Geocoding to Manage Missing Data in a Secondary Analysis of Community-Dwelling, Low-Income Older Adults

Abstract

Managing missing data in a secondary analysis is daunting, particularly if the data of interest were not included in the parent study design. The current study describes the use of geocoding to replace missing data from a parent study for a secondary analysis of socioeconomic and neighborhood characteristics in community-dwelling older adults who are dually eligible for Medicare and Medicaid. Geocoding was used to link participants’ addresses to data from the American Community Survey to replace missing income and neighborhood data. After geocoding, data completeness was 100% for neighborhood poverty and education composition, and 99.9% for income. Using geocoding provides the gerontological nurse researcher with a sample that is more reflective of the population. The current findings can be used to tailor neighborhood-centered interventions to promote health in low-income older adults.

Approximately 40% of psychosocial studies report missing data, which is a conservative estimate, with most studies either not adequately discussing or providing information regarding missing data (Bodner, 2006). Generalization of participant characteristics is challenging if data are missing from studies—particularly socioeconomic data, which can reveal many details about the health of the population of interest (Fuller-Thomson, Nuru-Jeter, Minkler, & Guralnik, 2009; Schulz et al., 2012; Szanton et al., 2015). Two key factors (i.e., income and neighborhood poverty) are considered to have a significant influence on the mental and physical health of older adults (House, 2002 ). Older adults in poorer neighborhoods are at greater risk for morbidity and mortality (Institute of Medicine, 2006 ; Link & Phelan, 1995 ; Phelan, Link, & Tehranifar, 2010), yet socioeconomic data such as income and neighborhood poverty are often incomplete or missing from psychosocial studies (Bodner, 2006).

Gerontological nurse researchers can benefit from exploring innovative new alternatives to managing missing socioeconomic data. For example, Wright et al. (2015) used Geographic Information System (GIS) technology to link addresses to missing income and neighborhood poverty data to test pathways to mental and physical health in low-income older adults. The purpose of the current article is to provide an exemplar of using GIS to geocode and link data from national datasets to interpolate missing socioeconomic variables in a sample of frail, Medicare-/Medicaid-eligible older adults who were part of a hospital-to-home nurse care management study. The intent is to provide a guide for gerontological nurse researchers to manage missing data from a parent dataset when conducting a secondary analysis.

BACKGROUND AND SIGNIFICANCE

When conducting a secondary analysis of a parent dataset, the gerontological nurse researcher is challenged with filling in the gaps of uncollected or missing data. Data can be missing for a variety of reasons, including participants’ unwillingness to respond, survey design, or human error (Squires & Tourangeau, 2009). In studies of older adults, missing data are more common compared to studies of younger adults due to problems with mobility, transportation, and multiple medical conditions—all of which can hinder research participation (DeCrane, Sands, Young, DePalma, & Leung, 2013; Hardy, Allore, & Studenski, 2009). Moreover, questions regarding income are sensitive for many participants and contribute to missing socioeconomic data (Krieger, 1992; Sterne et al., 2009). The scope of the problem of missing data is often underreported, which contributes to the complexity of identifying the magnitude of the problem. For example, in a large survey of 151 psychosocial research studies, 61.9% had no information about missing data or insufficient information. The extent of missing data is often a challenge to determine because most studies do not report details regarding missing data (Eekhout, de Boer, Twisk, de Vet, & Heymans, 2012).

Three categories of missing data exist: missing completely at random, missing at random, and missing not at random (El-Masri & Fox-Wasylyshyn, 2005; Tabachnick & Fidell, 2007). Data that are completely unrelated to other variables are categorized as missing completely at random. For example, if a participant does not return for a follow-up survey due to illness or lack of transportation, the data would be missing completely at random. Conversely, if data are missing because of participant characteristics such as gender, and participants were uncomfortable answering the questions, then the data are missing at random. However, if data are missing without any specific participant characteristics such as gender or age cohort, then data are missing not at random (El-Masri & Fox-Wasylyshyn, 2005; Tabachnick & Fidell, 2007).

The gerontological nurse researcher must determine how to manage missing data. The best approach to missing data is planning—prior to study implementation—a design that minimizes its occurrence. Effective strategies used to prevent missing data in studies of older adults include adjusting the study design to include transportation, frequent breaks during data collection, and assistive listening devices to reduce the chance of collecting incomplete data (Squires & Tourangeau, 2009; Tsai et al., 2009).

The most commonly used method of managing missing data is conducting a complete case analysis (Eekhout et al., 2012), which means participants without complete data are excluded from analysis. In a systematic review of 262 high impact studies, 81% used complete case analysis (Eekhout et al., 2012). The problem with using the complete case analysis approach is that results may be biased and include only participants who were willing and capable of completing the study (Eekhout et al., 2012; Wood, White, & Thompson, 2004). Imputation is another commonly used approach to manage missing data, which can be done by replacing the missing value with the case mean, group mean, or regression equation (El-Masri & Fox-Wasylyshyn, 2005; Tabachnick & Fidell, 2007). Imputation is preferred over complete case analysis because there is no reduction in sample size. In a secondary analysis of a parent dataset, complete case analysis and imputation can provide solutions to managing missing data. However, if data are completely missing because the variables were not included in the study design, then alternatives to data replacement must be considered.

When a parent study design does not include data needed for a secondary analysis (e.g., the current study), publicly available data from the American Community Survey (ACS) and U.S. Census Bureau can be used. The 2010 U.S. Census is a 10-item questionnaire that collected data regarding age, gender, race/ethnicity, households, housing units (e.g., apartment building), and group quarters (e.g., nursing home). Additional information not covered in the U.S. Census, but that can be found in the ACS, includes items about living arrangements, housing, disabilities, education, insurance coverage, occupation, and income (U.S. Census Bureau, 2010). Unlike the U.S. Census, which is administered every 10 years, ACS questionnaires are administered in 1-, 3-, or 5-year estimates. Sample sizes in the ACS are smaller than in the U.S. Census. Sample representation for Census 2010 was 1 in 1 household compared to 1 in 40 households for the ACS. Hence, the U.S. Census contains a larger number of cases, but the ACS has greater detailed information.

Public data from the U.S. Census and ACS are obtained through a procedure known as geocoding, which is the process of transforming addresses and places to coordinates using GIS technology (Cromley & McLafferty, 2012). Geocoded addresses are used to collect sociodemographic data from the ACS. Similar to navigational systems used in automobiles that integrate vehicle location with road network data and service locations, GIS integrates spatial data based on the data’s geographical location, permitting census data to be linked to geographic regions at a more refined level.

In GIS technology, using full addresses rather than zip codes is the most reliable source for geocoded data because participants are less likely to report an incorrect address than incorrect zip code. When addresses are used, additional measures are taken to secure the data under Health Insurance Portability and Accountability Act (HIPPA) regulations (U.S. Department of Health & Human Services [USDHHS], 2002). The de-identification of Protected Health Information (PHI) is maintained by covered entities such as hospitals, which require addresses and other identifiable data for billing purposes or other transactions (Croner, 2003). There are exemptions to requisite de-identification when PHI is used in research, payment, and/or public safety (Croner, 2003).

To comply with HIPPA regulations, the covered entities may enter into a business associate agreement with the noncovered entity to allow access to identifiable data while simultaneously securing the information from potential harm to individuals (Davenhall, 2002 ; USDHHS, 2002). The Summa Health System and University of Utah’s Department of Geography’s Digit Lab completed a business associate agreement for the current and parent study (Wright et al., 2015). In addition to binding contracts to protect individuals and data, the National Geospatial Program established standards of no greater than 1:100,000-scale maps (Cromley & McLafferty, 2012 ), which prevents the visible identification of exact locations on maps of the data source. Geocoding was used to link data from the ACS to augment income and neighborhood data from a parent study of community-dwelling, low-income older adults. The results of the path analysis describing factors that influence physical function and emotional well-being in Medicare–Medicaid enrollees were published previously (Wright et al., 2015).

METHOD

Sample

Cases from a parent study (Allen et al., 2011; Wright, Hazelett, Jarjoura, & Allen, 2007) of a randomized controlled trial of a hospital-to-home, interdisciplinary, nurse-led care management intervention of community-dwelling older adults were used in this exemplar of geocoding to augment a secondary data analysis. Enrollment criteria for the parent study included age 65 or older, dually enrolled in Medicare and Medicaid or Medicaid-eligible (individual income <$10,830 per year), and at least one deficit in activities of daily living or instrumental activities of daily living (Katz, Down, Cash, & Grotz, 1970). These individuals also had to have at least one of the following chronic conditions: chronic obstructive pulmonary disease, diabetes, cerebrovascular accident, coronary artery disease, hypertension, congestive heart failure, osteoporosis, or osteoarthritis. Recruitment for the parent study was conducted in the hospital prior to discharge to home. All cases from the parent dataset (N = 409) were used, and the parent study was a representative sample of the population in the secondary analysis study. The secondary analysis was approved by the Summa Health System and University of Utah Institutional Review Board following completion of a business associate agreement to protect identifiable information.

Measures

The current authors were interested in identifying data that provided missing information about income, poverty levels in neighborhoods occupied by participants, and education. Participant education was available from the parent dataset along with other demographic characteristics. Pre-retirement occupation reported in the parent dataset was used as a proxy to estimate participants’ yearly income (Herd, Goesling, & House, 2007; House, Lantz, & Herd, 2005 ; Kim & Durden, 2007 ; Zimmer & House, 2003). Yearly earnings based on pre-retirement occupation were grouped into five categories per census block: (a) service industry; (b) production, transportation, and material moving; (c) professional sales; (d) management, business, sciences, and arts; and (e) natural resources, construction, and maintenance.

Procedures

Participants’ addresses were used to identify their location within a census block. The tables containing information for income and neighborhood poverty were downloaded from American FactFinder, which contains U.S. Census community, city, and state data. TIGER/Line® Shapefile was then downloaded, which is a digital vector system that stores topographical information (Cromley & McLafferty, 2012). American FactFinder data and TIGER/Line Shapefile were joined in Esri® ArcGIS 10.0 desktop software, which was used to spatially link participants’ geocoded addresses to the corresponding census block with a 100% match rate. The aggregation of census blocks contained 1,200 to 1,800 residents (Cromley & McLafferty, 2012) identified by the unique Federal Information Processing Standard (FIPS) code assigned to each census block (FIPS Publications, 2010).

Geocoded addresses were assigned to census blocks, which are the smallest geographic unit used in the decennial census count. Using census block data provided an avenue to collect missing data from a primary dataset through national data. The U.S. government assigns a FIPS code to each census block, and that code was linked to the ACS data on poverty for the specific area. Once data were collected, participants’ addresses were removed and replaced by investigator-generated identification numbers. Data were collected to determine the median income by occupation and participants’ geographic location (census block).

Neighborhood poverty was defined as the percent of individuals living on ≤$10,830 per 1-person household in 2010, according to the 2010 poverty guidelines (USDHHS, 2010). Microsoft Excel® data were transferred to SPSS to analyze frequencies and percentages. Table 1 lists the variables that were collected and whether the data were collected from the parent study or ACS.

TABLE 1.

Variables and Data Sources

Variable	Data Source
Race	Parent study
Gender	Parent study
Age	Parent study
Years of education	Parent study
Address	Parent study
Occupation at retirement	Parent study
Percent of individuals whose income in the past 12 months was at or below the poverty line	American Community Survey
Occupation by median earnings in the past 12 months for civilian-employed populations by categories: (a) service industry; (b) production, transportation, and material moving; (c) professional sales; (d) management, business, sciences, and arts; and (e) natural resources, construction, and maintenance	American Community Survey
Neighborhood percent population (White or Black)	American Community Survey
Neighborhood percent population (65 or older)	American Community Survey
Neighborhood percent population (25 or older and graduated from high school)	American Community Survey
Neighborhood percent of population (age 25 or older with a Bachelor’s degree or higher)	American Community Survey

RESULTS

Of 409 participants, 346 were female (84.6%) and 63 were male (15.4%). The mean age of the participants was 74.16 years (SD = 7.3 years, range = 65 to 94 years). Most participants were White and had less than a high school education. Participants also retired early, and the average age at retirement was 57.42 years (SD = 11.18 years, range = 18 to 86 years). Table 2 shows the frequencies and percentages for income, neighborhood education and poverty, and degree of improvement of completeness of data after geocoding to account for missing data. Prior to data augmentation through geocoding, income and neighborhood poverty data were 100% missing.

TABLE 2.

Participant Characteristics from the Parent Study and Improvements in Data After Geocoding (N = 409)

Characteristic	n (%)	After Geocoding (n, %)
Race/ethnicity
White	289 (70.7)	Same
Black	116 (28.4)	Same
Other	4 (0.9)	Same
Income	Missing	381 (99.9)
Neighborhood poverty	Missing	28 (0.06 missing)
Education
Less than high school	181 (44.9)	Same
High school or GED	144 (35.5)	Same
Some college	73 (17.8)	Same
Bachelor’s degree or higher	7 (1.8)	Same
Missing	4 (0)	Same
Neighborhood education level	Missing	409 (100)
Insurance
Medicare and Medicaid	354 (86.8)	Same
Medicare only	1 (0.2)	Same
Medicaid only	54 (13.2)	Same

Participants’ income prior to retirement ranged from $9,983 to $113,542 per year, with a median income of $34,911. Most participants worked in the service industry, including housekeeping, food service, and child care. The remaining occupations included production (i.e., manufacturing) and transportation (i.e., truck driver); 6.6% of participants never worked. Neighborhood poverty levels were reflective of individuals living at a lower income in 2010. The average level of poverty per address was 22.33% (SD = 14.6%). Most participants were dually enrolled in Medicare and Medicaid. The percentage of the population that had a high school education was 83.9% (range = 56.1% to 98.7%), and 19.4% (range = 2.7% to 66.9%) for those with a Bachelor’s degree or higher.

The Figure depicts a geographical map of participants by census block. Each dot represents participants’ census block. The map also provides a visualization of sample distribution. Most participants were situated in urban areas and from the same county. Only four locations were outside of the map.

Figure.

Geographic Information System map of sample distribution by census block.

DISCUSSION

The percentage of missing data was improved using geocoding to link census blocks to ACS socioeconomic data. In the parent study used for the secondary analysis, income data were not part of the study design because only individuals eligible for in-home Medicaid services were enrolled. The primary focus of the parent study was to test the superiority of an interdisciplinary nurse-led care management intervention to reduce functional decline and improve health outcomes in low-income, frail older adults (Allen et al., 2011). Instead of collecting actual income data in the parent study, the data were dichotomized. Participants were asked if their income was at or above the cutoff point for Medicaid. Using GIS to geocode addresses, it was possible to store, manage, retrieve, and integrate 100% of the income, neighborhood poverty level, and education statuses from different sources to impute data that were not collected in the parent dataset.

Managing missing data by replacing it with U.S. Census and ACS data allowed comparison of the current sample to the U.S. dually enrolled Medicare–Medicaid population. Most of the current sample had less than a high school education, worked in service industries prior to retirement, retired at an early age, and were clustered in poorer urban areas. Participants’ occupations at retirement suggested a range of income that reflected the geographical settings. Neighborhood poverty for participants was slightly above the 15.1% poverty level for the U.S. population in 2010 (Proctor, Semega, & Kollar, 2016).

Level of education determines the life course that leads to occupation and income, predicting the health of older adults (Herd et al., 2007; Phelan et al., 2010). Participants’ average age at retirement was younger than 65, and the youngest retired at age 18. This young age at retirement may be the result of a disability (Proctor et al., 2016). Most Medicaid enrollees were disabled, and the second-largest subgroup in this population was older adults. Because most participants were enrolled in both Medicare and Medicaid, they may have been on Medicaid disability prior to receiving Medicare. Fourteen percent of Medicaid beneficiaries were eligible for Medicare and Medicaid (Paradise, 2015; Proctor et al., 2016).

By augmenting a parent dataset, the current findings provided insight into the socioeconomic and neighborhood characteristics of community-dwelling, low-income older adults. Gerontological nurses individualize health interventions based on the characteristics of these older adults. With geocoding, additional identified characteristics suggest that interventions should be tailored to a group that has limited education, worked in service industries, and was disabled at an early age.

The ACS contains additional environmental data that would be useful in research of aged populations, such as adequate plumbing facilities (e.g., flushing toilets, hot/cold running water) and individual data related to the functional status of residents residing in the community. Some of these individual data include hearing and vision impairment, difficulty dressing or bathing, need for assistance with errands or shopping, climbing stairs, and concentration and decision-making difficulties. Such information can be particularly useful to the nurse researcher interested in understanding health and wellness factors that influence accessibility and delivery of services for older adults in a given neighborhood.

LIMITATIONS

The sample of Medicare–Medicaid participants in the care management study had a diverse pre-retirement income, which can be explained by participants’ range of occupations (Holahan & Ghosh, 2005; Mandal & Roe, 2008). Although pre-retirement income can be used as a proxy, a limitation of the current study was that the number of years of retirement for each participant was not included. Only ACS 2010 data were used, which would not account for individuals who were in retirement longer than 10 years.

CONCLUSION

Use of GIS technology to geocode participants’ addresses offers an innovative approach to managing missing data in a secondary analysis. Rich data can be obtained about socioeconomic characteristics and neighborhoods from big datasets such as the ACS to design and test interventions to improve the health of community-dwelling, low-income older adults. Geocoding offers gerontological nurse researchers the opportunity to combine datasets by collaborating with others to leverage resources.