In n8thangreen/STIecoPredict: Multilevel Regression With Post-Stratification Using Natsal Data

In order to later-on do the mulit-level regression and poststratification (MRP) we need to have a joint distribution of the covariates for the population we are interested in. Unfortunately, most data are available as 1, 2 or maybe 3-way tables but the complete cross-classification is not available. What we do here is estimate a joint cross-classification for the risk factors of interest by combining several different data sets together. The principle is that they are, in the main, conditionally independent given the age and sex of an individual so we can kind-of use the age and sex distributions at the local level to link together the different data sets. This isn't perfect for for all intents and purposed its probably ok.

This is a special case of what Jackson et al does in his hierarchical related regression work (see the ecoreg package).

I also scale each LA's risk factor values by the relative overall value when compared to the national average.

I've written 2 separate function but they are based on the same idea and code. The first locallevel_pop_sim generates individual samples of a population with the required joint distribution. This can be pretty slow to run.

sim_indiv <- locallevel_pop_sim(NATSAL.dat, n=4)
head(sim_indiv[[3]])

save(sim_indiv, file="./data/sim_indiv.RData")

However, in the case of MRP we are only really interested in the joint distribution, in order to do the post-stratification step, so we end up just aggregating the individual level data anyway. Better to just estimate the joint probabilities in the first place and so thats what I do with locallevel_pop_props.

sim_prop <- locallevel_pop_props()
##TODO## LA~age+sex+smokenow)
head(sim_prop[[1]], 100)

save(sim_prop, file="./data/sim_prop.RData")

n8thangreen/STIecoPredict documentation built on June 7, 2020, 12:50 p.m.