ABSTRACT
In this chapter we estimate and forecast infant mortality rates for counties of Sweden.
We first examine the data. Since the number of infant deaths is low, random variation dominates the data for the smallest counties, and is visible even for the most populous county.
We then step through the components of a Bayesian model: binomial likelihood, a normal prior for the rate on the logit scale, whose mean is determined by an intercept, a region effect, and a time effect, an exchangeable prior for the region effect, a local trend model for the time effect, a weak prior for the intercept, and weakly informative half-t priors for the standard deviations.
We scrutinize results of the model, including those for the rates, and those for the other unknown quantities such as the intercept, region effects, and time effects. We do model checking with the replicate data technique. We infer the probability that the underlying infant mortality rate in each county in each year is less than 2.5 per thousand. We forecast underlying infant mortality rates by county, and discuss alternative stronger priors for standard deviation terms that lead to narrower credible intervals for forecasts.
