I never got around to polishing my Small Area Estimation (SAE) “101” tutorial materials that I promised a while ago. So here they are, though still unedited and not as clean / self-explanatory as I’d like.
The slides introduce a few variants of the simplest area-level (Fay-Herriot) model, analyzing the same dataset in a few different ways. The slides also explain some basic concepts behind Bayesian inference and MCMC, since the target audience wasn’t expected to be familiar with these topics.
- Part 1: the basic Frequentist area-level model; how to estimate it; model checking (pdf)
- Part 2: overview of Bayes and MCMC; model checking; how to estimate the basic Bayesian area-level model (pdf)
- All slides, data, and code (ZIP)
The code for all the Frequentist analyses is in SAS. There’s R code too, but only for a WinBUGS example of a Bayesian analysis (also repeated in SAS). One day I’ll redo the whole thing in R, but it’s not at the top of the list right now.
- “ByHand” where we compute the Prasad-Rao estimator of the model error variance (just for illustrative purposes since all the steps are explicit and simpler to follow; but not something I’d usually recommend in practice)
- “ProcMixed” where we use mixed modeling to estimate the model error variance at the same time as everything else (a better way to go in practice; but the details get swept up under the hood)
- “ProcMCMC” and “ProcMCMC_alt” where we use SAS to fit essentially the same model parameterized in a few different ways, some of whose chains converge better than others
- “R_WinBUGS” where we do the same but using R to call WinBUGS instead of using SAS
The example data comes from Mukhopadhyay and McDowell, “Small Area Estimation for Survey Data Analysis using SAS Software” [pdf].
If you get the code to run, I’d appreciate hearing that it still works 🙂
My SAE resources page still includes a broader set of tutorials/textbooks/examples.