Bench test with Probabilistic Inversion and Isotonic Regression Roger Cooke Resources for the Future & Dept. Math Delft Univ of Technology Oct. 22-23, 2007 Tentpole 1: Observational Uncertainty Give 49 rats dose 21[units] and observe 15 responses. Sample independently 49 fresh rats from the same population, what is our uncertainty on # responses? Binomial uncertainty Bayes Uncertainty With uniform prior (normal approximation) 1 Tentpole 2: Uncertainty via parameter distribution (doesn’t depend on dose) Prob_Response(Dose) = g + (1-g) × (1 – e-b1×dose) Distribution over {g, b1}, the same for all doses Tentpole 3: Probability of Response does NOT decrease as dose increases Remove Non-monotonicity noise with ISOTONIC REGRESSION: PAV (pair adjacent violators) # R e s p o n s e Sample Independent binomials, if increasing, then do nothing if NOT increasing, then average Build up Isotonic Uncertainty Distribution dose 2 Probabilistic Inversion (Assume # subjects is same at each dose) DR model: ProbResp(dose) = f(d,α, β, γ,…) # R e s p o n s e Observational Uncertainty: Isotonic Observational Uncertainty: binomial dose Probabilistic Inversion: an inverse DOES NOT always exist, and NOT always unique DR model: ProbResp(dose) = f(d,α, β, γ,…) # R e s p o n s e dose 3 How do we do Probabilistic Inversion? Iterative Proportional Fitting – variant ProbResp(dose) = f(d,α, β, γ,…) Start with wide distribution # R e s p o n s e Trim to Observ’l uncertainty Re-weight to fit percentiles dose What is the difference with Statistics as Usual? 4 Statistics as Usual Recompute MLE estimates of parameters, get (asymptotically) joint normal parameter distribution DR model: ProbResp(dose) = f(d,α, β, γ,…) Sample Parameters, Compute prob of response for each sample Compute dist’n over # responses # R e s p o n s e ASSUME model is correct; If we repeat experiments, dose DR model: We get binomials ProbResp(dose) = f(d,α, β, γ,…) # R e s p o n s e dose 5 Probabilistic Inversion: Equivalent Picture DR model: P r o b a b i l i t y ProbResp(dose) = f(d,α, β, γ,…) Dose 1 Dose 2 Dose 3 Dose 4 Nr Responses BMD Tech. Guidance Doc Perferred model, log logistic, β = 1: 6 BMD Tech. Guidance Doc Probability Dose 0 Dose 60 Dose 21 Number of responses Probabilistic Inversion (fitting 5, 50, 95 %-tiles) BMD preferred model is loglogistic, b=1 Fit not too good for preferred model 7 Frambozadrine Fambozadrine; males multistage: Nr*(gamma+(1-gamma)*(1-exp(-b1*dose))) Dose zero Dose 15 Dose 1.2 Dose 82 8 Fambozadrine; females multistage: Nr*(gamma+(1-gamma)*(1-exp(-b1*dose))) Dose zero Dose 1.8 Dose 109 Dose 21 Frambozadrine, males and females Only found good fit with threshold model: Nr*(g+(1-g)*i1{s,dose,∞}*(1-exp(-b1*max{t,dose}-b2*max2 {t,dose} -b3*max3{t,dose}))) Probability of Response, Threshold Model 1 probability 0.8 0.6 Prob_Resp 0.4 0.2 0 dose 0 20 40 60 80 100 120 Below threshold s, background rate g applies Above threshold t a multistage model applies Distributions for g, s, t, b1, b2, b3 found by PI 9 Frambozadrine, MF, doses 0, 1.2, 1.8, 15 Dose zero Dose 1.2 Dose 15 Dose 1.8 Frambozadrine, MF, doses 21, 82, 109 Dose 21 Dose 82 Dose 109 10 Distribution of s, and t = 21 - s Before PI After PI Distribution of g Before PI After PI 11 Distribution of B1 Before PI After PI Parameters’ Joint Distribution 12 Probability as function of dose Nectorine Prob(adenoma OR neuroblastom) = Prob(adenoma) + Prob(neurblastoma) – Prob(adenoma) × Prob(neurblastoma) 13 BMD: logistic. PI for multistage:NR*(g+(1-g)*(1-exp(-b1*dose))) Dose 10 Dose 30 Dose 60 Persimonate BARRIER MODEL: NR*(g+(1-g)*(1-exp(-b1*i1{0,dose,t}-b2*i1{t,dose,s}-b3*i1{s,dose,∞ }))) b1 < b2 < b3 14 Barrier Model NR*(g+(1-g)*(1-exp(-b1*i1{0,dose,t}-b2*i1{t,dose,s}-b3*i1{s,dose,∞ }))) b1 < b2 < b3 Probability of Response, Barrier Model probability 1 0.8 0.6 Prob_Resp 0.4 0.2 0 0 20 40 60 80 100 120 dose Distribution on intervals and coefficients from PI Persimonate; Dose 0, 1.8, 9 ppm Dose 0 Dose 1.8 Dose 9 15 Persimonate; Dose 18, 36, 45 ppm Dose 18 Dose 36 Dose 45 Distributions for b1,b2, b3 Before PI After PI 16 Probability as function of dose CONCLUSIONS PLUS • Use Isotonic regression gets rid of ‘non-monotonic noise’ • PI CAN recover isotonic observational uncertainty • Minus • These methods are new and unfamiliar • Many open questions 17