Assessment of automatic geocoding tools in Temuco, Chile

Abstract

TPS 651: Air pollution exposure modeling 1, Exhibition Hall, Ground floor, August 27, 2019, 3:00 PM - 4:30 PM Automatic geocoding process is carried out in many epidemiological studies. However, the quality of the process has been scarcely evaluated. The objective of this work is to compare the performance of three automatic online geocoding tools using addresses from hand-written hospital files. To do this, from an initial set of transcribed addresses a random sample was geocoded using a gold-standard approach, then the same addresses were geocoded using the three automatic approaches. Analyses were based on a retrospective pregnancy cohort study including 15,500 pregnant women who lived in two cities: Temuco and Padre Las Casas. A sampling design stratified by city was carried out. A sub-sample with 300 random addresses were selected: 200 from Temuco and 100 from Padre Las Casas. The geocoding process of the sub-sample was manually conducted. Every address was located in front of the house using Google Street View. If the address was not found, then a Global Positional System (GPS) receiver was used to obtain coordinates. Then the same addresses were geocoded using the three automatic approaches: Google using Qgis, Google Earth using R and Bing using R. Measures of assessment of performance included completeness, match agreement and positional error. Analyses showed 90% of manual gecoding success. Automated geocoding completeness were 90% for Google, 57% for Bing and 84% for Google Earth. It was observed differences between cities. Some neighborhoods have completeness lower than 75%. Bing was not able to find addresses in Padre Las Casas. Positional error was estimated 11m (SD: 13.9) for Bing; 80m (SD: 446.3) for Google Earth, and 83m (SD: 364.5) for Google. There are differences between cities: Temuco has the lower positional error. Analyses showed that quality of geocoding differs by city and neighborhood. Could be advisable to study locally geocoding performance before to apply it to epidemiological studies.