مقایسه کلان داده یک نظرسنجی Big Data versus a survey

1 Introduction Economists appear to be rapidly shifting much of their research time and attention from work involving surveys to work involving “Big Data.” In the process, there has been some discussion of the advantages and disadvantages of this transition, but little empirical exploration of the trade-offs (Einav and Levin, 2014; Cook, 2014; Sonka, 2014). This analysis will illuminate the discussion by estimating parallel models using data from the carefully designed, long-established Survey of Consumer Finances (SCF) and a sample from one of the oldest, most carefully maintained big-data data sets, the Equifax consumer credit records. The credit record sample is formally known as the Federal Reserve Bank of New York Consumer Credit Panel/Equifax (CCP). I estimate models of household debt using variables contained in both data sets as well as models with census-tract aggregate demographic data incorporated into the credit records. The SCF collects its own demographic measures. To maximize the chances of reaching comparable results, I take several steps to align the coverage and definitions in the two samples. Despite the adjustments, the corresponding coefficients in the models range from similar in magnitude and sign to starkly dissimilar. This example illustrates that while big data can offer frequencies and measures that surveys cannot match, we must be cautious about treating big data as a direct substitute for a carefully designed survey. If policy recommendations will hinge on the magnitude of parameters that econometricians estimate, then research based on big data could point in a different direction than research based on surveys. Although the term “Big Data” has been applied in a variety of situations, some concepts are commonly associated with it. Big data is the byproduct of our daily activities. It has arisen with the automation of nearly all economic transactions and a major portion of our personal interactions. Every communication, payment, and trip is facilitated by computer systems and recorded. The resulting big data sets are updated frequently or continuously and their observation and variable counts are orders of magnitude larger than the surveys that researchers are accustomed to working with. The enormous size and complexity of big data sets in most cases requires that the data be stored in relational databases on multiple servers and accessed with structured query language (SQL) (Varian, 2014). In contrast, most survey and administrative data sets can be stored in a “flat” or “rectangular” format on a single personal computer and processed with any statistical software.