Correlates of Representation Errors in Internet Data Sources for Real Estate Market

Abstract

New data sources, namely big data and the Internet, have become an important issue in statistics and for official statistics in particular. However, before these sources can be used for statistics, it is necessary to conduct a thorough analysis of sources of nonrepresentativeness.

In the article, we focus on detecting correlates of the selection mechanism that underlies Internet data sources for the secondary real estate market in Poland and results in representation errors (frame and selection errors). In order to identify characteristics of properties offered online we link data collected from the two largest advertisements services in Poland and the Register of Real Estate Prices and Values, which covers all transactions made in Poland. Quarterly data for 2016 were linked at a domain level defined by local administrative units (LAU1), the urban/rural distinction and usable floor area (UFA), categorized into four groups. To identify correlates of representation error we used a generalized additive mixed model based on almost 5,500 domains including quarters.

Results indicate that properties not advertised online differ significantly from those shown in the Internet in terms of UFA and location. A non-linear relationship with the average price per m² can be observed, which diminishes after accounting for LAU1 units.