statisticalrethinkin..

5.2. MASKED RELATIONSHIP 147a ~ dnorm( 0 , 100 ) ,bn ~ dnorm( 0 , 1 ) ,sigma ~ dunif( 0 , 10 )) ,data=dcc ,start=list(a=mean(d$kcal.per.g),bn=0,sigma=sd(d$kcal.per.g)) )You might get some NaNs produced warnings, but the model will now fit correctly. Take alook at the quadratic approximate posterior now:precis( m5.5 , digits=3 )R code5.20Mean StdDev 2.5% 97.5%a 0.353 0.471 -0.569 1.276bn 0.005 0.007 -0.009 0.018sigma 0.166 0.028 0.110 0.221First, note that I added digits=3 to the precis call. e reason is that the posterior meanfor bn is very small. So we need more digits to see that it’s not exactly zero—a change from thesmallest neocortex percent in the data, 55%, to the largest, 76%, would result in an expectedchange of only:coef(m5.5)["bn"] * ( 76 - 55 )R code5.210.09456748at’s less than 0.1 kilocalories. e kilocalories in the data range from less than 0.5 to morethan 0.9 per gram, so this effect isn’t so impressive. More importantly, it isn’t very precise.e 95% interval of the parameter extends a good distance on both sides of zero. You canplot the predicted mean and 95% interval for the mean to see this more easily:np.seq