Geocoding large population-level administrative datasets at highly resolved spatial scales

Abstract

Using geographic information systems to link administrative databases with demographic, social, and environmental data allows researchers to use spatial approaches to explore relationships between exposures and health. Traditionally, spatial analysis in public health has focused on the county, zip code, or tract level because of limitations to geocoding at highly resolved scales. Using 2005 birth and death data from North Carolina, we examine our ability to geocode population-level datasets at three spatial resolutions – zip code, street, and parcel. We achieve high geocoding rates at all three resolutions, with statewide street geocoding rates of 88.0% for births and 93.2% for deaths. We observe differences in geocoding rates across demographics and health outcomes, with lower geocoding rates in disadvantaged populations and the most dramatic differences occurring across the urban-rural spectrum. Our results suggest highly resolved spatial data architectures for population-level datasets are viable through geocoding individual street addresses. We recommend routinely geocoding administrative datasets to the highest spatial resolution feasible, allowing public health researchers to choose the spatial resolution used in analysis based on an understanding of the spatial dimensions of the health outcomes and exposures being investigated. Such research, however, must acknowledge how disparate geocoding success across subpopulations may affect findings.

1. Introduction

Geographic Information Systems (GIS) and spatial analysis are of growing importance in public health research, outreach, and policy. The increasing availability and evolving methodologies of GIS technology and spatial statistics enable researchers to explore the connections between public health endpoints and relevant demographic, social, and environmental conditions by integrating across previously disparate datasets (Bergquist and Rinaldi 2010; Comer et al. 2011; Eisen and Eisen 2011; Goldberg and Jacquez 2012; Krieger et al. 2005; Mindell and Barrowcliffe 2005; Miranda and Edwards 2011; Robinson et al. 2010).

Administrative datasets, such as birth certificate, immunizations, student enrollment, notifiable diseases, and death records, are an important resource for public health researchers, as these data often cover large populations and extended periods of time. Using GIS to display these data can reveal spatial patterns which may help generate hypotheses for future research, provide information for targeting community outreach, or motivate policy efforts and priorities. Using GIS to link these administrative databases with relevant demographic, social, and environmental data via shared geography can allow for spatial statistical approaches to explore relationships between exposures and health endpoints. With key administrative datasets containing information covering many years, these data may be even more valuable in that they can enable spatio-temporal analysis of health outcomes as populations shift and exposures change over time.

Geocoding, the process of converting address information into latitude and longitude coordinates, is the key to leveraging the valuable information already collected in administrative datasets for use in spatial analyses. Four key measures of geocoding quality have been identified: completeness, resolution, matching algorithm criteria, and positional accuracy (Goldberg and Jacquez 2012). Much of the research on geocoding has focused on the positional accuracy of different geocoding techniques (Bell et al. 2012; Bonner et al. 2003; Bow et al. 2004; Cayo and Talbot 2003; Duncan et al. 2011; Goldberg and Cockburn 2012; Healy and Gilliland 2012; Jacquez 2012; Krieger et al. 2001; Ratcliffe 2001; Ward et al. 2005). Errors in positional accuracy can lead to incorrect assignment of areal units such as Census tract or even county, leading to misclassification errors (Goldberg and Cockburn 2012; McLafferty et al. 2012). Completeness of geocoding and matching algorithm criteria are difficult to compare across many studies due to a lack of detail describing the geocoding process in many papers (Robinson et al. 2010).

At least three important advances have occurred that can lead to improvements in geocoding of administrative datasets: 1) address information in administrative datasets has become more complete and standardized; 2) reference layers have been updated, standardized, and more completely populated (Rushton et al. 2006); and 3) geocoding processes and methodologies have improved (Goldberg et al. 2007; Zandbergen and Chakraborty 2006). Inevitably, however, there are records in any dataset that cannot be geocoded, and understanding systematic issues in geocoding completeness is important for understanding the types of research that can be undertaken with a particular dataset and the bias to which such research may be subject.

In this paper, we focus on two of the key measures of geocoding affected by these advances: completeness and spatial resolution. Focusing on 2005 birth and death certificate data from the State of North Carolina, we explore our ability to construct complete spatial datasets for large population-level administrative datasets at highly resolved spatial scales. If we are able to geocode administrative datasets thoroughly and at refined spatial scales, this will have important implications for the spatial scale at which public health research can be undertaken. Using data for the entire state, we assess geocoding completeness, or match rates, at different spatial resolutions across various subpopulations and regions.

2. Methods

2.1. Data

We used two large administrative datasets from North Carolina (NC) – the detail birth record (DBR) and the detail death record (DDR). Through a data sharing agreement, the NC State Center for Health Statistics provided individually-identified birth and death record data with key demographic, health, and geographic variables for all births and deaths occurring in North Carolina in 2005. Each record provided geographic information that was collected at the time of the birth or death, including county of residence, zip code, and street address. Note that the street address provided in the DBR is actually the mailing address, with a variable to indicate if this was also the residential address. Quality of the DBR data has been studied and reported elsewhere (Buescher et al. 1993; Vinikoor et al. 2010), with studies indicating demographic and outcome data is fairly well recorded. No studies were available on the quality of the DDR data.

Race, age, education, marital status, and select health outcome variables provided in the DBR and DDR allowed us to examine trends in geocoding rates in the state overall, as well as within demographic subgroups. Births and deaths were subsetted by race into non-Hispanic white (NHW), non-Hispanic black (NHB), Hispanic (H), and Asian/Pacific Islander (A/PI); all records falling into other race categories were dropped from race-stratified analysis due to small sample size. Educational attainment of the decedent (DDR) and of the mother (DBR) was classified as less than 9^th grade, some high school, high school degree, some college, and college degree+. For the DBR, maternal age was classified into 5-year age groups for those women aged 15–44. For the DDR, the age of the decedents was classified as <20 years, 20–29 years, 30–39 years, 40–49 years, 50–59 years, 60–69 years, 70–79 years, 80–89 years, and ≥90 years. Preterm birth (<37 weeks gestational age at delivery) and diabetes-related death were selected as example public health outcomes to which geospatial analysis might be applied.

In addition, we used data from the National Center for Health Statistics (NCHS) to assign each birth and death an urbanization level based on the reported county of residence. The NCHS classifies all U.S. counties into one of six levels of urbanization: large central metro, large fringe metro, medium metro, small metro, micropolitan (nonmetro), and noncore (nonmetro) (U.S.Centers for Disease Control and Prevention et al. 2012). Figure 1 shows the distribution of these urbanization levels across NC.

Figure 1.

County urban-rural designations as classified by the National Center for Health Statistics (NCHS) and the 7 counties highlighted selected for parcel geocoding.

2.2. Geocoding

Residential addresses were geocoded using Esri ArcGIS 9.3 (Redlands, California) at three levels of spatial resolution—zip code, street, and parcel. Figure 2 displays how the same address would be spatially located using the three methods. Geocoding to an areal unit (e.g., zip code or parcel) typically assigns an address the latitude and longitude of the centroid of the areal unit with the corresponding address information. Street geocoding places the observation proportionally along a street segment, taking into account the address range of the street segment and the street number of the address. Rushton et al. (2006) provide a more in-depth description of how spatial location is assigned during the geocoding process.

Figure 2.

Differences in assigned location of an address using three different geocoding methods.

The U.S. Census Zip Code Tabulation Area 2000 geography (U.S.Census Bureau 2012) was used as the reference layer for zip code level geocoding. Only records with a zip code that exactly matched a zip code in the reference layer were geocoded.

Street level geocoding located residential street addresses based on the 2006 Topologically Integrated Geographic Encoding and Referencing system (TIGER) street centerlines file (U.S.Census Bureau and Geography Division 2007). Similar to the manual geocoding improvement approach (Goldberg et al. 2008), street level geocoding occurred in three batches. First, we geocoded addresses that exactly matched to the road reference layer’s street address components. Unmatched addresses from this first batch were standardized to be more compatible with the road reference layer by removing apartment, unit, suite, and lot numbers; formatting highways and routes numbers; and obtaining street addresses for common apartment complexes, housing developments, and mobile home parks. For the second batch, we geocoded those standardized addresses matching exactly to street address components in the reference layer. The final batch required a GIS technician to examine candidate matches for all remaining unmatched records and select candidate matches for geocoding when appropriate. An inexact match might be selected as the geocoded location for a DBR or DDR record due to, for example, spelling errors in an address; incorrect street type or direction; and units within housing developments, apartment complexes, or mobile home parks.

The parcel geocoding process referenced data provided by county tax assessor offices in order to assign records to individual tax parcels based on the physical address of each parcel (i.e., not the owner’s mailing address). Geocoding to the tax parcel level was done using the same three tiered batch process described above (exact matches with original address, exact matches with standardized address, and hand matching).

Zip code and street level geocoding were performed statewide for both the DBR and DDR data; however, parcel geocoding could not be undertaken statewide as not all NC counties have developed GIS layers of their tax parcel data and reporting of physical address within parcel data varies by county. The seven most rural counties in NC did not have any parcel GIS data, and rural counties also included physical address in parcel data less frequently than more urban counties. Thus, parcel geocoding of births and deaths was undertaken in seven select North Carolina counties: Buncombe, Cabarrus, Durham, Guilford, Pitt, Sampson, and Wilson. These counties are highlighted in Figure 1. While additional counties have developed tax parcel GIS layers, these seven counties were chosen for their distribution across the state and as examples of geographic areas with different levels of urbanization. In order from most urban to most rural, these counties are classified as large fringe metro (Cabarrus), medium metro (Buncombe, Durham, Guilford), small metro (Pitt), micropolitan (Wilson), and noncore (Sampson).

While zip code geocoding was undertaken for all records, street and parcel level geocoding were only undertaken for those records for which a residential street address was reported. Records were considered to be missing residential street address if the address field was blank, only a mailing address was provided, a partial street name was provided, a post office box was reported, or rural route box was reported. Missing addresses were identified manually. The inclusion of a residential street address in a record does not guarantee the address can be geocoded. Records may not be able to be geocoded due to incomplete residential address (e.g., missing street direction or type), incomplete reference layer data, or inability to determine the address of a multi-residence facility. The geocoding match rate is defined as the proportion of all records (both those reporting and those not reporting an address) that were able to be successfully assigned a spatial location at a specific spatial resolution.

2.3. Statistical analysis

We tested for differences in reporting of residential street address and geocoding match rates across demographic (race, age, education, marital status, and urbanization) and outcome (preterm birth for births and diabetes-related death for deaths) subgroups using Chi-squared tests. For both births and deaths, differences in address reporting and geocoding rates were assessed for each level of geocoding—zip code, street, and parcel (select counties only). Since this analytic approach required conducting a large number of tests, we would expect to find significant results purely by chance. To address this issue and be more confident that our results highlight true differences in reporting and geocoding rates, we calculated adjusted p-values using the false discovery rate (FDR) method of controlling error under multiple comparisons. This method controls for the expected proportion of tests that are incorrectly identified as significant, and attempts to balance the desire for power to identify true difference and the desire to be conservative in falsely identifying tests as significant. For the DBR and DDR analyses separately, the false discovery rate was controlled at 0.05, thus tests with adjusted p-values less than 0.05 were considered significant and we could expect that, on average, 5% of the differences found may not actually be significant.

3. Results

In 2005, there were 121,248 births and 73,016 deaths reported in North Carolina’s detailed birth and death records, respectively. Tables 1 and 2 present the distribution of births and deaths, respectively, across demographic subgroups, as well as residential street address reporting rates within each group. Most births were to mothers who were NHW, less than 30 years old, married, and with at least a high school education. Most deaths were to individuals who were 60 years of age or older, not married, and with at least a high school education. Both births and deaths were most commonly among residents of medium metro and micropolitan counties.

Table 1.

Counts of records and percent reporting a residential street address for 2005 births in NC and select counties.^a

	Statewide		Buncombe County		Cabarrus County		Durham County		Guilford County		Pitt County		Sampson County		Wilson County
	N	% with address	N	% with address	N	% with address	N	% with address	N	% with address	N	% with address	N	% with address	N	% with address
All	121,248	92.7	2,587	90.5	2,365	98.4	4,038	98.3	5,956	99.1	2,016	94.5	929	81.0	1,043	90.9
Maternal race/ethnicity
NHW	68,856	93.8	2,013	89.5	1,516	98.4	1,454	98.1	2,629	99.1	978	95.4	400	92.5	399	93.2
NHB	27,777	91.4	192	93.8	327	97.3	1,368	97.9	2,189	99.2	818	92.8	226	84.1	424	89.6
Hispanic	19,437	91.4	331	94.3	463	98.9	967	99.1	822	99.3	180	96.1	279	60.6	208	88.5
Other	5,039	91.0	46	95.7	57	98.3	237	98.7	314	99.4	40	100.0	23	95.7	12	100.0
Maternal age(yrs)
15–19	13,684	89.9	284	88.0	202	97.5	376	98.9	559	99.1	235	90.6	115	84.4	166	92.8
20–24	32,165	91.6	696	90.1	575	99.0	828	98.8	1,427	99.3	516	93.8	300	82.3	293	87.4
25–29	32,867	93.1	669	90.7	671	99.0	1,126	98.9	1,603	99.1	609	94.9	268	81.7	284	94.4
30–34	27,263	94.2	581	90.9	588	97.6	1,110	97.1	1,497	99.1	444	95.3	173	76.9	187	87.7
35–39	12,515	94.7	274	92.0	286	97.6	488	98.4	727	99.3	169	97.6	61	77.1	97	93.8
40–44	2,381	92.8	68	92.7	36	100.0	92	97.8	131	97.0	35	97.1	11	72.7	14	92.9
Maternal education
<9th grade	8,956	89.3	148	90.5	172	100.0	456	98.5	404	99.5	128	92.2	149	63.8	115	87.8
Some high school	19,206	89.8	441	87.8	342	98.0	606	99.3	816	99.8	326	92.9	163	75.5	218	92.2
High school	34,506	91.3	737	87.9	647	97.8	755	98.3	1,462	99.1	410	93.4	314	83.4	329	87.5
Some college	26,272	93.2	502	91.6	563	98.6	543	97.1	1,364	98.6	548	93.6	207	89.4	192	92.7
College degree	32,093	96.6	754	93.9	640	98.4	1,661	98.4	1,905	99.2	602	97.3	92	91.3	189	95.2
Maternal marital status
Married	74,755	94.3	1,616	91.0	1,615	98.7	2,351	98.1	3,511	99.2	1,090	95.2	494	84.4	505	92.5
Not married	46,489	90.2	970	89.7	750	97.6	1,687	98.6	2,445	99.0	926	93.6	435	77.0	538	89.4
Gestational age at delivery
Term birth	105,507	92.9	2,260	90.9	2,066	98.4	3,545	98.4	5,166	99.2	1,705	94.8	799	79.9	895	91.0
Preterm birth	15,678	92.0	326	87.7	298	98.3	492	98.0	785	99.1	311	92.9	130	87.7	147	91.2
Urban-rural classification
Large central metro	13,428	98.9
Large fringe metro	7,902	95.2
Medium metro	55,449	95.9
Small metro	10,937	93.6
Micropolitan	25,174	87.8
Noncore	8,358	73.5

^aGray shading highlights reporting rates that were significantly different (false discovery rate adjusted p-value < 0.05) across demographic or outcome subgroups.

Table 2.

Counts of records and percent reporting a residential street address for 2005 deaths in NC and select counties.

	Statewide		Buncombe County		Cabarrus County		Durham County		Guilford County		Pitt County		Sampson County		Wilson County
	N	% with address	N	% with address	N	% with address	N	% with address	N	% with address	N	% with address	N	% with address	N	% with address
All	73,016	99.2	2,195	99.3	1,277	99.8	1,653	99.4	3,618	99.8	1,099	99.0	587	99.3	772	99.6
Race/ethnicity
NHW	55,922	99.3	2,004	99.3	1,100	99.8	956	99.9	2,611	99.9	634	99.2	376	99.5	472	99.8
NHB	15,461	98.9	171	99.4	168	100.0	661	99.1	944	99.7	457	98.7	188	98.9	285	99.3
Hispanic	573	96.0	10	100.0	5	100.0	17	94.1	20	100.0	3	100.0	10	100.0	13	100.0
Other	1,040	98.5	9	100.0	4	100.0	18	94.4	43	100.0	5	100.0	13	100.0	2	100.0
Age(yrs)
<20	1,758	99.0	49	98.0	23	100.0	46	97.8	89	100.0	40	97.5	14	100.0	19	94.7
20–29	1,216	98.3	36	100.0	14	100.0	35	94.3	59	100.0	15	93.3	12	100.0	14	100.0
30–39	1,764	98.9	43	97.7	27	100.0	39	97.4	78	100.0	33	100.0	13	100.0	19	100.0
40–49	4,093	98.9	127	98.4	69	100.0	98	98.0	185	100.0	66	100.0	33	97.0	44	100.0
50–59	7,421	99.1	193	97.9	126	100.0	181	100.0	352	99.7	129	98.5	63	98.4	94	100.0
60–69	10,429	99.2	275	99.6	173	100.0	212	99.1	472	99.6	179	100.0	77	100.0	110	100.0
70–79	16,568	99.4	456	99.8	292	100.0	333	100.0	768	100.0	250	98.8	145	100.0	172	100.0
80–89	20,707	99.2	683	99.4	385	99.5	501	99.8	1,079	99.8	278	98.9	161	100.0	208	99.0
>89	9,051	99.1	333	99.4	168	100.0	207	100.0	536	99.6	109	99.1	69	97.1	92	100.0
Education
<9th grade	16,319	98.9	404	99.5	276	99.6	265	99.3	641	99.5	270	98.5	166	99.4	221	99.1
Some high school	15,189	99.1	440	99.6	310	100.0	314	98.7	672	99.9	241	97.9	142	100.0	174	100.0
High school	22,037	99.3	620	98.6	416	100.0	486	99.8	1,120	100.0	305	99.3	165	98.8	221	100.0
Some college	9,861	99.5	336	99.4	171	100.0	257	100.0	559	100.0	140	100.0	67	100.0	84	100.0
College degree	8,276	99.5	375	100.0	89	100.0	297	100.0	581	99.8	130	100.0	32	96.9	54	98.2
Marital status
Married	28,587	99.4	813	99.3	521	100.0	620	99.8	1,357	100.0	380	99.0	207	100.0	276	100.0
Not married	44,294	99.0	1,379	99.3	754	99.7	1,030	99.3	2,257	99.7	717	99.0	379	98.9	495	99.4
Cause of death
Not diabetes-related	64,846	99.2	2,033	99.3	1,161	99.8	1,496	99.4	3,250	99.8	942	99.0	520	99.2	668	99.6
Diabetes-related	8,170	99.2	162	99.4	116	100.0	157	99.4	368	99.7	157	98.7	67	100.0	104	100.0
Urban-rural classification
Large central metro	4,723	99.6
Large fringe metro	4,486	99.1
Medium metro	30,821	99.5
Small metro	6,018	98.7
Micropolitan	19,283	99.1
Noncore	7,685	98.0

^aGray shading highlights reporting rates that were significantly different (false discovery rate adjusted p-value < 0.05) across demographic or outcome subgroups.

3.1. Address reporting

The proportion of records that included a residential street address represent an upper limit on geocoding match rates for street and parcel level geocoding efforts (zip code geocoding only requires a zip code, not a full street address). Statewide, 92.7% of birth records and 99.2% of death records included a residential street address. In the DBR, missing addresses were predominantly due to reporting of mailing address rather than residential address (95%). In the death data, post office and rural route boxes accounted for almost half of the records missing addresses, with partial addresses accounting for 42% and no street address information for 8% of these records.

Birth and death records from more rural areas were less likely to include a street address (FDR-adjusted p<0.05); however, while address reporting rates in the death record remained above 98% in all urban-rural classifications, the reporting rate in the birth record dropped to as low as 73.5% in noncore counties. Address reporting among births varied by maternal race, age, education, and marital status (FDR-adjusted p<0.5). Generally, births to women at higher risk for adverse pregnancy outcomes –minority women, extremely young or old mothers, women with low educational attainment, and single women – were less likely to report a street address in the DBR. Preterm births were also slightly less likely to report a street address compared to term births (92.9% for term versus 92.0% for preterm births; FDR-adjusted p<0.05). In the death record data, although there were differences in residential street address reporting by race, age, marital status, and education (FDR-adjusted p<0.5), differences across subgroups were small.

Among the seven counties selected for parcel geocoding, residential street address reporting in the birth data generally followed the statewide pattern of more urban areas having higher reporting rates than more rural areas; however, address reporting among births in Buncombe County, a medium metro area, was more on par with Wilson County, a micropolitan area, than with other medium metro counties. Among births in Sampson and Buncombe Counties, differences in address reporting across race/ethnicity were observed, with Hispanic women in Sampson County and NHW women in Buncombe County having lower levels of address reporting (FDR-adjusted p<0.5). In Durham County, births to older mothers were less likely to report a street address (FDR-adjusted p<0.5). In Buncombe, Pitt, and Sampson Counties, reporting rates in the DBR generally increased with education (FDR-adjusted p<0.5). In all seven of these counties, address reporting rates for the death record were at least 99%. Reporting rates among younger decedents and those of Hispanic ethnicity were lower in Durham County (FDR-adjusted p<0.5), but no other differences in address reporting across demographic and outcome subgroups within these counties were noted.

3.2. Zip code geocoding

Zip code geocoding rates were high (see Tables 3 and 4), with 96.8% of births and 97.7% of deaths able to be geocoded at this level of spatial resolution. While geocoding rates did vary by race, age, education (in DDR only), marital status, and urbanization (FDR-adjusted p<0.5), differences across subgroups were small, within ±3 percentage points. Among the seven counties selected for closer examination, Buncombe County had the lowest zip code geocoding rate for births at 96.5% and Durham County had the lowest rate for deaths at 97.5%. In Durham County, small differences in zip code geocoding rates across race, age, and education were observed (FDR-adjusted p<0.5).

Table 3.

Geocoding match rates (percent of records geocoded) at the zip code (zip), street (str), and parcel (parc) level among 2005 births.^a

	Statewide		Buncombe County			Cabarrus County			Durham County			Guilford County			Pitt County			Sampson County			Wilson County
	Zip	Str	Zip	Str	Parc	Zip	Str	Parc	Zip	Str	Parc	Zip	Str	Parc	Zip	Str	Parc	Zip	Str	Parc	Zip	Str	Parc
All	96.8	88.0	97.5	83.5	83.6	99.6	89.9	92.1	97.8	93.4	95.5	99.4	97.5	93.5	98.6	86.9	80.2	98.7	76.5	39.8	99.4	82.1	82.9
Maternal race/ethnicity
NHW	96.5	89.7	97.5	82.9	83.4	99.6	91.0	92.4	96.1	94.5	94.6	99.5	97.8	94.0	99.0	89.6	82.2	100.0	88.5	44.8	99.8	86.0	84.0
NHB	97.2	86.2	97.9	87.5	88.5	99.1	85.6	92.4	98.5	92.0	95.8	99.2	97.6	93.2	98.5	83.4	78.9	98.2	77.9	46.0	99.1	80.9	83.7
Hispanic	97.6	85.5	98.2	84.6	81.0	99.8	90.1	90.7	99.2	94.5	95.8	99.4	96.6	92.8	96.7	86.1	74.4	97.1	56.6	26.2	99.5	76.9	78.4
Other	95.6	84.4	97.8	87.0	93.5	100.0	86.0	93.0	97.9	90.7	98.3	99.4	96.8	93.6	100.0	95.0	85.0	100.0	95.7	56.5	100.0	83.3	100.0
Maternal age(yrs)
15–19	97.0	84.4	96.5	79.9	78.2	99.5	91.6	88.1	99.7	93.9	95.0	99.1	97.1	94.5	97.9	81.3	75.7	97.4	79.1	38.3	100.0	78.3	83.1
20–24	97.5	86.1	98.3	81.2	81.5	99.8	91.8	91.1	98.9	93.7	94.7	99.3	97.3	93.5	98.1	86.2	74.6	99.3	77.7	38.3	99.7	81.2	78.2
25–29	97.0	88.3	97.3	81.9	83.0	99.9	89.3	93.1	98.1	92.7	97.0	99.6	97.6	92.9	98.9	87.5	81.9	98.9	78.7	42.9	98.9	85.9	85.6
30–34	96.0	90.2	97.3	86.4	87.1	99.2	87.1	92.4	96.1	92.5	94.8	99.3	97.5	93.6	98.9	88.3	84.5	98.8	73.4	35.3	99.5	80.8	84.5
35–39	96.1	91.0	97.5	88.0	87.6	99.3	92.7	94.1	97.1	95.7	95.9	99.6	98.2	93.8	99.4	90.5	85.2	98.4	65.6	49.2	99.0	86.6	87.6
40–44	95.8	89.0	100.0	92.7	86.8	100.0	83.3	94.4	97.8	95.7	93.5	99.2	97.0	95.4	100.0	85.7	82.9	90.9	72.7	45.5	100.0	57.1	78.6
Maternal education
<9th grade	97.7	82.3	99.3	79.7	76.4	100.0	92.4	91.3	99.8	94.1	94.1	99.3	95.3	92.6	96.9	79.7	71.1	98.7	60.4	30.9	100.0	78.3	78.3
Some high school	97.3	84.0	96.8	78.7	77.3	99.7	91.5	90.6	99.3	94.1	95.5	99.5	98.2	95.8	98.2	85.3	76.1	96.3	70.6	33.1	99.5	78.9	80.3
High school	97.2	86.3	97.0	80.7	79.8	99.1	90.7	89.0	98.5	93.6	94.8	99.1	97.0	93.0	98.8	85.4	74.6	98.7	78.3	39.2	99.4	81.8	83.0
Some college	97.0	88.9	97.6	84.3	87.7	99.8	90.8	92.9	98.2	91.7	94.7	99.5	97.7	92.2	98.2	85.6	79.6	100.0	86.5	44.4	99.0	83.9	83.9
College degree	95.7	93.1	98.1	89.1	89.7	99.7	86.9	95.5	96.3	93.5	96.5	99.5	98.0	94.1	99.5	91.4	88.9	100.0	85.9	57.6	99.5	86.8	87.8
Maternal marital status
Married	96.4	90.0	97.7	84.2	84.7	99.7	89.4	92.8	96.8	93.5	95.3	99.5	97.6	93.2	98.8	89.0	83.3	99.4	80.4	42.1	99.6	84.6	84.8
Not married	97.4	84.7	97.4	82.4	81.9	99.3	90.9	90.5	99.2	93.3	95.7	99.2	97.3	94.0	98.4	84.3	76.6	97.9	72.2	37.2	99.3	79.7	81.2
Gestational age at delivery
Term birth	96.8	88.1	97.7	83.9	84.3	99.6	89.7	92.0	97.8	93.7	95.6	99.4	97.4	93.7	98.6	87.2	80.1	98.6	75.6	39.3	99.3	81.9	82.4
Preterm birth	96.6	87.1	96.6	81.0	79.1	99.7	91.3	92.6	97.6	91.7	94.7	99.4	97.8	92.6	98.7	84.9	80.7	99.2	82.3	43.1	100.0	83.7	87.1
Urban-rural classification
Large central metro	99.5	97.5
Large fringe metro	99.3	89.2
Medium metro	95.4	91.6
Small metro	98.8	88.3
Micropolitan	96.2	81.9
Noncore	98.8	65.6

^aGray shading highlights geocoding rates that were significantly different (false discovery rate adjusted p-value < 0.05) across demographic or outcome subgroups.

Table 4.

Geocoding match rates (percent of records geocoded) at the zip code (zip), street (str), and parcel (parc) level among 2005 deaths.

	Statewide		Buncombe County			Cabarrus County			Durham County			Guilford County			Pitt County			Sampson County			Wilson County
	Zip	Str	Zip	Str	Parc	Zip	Str	Parc	Zip	Str	Parc	Zip	Str	Parc	Zip	Str	Parc	Zip	Str	Parc	Zip	Str	Parc
All	97.7	93.2	99.5	94.8	82.9	99.1	94.4	89.4	97.5	95.2	91.8	99.5	97.0	89.4	98.4	93.3	80.6	99.3	94.6	46.5	99.4	92.2	86.1
Race/ethnicity
NHW	97.7	93.9	99.5	94.7	83.3	99.3	95.5	89.6	97.2	95.4	92.6	99.7	97.3	89.7	99.2	95.1	84.7	99.5	94.7	44.7	99.4	92.4	86.2
NHB	97.7	91.3	100.0	94.7	80.1	98.2	88.7	88.1	97.9	95.3	91.5	99.2	96.2	88.2	97.2	91.0	74.8	98.9	93.6	52.7	99.3	91.9	88.1
Hispanic	97.4	86.2	100.0	100.0	70.0	100.0	60.0	100.0	100.0	88.2	70.6	95.0	100.0	75.0	100.0	66.7	66.7	100.0	100.0	10.0	100.0	92.3	38.5
Other	98.5	86.6	100.0	100.0	44.4	100.0	100.0	100.0	100.0	94.4	88.9	100.0	100.0	97.7	100.0	80.0	100.0	100.0	100.0	38.5	100.0	100.0	100.0
Age(yrs)
<20	96.9	91.3	98.0	89.8	75.5	100.0	91.3	95.7	95.7	95.7	80.4	100.0	97.8	82.0	100.0	87.5	67.5	100.0	85.7	35.7	100.0	89.5	68.4
20–29	96.3	89.3	100.0	97.2	69.4	92.9	85.7	85.7	94.3	85.7	82.9	100.0	98.3	81.4	100.0	93.3	73.3	100.0	91.7	41.7	100.0	92.9	64.3
30–39	97.2	91.5	100.0	90.7	76.7	100.0	85.2	92.6	100.0	84.6	92.3	100.0	96.2	84.6	100.0	90.9	72.7	100.0	92.3	15.4	100.0	94.7	79.0
40–49	97.7	92.4	100.0	96.1	81.1	100.0	92.8	81.2	98.0	94.9	88.8	99.5	94.6	83.8	100.0	93.9	75.8	100.0	93.9	39.4	100.0	90.9	79.6
50–59	97.6	92.5	99.5	90.2	74.1	100.0	94.4	87.3	98.9	96.1	89.5	99.4	96.6	90.1	97.7	90.7	71.3	98.4	96.8	47.6	100.0	95.7	85.1
60–69	97.6	92.9	99.6	95.6	84.7	98.8	94.2	92.5	96.7	95.3	89.2	99.6	96.4	89.2	98.9	96.7	79.9	100.0	94.8	42.9	100.0	91.8	84.6
70–79	97.7	93.4	99.3	95.6	86.8	99.0	93.2	87.3	97.3	96.1	94.3	99.2	97.1	89.6	98.0	92.4	85.2	100.0	95.9	49.7	100.0	91.3	90.1
80–89	98.0	93.9	99.6	94.4	84.5	98.7	96.1	90.9	97.4	96.4	92.8	99.5	97.4	90.6	98.9	93.9	84.5	99.4	96.3	53.4	97.6	91.8	85.6
>89	97.9	93.4	99.4	96.7	81.7	100.0	96.4	90.5	98.1	94.2	96.1	99.6	97.8	91.0	95.4	93.6	83.5	97.1	88.4	39.1	100.0	92.4	94.6
Education
<9th grade	98.0	92.0	99.5	94.3	83.2	98.9	95.3	90.6	98.9	96.2	90.6	99.4	97.7	88.8	97.8	93.0	78.5	100.0	94.0	41.6	98.6	90.1	81.0
Some high school	97.7	92.8	99.8	95.2	83.6	99.0	94.8	89.0	99.0	94.6	90.5	99.4	97.8	90.3	98.3	90.0	73.0	99.3	97.9	52.8	99.4	90.2	86.8
High school	97.7	93.3	99.5	93.4	81.6	99.0	95.2	87.5	97.9	96.1	92.0	99.6	96.9	89.5	99.3	92.8	80.7	99.4	92.1	41.8	99.6	95.0	87.8
Some college	97.9	94.3	99.4	95.2	81.6	100.0	95.9	91.8	97.3	97.3	93.8	99.6	95.2	90.0	96.4	96.4	89.3	97.0	98.5	52.2	100.0	94.1	94.1
College degree	97.2	95.1	99.7	96.5	85.6	98.9	91.0	91.0	94.6	92.9	93.6	99.8	97.9	88.5	99.2	96.9	90.0	100.0	93.8	59.4	100.0	94.4	90.7
Marital status
Married	97.7	94.1	99.5	95.9	86.5	99.6	94.6	87.9	97.3	95.7	93.1	99.6	97.2	90.6	99.2	93.4	83.4	99.0	96.1	41.6	100.0	94.6	86.6
Not married	97.8	92.6	99.5	94.1	80.8	98.8	94.6	90.5	97.9	95.2	91.3	99.5	96.9	88.6	97.9	93.2	79.1	99.5	93.7	49.3	99.0	90.9	85.9
Cause of death
Not diabetes-related	97.7	93.3	99.5	94.6	82.7	99.1	94.1	89.4	97.3	95.3	91.6	99.5	97.1	88.9	98.2	93.1	81.5	99.4	94.4	45.8	99.3	92.1	86.7
Diabetes-related	98.0	92.3	99.4	96.3	84.6	100.0	98.3	89.7	98.7	94.9	93.6	99.2	96.5	93.8	99.4	94.3	75.2	98.5	95.5	52.2	100.0	93.3	82.7
Urban-rural classification
Large central metro	99.4	95.9
Large fringe metro	99.4	94.0
Medium metro	96.8	94.8
Small metro	98.9	92.8
Micropolitan	97.4	92.2
Noncore	99.3	87.2

^aGray shading highlights geocoding rates that were significantly different (false discovery rate adjusted p-value < 0.05) across demographic or outcome subgroups.

Over 500 zip codes reported in either the DBR or DDR were not able to be geocoded. Zip codes were not geocoded because they were not in NC (42% of the unmatched zip codes), not real zip codes (16% of the unmatched zip codes), or not found in the NC ZCTA data (42% of the unmatched zip codes). Non-geocoded zip codes that actually exist in NC were distributed across the state, occurring in 67 of the 100 NC counties; however, there were clusters of these zip codes in the most urban areas of NC, including Mecklenburg, Wake, Forsyth, and Guilford Counties. These zip codes may have been established since the 2000 ZCTA data were created.

Births and deaths in counties with parcel data that included data on the physical address of each parcel (i.e., could potentially be used for geocoding) were slightly less likely to be geocoded at the zip code level compared to births and deaths in counties lacking parcel data. For example, 96.3% of births in counties with parcel data that included an address were geocoded, while 97.1% of births in counties without such data were geocoded. This may reflect the creation of new zip codes in urban areas that typically have more complete parcel data.

3.3. Street geocoding

County street level geocoding rates for both the DBR and DDR data are displayed in Figure 3a–b. The maps are drawn on the same legend using a monochromatic chloropleth in which higher geocoding rates are represented by darker shading and lower geocoding rates are represented by lighter shading. As highlighted by the darker overall color of the map in Figure 3b (death data) compared to the map in Figure 3a (birth data), street geocoding rates were higher for the death record (93.2%) than the birth record (88.0%). Both maps also reveal geographic variation in street level geocoding rates across the state. This geographic variation is at least partially due to urbanization, which was significantly associated with geocoding rates for both datasets (FDR-adjusted p<0.5). Similar spatial patterning was not observed for county zip code level geocoding rates; county street and zip code level geocoding rates were not correlated.

Figure 3.

County street geocoding rates for 2005 NC births (a) and deaths (b).

Street geocoding rates for the entire state also varied by race, age, education, and marital status (FDR-adjusted p<0.5). Generally, minority race, lower educational attainment, younger age, not being married, and residence in more rural areas were associated with lower geocoding rates. Street geocoding rates for the DBR data were also slightly higher for term (88.1%) than for preterm (87.1%) births (FDR-adjusted p<0.5), and rates for the DDR data were slightly higher for not diabetes-related (93.3%) than for diabetes-related (92.3%) deaths (FDR-adjusted p<0.5). Across the state, street geocoding rates for both births and deaths were slightly lower in counties with parcel data that included data on the physical address of parcels compared to births and deaths in counties without such parcel data.

Although street geocoding rates for the seven select counties generally followed the statewide pattern of higher geocoding rates in more urban areas and lower geocoding rates in more rural areas, Cabarrus County, the most urban of the seven counties, had only the third highest geocoding rate for births and fourth highest geocoding rate for deaths. Street geocoding rates for the birth data were quite disparate across the seven counties, ranging from 97.5% in Guilford County to 76.5% in Sampson County. The range of geocoding rates for the death data, on the other hand, was much smaller, ranging from 97.0% in Guilford County to 92.2% in Wilson County.

Within the counties, differences in geocoding rates were observed across each of the demographic variables considered. Geocoding rates for Hispanics were significantly lower than other race/ethnicity groups for deaths in Cabarrus and Pitt Counties, as well as births in Sampson County (FDR-adjusted p<0.05). On the other hand, geocoding rates were lowest for NHBs among births in Pitt and Durham Counties (FDR-adjusted p<0.05). Street geocoding rates increased with maternal age among Buncombe County births (FDR-adjusted p<0.05). Lower educational attainment was associated with lower geocoding rates for births in Buncombe, Guilford, Pitt, and Sampson Counties (FDR-adjusted p<0.05). Births to married mothers were more likely to be street geocoded in Pitt and Sampson Counties (FDR-adjusted p<0.05).

3.4. Parcel geocoding

Since tax parcel GIS layers with address information are not available for the entire State of North Carolina, parcel geocoding was only undertaken in the seven select counties. As was the case with street geocoding rates, parcel geocoding rates for each of the seven counties followed the general pattern of geocoding rates increasing with urbanization. In Sampson County, the most rural of the seven counties, parcel geocoding rates lagged far behind the other counties at 39.8% for births and 46.5% for deaths. Parcel geocoding rates in all other counties were at least 80% for both datasets.

Parcel geocoding was far less successful among Hispanics compared to NHWs and NHBs for Sampson County births and Buncombe, Pitt, and Wilson County deaths (FDR-adjusted p<0.05). Geocoding rates generally increased with age for deaths to residents of Buncombe, Durham, Guilford, Pitt, and Wilson Counties, and with maternal age for births to residents of Buncombe and Pitt Counties (FDR-adjusted p<0.05). Geocoding rates varied by education among births in Buncombe, Cabarrus, Guilford, Pitt, and Sampson Counties, as well as among deaths in Pitt County (FDR-adjusted p<0.05). Marital status was associated with geocoding rate for Buncombe County deaths and Pitt County births (FDR-adjusted p<0.05). Finally, in Guilford County, records for deaths related to diabetes were more likely to be parcel geocoded than records for deaths not related to diabetes (93.8% versus 88.9%, respectively; FDR-adjusted p<0.05).

4. Discussion

Population-level administrative datasets contain vast amounts of important public health information. Converting these databases into spatial datasets through geocoding also allows administrative data to be linked with key environmental and social exposure data that could not be linked using traditional aspatial data architectures. However, public health researchers can only take advantage of the advances in GIS and spatial statistics if key datasets can be successfully geocoded at appropriately refined spatial scales.

In this paper, we investigate geocoding completeness of public health data at the population scale and highlight issues related to geocoding completeness that may impact research relying on geocoded data. We find that population-level geocoding of administrative birth and death records for North Carolina can be undertaken at highly resolved spatial scales with high geocoding completeness. The minimum geocoding rate for producing a reliable spatial dataset has been established at 85% (Ratcliffe 2004), and our zip and street geocoding rates exceed or approach this mark in nearly all cases. Our geocoding rates at the zip code and street level were on par with previous studies (Krieger et al. 2002; McElroy et al. 2003; Zandbergen 2008), and our parcel rates, often over 80% and even 90%, were quite high compared to previous work in which parcel geocoding rates rarely exceeded 70% (Zandbergen 2008). Our parcel geocoding rates did, however, fall below the 85% mark in more rural counties where parcel reference layers are less fully developed.

These findings have implications for the spatial resolution at which population-level public health research is undertaken. Traditionally, public health research has focused on large geographic scales such as county, zip code, or Census tract (Rushton et al. 2006; Zandbergen and Chakraborty 2006) because geocoding to more highly resolved scales was difficult. We know, however, that the spatial resolution at which health research is conducted has important implications for our understanding of health and place (Dolinoy and Miranda 2004; Krieger et al. 2002; Leonard et al. 2011; Moore and Carpenter 1999; Root 2012; Rushton et al. 2006). Analysis at coarse spatial resolution may obscure within-unit spatial heterogeneity, hiding relationships that become apparent when analyses are conducted at a finer spatial resolution (Billaudeau et al. 2010; Dolinoy and Miranda 2004; Krieger et al. 2002; Leonard et al. 2011; Root 2012). In addition, exposures of interest may not coincide with the boundaries defined by governmental organizations and may vary at short distances (Root 2012; Rushton et al. 2006). The high geocoding match rates we achieved for population-level birth and death data indicate that it is possible to create high quality, highly resolved geocoded data layers from large administrative datasets. Such datasets can support research in which analytical approaches are determined based on the research question of interest rather than driven by limitations of the spatial data available. We thus support routinely geocoding to the highest spatial resolution feasible in order to provide researchers with the flexibility to aggregate data in a variety of ways to support the broadest range of analyses (Leonard et al. 2011; Rushton et al. 2006).

While we were able to geocode at highly resolved spatial scales, which can support spatial analysis at more refined geography, it is important to note that we did observe differences in geocoding success across subpopulations, which can affect the interpretation and generalizability of analytical findings based on the resulting datasets (Krieger et al. 2001; Oliver et al. 2005). Importantly, we noted differences in our ability to geocode data across health outcomes. In the birth data, address reporting was slightly lower for preterm births than for term births (92.0% versus 92.9%). Street geocoding rates were slightly higher for term births than preterm births (88.1% versus 87.1%), although there was considerable variability across counties (e.g., street geocoding match rates were higher for preterm births in four of the seven counties we investigated in detail). Generally, parcel geocoding rates were higher for term births than for preterm births, although not so in two of the seven detailed counties. Results varied similarly in geocoding match rates for non-diabetes-related deaths and diabetes-related deaths.

In the death data, statewide street geocoding rates were 1 percentage point higher for deaths not related to diabetes than for deaths related to diabetes. In addition, geocoding rates were lower for disadvantaged and/or higher risk demographic subgroups. The largest differences we observed were in our ability to geocode at the street level across urbanization categories. For example, statewide, just 65.5% of births in noncore counties were able to be street geocoded, compared to 81.9% of births in the next most rural counties (micropolitan areas) and 95.7% of births in the most urban counties (large central metros). This difference in geocoding success by urbanization, which has also been noted in previous studies (Bonner et al. 2003; Cayo and Talbot 2003; McElroy et al. 2003), reflects the more complete reference layers in more urban areas and indicates a need to help lower-resourced rural counties improve their GIS infrastructure. For other demographics, the magnitude of the differences in street geocoding rates across demographic subgroups was less than 10 percentage points for statewide data; although within some of the seven counties considered in more detail, we did see some large differences in geocoding across race groups at the street and parcel levels, with births and deaths to Hispanics proving especially difficult to geocode, particularly in rural areas.

Research using geocoded data must be careful to note how variations in geocoding rates across key demographic groups and health outcomes may impact analytical findings – and how these differences in geocoding rates may be pertinent to the outcomes under investigation (e.g. impact of residential instability, access to care, etc.)

We also note that geocoding rates were lower for the birth record than the death record data (e.g., street geocoding match rates of 88.0% and 93.2%, respectively). This difference, however, does not reflect a true difference between the two datasets in our ability to locate addresses. Rather, the difference in the geocoding rates is primarily due to the higher rate of residential street address reporting in the death record. For example, if we only consider records that included a residential street address, than the proportion of NC birth and death records that could be street geocoded are actually quite similar at 94.9% and 93.9%, respectively. Since we are interested in our ability to geocode these datasets at the population-level, we cannot ignore such records. This suggests the importance of improving reporting of residential street addresses in administrative datasets.

This study is not without limitations. For small subpopulations, such as very young decedents, certain minority groups, or very rural communities, small changes in the number of records may translate to large changes in address reporting rates and/or geocoding rates. Additionally, we only studied birth and death certificate data in North Carolina. The quality and consistency of reporting of address information in corresponding datasets may vary by state, and thus have important implications for geocoding these data. Other administrative datasets, even those covering North Carolina, such as immunizations, blood lead screening, disease registries, and educational outcomes may also vary in the ability to be successfully geocoded. Finally, we assess geocoding completeness at various spatial scales, but are not able to comment on the positional accuracy of geocoded locations. Positional accuracy is another important metric of geocoding quality, and may vary across the urban-rural spectrum, thus potentially compounding the differences in completeness we note across this same spectrum.

An additional caveat of this work is that we did not utilize some recent advances in geocoding technology; however, we believe many health researchers may have to make a similar choice. Advances in geocoding in the ArcGIS platform, including both web-based and desktop-based services, may produce high quality geocoded datasets more quickly than the geocoding methodology we used here, but this may be at the expense of confidentiality and spatial resolution. While web-based geocoding services can be both fast and accurate, these technologies require data be passed over a web connection to a remote server. Data sharing agreements and security protocols may not permit such sharing of confidential health data over the internet. Data confidentiality concerns lead us to forego the web-based geocoding options in favor of traditional, locally-based geocoding methods. We acknowledge, however, that in instances when data confidentiality is not of concern, the recently developed web-based geocoding options and other geocoding products may produce quality geocoding results. In addition, advances have also been made in the desktop-based address locators included with the ArcGIS platform; however, these locators may geocode at lower levels of resolution, for example zip code or city level, if they cannot match a street address. Thus, use of these locators requires careful attention to their underlying methodology and to the interpretation of the resulting geocoded datasets. For this analysis, we selected our geocoding methodology, including the use of TIGER streets as our address reference layer, based on our desire to protect confidentiality and control spatial resolution. Furthermore, TIGER streets align spatially with Census boundaries, making it easier to spatially join our geocoded births and deaths with demographic and socioeconomic data for future research.

While we demonstrate high geocoding rates for both births and deaths at the population level and within key demographic subgroups, we were not able to geocode all records in either administrative dataset. In most datasets, there will inevitably be records that cannot be geocoded to the individual address level. Methods are being developed to combine data geocoded at different levels in order to allow researchers to use all records (Goovaerts 2012; Hibbert et al. 2009). Simultaneously pursuing the further development of such methodologies and efforts to improve our ability to geocode administrative data at highly resolved spatial scales through improved reference layers and address reporting will create a multitude of new research opportunities in spatial public health analysis.

5. Conclusion

Overall, we find that we are able to successfully geocode population-level birth and death certificate data in North Carolina at highly resolved spatial scales. The spatial scales at which we were able to achieve high geocoding completeness allow for analysis at a more refined spatial scale than has traditionally been used for public health research. More spatially refined analyses allows for better characterization of associations between exposures and outcomes exhibiting spatial heterogeneity within larger areal units such as zip codes.

This work affirms that highly resolved spatial data architectures for population-level administrative public health datasets are viable through geocoding of street address information. We concur with previous work that has recommended geocoding to the highest spatial resolution feasible given consideration of the other dimensions of geocoding quality (completeness, accuracy, and matching algorithm criteria) (Leonard et al. 2011; Rushton et al. 2006). We recommend routinely geocoding administrative health data to the street level and, even more ideally, the parcel level where good quality reference data are available. As more administrative datasets are geocoded with high levels of completeness at highly resolved scales, the decision of the spatial resolution at which research is conducted need not be dictated by limitations of geocoding, but can instead be determined by our understanding of the biological processes and social and environmental exposures being investigated. Such research, however, must acknowledge how disparate geocoding success across subpopulations may affect findings.