Molina Marhuenda 2015 - sae Small Area Estimation - Overview
Summary
Molina & Marhuenda (2015, The R Journal 7(1): 81-98) describe the
saepackage and, with it, the standard frequentist toolkit for small area estimation (SAE): the Fay-Herriot (FH) area-level model, its spatial and spatio-temporal extensions, the Battese-Harter-Fuller (BHF) unit-level model, and empirical-best estimators for nonlinear targets such as poverty rates. The shared idea is shrinking noisy direct survey estimates toward a regression prediction in proportion to their sampling variance.
Overview
Problem. A survey designed for national or regional figures leaves some areas with tiny samples. A direct estimator from the area’s own data is unbiased but has huge variance. Model-based indirect estimators “borrow strength” through a regression common to all areas, improving efficiency as long as the model holds.
Contribution. A single reference with the model equations (area-level 1-5, SAR(1) spatial 6, AR(1) time, nested-error 7-8, best predictor 9-12), the fitting and MSE methods (Fay-Herriot, ML, REML; Prasad-Rao analytic MSE; parametric and non-parametric bootstrap), and four worked examples (milk expenditure, Tuscany grapes, a synthetic spatio-temporal set, Iowa corn/soybean).
Key result. FH EBLUPs track direct estimates but are less volatile, and cut the coefficient of variation (CV) below the 20% publication limit for all 43 milk areas where several direct CVs exceeded it; gains are largest where samples are smallest.
Motivating PyMC examples
E112 (“How many children in your county are poor? Small-area estimates and a surprise map”: Fay-Herriot from scratch for 3,142 US counties with ACS margins of error, logit classic model, generalised variance functions, censored zeros). The Bayesian version and the map idea are in Molina 2015 - Bayesian Fay-Herriot, Surprise Maps and Variance Functions (Concept Note).
Source Map
| Paper section | Pages | Covered in |
|---|---|---|
| Abstract, package at a glance, Introduction | 81-82 | this note |
| Design vs model-based inference; Notation | 82-83 | Molina 2015 - Fay-Herriot Area-Level Model and EBLUP |
| EBLUPs based on a FH model; fitting methods; MSE; Example 1 (milk) | 83-86 | Molina 2015 - Fay-Herriot Area-Level Model and EBLUP |
| Spatial FH (SAR(1)); Example 2 (grapes) | 86-87 | Molina 2015 - Spatial and Spatio-Temporal Fay-Herriot Models |
| Spatio-temporal FH; Example 3 | 87-89 | Molina 2015 - Spatial and Spatio-Temporal Fay-Herriot Models |
| Unit-level BHF model; Example 4 (Iowa) | 89-91 | Molina 2015 - Unit-Level BHF Model and EB for Nonlinear Parameters |
| EB for nonlinear parameters (poverty), Box-Cox | 91-94 | Molina 2015 - Unit-Level BHF Model and EB for Nonlinear Parameters |
| Other software, summary | 94-98 | this note (list only) and the concept note |
Connections
SAE is partial pooling: the FH shrinkage factor is the classical form of the pooling weight in Hierarchical Models and Partial Pooling as Multiple Comparisons Correction. Poststratifying a model to small cells is the unit-level cousin (Poststratification).
Relevance to marketing measurement
Geo-level or DMA-level response with few observations per cell is the same problem: noisy per-geo lift estimates should be shrunk toward a covariate-driven mean by precision. Known sampling variances play the role of per-geo standard errors in a geo-experiment meta-analysis.