Probability-based samples backed by theory and literature on best practices are becoming more difficult to implement with higher costs and lower response rates. Meanwhile, “Big Data” (e.g., administrative records, social media data) are more accessible than ever. A two-step hybrid design offers promise by allowing estimation from a combined set of probability and nonprobability samples. First, the weighted probability sample is used to calibrate the nonprobability sample to improve representativeness of the nonprobability sample. Second, the probability and nonprobability samples are blended to reduce sampling and nonsampling errors. Open questions remain on the best practices for both of these steps.
In this course, Berzofsky, the instructor, will briefly discuss the current issues researchers face with probability samples and Big Data; define hybrid or blended sample surveys in the context of a few examples; and, through review of the current literature, discuss the estimation challenges and solutions currently in use.The course will consider the practical implications of hybrid sampling on survey estimates including sampling precision, nonresponse bias, coverage bias, measurement error, processing error and specification errors in the integration process. The instructor will use examples from his research to help illustrate and demonstrate each of the methods discussed in the course.
Marcus Berzofsky, DrPH
Senior Research Statistician