Abstract:
|
This paper assesses weighting methods for a national panel sample data selected with non-probability sampling from a mobile panel. In addition to a raking method, we use a propensity model method which relies on a large probability sample as a reference sample. The non-probability sample for the study, the Child Immunization Panel Survey (ChIMPS), was selected using a very large mobile panel which is representative of the US population along some key demographics. The probability sample is the 2014 National Immunization Survey (NIS), a large probability sample of US households with children in the target age range. We match on demographics, geography and telephone characteristics, and use a propensity model for the combined sample. The logistic regression models include the survey weights for the NIS; weight adjustments are computed as the inverse of propensity scores. We compare weighted estimates of immunization outcomes with those computed with raking for the ChIMPS data. The comparisons also look at variances and weight variability. We address the challenge that the NIS is already a dual frame RDD sample where the cell component is more comparable with the ChIMPS sample.
|