qualification Prior Calibration ASTM E2709 Number of Batches? USP905 Sample Data Bayesian Approach Summary USP<905> Uniformity of Dosage Units: Cost Let • Ct = Cost of testing 10 tablets ( ~ $3K ) • Cm = Cost of manufacturing 1 batch ( ~ $200K ) • Ci = Cost of a failure investigation ( ~ $2K ) Then E[total cost] = Ct + 2·Ct·Ps + Cm +Cm· (1-Pa) + Ci· (1-Pa) • Stage 2 triples testing cost • Failure doubles manufacturing cost and requires a failure investigation. 31 Posterior Probability of Passing a Compendial Test Dave LeBlond May 10, 2012 Process Qualification Prior Calibration ASTM E2709 Number of Batches? USP905 Sample Data Bayesian Approach Summary Posterior Distribution of Future Costs for Product C Pa Mean = 0.97 40 20 50 0 Ps Mean = 0.05 40 30 20 10 Percent of Total 60 60 Percent of Total Percent of Total • Estimation after first 6 batches 40 20 0 0 0.6 0.7 0.8 0.9 0.1 Pa Posterior Total Cost Mean = $209K 60 0.2 0.3 200 0.4 • Production costs per batch (posterior expectation): – Testing: $3.3K – Manufacturing: $206K – Investigation: $60 32 Dave LeBlond 300 350 Total Cost ($1K) Posterior Ps Posterior Posterior Probability of Passing a Compendial Test 250 May 10, 2012 Process Qualification Prior Calibration ASTM E2709 Number of Batches? USP905 Sample Data Bayesian Approach Summary Future Work • Compare Stringency of Proposed Bayesian Criterion with CUDAL.10, 11 • Dependence of OC curve on θ ? • Non-normal populations? • Excel tool? • Extension to any compendial test – Same principles – Most will be trivial extensions, compatible with Excel • Dissolution & disintegration multi-stage tests – USP<711>, <724>, <701> 33 Posterior Probability of Passing a Compendial Test Dave LeBlond May 10, 2012 Process Qualification Prior Calibration ASTM E2709 Number of Batches? USP905 Sample Data Bayesian Approach Summary Conclusions • Pa, is a key quality metric in PQ. • Ps may also be of interest to manufacturers. • ASTM E2709 is a breakthrough in PQ thinking. • PQ requires a model for between-batch variance. • Bayesian hierarchical modeling provides – Direct inference on Pa and Ps – Basis for sampling plan choice – Manufacturing cost projections 34 Posterior Probability of Passing a Compendial Test Dave LeBlond May 10, 2012 References 1. 2. American Society for Testing and Materials (ASTM) Standard E11 2709-10, May, 2010 Torbeck, LD (May 2, 2010) Statistical solutions: Bergum’s method recognized, Pharmaceutical Technology. 3. Jon Clark (September 29, 2010) Confidence- critical to batch release: Application of ASTM E2709, presented at QbD/PAT Conference, University of Heidelberg. 4. Lunney, P.D., Anderson, C.A., “Investigation of the Statistical Power of the Content Uniformity Tests Using Simulation Studies”, Journal of Pharmaceutical Innovation, pp 24-35, 13March2009. 5. US Pharmacopoeia 34 (2011) General Chapter <905> Uniformity of Dosage Units (harmonized with JP and EP). 6. LeBlond, DJ (Spring, 2005) Methodology for predicting batch manufacturing risk. MS Thesis, Colorado State University. 7. LeBlond DJ (August, 2009) Risk Assessment of Drug Product Content Uniformity Release Failure: A Bayesian Approach, Joint Statistical Meetings, Washington DC 8. FDA CDER, CBER, CVM (January 2011) Guidance for Industry, Process Valdiation: General Principles and Practices, rev 1. 9. Gelman, A (2006) Prior distributions for variance parameters in hierarchical models, Bayesian Analysis 1(3), 515-533 10. J.S. Bergum and L. Hua (October 2, 2007), Acceptance Limits for the New ICH USP 29 ContentUniformity Test, Pharm. Technol. Online http://pharmtech.findpharma.com/pharmtech/article/ articleDetail.jsp?id=463577), accessed Apr. 4, 2012, October 2007. 11. P. Cholayudth (2009), Establishing Acceptance Limits for Probability of Passing Multiple Stage Tests in Proces Validation through a Process Capability Approach, Jrnl. of Validat. Technol. 15 (4), 77–90. 12. Y. Hu and D. LeBlond (2011) Assessment of Large-Sample Unit-Dose Uniformity Tests, Pharmaceutical Technology 35(10) 82-92. 35 Posterior Probability of Passing a Compendial Test Dave LeBlond May 10, 2012 Backup Slides 36 Posterior Probability of Passing a Compendial Test Dave LeBlond May 10, 2012 Full Conditionals for Gelman’s Half-t Prior9 ( p β | y, η, ξ , σ T2 ) B Ti ⎛ ⎞ 2 2 β σ + σ ⎜ ⎟ 0 T 0 ∑∑ ( yij − ξηi ) 2 2 σ 0 σ T ⎟ i =1 j =1 = N ⎜ β | , ⎜ Nσ 02 + σ T2 Nσ 02 + σ T2 ⎟ ⎜ ⎟ ⎝ ⎠ N S ⎛ ⎞ p σ T2 | y, η, β , ξ = IG ⎜ σ T2 | + aT , + bT ⎟ 2 2 ⎝ ⎠ where ( ) B N = ∑ Ti i =1 B 2 Ti S = ∑∑ ( yij − ui ) i =1 j =1 ui = β + ξηi ( p ξ | y, η, β , σ ( 2 T ) Ti B ⎛ ⎜ ∑ηi ∑ ( yij − β ) σ T2 i =1 j =1 = N ⎜ ξ | B , B ⎜ σ T2 σ T2 2 2 T η + T η + ∑ ∑ ⎜ i i i i A2 A2 i =1 i =1 ⎝ p ηi | y , β , ξ , σ η2 , σ T2 ) ⎞ ⎟ ⎟ ⎟ ⎟ ⎠ Ti ⎛ ⎞ 2 ξσ 2 2 ⎜ ⎟ η ∑ ( yij − β ) ση σ T i =1 ⎟ = N ⎜ηi | , σ T2 + Tiξ 2σ η2 σ T2 + Tiξ 2σ η2 ⎟ ⎜ ⎜ ⎟ ⎝ ⎠ • m σ B2 = ξ 2 ⋅ ση2 B ⎛ ⎞ η i2 + 1 ⎟ ∑ ⎜ B + 1 i =1 ⎟ p σ η2 | η = IG ⎜ σ η2 | , 2 2 ⎜ ⎟ ⎜ ⎟ ⎝ ⎠ ( ) 37 Posterior Probability of Passing a Compendial Test Dave LeBlond May 10, 2012 %Bias in β posterior mean More Most 5 ^ % Bias in β β : 102 σB : 5 σT : 2 Least None 10 15 β : 107 σB : 5 σT : 2 β : 102 σB : 2 σT : 5 0.0 15 20 β : 107 σB : 5 σT : 5 -0.5 -0.5 -1.0 β : 107 σB : 2 σT : 2 β : 107 σB : 2 σT : 5 0.0 -0.5 -1.0 -1.0 15 10 0.0 -0.5 10 None 0.0 -1.0 5 Most 5 β : 102 σB : 5 σT : 5 β : 102 σB : 2 σT : 2 More 20 ^ % Bias in β Least 5 20 10 15 20 Number of Batches Number of Batches • Non-informative prior is least biased at lower number of batches • Bias may be induced by prior mean 38 Posterior Probability of Passing a Compendial Test Dave LeBlond May 10, 2012 %RMSE in β posterior mean Least More Most 5 β : 102 σB : 5 σT : 2 None 10 15 20 β : 102 σB : 5 σT : 5 ^ % RMSE in β 3 2 1 β : 102 σB : 2 σT : 2 β : 102 σB : 2 σT : 5 3 2 1 5 10 15 20 Number of Batches • Virtually identical for β=107 • Non-informative prior has lowest loss based on MSE 39 Posterior Probability of Passing a Compendial Test Dave LeBlond May 10, 2012 90% Interval Coverage of beta Least More Most 90% Interval Coverage for β 5 β : 102 σB : 5 σT : 2 None 10 15 β : 102 σB : 5 σT : 5 20 1.00 0.95 0.90 0.85 β : 102 σB : 2 σT : 2 1.00 β : 102 σB : 2 σT : 5 0.95 0.90 0.85 5 10 15 20 Number of Batches • Coverage is nearly nominal for Non-informative prior regardless of number of batches. 40 Posterior Probability of Passing a Compendial Test Dave LeBlond May 10, 2012 % Bias in sB and sT Posterior Means More Most 5 ^B % Bias in σ β : 102 σB : 5 σT : 2 Least None 10 15 More β : 102 σB : 5 σT : 2 β : 102 σB : 5 σT : 5 500 400 300 200 100 0 β : 102 σB : 2 σT : 2 Most 5 20 β : 102 σB : 2 σT : 5 500 400 300 200 100 0 None 10 15 20 β : 102 σB : 5 σT : 5 6 4 ^T % Bias in σ Least 2 β : 102 σB : 2 σT : 2 β : 102 σB : 2 σT : 5 6 4 2 0 5 10 15 5 20 10 15 20 Number of Batches Number of Batches • Virtually identical for beta = 107 • Non-informative prior has lowest bias. Bias may be induced by prior mean. 41 Posterior Probability of Passing a Compendial Test Dave LeBlond May 10, 2012 0 % RMSE in sB and sT Posterior Means Least More Most 5 β : 102 σB : 5 σT : 2 None 10 15 Least More 20 Most 5 β : 102 σB : 5 σT : 5 β : 102 σB : 5 σT : 2 None 10 15 20 β : 102 σB : 5 σT : 5 400 200 β : 102 σB : 2 σT : 2 0 β : 102 σB : 2 σT : 5 600 400 ^T % RMSE in σ ^B % RMSE in σ 600 200 0 5 10 15 20 14 12 10 8 6 β : 102 σB : 2 σT : 2 β : 102 σB : 2 σT : 5 14 12 10 8 6 5 Number of Batches 10 15 20 Number of Batches • Virtually identical for beta = 107 • For sigB, Non-informative prior has least RMSE. For sigT, prior choice is irrelevant. 42 Posterior Probability of Passing a Compendial Test Dave LeBlond May 10, 2012 Coverage of 90% interval estimates of sB and sT More Most 90% Interval Coverage for σB 5 β : 102 σB : 5 σT : 2 None 10 15 Least 0.95 0.90 0.85 0.80 β : 102 σB : 2 σT : 5 0.95 0.90 0.85 0.80 5 10 15 20 Most 5 β : 102 σB : 5 σT : 5 β : 102 σB : 2 σT : 2 More 20 90% Interval Coverage for σT Least β : 102 σB : 5 σT : 2 None 10 15 20 β : 102 σB : 5 σT : 5 0.92 0.91 0.90 0.89 0.88 β : 102 σB : 2 σT : 2 β : 102 σB : 2 σT : 5 0.92 0.91 0.90 0.89 0.88 5 Number of Batches 10 15 20 Number of Batches • Virtually identical for beta = 107 • For sigB, Non-informative prior coverage is a little low in some cases. For sigT, prior has little effect on coverage. 43 Posterior Probability of Passing a Compendial Test Dave LeBlond May 10, 2012 % Bias and % RMSE in the Ps Posterior Mean More Most 5 % Bias in Ps estimate β : 102 σB : 5 σT : 2 None 10 15 Least β : 102 σB : 5 σT : 5 β : 102 σB : 2 σT : 2 β : 107 σB : 5 σT : 2 β : 102 σB : 2 σT : 5 250 200 150 100 50 0 10 15 Most 5 250 200 150 100 50 0 5 More 20 % Bias in Ps estimate Least 20 None 10 15 β : 107 σB : 5 σT : 5 250 200 150 100 50 0 β : 107 σB : 2 σT : 2 β : 107 σB : 2 σT : 5 250 200 150 100 50 0 5 10 Number of Batches 15 20 Number of Batches • The % bias for 102, 2, 2 was extremely high because the population Ps is so close to zero. • The % RMSE is nearly identical because the bias is the largest contributor to MSE • Non-informative prior has lowest % Bias 44 Posterior Probability of Passing a Compendial Test 20 Dave LeBlond May 10, 2012 Coverage of the 90% Credible Interval for Ps More Most 90% Interval Coverage for Ps 5 β : 102 σB : 5 σT : 2 None 10 15 β : 102 σB : 2 σT : 5 15 20 Number of Batches Most 5 0.9 0.8 0.7 0.6 0.5 0.4 0.9 0.8 0.7 0.6 0.5 0.4 10 More 20 β : 102 σB : 5 σT : 5 β : 102 σB : 2 σT : 2 5 Least 90% Interval Coverage for Ps Least β : 107 σB : 5 σT : 2 None 10 15 β : 107 σB : 5 σT : 5 β : 107 σB : 2 σT : 2 β : 107 σB : 2 σT : 5 0.9 0.8 0.7 0.6 0.5 0.4 5 10 15 20 Number of Batches • Non-informative prior has coverage closest to nominal 45 Posterior Probability of Passing a Compendial Test Dave LeBlond 20 May 10, 2012 0.9 0.8 0.7 0.6 0.5 0.4 OC of Proposed Approach: Effect of model parameter (one at a time, Non-informative prior) Varied beta sigB • Population parameters varied sigT – beta: 100-120 with sigT=sigB=3 – sigB: 2-15 with sigT=2 and beta=102 – sigT: 2-15 with sigB=2 and beta=102 B:3 Prob[median(Pa) >= 0.9] 1.0 0.8 • In principle, the median(Pa) and its variability may depend on: 0.6 0.4 – Population Pa – Population Parameters. 0.2 • This graph shows that model 0.0 0.0 0.2 0.4 0.6 0.8 1.0 Population Pa parameter has a minor affect on the OC curve, the OC curve is controlled largely by the population Pa. 46 Posterior Probability of Passing a Compendial Test Dave LeBlond May 10, 2012 Operating Characteristics of Proposed Approach: Effect of prior, number of batches, choice of statistic Number of Batches 3 4 6 12 Number of Batches 24 0.5 0.6 Prior : Most 0.7 0.8 3 4 0.9 6 12 1.0 0.5 Prior : None 0.6 Prior : Most 1.0 0.7 0.8 0.6 0.4 0.2 0.0 Prior : Least Prior : More 1.0 0.8 0.6 0.4 0.8 0.6 0.4 0.2 0.0 Prior : Least 0.8 0.6 0.4 0.2 0.0 0.7 0.8 0.9 1.0 Prior : More 1.0 0.0 0.6 0.5 0.6 Population Pa 0.7 0.8 0.9 1.0 Population Pa • Non-Informative Prior least sensitive to sample size • Mean(Pa) is more conservative than median(Pa) 47 Posterior Probability of Passing a Compendial Test 1.0 1.0 0.2 0.5 0.9 Prior : None 0.8 Prob[mean(Posterior Pa)>0.9] Prob[median(Posterior Pa)>0.9] 24 Dave LeBlond May 10, 2012
© Copyright 2024 Paperzz