> > # entering data (p. 434) > id<-(1:15) > libido<-c(3,2,1,1,4,5,2,4,2,3,7,4,5,3,6) > dose<-gl(3,5, labels = c("Placebo", "Low Dose", "High Dose")) > viagraData<-data.frame(dose, libido) > > > ###################### regular, non Bayesian analyses of the data ###################### > > # conventional ANOVA (p. 439) > viagraModel <- aov(libido ~ dose, data = viagraData) > summary.lm(viagraModel) Call: aov(formula = libido ~ dose, data = viagraData) Residuals: Min 1Q Median 3Q Max -2.0 -1.2 -0.2 0.9 2.0 Coefficients: Estimate Std. Error t value Pr(>|t|) (Intercept) 2.2000 0.6272 3.508 0.00432 ** doseLow Dose 1.0000 0.8869 1.127 0.28158 doseHigh Dose 2.8000 0.8869 3.157 0.00827 ** --- Signif. codes: 0 Ô***Õ 0.001 Ô**Õ 0.01 Ô*Õ 0.05 Ô.Õ 0.1 Ô Õ 1 Residual standard error: 1.402 on 12 degrees of freedom Multiple R-squared: 0.4604, Adjusted R-squared: 0.3704 F-statistic: 5.119 on 2 and 12 DF, p-value: 0.02469 > > > > ###################### Bayesian ANOVA using MCMCglmm ###################### > > model1 = MCMCglmm(libido ~ dose, data=viagraData, verbose=F, + nitt=50000, burnin=3000, thin=10) > summary(model1) Iterations = 3001:49991 Thinning interval = 10 Sample size = 4700 DIC: 57.58071 R-structure: ~units post.mean l-95% CI u-95% CI eff.samp units 2.329 0.8284 4.453 4700 Location effects: libido ~ dose post.mean l-95% CI u-95% CI eff.samp pMCMC (Intercept) 2.2000 0.8032 3.4958 4700 0.00553 ** doseLow Dose 0.9949 -0.9133 2.8483 4700 0.28553 doseHigh Dose 2.7779 0.9522 4.7755 4700 0.01064 * --- Signif. codes: 0 Ô***Õ 0.001 Ô**Õ 0.01 Ô*Õ 0.05 Ô.Õ 0.1 Ô Õ 1 > autocorr(model1$VCV) , , units units Lag 0 1.0000000000 Lag 10 0.0127178114 Lag 50 -0.0047286002 Lag 100 -0.0001693583 Lag 500 -0.0277471522 > plot(model1) Hit to see next plot: > > # running the same analyses using different priors, & then plotting the various posteriors for "dose" > > estimateNum = 2 # the number of regression estimate to be used for the analyses e.g., the 1st (intercept), or 2nd, etc > > Ndatasets = 5 # the number of additional analyses using different priors > > # plot data for the first Bayes model > densdat <- density(model1$Sol[,estimateNum]) > plotdatx <- densdat$x > plotdaty <- densdat$y > > # for MCMCglmm priors, B is for fixed effects. B is a list containing the expected value (mu) > # and a (co)varcv bgiance matrix (V) representing the strength of belief: the defaults are B$mu=0 > # and B$V=I*1e+10, where where I is an identity matrix of appropriate dimension > > # the different priors will be derived from the lm, non-Bayesian regression coefficients > # the mu values will be random values from a 4 SE range around the lm estimates > # the V values will be set at (4 * the SEs)**2 > > fit.list <- summary.lm(viagraModel) > coeffs <- fit.list$coefficients > coefnames <- rownames(coeffs) > > # ranges around the lm estimates > MUranges <- cbind( (coeffs[,1] - (4 * coeffs[,2])), (coeffs[,1] + (4 * coeffs[,2])) ) > > # random values from the MUranges for each coefficient > MUs <- matrix(-9999,nrow(coeffs),Ndatasets) > for (lupe in 1:nrow(coeffs)) { MUs[lupe,] <- runif(Ndatasets, min=MUranges[lupe,1], max=MUranges[lupe,2]) } > dimnames(MUs) <-list(rep("", dim(MUs)[1])) > colnames(MUs) <- seq(1:Ndatasets) > cat('\n\nPrior regression coefficient estimates for each additional analysis:') Prior regression coefficient estimates for each additional analysis: > print(round(MUs,2)) 1 2 3 4 5 3.04 1.82 4.40 1.49 4.20 1.61 -0.24 2.70 3.29 2.54 2.83 2.21 4.86 1.32 -0.47 > > # (co)variance matrix (V) representing the strength of belief - which will be (2 * the lm SEs)**2 > Vlmx2 <- diag((4 * coeffs[,2])**2) > colnames(Vlmx2) <- seq(1:nrow(MUs)) > cat('\n\nPrior variance (strength of belief) value used for each regression estimate in the additional analyses:\n', + round(diag(Vlmx2),5), sep="\n") Prior variance (strength of belief) value used for each regression estimate in the additional analyses: 6.29333 12.58667 12.58667 > > > for (lupeSets in 1:Ndatasets) { + prior.list <- list(B = list(mu = MUs[,lupeSets], V = Vlmx2)) + model2 = MCMCglmm(libido ~ dose, data=viagraData, prior = prior.list, + nitt=50000, burnin=3000, thin=10, verbose=F) + + densdat <- density(model2$Sol[,estimateNum]) + plotdatx <- cbind( plotdatx, densdat$x) + plotdaty <- cbind( plotdaty, densdat$y) + } > > matplot( plotdatx, plotdaty, main="", xlab=paste('Estimate for', coefnames[estimateNum]), ylab='Density', + font.lab=1.8, type='l', cex.lab=1.2, col=c(2,rep(1,ncol(plotdatx))), lty=1, lwd=1 ) > title(main=paste('Density Plots of the Posteriors for', coefnames[estimateNum], '\nProduced Using Differing Priors')) > legend("topright", c("broad priors","random priors"), bty="n", lty=c(1,1), + lwd=2, cex=1.2, text.col=c(2,1), col=c(2,1) ) > > > ################ posterior predictive checks using the bayesplot package ###################### > > # from the bayesplot package documentation: > # The bayesplot package provides various plotting functions for graphical posterior predictive checking, > # that is, creating graphical displays comparing observed data to simulated data from the posterior > # predictive distribution. The idea behind posterior predictive checking is simple: if a model is a > # good fit then we should be able to use it to generate data that looks a lot like the data we observed. > > y <- viagraData$libido # the DV that was used for the MCMCglmm model > > # generating the simulated data using the simulate.MCMCglmm function from the MCMCglmm package > yrep <- simulate.MCMCglmm(model1, 25) > > yrep <- t( yrep ) # transposing yrep for the bayesplot functions > > > # ppc_stat: A histogram of the distribution of a test statistic computed by applying stat to each > # dataset (row) in yrep. The value of the statistic in the observed data, stat(y), is > # overlaid as a vertical line. > ppc_stat(y, yrep, binwidth = 1) > > > # ppc_dens_overlay: Kernel density or empirical CDF estimates of each dataset (row) in yrep are > # overlaid, with the distribution of y itself on top (and in a darker shade). > ppc_dens_overlay(y, yrep[1:25, ]) > > > # ppc_scatter_avg: A scatterplot of y against the average values of yrep, i.e., the > # points (mean(yrep[, n]), y[n]), where each yrep[, n] is a vector of length equal > # to the number of posterior draws. > ppc_scatter_avg(y, yrep) > > > # ppc_hist: A separate histogram, shaded frequency polygon, smoothed kernel density estimate, > # or box and whiskers plot is displayed for y and each dataset (row) in yrep. > # For these plots yrep should therefore contain only a small number of rows. > ppc_hist(y, yrep[1:8, ], binwidth = .3) > > > > ###################### now conduct the Bayesian analyses using the rstanarm package ###################### > > model10 <- stan_glm(libido ~ dose, data = viagraData, + warmup = 1000, iter = 2000, sparse = FALSE, seed = 123) SAMPLING FOR MODEL 'continuous' NOW (CHAIN 1). Gradient evaluation took 1.4e-05 seconds 1000 transitions using 10 leapfrog steps per transition would take 0.14 seconds. Adjust your expectations accordingly! Iteration: 1 / 2000 [ 0%] (Warmup) Iteration: 200 / 2000 [ 10%] (Warmup) Iteration: 400 / 2000 [ 20%] (Warmup) Iteration: 600 / 2000 [ 30%] (Warmup) Iteration: 800 / 2000 [ 40%] (Warmup) Iteration: 1000 / 2000 [ 50%] (Warmup) Iteration: 1001 / 2000 [ 50%] (Sampling) Iteration: 1200 / 2000 [ 60%] (Sampling) Iteration: 1400 / 2000 [ 70%] (Sampling) Iteration: 1600 / 2000 [ 80%] (Sampling) Iteration: 1800 / 2000 [ 90%] (Sampling) Iteration: 2000 / 2000 [100%] (Sampling) Elapsed Time: 0.034771 seconds (Warm-up) 0.032755 seconds (Sampling) 0.067526 seconds (Total) SAMPLING FOR MODEL 'continuous' NOW (CHAIN 2). Gradient evaluation took 1e-05 seconds 1000 transitions using 10 leapfrog steps per transition would take 0.1 seconds. Adjust your expectations accordingly! Iteration: 1 / 2000 [ 0%] (Warmup) Iteration: 200 / 2000 [ 10%] (Warmup) Iteration: 400 / 2000 [ 20%] (Warmup) Iteration: 600 / 2000 [ 30%] (Warmup) Iteration: 800 / 2000 [ 40%] (Warmup) Iteration: 1000 / 2000 [ 50%] (Warmup) Iteration: 1001 / 2000 [ 50%] (Sampling) Iteration: 1200 / 2000 [ 60%] (Sampling) Iteration: 1400 / 2000 [ 70%] (Sampling) Iteration: 1600 / 2000 [ 80%] (Sampling) Iteration: 1800 / 2000 [ 90%] (Sampling) Iteration: 2000 / 2000 [100%] (Sampling) Elapsed Time: 0.031963 seconds (Warm-up) 0.045802 seconds (Sampling) 0.077765 seconds (Total) SAMPLING FOR MODEL 'continuous' NOW (CHAIN 3). Gradient evaluation took 1.3e-05 seconds 1000 transitions using 10 leapfrog steps per transition would take 0.13 seconds. Adjust your expectations accordingly! Iteration: 1 / 2000 [ 0%] (Warmup) Iteration: 200 / 2000 [ 10%] (Warmup) Iteration: 400 / 2000 [ 20%] (Warmup) Iteration: 600 / 2000 [ 30%] (Warmup) Iteration: 800 / 2000 [ 40%] (Warmup) Iteration: 1000 / 2000 [ 50%] (Warmup) Iteration: 1001 / 2000 [ 50%] (Sampling) Iteration: 1200 / 2000 [ 60%] (Sampling) Iteration: 1400 / 2000 [ 70%] (Sampling) Iteration: 1600 / 2000 [ 80%] (Sampling) Iteration: 1800 / 2000 [ 90%] (Sampling) Iteration: 2000 / 2000 [100%] (Sampling) Elapsed Time: 0.033683 seconds (Warm-up) 0.047593 seconds (Sampling) 0.081276 seconds (Total) SAMPLING FOR MODEL 'continuous' NOW (CHAIN 4). Gradient evaluation took 9e-06 seconds 1000 transitions using 10 leapfrog steps per transition would take 0.09 seconds. Adjust your expectations accordingly! Iteration: 1 / 2000 [ 0%] (Warmup) Iteration: 200 / 2000 [ 10%] (Warmup) Iteration: 400 / 2000 [ 20%] (Warmup) Iteration: 600 / 2000 [ 30%] (Warmup) Iteration: 800 / 2000 [ 40%] (Warmup) Iteration: 1000 / 2000 [ 50%] (Warmup) Iteration: 1001 / 2000 [ 50%] (Sampling) Iteration: 1200 / 2000 [ 60%] (Sampling) Iteration: 1400 / 2000 [ 70%] (Sampling) Iteration: 1600 / 2000 [ 80%] (Sampling) Iteration: 1800 / 2000 [ 90%] (Sampling) Iteration: 2000 / 2000 [100%] (Sampling) Elapsed Time: 0.035613 seconds (Warm-up) 0.053674 seconds (Sampling) 0.089287 seconds (Total) > > summary(model10, digits = 3) Model Info: function: stan_glm family: gaussian [identity] formula: libido ~ dose algorithm: sampling priors: see help('prior_summary') sample: 4000 (posterior sample size) observations: 15 predictors: 3 Estimates: mean sd 2.5% 25% 50% 75% 97.5% (Intercept) 2.273 0.681 0.933 1.842 2.271 2.695 3.610 doseLow Dose 0.909 0.954 -0.991 0.302 0.906 1.513 2.827 doseHigh Dose 2.666 0.926 0.778 2.089 2.681 3.264 4.424 sigma 1.496 0.329 1.004 1.263 1.442 1.662 2.296 mean_PPD 3.456 0.568 2.306 3.112 3.460 3.814 4.572 log-posterior -33.898 1.624 -37.995 -34.698 -33.534 -32.699 -31.893 Diagnostics: mcse Rhat n_eff (Intercept) 0.013 1.001 2624 doseLow Dose 0.018 1.001 2728 doseHigh Dose 0.018 1.000 2525 sigma 0.007 1.003 2319 mean_PPD 0.009 1.000 3754 log-posterior 0.047 1.003 1218 For each parameter, mcse is Monte Carlo standard error, n_eff is a crude measure of effective sample size, and Rhat is the potential scale reduction factor on split chains (at convergence Rhat=1). > plot(model10) > > >