Federal, state, and local public health surveillance systems often collect residential address information as part of their surveillance activities. Prior to spatial statistical analyses, residential address information must be geocoded (e.g., latitude and longitude coordinates), a process typically accomplished through the use of electronic street databases 12. Public health applications of geocoded data include defining a study population, linking health outcomes with environmental hazards, and investigating disease clusters. Although the hope is that all geocodes correctly reflect the true geographic location of the addresses, some geocodes are likely inaccurate due to errors in street databases, errors in residential address information, algorithms that permit imperfect address matches (i.e., the "match rate," or how similar the submitted address must be to the address in the database), and the distance geocodes are placed from the street centerline 2345. There is generally a trade-off between the proportion of missing geocodes and geocode accuracy; lenient match rates tend to increase the proportion of successfully geocoded addresses at the expense of geocode accuracy 6. In addition to carefully collected residential address information, street databases that are current, free of errors, and spatially accurate should help reduce location error 2.

Because geocode inaccuracies can affect spatial analyses, 7 understanding the magnitude of location error in geocoded data is desirable. One approach is to travel to the address location and verify coordinates using a global positioning system (GPS) 128910. Although this approach is accurate it is also resource-intensive, particularly when the geographic area of interest is large. Whereas a GPS may be the only viable option for geocoding in remote settings [e.g., 11], in the U.S. alternative options for geocode validation are generally available, and those overseeing surveillance systems may not wish to, or have the resources to, verify large numbers of addresses using a GPS. In this paper we describe an alternative computer-based method 5 to verify address locations for a sample of birth defect records in metropolitan Atlanta.

The commercial vendor and OHIP created address-level geocodes for 96.0% and 91.7% of the sample, respectively (Table 1). Although 435 addresses included in the sample were located (72.6%), gold standard points were placed for only 376 addresses (62.8%). Points were not placed for 59 addresses (9.8%) because the parcels contained large, multi-unit housing complexes.

Selected percentiles from the empirical cumulative distributions of location error, stratified by county, are presented for both OHIP and commercial vendor geocodes in Table 2. Median location error was 71 meters for the commercial vendor and 91 meters for OHIP geocodes. Median location error was approximately 35% greater in Gwinnett County than in Fulton County. This finding was anticipated, as Gwinnett County is predominantly suburban whereas Fulton County has a mix of urban and suburban areas. Rose plots (Figure 2) were constructed by placing each record into one of 24 15° bins according to the angle of the location error. Inspection of the rose plots did not suggest systematic bias in the direction of the geocode relative to the gold standard. There was no strong evidence to reject the null hypothesis of a uniform circular distribution. Rayleigh tests, which were performed for each geocoding agency (data pooled over counties) as well as for each combination of county and geocoding agency, were not significant (all p-values > 0.2).

The magnitude of location error reported in Table 2 occasionally causes geocodes to be placed into incorrect census tracts. The point-in-polygon simulations for Fulton County using the commercial vendor location error caused 4.5% (4.0%, 5.0%) of the randomly generated geocodes to be placed into incorrect census tracts. OHIP location error caused incorrect census tract assignment for 5.3% (4.8%, 5.9%) of the geocodes in Fulton County. Results were similar for Gwinnett County; 4.8% (4.3%, 5.4%) of geocodes were placed into incorrect census tracts because of commercial vendor location error, and OHIP location error caused 5.2% (4.7%, 5.8%) of geocodes to be assigned the incorrect census tract.

Location error is intrinsic to both the OHIP and commercial vendor geocodes. Interpolation is likely a major component of this error, as interpolation is necessary whenever the submitted address is not demarcated on the street map. Because only a small proportion of addresses are demarcated, interpolation occurs frequently. Observed differences in geocoding success and in location error magnitude between the commercial vendor and OHIP geocodes may be due to a number of factors, including the quality of the street database, the correctness of the submitted addresses, the ability of the software to match submitted addresses with addresses in the database (e.g., recognize that "Cir" is short for "Circle"), the tolerance for geocoding imperfect matches (i.e., the match rate), and the methodology used to geocode addresses that were not geocoded in batch mode. Although we were unable to quantify the relative contribution of each of these factors, it is likely that much of the difference in the percentage of addresses successfully geocoded is attributable to the manual address matching performed by the commercial vendor.

Although the aim of our study was to estimate the distributions of geocode location error, there are additional errors to consider when analyzing geocoded address data. The addresses unsuccessfully geocoded by the commercial vendor and/or OHIP (and therefore excluded from analyses) may result in selection bias. If the probability that a geocode is missing is differential across space then this can bias the relationship between spatially-varying covariates and disease incidence 1718. Furthermore, a high proportion of successfully geocoded addresses, although reassuring, does not preclude this selection bias. An additional source of error arises when spurious geocodes are created from fictitious addresses. This can occur when an imperfectly recorded address happens to fall within a range of viable addresses.

Our study design excluded discordant addresses in the linked MACDP and OHIP database. This approach ensured that comparisons between the two sets of geocodes were fair and presumably reduced the number of low quality addresses that were validated (an address that is identical in both the MACDP and OHIP databases is likely to be correct). This design, however, may have underestimated the true distribution of geocode location error for both commercial vendor and OHIP geocodes. Many imperfect addresses, which were excluded from our study (because the MACDP and OHIP addresses were discordant), were nevertheless geocoded to the address-level by both agencies. The distribution of geocode location error for these addresses may be large relative to the distribution for the set of addresses selected for validation. Additionally, the 164 addresses selected for validation that we were unable to locate in the tax parcels (Table 1) frequently had address-level geocodes. Although it is probable that many of these addresses correspond to apartment complex roads not delineated in the shapefile, the true distribution of geocode location error for these geocodes may be larger than the distributions presented in Table 2.

We used a simulation-based approach to evaluate the potential impact of location error on census tract assignment. The percentage of geocodes displaced into incorrect census tracts was similar for Fulton County and Gwinnett County, even though median location error was approximately 35% larger in Gwinnett County (Table 2). This finding is likely due to census tract size (tracts tend to be larger in the predominantly suburban Gwinnett County). A larger location error is needed to displace a geocode outside of its original census tract in Gwinnett.

Presumably, the ramifications of geocode location error will vary depending upon the study design. For example, in air pollution epidemiology, designs using ambient air quality monitors to assign pollution levels to cohort members [e.g., 19] may not be strongly impacted by the magnitude of geocode location error reported in Table 2, whereas fine-scale studies of traffic proximity [e.g., 20] may be strongly impacted by this magnitude of location error. In both of these settings, however, an empirical estimate of geocode location error, specific to the local setting in which the study was conducted, can be used to formally evaluate the consequences of this source of measurement error on epidemiologic results. Tenuous speculations about measurement error can be replaced with inferences from more rigorous statistical approaches.

Relative to street databases, tax parcel shapefiles offer two main advantages: 1) fictitious addresses that happen to fall within ranges of legitimate addresses are not geocoded, and 2) there is no need to interpolate between address ranges. Tax parcels, however, have certain disadvantages as well. They are created to assist the county tax assessor rather than to geocode addresses. Accordingly, an apartment complex encompassing numerous roads appears as one polygon because this is the taxable parcel. We were unable to locate 164 addresses (Table 1), and it is probable that many of these addresses correspond to apartment complex roads not delineated in the shapefile. An additional disadvantage is the limited availability of tax parcel shapefiles (as of June 2005 only two of the five counties covered by MACDP had tax parcel shapefiles). Some GIS software packages offer capabilities for batch parcel geocoding (e.g., the "One Field" style of locator in ArcGIS); as tax parcel shapefiles become increasingly available parcel-based geocoding may become more feasible. Building "footprints," where polygons in the shapefile correspond to building dimensions, also offer possibilities for geocoding and geocode validation.

Cost is also an important consideration – whereas 25–30 addresses per hour can be validated using tax parcels, online commercial geocoding services offer near real-time geocoding for less than two cents per address. Tax parcels are therefore not a viable alternative to batch street database geocoding. Tax parcel validation, however, is more efficient than GPS field validation. Although the number of addresses per hour that can be field validated will vary greatly depending upon address proximity, it would be nearly impossible to field validate 25 addresses per hour in Atlanta. Past experiences at MACDP suggest rates of 5–10 addresses per hour are more typical.

Applications of tax parcel datasets in environmental health extend beyond geocode generation and validation. For example, tax parcel datasets and housing characteristics have been combined to identify high priority regions of lead poisoning risk 21. The 2004 Olympic and Para Olympic environmental health inspection program 22 utilized tax parcel data in their GIS applications. Tax parcel datasets are also routinely used in urban planning, and some investigators have used tax parcels to model the environmental impact of urban development 23.

Geocode uncertainty can be quantified using tax parcel datasets, high resolution orthoimagery, and GIS. In metropolitan Atlanta, the median geocode location error was less than 100 meters for both the OHIP and commercial vendor geocodes, and there was no evidence of systematic bias in the angle of the location error. Geocode location error caused approximately 5% of the randomly generated geocodes to be placed into the incorrect census tract. We contend that the motivation for understanding the distribution of geocode location error parallels the motivation for assessing disease misclassification or exposure measurement error in epidemiological studies. Geocodes have an important role in environmental health research and surveillance, as they are frequently used to define the study population and to link health data with environmental hazards. Furthermore, many spatial statistical methods use geocodes, and the validity of these approaches may be compromised by location error. Further work is needed to evaluate the impact of location error on statistical methods and surveillance applications.

The author(s) declare that they have no competing interests.

MJS conducted the addresses validation and drafted the manuscript. CS helped design the study, draft the manuscript, and is responsible for the maintenance of MACDP geocoded data. BRG examined a subset of records to verify whether the addresses fell within the counties and assisted with the creation of Figures. AKB was responsible for obtaining the tax parcel datasets, working with the dataset owners to ensure appropriate use of the datasets, and assisting MJS and CS with GIS activities. AC helped design the study and participated in its coordination. All authors read and approved the final manuscript.