kpop: a kernel balancing approach for reducing specification assumptions in survey weighting

Area of Science:

• Survey Methodology
• Statistical Modeling
• Data Science

Background:

• Declining survey response rates lead to nonrepresentative samples.
• Traditional weighting methods rely on assumptions about variable relationships.
• Expert knowledge in variable selection does not guarantee accurate functional forms.

Purpose of the Study:

• Introduce kernel balancing for population weighting (kpop) as an alternative to traditional calibration weights.
• Alleviate functional form dependency in survey weighting.
• Improve representativeness of sample data without explicit functional form specification.

Main Methods:

• Replace the design matrix X with a kernel matrix K encoding high-order information.
• Calculate weights to match the weighted average of the kernel matrix in the sample to the population.
• Utilize kernel balancing for population weighting (kpop) for robust calibration.

Main Results:

• Kernel balancing achieves good calibration on a wide range of smooth functions of X.
• The kpop method reduces reliance on user-specified functional forms.
• Demonstrated effectiveness using 2016 US presidential election polling data.

Conclusions:

• Kernel balancing for population weighting (kpop) offers a more flexible and robust approach to survey weighting.
• This method enhances the representativeness of nonprobability samples.
• kpop is a valuable tool for researchers facing declining response rates.