Part 05: Basic Bayes
480 likes | 622 Vues
Part 05: Basic Bayes. TOXOPLASMOSIS RATES (centered). The essential Bayes/Frequentist difference. As a Bayesian you can say, “Conditional on the data, the probability that the parameter is in the interval is 0.95” (you’ve always wanted to be able to say this!). SURGICAL
Part 05: Basic Bayes
E N D
Presentation Transcript
Part 05: Basic Bayes BIO656--Multilevel Models
TOXOPLASMOSIS RATES (centered) BIO656--Multilevel Models
The essential Bayes/Frequentist difference • As a Bayesian you can say, “Conditional on the data, • the probability that the parameter is in the interval is 0.95” • (you’ve always wanted to be able to say this!) BIO656--Multilevel Models
SURGICAL Hospital # of ops # of deaths A [1] 47 0 B 148 18 C 119 8 D 810 46 E 211 8 F 196 13 G 148 9 H 215 31 I 207 14 J 97 8 K 256 29 L 360 24 BIO656--Multilevel Models
“Surgical” Beta-Binomial Model (no combining; stand alone) model { for( i in 1 : N ) { p[i] ~ dbeta(1.0, 1.0) #need to specify the prior r[i] ~ dbin(p[i], n[i]) } righttail<-step(p[1]-3/n[1]) } # Also run with p[i] ~ dbeta(0.25,0.25) BIO656--Multilevel Models
“Surgical” Results for p[1] (no combining; stand alone) Beta mean sd 2.5% median 97.5% (1,1) 0.020 0.019 0.0003 0.0006 0.014 (.25,.25) 0.005 0.010 0.0002 0.0010 0.034 MLE 0 0.078 Beta(.25, .25) Beta(1,1) BIO656--Multilevel Models
Allows learning about the prior BIO656--Multilevel Models
“Surgical” Beta-binomial model (combine evidence; “estimate” the prior) model {for( i in 1 : N ) { b[i] ~ dnorm(mu,tau) # tau = 1/var r[i] ~ dbin(p[i],n[i]) logit(p[i]) <- b[i] } popmn <- exp(mu) / (1 + exp(mu)) mu ~ dnorm(0.0,1.0E-6) sigma <- 1 / sqrt(tau) tau ~ dgamma(alphatau, betatau) mutau<-1 alphatau<-.001 betatau<-alphatau/mutau } BIO656--Multilevel Models
“Surgical” Results (combine evidence) node mean sd 2.5% median 97.5% popmn 0.073 0.010 0.053 0.073 0.095 p[1] 0.053 0.020 0.018 0.052 0.094 BIO656--Multilevel Models
“Surgical” Results for p[1] (stand alone & combine) Beta mean sd 2.5% median 97.5% Comb 0.053 0.020 0.0180 0.0520 0.094 (1,1) 0.020 0.019 0.0003 0.0006 0.014 (.25,.25) 0.005 0.010 0.0002 0.0010 0.034 MLE 0 0.078 .25, .25 1,1 Comb BIO656--Multilevel Models
BACK TO HISTORICAL CONTROLS BIO656--Multilevel Models
BLOCKER Study deaths/n Treated Control 1 3/38 3/39 2 7/114 14/116 3 5/69 11/93 4 102/1533 27/1520 ..... 20 32/209 40/218 21 27/391 43/364 22 22/680 39/674 BIO656--Multilevel Models
delta = BIO656--Multilevel Models
Meta Analysis on the control probability (pc) (Historical Controls) model { for( i in 1 : Num ) { rc[i] ~ dbin(pc[i], nc[i]) rt[i] ~ dbin(pt[i], nt[i]) logit(pc[i]) <- mu[i] logit(pt[i]) <- mu[i] + delta[i] mu[i] ~ dnorm(d, tau) delta[i] ~ dnorm(0.0,1.0E-5) } d ~ dnorm(0.0,1.0E-6) mutau<-1 alphatau<-.0001 betatau<-alphatau/mutau tau ~ dgamma(alphatau, betatau) delta.new ~ dnorm(d, tau) sigma <- 1 / sqrt(tau) } BIO656--Multilevel Models
Meta Analysis on delta = log(OR) model { for( i in 1 : Num ) { rc[i] ~ dbin(pc[i], nc[i]) rt[i] ~ dbin(pt[i], nt[i]) logit(pc[i]) <- mu[i] logit(pt[i]) <- mu[i] + delta[i] mu[i] ~ dnorm(0.0,1.0E-5) delta[i] ~ dnorm(d, tau) } d ~ dnorm(0.0,1.0E-6) mutau<-1 alphatau<-.0001 betatau<-alphatau/mutau tau ~ dgamma(alphatau, betatau) delta.new ~ dnorm(d, tau) sigma <- 1 / sqrt(tau) } BIO656--Multilevel Models
Blocker MLE: pc = 0.079, pt = 0.077, delta = -0.28 Meta-analysis on delta node mean sd 2.5% median 97.5% pc[1] 0.08684 0.03338 0.03255 0.08207 0.1623 pt[1] 0.06883 0.02772 0.0256 0.06517 0.1321 delta[1] -0.2441 0.126 -0.4995 -0.2479 0.03553 Meta-analysis on pc node mean sd 2.5% median 97.5% pc[1] 0.09164 0.03144 0.04031 0.08749 0.164 pt[1] 0.07797 0.04282 0.01743 0.07022 0.1779 delta[1] -0.2504 0.762 -1.846 -0.2192 1.13 BIO656--Multilevel Models
Summary Carefully specified and applied, the Bayesian approach is very effective in • Structuring designs and analyses • Structuring complicated models and goals • we’ll see more of this in ranking • Incorporating all relevant uncertainties • Improving estimates • Communicating in a more “scientific” manner However, • The approach is no panacea and must be used carefully • Traditional values still apply Space-age methods will not rescue stone-age data BIO656--Multilevel Models