A1968
Title: Bayesian integration of survey data sources: Costs and quality
Authors: Joseph Sakshaug - (Germany) [presenting]
Abstract: Non-probability web surveys (NPS) have become increasingly popular as they are convenient and inexpensive. At the same time, traditional probability-based surveys (PS) are experiencing steadily declining response rates that drive up costs. Combining the two approaches to offset their respective weaknesses has become one of the central challenges in survey research. A new methodology is presented that fuses probability and non-probability samples to produce more accurate analytic inferences. The method uses a Bayesian framework that optimally weights the two data sources according to estimated measures of bias. The approach is evaluated through extensive simulations and a real-data application that pairs a large probability-based survey with several parallel non-probability web surveys fielded by different vendors, each representing distinct selection mechanisms. Results show that the combined estimator consistently achieves lower mean squared error (MSE) than either source used alone, and in the worst case produces no increase in MSE. A cost analysis further demonstrates that the approach can substantially reduce overall survey expenditure, making it especially attractive for studies seeking to lower data-collection costs while maintaining or improving data quality.
