Spatial and Spatio-temporal Epidemiology 3 (2012) 17–29

Contents lists available at SciVerse ScienceDirect

Spatial and Spatio-temporal Epidemiology

journal homepage: www.elsevier .com/locate /sste

Investigating impacts of positional error on potential healthcare accessibility

Scott Bell a,⇑, Kathi Wilson b,1, Tayyab Ikram Shah a,2, Sarina Gersher a,3, Tina Elliott a,3

a Department of Geography and Planning, University of Saskatchewan Saskatoon, Canada SK S7M 5C8b Department of Geography, University of Toronto Mississauga Mississauga, Canada ON L5L 1C6

a r t i c l e i n f o a b s t r a c t

Article history:Available online 23 February 2012

Keywords:GISGeocodingPositional errorHealth care accessibility

1877-5845/$ - see front matter � 2012 Elsevier Ltddoi:10.1016/j.sste.2012.02.003

⇑ Corresponding author. Tel.: +1 306 966 5676; faE-mail addresses: smb332@campus.usask.ca (S.

utoronto.ca (K. Wilson), tayyab.shah@usask.ca (T.usask.ca (S. Gersher), tme021@mail.usask.ca (T. Elli

1 Tel.: +1 905 828 3864; fax: +1 905 828 5273.2 Tel.: +1 306 341 4145; fax: +1 306 966 5680.3 Tel.: +1 306 966 5133; fax: +1 306 966 5680.

Accessibility to health services at the local or community level is an effective approach tomeasuring health care delivery in various constituencies in Canada and the United States.GIS and spatial methods play an important role in measuring potential access to health ser-vices. The Three-Step Floating Catchment Area (3SFCA) method is a GIS based proceduredeveloped to calculate potential (spatial) accessibility as a ratio of primary health care(PHC) providers to the surrounding population in urban settings. This method uses PHCprovider locations in textual/address format supplied by local, regional, or national healthauthorities. An automated geocoding procedure is normally used to convert such addressesto a pair of geographic coordinates. The accuracy of geocoding depends on the type of ref-erence data and the amount of value-added effort applied. This research investigates thesuccess and accuracy of six geocoding methods as well as how geocoding error affectsthe 3SFCA method. ArcGIS software is used for geocoding and spatial accessibility estima-tion. Results will focus on two implications of geocoding: (1) the success and accuracy ofdifferent automated and value-added geocoding; and (2) the implications of these geocod-ing methods for GIS-based methods that generalise results based on location data.

� 2012 Elsevier Ltd. All rights reserved.

1. Introduction

Accessibility to health services at local and communityscales is an important metric for measuring health caredelivery in Canada and the United States. The concept ofaccess to health care is multifaceted; it builds links be-tween populations at risk (clients) and the delivery system(service providers) which vary across both space and place(Penchansky and Thomas, 1981). In measuring potentialaccess to health services Geographical Information Sys-

. All rights reserved.

x: +1 306 966 5680.Bell), kathi.wilson@

Shah), sag221@mail.ott).

tems (GIS) and spatial methods provide powerful analytictools. The Three-Step Floating Catchment Area (3SFCA)method is a GIS-based procedure developed by Bell (forth-coming) to calculate potential (spatial) accessibility at theneighbourhood level as a ratio of primary health care(PHC) providers to population in urban settings.

Like other GIS based methods, measuring potential(spatial) access to health care requires locations of PrimaryHealth Care (PHC) providers in global absolute geographiccoordinates (Latitude/Longitude, Universal TransverseMercator (UTM), etc.) and population information associ-ated with enumeration areas (census areas or local neigh-bourhoods) (Bell et al., forthcoming; Luo, 2004; Luo andWang, 2003; McGrail and Humphreys, 2009; Paez et al.,2010; Schuurman and BÉRubÉ, 2010). In Canada, censusbased population data is gathered by Statistics Canadaevery five years and is available at a variety of enumerationlevels. One such enumeration unit, and the unit used in

18 S. Bell et al. / Spatial and Spatio-temporal Epidemiology 3 (2012) 17–29

this study, is the dissemination area (DA) which providesgood spatial resolution.

Geocoding is increasingly used in health studies to mapsites of service providers and participants. There are manyaspects of geocoding that require attention in order to en-sure a sufficient match rate to generate location data thatis reliable. For instance, the positional accuracy of geocod-ed locations depends on the geocoding techniques em-ployed and reference data used. In general, an automatedaddress match geocoding procedure is used to converteach address to a pair of geographical coordinates. Theautomated part of the geocoding process can be accom-plished in several ways, these include range interpolation,areal unit interpolation, and rooftop geocoding; in this re-search we use range interpolation as the primary auto-mated method. Range interpolation involves interpolatingaddress locations along a street segment; such estimationsof position can introduce positional error in the geocodedpoint (Goldberg, 2008; Zandbergen, 2009). This is tradi-tionally followed by a manual intervention/interactivegeocoding process that involves examining possible loca-tion candidates for an address from a digital street file(Goldberg et al., 2008). In this research, the term value-added refers to those manual interventions or measures ta-ken after the initial automated geocoding process; suchmethods are generally accepted as increasing the validityand accuracy of the output. The primary objective of thisresearch is to investigate the positional error that resultsfrom geocoding PHC provider addresses using six differentgeocoding methods. These methods include: (1) auto-mated Postal Code geocoding with Desktop Mapping Tech-nologies Inc. (DMTI) data; (2) value-added matching withPostal Code data; (3) automated range interpolation usingDMTI street data; (4) value-added matching with DMTIdata; (5) automated range interpolation using ESRI Tele At-las street data (bundled with ArcGIS 10); and (6) value-added matching with ESRI Tele Atlas/street data. The sec-ondary objective of our study is to investigate the impactof positional error on measures of accessibility estimatedusing the three-Step Floating Catchment Area (3SFCA)method (Bell et al., forthcoming).

2. Background

Geocoding has different meanings depending on itsapplication (Goldberg et al., 2007). In health research, geo-coding is used as a means of transforming textual geo-graphic descriptions into explicitly georeferenced datathat can be used for spatial analyses (Goldberg, 2008). Geo-coding has an established role in health research, whetherestimating supply and demand of various health services(Schuurman and BÉRubÉ, 2010) or dealing with diseasepatterns and distribution (Bruneau et al., 2008). Addressmatch geocoding is the process of transforming addressesin local and relative coordinate systems (such as street ad-dresses or postal codes), which are not themselves amena-ble to GIS-based spatial analysis, into a format whichassigns coordinates in an global absolute coordinate sys-tem (such as latitude and longitude or UTM). Geocodingis largely an automated process and is at least partially

dependent on the quality of the reference data used to esti-mate an addresses location (Zandbergen, 2009, 2011);therefore, uncertainty exists in all geocoding output. Inthe context of health research, there are important impli-cations of such errors. What is little understood is the ex-tent to which such geocoding errors manifest themselvesin higher order applications of the geocoded results.

2.1. Common geocoding errors

There are four main types of error related to addressmatch geocoding (Zandbergen, 2009). The four types of er-ror include:

1. Error arising from geocoding an incorrect address. Anaddress can be incorrect in a variety of ways, two com-mons errors include: (1) a typo in the number compo-nent of the address, and (2) an error in the streetdesignation (wrong designation or an incorrect inter-preted abbreviation).

2. Errors related to inaccurate interpolation along a streetsegment. Since it is rare that street addresses are evenlydistributed across a street segment such errors are com-mon but tend to be smaller in urban areas where streetsegments are shorter (Bakshi et al., 2004; Cayo and Tal-bot, 2003); even within urban areas there can be vari-ability in such errors, for instance, in areas dominatedby multi-family housing units, condominiums, or apart-ment buildings (Ward et al., 2005; Zimmerman and Li,2010). Interestingly, commercial areas are also suscep-tible to higher error resulting from this type as theretend to be fewer commercial entities per street segmentthan in residential areas, resulting in sometimes arbi-trary address numbering along the segment (Zandber-gen, 2008).

3. Error resulting from the geocoded point being placed atan incorrect perpendicular distance from the street seg-ment (i.e. incorrect side offsets); again, this type orerror is generally minimal.

4. Error in the placement of the reference data’s street seg-ments within the road network can produce misplacedlocation that are difficult to reconcile without localknowledge; in this type of error the address is correctlylocated along the street segment but the street segmentis not in a location congruous with its position in thestreet network.

2.2. Geocoding with areal units

A secondary geocoding method involves the geocodingof points using areal unit reference data. Examples of suchgeocoding include zip codes, postal codes, and other landallocation systems. While some of the potential errorslisted above can extend to areal unit geocoding, such geo-coding has its own set of challenges. In Canada, postalcodes can either be used to geocode independently (Bru-neau et al., 2008) or to supplement missing or erroneousstreet address information (Schuurman and BÉRubÉ,2010). Here, it is important to describe the difference be-tween Canadian postal codes and US Zone ImprovementProgram (ZIP) Codes. In the United states a ZIP Code

S. Bell et al. / Spatial and Spatio-temporal Epidemiology 3 (2012) 17–29 19

consists of five numerical digits. The first three digits rep-resents the mail sorting and distribution center for an areawhile the final two digits represents the area within a city(in case of metropolitan area) or a village/town (outsidemetro areas).In Canada, postal code consists of six alpha-numeric characters, divided into two parts. The first threerepresent the Forward Sortation Area (FSA) while the sec-ond three indicate the Local Delivery Unit (LDU) (CanadaPost, 2011). The FSA, recorded in number-letter-numberorder, represents the major geographic region in an urbanor a rural location (for example, M9L). The LDU, recorded inletter-number-letter order, indicates the local, or smallest,mail delivery unit (for example, 0J3) (Canada Post, 2011).The LDU boundaries represent the geographic location ofthe six digits postal code. LDU geocoding allows for bothone-to-many and many-to-one address matching. A singleLDU boundary can be associated with multiple postalcodes, such as in the case of a large office or residentialtower within which individual postal codes are associatedwith single or multiple floors; therefore, more than onepostal codes is assigned to the building. On the other hand,more than one LDU area could be associated with a singlepostal code, as in the case of a small town or rural areabeing composed of disjointed parcels; in this case the sin-gle postal code is associated with many potential geo-graphic places. Postal code locators often place the pointat the centroid of the postal code area; in postal code-onlygeocoding accuracy decreases as postal code area increases(Goldberg et al., 2007; Zandbergen, 2009). However, in acomparison between postal code and street address geo-coding 87.9% of postal code locations were within 200 mof the true street address and 96.5% were within 500 m(Bow et al., 2004). In this research, a Multiple EnhancedPostal code (MEP) product is used. The MEP provides mul-tiple points within a single postal code to which addresscan be matched; this matching is dependent on the avail-ability of viable building address information. If addressinformation is missing or incorrect a single postal codecould be associated with multiple points (resulting in tiedresults).

2.3. Value-added geocoding

With both street segment interpolation and areal unitgeocoding there are two components: the initial, auto-mated process (non-interactive) and the interactive man-ual review that we call the value-added step (Goldberg,2008). The term value-added refers to those measures ta-ken after the initial automated geocoding process whichincrease the validity and accuracy of the output (Goldberget al., 2008). Strategies involved in this process includechecking for spelling or format errors in the input data;bringing previous knowledge of the study area to bear onthe unmatched addresses; or investigating individual ad-dresses for anomalies. It may also require choosing oneof two or more tied options for a given address. For exam-ple, in one study interactive rematching was performedafter automated geocoding: addresses that were initiallymatched with a low match score or addresses which weretied were reviewed on a case by case basis (Duncan et al.,2011).

These geocoding methods have been compared in vari-ous ways in the geocoding literature. Some studies havecompared one method of street address geocoding to an-other (Ward et al., 2005), while others have comparedpostal code geocoding methods to street address geocod-ing (Bow et al., 2004). Still others have sought to comparegeocoding results to ‘‘true locations’’ determined via satel-lite imagery and/or GPS data (Cayo and Talbot, 2003; Zhanet al., 2006). Some studies have used commercial geocod-ing firms (Ward et al., 2005), while others have used freeonline geocoding tools (e.g. the use of batchgeocode.comby Duncan el al. (2011)). As indicated earlier, a relativelyunexamined aspect of geocoding in the Canadian context,particularly in urban areas is the implication of differentgeocoding methods and different accuracy rates on analy-sis that uses the output from the geocoding process.

2.4. Accessibility and the 3 step floating catchment areamethod

Recent research supports the notion that PHC accessi-bility and availability directly impact health; for instance,distance to PHC is inversely related to both utilization ofservices and area-based equity in health status (Hiscocket al., 2008; Korda et al., 2007).Although access is multivar-iate, an important component is the capacity to obtain ser-vices whenever needed (Humphreys and Smith, 2009).Two distinct approaches to health care accessibility haveemerged: the first is revealed accessibility (the actual useof health care services) and the second is potential accessi-bility (characteristics of delivery system and of populationat risk) (Andersen et al., 1983; Joseph and Phillips,1984).Finally, each approach is further distinguished aseither spatial (focused on factors such as location or dis-tance) or non-spatial (focused on factors such as age, sex,income etc.) (Khan, 1992; Luo, 2004).

As geographic technologies have evolved so too havethe tools used to measure accessibility. Early tools werelimited due to their lack of sensitivity to supply and de-mand (Brabyn and Barnett, 2004; Charreire and Combier,2009; Hiscock et al., 2008; Lovett et al., 2002; Pearceet al., 2006) as well as problems related to the mismatchbetween units of analysis and enumeration areas for whichpopulation data is available (Brabyn and Barnett, 2004;Kindig and Movassaghi, 1989; Rosenthal et al., 2005). Incontrast, the floating catchment area (FCA) method takesinto consideration both of these issues (Luo, 2004). Thismethod moves (or floats) a window across a study area;the density of events within the window becomes repre-sentative of the center point of the window. More recentlythe Two Step Floating Catchment Area (2SFCA) method hasbeen developed to examine potential geographical accessto health care (Bagheri et al., 2005; Luo and Wang, 2003).This method provides more accurate measures of accessi-bility because it uses fluid units of analysis. It does not pro-duce ratios of physicians to population within aneighbourhood, but rather it acknowledges that individu-als may seek care in a neighbourhood other than theirown. The 2SFCA first uses a buffer (or catchment) aroundPHC locations to calculate a provider-to-population ratio.A second buffer is then placed around centroids of

20 S. Bell et al. / Spatial and Spatio-temporal Epidemiology 3 (2012) 17–29

population enumeration areas, and the ratios from all pro-vider points within the second buffer are added together.These two steps provide a more accurate assessment ofcross border health care service provision (as explained inthe following section). While our 3rd step extension of thismethod is relatively straightforward (averages of 2SFCA val-ues are calculated for all enumeration areas falling with aneighbourhood) it supports the ability to use units of analy-sis sensitive to the local environment as well as disentanglesthe interpretation of results from units of analysis that aregenerally set at a national or regional level (Census units,for instance); such units do not reflect the local social, cul-tural, political, economic, or health landscape (Bell et al.,forthcoming). In our application we use neighbourhoodboundaries defined at the local government level; other po-tential units of analysis might include health regions, healthdelivery/service areas, or planning areas within a city.

3. Materials and methods

3.1. Overview

This research examines access to PHC in three Canadiancities: Saskatoon, Saskatchewan; Edmonton, Alberta; and

Fig. 1. Study area map: Areas in grey represent neighbour

Mississauga, Ontario (see Fig. 1). These three cities havebeen selected to compare geocoding results in cities withdifferent neighbourhood structures, size, and situations.

Overall, the research was conducted in two steps: Geo-coding PHC locations and using these locations in the3SFCA. The first step consisted of creating the following se-ven layers for each city:

1. A layer of physicians’ verified locations (VL) created withDMTI based automated geocoding followed by extensivepost-geocoding address validation. Validation was doneusing a suite of Google products (Google Earth, Google EarthPro, and Street View). Validation included a visual search inStreet View for actual clinic and office locations where pos-sible. The purpose of theVL method was toprovide abase forcalculating accuracy of different geocoding methods(Cayo and Talbot, 2003; Schootman et al., 2007);

2. a geocoded layer using only postal code (similar to USzip codes) reference data from Canadian spatial datacompany DMTI with NO interactive matching (onlyautomated geocoding);

3. a geocoded layer using only postal code reference datafrom DMTI with Value-added matching (often calledinteractive or manual matching);

hood boundaries used in the final step of the 3SFCA.

S. Bell et al. / Spatial and Spatio-temporal Epidemiology 3 (2012) 17–29 21

4. a layer using DMTI reference data with both streetaddresses and postal codes, with NO interactive match-ing (only automated geocoding);

5. a layer using DMTI reference data with both streetaddresses and postal codes, with Value-added matching;

6. a layer using ESRI Tele Atlas reference data with streetaddresses, with NO interactive matching (only auto-mated geocoding); and

7. a layer using ESRI Tele Atlas reference data with streetaddresses with Value-added matching.

(Additional detail on the geocoding process used foreach of the above layers can be found below).

3.2. Data sourcing and preparation

All data collection, coding, and analysis were carried outin the Spatial Analysis for Innovation in Health Research(SAFIHR, pronounced Sapphire) lab at the University of Sas-katchewan. The input data consisted of lists of physicianscollected via Provincial Colleges of Physician and Surgeons(see Table 1). The selection of physicians was limited tothose classified as Family Doctors, Family Physicians, GeneralPractitioners or Non-Specialists. Each address file (created inMS Excel) was converted into a common format for effectivegeocoding and manually inspected for errors. This includedremoving addresses outside the city, changing Street to St,and removing most punctuation. Post Office box addressespresent an distinct challenge in geocoding; inclusion ofthese addresses can increase error due to geocoding basedon the postal code centroids while exclusion can introduceselection bias (Zandbergen, 2009); in this context, addresseswhich had only P.O. box information were removed. In Sas-katoon and Mississauga the number of P.O. boxes removedwere negligible; however, the 63 address removed inEdmonton would pose a significant methodological issue ifthe purpose of this study were to compare access to PHCacross cities, or even within Edmonton. In this context, how-ever, the removal is justified to most effectively comparegeocoding methods. Due to the relatively small size of this

Table 1Physician totals for three cities.

City Total physicians Total physician practice locations

Saskatoon 264 70Edmonton 1026 283Mississauga 665 203

Table 2Geocoding data and manual settings.

Geocoding method Min match score (%) Spelling sensitivity (%)

Postal code 85 60Postal code (value-added) 85 60DMTI 85 60DMTI (value-added) 85 60Tele Atlas 85 60Tele Atlas (Value-added) 85 60

a DMTICanMapStreetfile dates: Saskatoon = 2011, Edmonton = 2011, Mississau

dataset such preprocessing was reasonable, in more exten-sive studies (nationwide, larger urban areas, etc.) such datacleaning might not be possible.

The reference data for geocoding included three datasources to generate point locations for PHC providers inthe study area: DMTI Platinum postal code (DMTI Spatial,2010b); CanMapStreetfile (DMTI Spatial, 2010a); and the10.0 North America GeocodeService (ArcGIS Online usingTeleAtlas) provided in ArcGIS 10 software. Both DMTIand TeleAtlas are value-added streetfile products. Bothfiles are based on Statistics Canada street file data createdto support the collection of census data (similar to TIGERfiles in the United States). Each private entity (DMTI andTeleAtlas) performs their own value-added methodsresulting in improved but somewhat different products.Our PHC data was the result of collecting or requesting ad-dress information for doctors practicing in our three studyareas from the respective provincial Colleges of Physiciansand Surgeons. Each College maintains an up-to-date data-base of practicing doctors (doctors are required to be reg-istered with their provincial College and update theirinformation when it changes). For Alberta (Edmonton)and Saskatchewan (Saskatoon) we submitted requests fordata and received an electronic file of all doctors in ourstudy area. For Ontario (Mississauga), data was down-loaded from an up-to-date webpage of doctor addresses.We believe this data represents a complete and accuraterecord of PHC providers in each city.

3.3. The Value-added method

The following steps were applied to both tied and un-matched addresses after automatic geocoding wascompleted:

1. The closest point on the correct side of the street (i.e. Ror L) from the closest street segment was chosen ifstreet segment (From-To) did not exist for that exactaddress. For example, if there was no From-To streetsegment the closest suggested interactive match was

Total P.O. box addresses (removed) Date of data collection

3 200863 20101 2008

Min candidate score (%) Side offset End offset Reference data

40 0 0 DMTI Streetfilesa

40 0 0 DMTI Streetfilesa

40 20 feet 3% DMTI Streetfilesa

40 20 feet 3% DMTI Streetfilesa

40 n/a n/a Tele Atlas 201040 n/a n/a Tele Atlas 2010

ga = 2010.

Fig. 2. Three step floating catchment method; (a) Step 1 – Buffer around PHC Location. In this example 67,352 is the total population of 90 DAs that fallwithin the buffer of a PHC location where two physicians are working; (b) Step 2 – Buffer around around DA centriod. In this example sum the 13 physiciansto population ratios which fall within the DA buffer; (c) Step 3 – Average DA access score within the neighbourhood based on DA centroids.

22 S. Bell et al. / Spatial and Spatio-temporal Epidemiology 3 (2012) 17–29

accepted as the matched location. Note: Tele Atlas doesnot provide detailed information regarding to whichside of the street the addresses belong.

2. Address records were reviewed for human errors andcorrected where possible; the street file data and onlinesources were consulted to check for correct spelling,punctuation, and abbreviations. While this step is sim-ilar to preprocessing described above certain correc-tions are only possible with the additional context ofpartial geocoding results.

3. Tele Atlas associates the Forward Sortation Area (FSA)Characters (the first three characters of a Canadianpostal code) with the address input data; therefore,the FSA was used to confirm a match candidate againstthe actual postal code.

4. When a postal code was returned as tied the first postalcode match candidate was selected. This step is postalcode specific.

5. The address was left unmatched if none of the afore-mentioned steps return results.

3.4. Geocoding methods: 1. verified locations (VL) layer

In order to quantify the success (match rate) and accu-racy (positional error) of each geocoding method we firstestablished a set of verified locations to which each meth-od can be compared. In addition to calculating a MatchRate for each method we also calculated the distance be-tween each geocoded location to the verified location forthat address. Creating the verified locations was time andlabour intensive and is not meant to represent a viablegeocoding option; instead the VL layer was created todetermine the positional accuracy of each subsequentmethod. Initially, physician data was geocoded using Arc-GIS with DMTI streetfiles as the reference data. To ensurethe positional accuracy of the VL layer, wherever possibleeach address was verified by consulting various onlinesources including ArcGIS basemap imagery, and variousGoogle products. Where verification was not possible, thelocation determined by the value-added method wasdeemed sufficient.

Table 4Edmonton match rate (N = 283).

Matched Tied Unmatched

Postal code 247 35 1Postal code with value added 282 0 1DMTI 282 1 0DMTI with value added 283 0 0Tele atlas 278 1 4Tele atlas with value added 282 1 0

Table 5Mississauga match rate (N = 203).

Matched Tied Unmatched

Postal code 247 35 1Postal code with value added 282 0 1DMTI 282 1 0DMTI with value added 283 0 0Tele atlas 278 1 4Tele atlas with value added 282 1 0

S. Bell et al. / Spatial and Spatio-temporal Epidemiology 3 (2012) 17–29 23

3.5. Geocoding methods: 2. postal code layer (DMTI)

The first geocoding method used only Postal Code inputdata and DMTI Platinum postal code suite 2010.3 as thereference data. This data is a postal code product calledMultiple Enhanced Postal Codes (MEP), a precision pointfile that is designed to reduce unmatched results that arethe product of multiple polygons in the reference data hav-ing the same postal code. Settings for this and all subse-quent geocoding methods are outlined in Table 2(below). The address locator was custom built in ArcGISto read postal codes only. Only the automatically gener-ated results with match scores at or above 85% wereaccepted.

3.6. Geocoding methods: 3. postal code layer (method 2 withvalue-added)

This method differs from the previous only in its value-added process. Additional manual steps were taken to in-crease accuracy and produce a higher match rate. Theseadditional steps are described in detail below (see The Va-lue-added Method below).

3.7. Geocoding methods: 4. DMTI composite layer (DMTI andpostal code)

The DMTI Composite method used DMTI Streetfiles asthe reference data. The address locator was custom builtin ArcGIS to utilize city, province, street address, and postalcode to determine a location. Only the automatically gen-erated results with match scores at or above 85% wereaccepted.

3.8. Geocoding methods: 5. DMTI composite layer (method 4with value-added)

This method differs from the previous only in its value-added process. These additional steps were taken to in-crease accuracy and produce a higher match rate. Addi-tional manual steps are described in detail below (seeThe Value-added Method below).

3.9. Geocoding methods: 6. Tele Atlas (ESRI) online layer

This method employed the online geocoding tool fromESRI using Tele Atlas reference data and used the ArcGIS10.0 online North America Geocode Service AddressLocator which supports street address, street name,city/province, and country. Only the automatically gener-

Table 3Saskatoon match rate (N = 70).

Matched Tied Unmatched

Postal code 69 1 0Postal code with value added 70 0 0DMTI 70 0 0DMTI with value added 70 0 0Tele atlas 69 0 1Tele atlas with value added 70 0 0

ated results with match scores at or above 85% wereaccepted.

3.10. Geocoding methods: 7. Tele Atlas (ESRI) online layer(method 6 with value-added)

This method differs from the previous only in its value-added process. Additional manual steps were taken to in-crease accuracy and produce a higher match score. Theseadditional steps are described in detail below (see The Va-lue-added Method below).

3.11. Measuring accessibility to primary health care

In the second analytic step, the 3SFCA method was appliedto all 21 (three cities, 7 geocoding methods) geocoding resultsto produce a ratio of primary health care providers to popula-tion. For this process, the supporting datasets for each city in-cluded a digital geographic neighbourhood boundary file,demographic data of the residents, and a digital geographicfile of the 2006 Canadian Census at the DA level. We chosethe 3SFCA method introduced by Bell (forthcoming) to pro-duce a ratio of PHC providers to population (accessibilityscore) using neighbourhoods as the unit of analysis (In Press).This method requires locations of health care providers, Cen-sus Dissemination Area (DA) centroids, and associated DApopulations, and polygon buffers around each to estimateaccessibility (Bell et al., forthcoming).

There are several steps in the 3SFCA method. First, ser-vice areas of 3 km are calculated around all locations ofPHC providers and census DA centroids using the bufferfunction in ArcGIS (results reported here are based on roadnetwork buffers). In previous and ongoing research wehave employed several buffer sizes (specifically 800 mand 1400 m) that are associated with different health seek-ing behaviours (walking, extended walking/short drive,drive to a nearby/local provider). A 3 km buffer is the larg-est buffer used and is defined as a distance that most peo-ple would drive but is still within a range that one might

Fig. 3. City-level average positional error (in meters) using different geocoding methods.

Table 6City-level results of positional error – Saskatoon.

Positional error(m)

Postal code(N = 70)

Postal code value added(N = 70)

DMTI(N = 70)

DMTI value added(N = 70)

Tele atlas(N = 69)

Tele atlas value added(N = 69)

Mean 91.8 93.4 72.9 72.9 92.5 91.4Median 54.6 59.9 33.7 33.7 63.4 62.7Minimum 0 0 0 0 11.7 11.7Maximum 1103.2 1103.2 1103.2 1103.2 860.4 860.4Standard

deviation91.8 93.4 144.3 144.3 117.8 117.2

Table 7City-level results of positional error - Edmonton.

Positional error(m)

Postal code(N = 282)

Postal code value added(N = 282)

DMTI(N = 282)

DMTI value added(N = 282)

Tele atlas(N = 280)

Tele atlas value added(N = 280)

Mean 249.8 250.5 49.3 49.6 314.9 314.9Median 42.3 43.8 18.5 18.5 51.8 51.8Minimum 0 0 0 0 0.2 0.2Maximum 13360.4 13360.4 2470.8 2570.8 714.8 7143.8Standard

deviation1291.1 1291.7 160.6 166 1126 1126

Table 8City-level results of positional error – Mississauga.

Positional error(m)

Postal code(N = 201)

Postal code value added(N = 201)

DMTI(N = 203)

DMTI value added(N = 203)

Tele atlas(N = 203)

Tele atlas value added(N = 203)

Mean 154.4 160.8 80.3 80.3 107.2 107.2Median 81.9 81.9 56.5 56.5 67.1 67.1Minimum 0 0 0 0 1.3 1.3Maximum 3518.3 3518.3 495.6 495.6 2002.8 2002.8Standard

deviation415.7 424.7 85.9 85.9 179.6 179.6

24 S. Bell et al. / Spatial and Spatio-temporal Epidemiology 3 (2012) 17–29

consider ‘‘local.’’ The second step involves calculating pro-vider-to-population ratios for each service area around thehealth care providers (see Fig. 2a). These ratios aresummed for each DA based on 3 km network buffers

(around DA centroids) and assigned to that DA’s centroid(see Fig. 2b). Our final unit of analysis is the local neigh-bourhood; neighbourhoods are defined by each city(Fig. 1) and have quite unique size and structure depending

Fig. 4. Accessibility Scores for Saskatoon, Edmonton, and Mississauga- Physical (spatial) accessibility to PHC calculated using the following layers of PHCproviders using Verified Locations and Tele Atlas value-added.

S. Bell et al. / Spatial and Spatio-temporal Epidemiology 3 (2012) 17–29 25

on the city in question. In the final step access ratios for allDAs within each neighbourhood were averaged within lo-cal neighbourhoods (Fig. 2c). Averaging produces a rela-tively comparable value that is not sensitive toneighbourhood size (and subsequently more or fewerDAs). All steps were performed using ArcGIS 10 (including

Network Analyst extension) software. All maps were visu-alized using classification cut points from a quintile classi-fication scheme applied to the VL locations. Thisclassification method is well suited to comparing a seriesof choropleth maps; we found that using a single set ofcut points facilitated comparison among applications of

Fig. 5. Scatterplots of 3SFCA accessibility scores – (a) Saskatoon; (b) Edmonton; (c) Mississauga.

26 S. Bell et al. / Spatial and Spatio-temporal Epidemiology 3 (2012) 17–29

the 3SFCA to different geocoding results (Brewer andPickle, 2002). In all cases, darker shading represents higheraccessibility ratios (and hence higher levels of health careaccessibility) and lighter shading represents lower ratios(i.e., lower levels of accessibility).

4. Results

4.1. Match rates

Tables 3–5 (below) show match rates for Saskatoon,Edmonton, and Mississauga. Matched addresses are thosethat returned a minimum match score of 85%. Tied ad-dresses returned a minimum match score of 85% but hadmore than one possible candidate. Once the value-addedmethod was performed, all but one tied record in Edmon-ton was matched. The highest match rate for automatedgeocoding was based on DMTI reference data, at almost100%. The lowest match rate for the automated methodswas from Postal Code (nearly 90%). The highest match ratefor the value-added methods was DMTI at 100% and thePostal Code method had the lowest match rate at just few-er than 100%. These high match rates can be primarilyattributed to the commercial nature of the addresses andthe location of such addresses in well-established sectionsof urban areas. An interesting result, and one unique to theMultiple Enhanced Postal code (MEP) data, is the largenumber of tied records for postal code only geocoding. Un-like ZIP codes in the United States, MEP data provides morethan one location within a postal code to which an address

can be matched. While this supports more accurate geo-coding results when measured as distance from actuallocation, it also results in potentially lower match rates.

4.2. Positional error

Two types of outliers were removed from the data; thefirst was removed before calculating positional errors.These outliers represented address that were erroneouslyincluded in the data and included addresses were not inour study area, there were fewer than 10 of these in all cit-ies combined. The positional errors were calculated and re-corded as Euclidean distances.

Outliers were also removed after positional errors werecalculated. Locations with error greater than 14000 m wereconsidered outliers and were excluded. Such errors are theresult of records not including street address or postal codeinformation. There were fewer than 5 of these in all cities.

Statistical analysis was performed on the positional er-ror results for each geocoding method and city (see Fig. 3for a graphical representation of the error results). Tables6–8 (below) show descriptive statistics associated witheach city and geocoding method. A one-way ANOVA indi-cated no differences among the geocoding methods forSaskatoon. However, following significant ANOVA resultsfor both Edmonton and Mississauga, Tukey Post Hoc anal-ysis identified the following significant differences inEdmonton:DMTI and Tele Atlas (a = 0.019);DMTI and TeleAtlas Value-added (a = 0.019); DMTI Value-added methodand Tele Atlas (a = 0.020); and DMTI Value-added andTeleAtlas Value-added (a = 0.020). In Mississauga there was a

Fig. 6. Change in 3SFCA accessibility score; (8-a, d, g) change between the estimates based on verified locations vs. estimates based on DMTI value-addedlayer in all urban areas; (8-b, e, h) change between the estimates based on verified locations vs. estimates based on postal code added layer in all urbanareas; (8-a, d, e) change between the estimates based on verified locations vs. estimates based on Tele Atlas value-added layer in all urban areas.

S. Bell et al. / Spatial and Spatio-temporal Epidemiology 3 (2012) 17–29 27

significant difference in the average positional error of:DMTI and Postal Code (a = 0.031); DMTI and Postal CodeValue-added (a = 0.031) methods. In all cases DMTI dataproduced more accurate positional geocoding.

The lower accuracy of postal codes is somewhat ex-pected, but a potentially more interesting result is the de-creased accuracy of postal codes with value-addedgeocoding. Recall that for tied postal codes (geocodingcould not automatically select which point with MEP post-al codes to use for geocoding) we selected the first record.The distance results indicate that this value added step in-creases distance error suggesting that these value-addedlocations were more inaccurate than the automaticallygeocoded postal code locations.

4.3. Results of the 3SFCA method

The 3SFCA accessibility scores were calculated usingeach of the value-added geocoding methods (only value-added methods were used to maximize the number ofmatched PHC providers included). Mapped results of allthree methods (plus the VL) are displayed in Figs. 4–6.For all measures higher numbers indicate higher access.In Figs. 4–6, the first map is classified using a quintile clas-sification scheme with five classes. The following threemaps are visualized based on the same cut points of firstmap (to support comparison).

In Saskatoon, access ratios for PHC physicians rangefrom a low of 0.00 per 1000 people in all cases to a high

28 S. Bell et al. / Spatial and Spatio-temporal Epidemiology 3 (2012) 17–29

of 4.30, 4.19, 4.18, 4.12 for the VL, DMTI value-added, Post-al Code value-added and Tele Atlas value-added maps,respectively (see Fig. 4 for a subset of these results). InEdmonton, access ratios for PHC physicians range from alow of 0.00 per 1000 people in all cases to a high of 7.29,7.72, 7.82, 8.07 for the VL, DMTI value-added, Postal Codevalue-added and Tele Atlas value-added maps, respectively(see Fig. 4). In Mississauga, access ratios for PHC physiciansrange from a low of 0.00 per 1000 people in all cases to ahigh of 2.16, 2.20, 2.18, 2.19 for the VL, DMTI value-added,Postal Code value-added and Tele Atlas value-added maps,respectively (see Fig. 4).

Bivariate correlations within cities and between geo-coding methods showed no substantial variation amongmethods (all methods were highly correlated with one an-other). The scatterplots of 3SFCA accessibility scores of ver-ified locations layer versus the Tele Atlas valued addedlayer in all cities reveal linear relationships and are typicalof all other comparisons (see Fig. 5. However, maps inFig. 6 show interesting variation in the arrangement ofaccessibility values in each city. Fig. 6 displays differencemaps for each method compared to the VL results for eachcity (Griffin et al., 2006). Each map shows the amount ofdifference in 3SFCA accessibility between pairs of geocod-ing methods and accessibility calculated using the vali-dated location data. Blue neighbourhoods representplaces with a greater than 0.2 decline in accessibility whilered neighbourhoods are those with a greater than 0.2 in-crease in accessibility. Edmonton shows the strongest pat-tern of change in accessibility rates. Change in Edmonton isprimarily isolated to the core area of the city and a closeexamination shows that bounding neighbourhoods haveeither similar changes or changes in the opposite direction.The latter case is the result of an incorrect location fallingfar enough away from the validated location to affect theoutcome of the accessibility calculation. Edmonton hadthe highest geocoding error, suggesting that incorrectlygeocoded location are contributing to the accessibilityscore of a different neighbourhood when compared withresults based on the validated locations.

5. Conclusion

This study contributes to our understanding of posi-tional error using common geocoding methods and the de-gree to which these errors affect measures of accessibilityto health care services (specifically the 3SFCA method).We first compared six geocoding methods to a verifiedlayer to determine the variation in success (match rate)and accuracy (positional errors). We found the DMTI meth-od performed best; it scored highest in match rate as wellas accuracy. This may be due to the fact that the DMTImethod is a composite of the other two methods, employ-ing both postal code and street address to match locations.Perhaps most interesting is the impact that using MEPpostal code data has on results. This product is intendedto increase geocoding accuracy (this should be reflectedin the Euclidean distance between actual and geocodedlocation). This is true when we compare the automaticMEP geocoding with value-added MEP geocoding (we se-

lected the first location from the tied options). These re-sults indicate that if there is no street address that can beused to select among tied points within a postal code thenthe MEP product is not beneficial to the geocoding process.

Further, there are patterns among the three cities; aone-way ANOVA indicated no significant differences be-tween the six methods in the City of Saskatoon, whereasthere were significant differences in Edmonton and Missis-sauga. It should be noted that the research was conductedin the City of Saskatoon by residents of this city; the levelof familiarity with a region can impact the ability to vali-date and investigate addresses (Goldberg, 2008) althoughthe same procedures were preformed (as outlined) for allcities. While the postal code method did not perform aswell as the DMTI method, it was surprisingly effectiveand accurate, especially used to measure accessibility. Thissupports previous findings regarding the efficacy of postalcode geocoding in urban settings which suggest that postalcode geocoding is reasonably accurate relative to street ad-dress geocoding in an urban context (Bow et al., 2004).

The second component of our study was to investigatethe impact of positional error on measures of accessibility.Measuring health care delivery and accessibility in variousCanadian and American constituencies at the local or com-munity level is essential to understanding the barriers tohealth care services and can only be executed effectivelywith accurate point locations for PHC providers. We foundthat while the variation in the 3SFCA method results usingdifferent geocoding methods were not statistically signifi-cant, there are noteworthy differences. In other contextswhere the positional error is higher, or the dataset of loca-tions being geocoded is larger, this could have a greater im-pact on the assessment of access to primary health careusing the 3SFCA method. The Tele Atlas and postal codemethods show greater variability in the results of accessi-bility, whereas the DMTI value-added performed compara-tively better than all others.

In our geocoding experience there was very little differ-ence in workload among the methods used, with theexception of the validated locations which were onlyestablished for the purpose of comparative analysis ofthe geocoding results of various common methods inhealth studies. The volume of the workload is commonamong all of the methods tested and is primarily associ-ated with data preparation, selecting reference data, estab-lishing the address locator, and performing value-addedmatching. In this study automated geocoding was quitesuccessful and resulted in few records needing value-added attention (manual or interactive matching); we ex-pect with a greater number of records or a larger numberof unmatched records workload would increase. Validatedgeocoding was labor intensive with several hours being de-voted to verifying the addresses for a single city. We wouldnot recommend this method for a large number of recordsor in a setting where accuracy reported here is satisfactory.It is our opinion that the range of accuracy generated witheach of the street address locators, with value added, is sat-isfactory for the vast majority of applications. The latterstep might seem less relevant given the success and accu-racy of the automated method using both DMTI and Tele-Atlas reference data; however, it is our strong opinion

S. Bell et al. / Spatial and Spatio-temporal Epidemiology 3 (2012) 17–29 29

that the additional time spent interactively matching re-cords has two important benefits. First, it increases thesuccess and accuracy of the geocoding, approaching a100% match rate in this examination. Second, it providesadditional peace of mind to the researcher when applyinggeocoding results in further analysis. In fact, our results aremost likely the result of geocoding a specific type of ad-dress (PHC are businesses, so these are business addresses)in a specific area of our study sites (most doctors locationin well established commercially zoned urban areas). Geo-coding addresses in new areas of a city or a mixed set ofacres (residential, commercial, etc.) would likely benefitmore from the value added steps.

Geocoding is a process with benefits beyond academia.Results reported here (both geocoding and accessibilityoutcomes) could be reported in scientific literature, usedto guide policy development and implementation, as wellas guide decision making associated with health servicedelivery and resource deployment. While it might not al-ways be possible to ensure that geocoded locations areaccurate, it is important to establish strict quality controlon the geocoding process to ensure error and uncertainlyare minimized. We believe the above results provide somesupport for the adoption of value-added geocoding in allhealth research applications in certain circumstances, par-ticularly when using MEP geocoding with unreliable streetaddress information. However, we do not believe that therelatively small amount of actual value that was added inthis application is justification for accepting automatedgeocoding results without post-processing inspection andintervention.

References

Andersen RM, McCutcheon A, Aday LA, Chiu GY, Bell R. Exploringdimensions of access to medical care. Health Serv Res1983;18(1):49–74.

Bagheri N, Benwell GL, Holt A. Measuring spatial accessibility to primaryhealth care. In: Annual Colloquium of the Spatial InformationResearch Centre; 2005; University of Otago, New Zealand; 2005.

Bakshi R, Knoblock CA, Thakkar S. Exploiting online sources to accuratelygeocode addresses. In: Proceedings of the 12th annual ACMinternational workshop on Geographic information systems.Washington DC, USA: ACM; 2004. p. 194–203.

Bell S, Wilson K, Bissonette L, Shah TI. Access to primary health care: Doesneighbourhood of residence matter? Annal of the Association ofAmerican Geographers forthcoming.

Bow CJ, Waters N, Faris P, Seidel J, Galbraith PD, Knudtson M, et al.Accuracy of city postal code coordinates as a proxy for location ofresidence. Int J Health Geogr 2004;3(1):5.

Brabyn L, Barnett AR. Population need and geographical access to generalpractitioners in rural new zealand. J N Z Med Assoc 2004;117:1199.

Brewer CA, Pickle L. Evaluation of methods for classifying epidemiologicaldata on choropleth maps in series. Ann Assoc Am Geogr2002;92(4):662–81.

Bruneau J, Daniel M, Kestens Y, Zang G, Généreux M. Associationsbetween hiv-related injection behaviour and distance to and patternsof utilisation of syringe-supply programmes. J Epidemiol CommunityHealth 2008;62(9):804.

Canada Post. Canada postal guide – addressing guidelines. In: 2011.Cayo M, Talbot T. Positional error in automated geocoding of residential

addresses. Int J Health Geogr 2003;2(1):10.Charreire H, Combier E. Poor prenatal care in an urban area: a geographic

analysis. Health & Place 2009;15(2):412–9.DMTI Spatial. Canmap routelogistics v2010.3 [canmap streetfile]. In: Inc.

DS, editor. Markham, Ontario; 2010a.

DMTI Spatial. Platinum postal code suite v2010.3 [multiple enhancedpostalcodes (mep)]. In: Inc. DS, editor. Markham, Ontario; 2010b.

Duncan DT, Castro MC, Blossom JC, Bennett GG, Gortmaker SL. Evaluationof the positional difference between two common geocodingmethods. Geospatial Health 2011;5(2):265–73.

Goldberg D, Wilson J, Knoblock C, Ritz B, Cockburn M. An effective andefficient approach for manually improving geocoded data. Int J HealthGeogr 2008;7(1):60.

Goldberg DW. A geocoding best practices guide. Springfield, IL: NorthAmerican Association of Central Cancer Registries; 2008.

Goldberg DW, Wilson JP, Knoblock CA. From text to geographiccoordinates: the current state of geocoding. URISA J 2007;19(1):33.

Griffin AL, MacEachren AM, Hardisty F, Steiner E, Li B. A comparison ofanimated maps with static small-multiple maps for visuallyidentifying space-time clusters. Ann Assoc Am Geogr2006;96(4):740–53.

Hiscock R, Pearce J, Blakely T, Witten K. Is neighborhood access to healthcare provision associated with individual-level utilization andsatisfaction? Health Serv Res 2008;43(6):2183–200.

Humphreys JS, Smith KB. Healthcare accessibility. In: Nigel T, editor. IntEncyclopedia Human Geogr. Oxford: Elsevier; 2009. p. 71–9.

Joseph A, Phillips D. Accessibility & utilization: geographical perspectiveson health care delivery. New York: Harper and Row; 1984.

Khan AA. An integrated approach to measuring potential spatial access tohealth care services. Socio-Economic Plan Sci 1992;26(4):275–87.

Kindig DA, Movassaghi H. The adequacy of physician supply in small ruralcounties. Health Aff 1989;8(2):63–76.

Korda RJ, Butler JR, Clements MS, Kunitz SJ. Differential impacts of healthcare in Australia: trend analysis of socioeconomic inequalities inavoidable mortality. Int J Epidemiol 2007;36(1):157–65.

Lovett A, Haynes R, Sünnenberg G, Gale S. Car travel time and accessibilityby bus to general practitioner services: a study using patient registersand gis. Social Science &amp. Medicine 2002;55(1):97–111.

Luo W. Using a gis-based floating catchment method to assess areas withshortage of physicians. Health & Place 2004;10(1):1–11.

Luo W, Wang F. Measures of spatial accessibility to health care in a gisenvironment: synthesis and a case study in the chicago region.Environ Plan B 2003;30(6):865–84.

McGrail MR, Humphreys JS. Measuring spatial accessibility to primarycare in rural areas: improving the effectiveness of the two-stepfloating catchment area method. Appl Geogr 2009;29(4):533–41.

Paez A, Mercado RG, Farber S, Morency C, Roorda M. Accessibility tohealth care facilities in montreal island: an application of relativeaccessibility indicators from the perspective of senior and non-seniorresidents. Int J Health Geogr 2010;9:52–62.

Pearce J, Witten K, Bartie P. Neighbourhoods and health: a gis approach tomeasuring community resource accessibility. J Epidemiol CommunityHealth 2006;60(5):389–95.

Penchansky R, Thomas JW. The concept of access: definition andrelationship to consumer satisfaction. Med Care 1981;19(2):127–40.

Rosenthal MB, Zaslavsky A, Newhouse JP. The geographic distribution ofphysicians revisited. Health Serv Res 2005;40(6 Pt 1):1931.

Schootman M, Sterling DA, Struthers J, Yan Y, Laboube T, Emo B, et al.Positional accuracy and geographic bias of four methods of geocodingin epidemiologic research. Ann Epidemiol 2007;17(6):464–70.

Schuurman N. BÉRubÉ M, Crooks VA. Measuring potential spatial accessto primary health care physicians using a modified gravity model. CanGeogr 2010;54(1):29–45.

Ward MH, Nuckols JR, Giglierano J, Bonner MR, Wolter C, Airola M, et al.Positional accuracy of two methods of geocoding. Epidemiology2005;16(4):542–7.

Zandbergen PA. A comparison of address point, parcel and streetgeocoding techniques. Comput Environ Urban Syst 2008;32(3):214–32.

Zandbergen PA. Geocoding quality and implications for spatial analysis.Geography Compass 2009;3(2):647–80.

Zandbergen PA. Influence of street reference data on geocoding quality.Geocarto Int 2011;26(1):35–47.

Zhan FB, Brender JD, De Lima I, Suarez L, Langlois PH. Match rate andpositional accuracy of two geocoding methods for epidemiologicresearch. Ann Epidemiol 2006;16(11):842–9.

Zimmerman D, Li J. The effects of local street network characteristics onthe positional accuracy of automated geocoding for geographic healthstudies. Int J Health Geogr 2010;9(1):1–11.