IASS Webinar 39: Model-Based Optimal Designs for a Multipurpose Farm Survey

WEBINAR RECORDING

Webinar Abstract 

We consider model-based optimal sampling designs for multipurpose surveys with multiple measures of size. The problem is motivated by surveys conducted by the United States Department of Agriculture’s National Agricultural Statistics Service (NASS), in which estimates of planted and harvested acres of different crops are of interest, and historical acreages are available on the frame as measures of size. We use convex optimisation to find the inclusion probabilities that minimise expected sample size subject to target precision requirements for different study variables, along with other inequality constraints. The precision requirements are computed as anticipated coefficients of variation under models relating study variables to frame measures of size. These same models are used in established NASS strategies for the multipurpose survey problem to obtain Multivariate Probability Proportional to Size (MPPS) inclusion probabilities. MPPS uses the measures of size to determine optimal inclusion probabilities for each model, then maximises over models. This solution is practical but not optimal. We compare the use of the MPPS and optimal inclusion probabilities under different designs (Poisson sampling and balanced sampling) and different estimators (calibrated and uncalibrated) via a Monte Carlo experiment using a simulated population of farms with realistic size and complexity.  

This is joint work with Benjamin Reist (NORC), Ruochen Ma (NORC), and Lu Chen (NISS/NASS).