Materials for Short Course

Location: Glasgow Caledonian University, Annie Lennox Building Lantern Room

Date: Mon, Jun 17, 2024

Time: 1:30 PM - 5:30 PM BST

Instructors

• Wesley Yung, Statistics Canada
• Kelly McConville, Harvard University

Description

National statistical offices are actively integrating data science methods into many aspects of producing official statistics, including data collection, editing and imputation, and survey estimation. As technological and statistical advances provide new data sources and modeling techniques, procedures must adapt to accommodate them. This short course will provide participants with examples of how data science methods are used in producing official statistics and the challenges faced. It will also introduce participants to a model-assisted approach for incorporating machine learning into survey estimation. The machine learning models will include generalized linear models, regularized (elastic net) regression, and regression trees. The course will also include demonstrations of how to fit these estimators using the statistical software R. R Markdown files with the relevant code will be provided so participants can actively follow along with the demonstrations. Prior R experience is encouraged but not required.

Course Objectives

• To be able to identify scenarios where data science methods would enhance a project in official statistics.
• To become familiar with several common machine learning models.
• To be able to compute model-assisted estimators in R.

Outline

• 1:30 - 1:45pm: Introductions
• 1:45 - 3:00pm: Data science in official statistics: Opportunities and challenges
• 3:00 - 3:30pm: Snack Break
• 3:30 - 4:30pm: Model-assisted estimation via machine learning models
• 4:30 - 5:00pm: Survey estimation in R
• 5:00 - 5:30pm: Recap

Slide Decks & R Code

• Data Science Opportunities and Challenges
• Model-Assisted Estimation Case Studies
• Applications in R
  • Quarto file of R Code

R/RStudio Instructions

During the workshop, we will cover examples of how to fit model-assisted survey estimators in the statistical software, R. You can optionally follow along by running these examples in R as we cover them. If you’d like to do this, we recommend either installing R (install page: https://cran.r-project.org/) and RStudio (install page: https://posit.co/download/rstudio-desktop/) locally on your computer or setting up a free account on https://posit.cloud/, a cloud-based RStudio server. Then you will want to install the following R packages: mase, recipes, survey, tidyverse.

Further Readings and Resources

• Model-Assisted Survey Estimation with Modern Prediction Techniques by Jay Breidt and Jean Opsomer

• Model Assisted Survey Sampling by Carl-Erik Särndal, Bengt Swensson, and Jan Wretman

Bonus Content

• Kelly’s slides for her Wed talk “Random Forest Models for Complex Survey Data”