Social Statistics Seminar Series - Prof Michael Elliott
| Dates: | 7 May 2025 |
|---|
| Times: | 15:00 - 16:00 |
|---|
| What is it: | Seminar |
|---|
| Organiser: | School of Social Sciences |
|---|
| Who is it for: | University staff, External researchers, Adults, Alumni, Current University students, General public |
|---|
| Speaker: | Prof Michael Elliott |
|---|
|
|
Please join us for the last Social Statistics seminar of the semester. We have coffee and cake!
Michael Elliott
Professor of Biostatistics
University of Michigan
Title: The Role of Probability Samples in the 21st Century
In this talk I will review the rise, fall, and rise(?) of probability sampling for human populations in the 20th and 21st Centuries, focusing on the problems of quota sampling that lead to the rise of probability sampling as a “gold standard” for population inference, followed by the increasing difficulties that have plagued probability sampling, including increasing cost, declining response rates, and the collapse of convenient sampling frames. At the same time the rise of easily accessible administrative datasets and other forms of “big data” have opened opportunities for research outside the probability sampling paradigm. This had led to a situation where probability samples minimize selection bias but are expensive and may have limited data, and where non-probability sample provide rich data less expensively but may be subject to selection bias. Thus a logical approach is to develop methods that combine information from probability and non-probability samples in an attempt to leverage the strength of each of these approaches. These will include quasi-randomization approaches that use weights to adjust the distribution of common covariates in the non-probability sample to those in the probability sample, post-stratification methods that use prediction models to obtain population-level inference of variables present in only the non-probability sample, and double-robust approaches that combine use of model prediction and weighting and can provide correct inference as long as either the model to account for selection bias in developing the weights or the model used to predict the outcome of interest is correct (but not necessarily both). I will conclude with a discussion about the need for high-quality probability samples for calibration/integration purposes, and briefly touch on the prospects and perils of large language models in the survey setting. I will leave time for discussion about experiences outside the United States, as well as the recent impact of degradation of official statistics in the United States.
Speaker
Prof Michael Elliott
Role: Professor of Biostatistics
Organisation: University of Michigan
Travel and Contact Information
Find event
Room 3.204
Uni Place
