Leveraging Predictive Modelling from Multiple Sources of Big Data to Improve Sample Efficiency and Reduce Survey Nonresponse Error

Abstract

Big data has been fruitfully leveraged as a supplement for survey data—and sometimes as its replacement—and in the best of worlds, as a “force multiplier” to improve survey analytics and insight. We detail a use case, the big data classifier (BDC), as a replacement to the more traditional methods of targeting households in survey sampling for given specific household and personal attributes. Much like geographic targeting and the use of commercial vendor flags, we detail the ability of BDCs to predict the likelihood that any given household is, for example, one that contains a child or someone who is Hispanic. We specifically build 15 BDCs with the combined data from a large nationally representative probability-based panel and a range of big data from public and private sources, and then assess the effectiveness of these BDCs to successfully predict their range of predicted attributes across three large survey datasets. For each BDC and each data application, we compare the relative effectiveness of the BDCs against historical sample targeting techniques of geographic clustering and vendor flags. Overall, BDCs offer a modest improvement in their ability to target subpopulations. We find classes of predictions that are consistently more effective, and others where the BDCs are on par with vendor flagging, though always superior to geographic clustering. We present some of the relative strengths and weaknesses of BDCs as a new method to identify and subsequently sample low incidence and other populations.