Impact of residential mobility on estimated environmental exposures in a prospective cohort of older women

Supplemental Digital Content is available in the text.

Background:

Longitudinal studies of environmental hazards often rely on exposure estimated at the participant’s enrollment residence. This could lead to exposure misclassification if participants move over time.

Methods:

We evaluated residential mobility in the Iowa Women’s Health Study (age 55–69 years) over 19 years of follow-up (1986–2004). We assessed several environmental exposures of varying spatial scales at enrollment and follow-up addresses. Exposures included average nitrate concentrations in public water supplies, percent of agricultural land (row crops and pasture/hay) within 750 m, and the presence of concentrated animal feeding operations within 5 km. In comparison to gold standard duration–based exposures averaged across all residences, we evaluated the sensitivity and specificity of exposure metrics and attenuation bias for a hypothetical nested case–control study of cancer, which assumed participants did not move from their enrollment residence.

Results:

Among 41,650 participants, 32% moved at least once during follow-up. Mobility was predicted by working outside the home, being a former/current smoker, having a higher education level, using a public drinking water supply, and town size of previous residence. Compared with duration-based exposures, the sensitivity and specificity of exposures at enrollment ranged from 94% to 99% and 97% to 99%, respectively. A hypothetical true odds ratio of 2.0 was attenuated 8% for nitrate, 9%–10% for agricultural land, and 6% for concentrated animal feeding operation exposures.

Conclusions:

Overall, we found low rates of mobility and mobility-related exposure misclassification in the Iowa Women’s Health Study. Misclassification and attenuation of hypothetical risk estimates differed by spatial variability and exposure prevalence.

What this study adds.

Our findings that mobility was predicted by smoking status, place of work, education level, drinking water source, and town size in the Iowa Women’s Health Study add to limited previous knowledge about mobility patterns in older, rural populations. We showed modest attenuation bias due to mobility-related exposure misclassification and demonstrated that such misclassification can increase with duration of the exposure period, greater spatial variability in the exposure, and lower exposure prevalence. These findings suggest that misclassification and subsequent attenuation bias in epidemiologic studies of long-term exposure to environmental factors can be mitigated by incorporating residential moves.

Introduction

Many prospective cohort studies use the participants’ address at study enrollment to estimate environmental exposures over time.^1–9 However, if participants change residences during the follow-up period and these changes are not accounted for, epidemiologic analyses may be subject to exposure misclassification and subsequent bias. Mobility-related misclassification may be greater for studies that require estimation of exposures over long periods of time, as needed in evaluating risk of chronic diseases like cancer. The rate of mobility will also depend on the population of interest and can vary by demographics, lifestyle, and other characteristics.^10–15 The effects of residential mobility on epidemiologic inference can be magnified for exposures that have low prevalence and high spatial variability.^{10,12,16–18}

A few cohort studies estimating exposure to outdoor air pollution have investigated the impact of residential mobility on epidemiologic associations and reported varying degrees of exposure misclassification and resultant bias.^11,19–23 A retrospective (Canada) and prospective (Sweden) exposure assessment of various air pollutants found high correlations (r = 0.70–0.80) between estimated exposure at enrollment and across all addresses for the exposure period.^19,22 However, Hystad et al.¹⁹ showed that despite high correlation, about half of individuals were misclassified into different exposure quintiles. Only some of the aforementioned studies quantified the impact of mobility on the exposure–disease relationship.^11,21,23 Additional studies are needed to evaluate mobility-related misclassification of environmental exposures beyond air pollution, and in rural areas, as previous studies are limited to urban settings.

The Iowa Women’s Health Study (IWHS) is a cohort of postmenopausal women in which several investigations on cancer have used the enrollment address as the basis for historical exposure assessment of environmental risk factors.^{1,4–6,24–27} Since residential histories were not collected in this study, we aimed to describe residential mobility and population characteristics related to mobility using address data collected over 19 years of follow-up. We estimated mobility-related exposure misclassification and subsequent bias in risk estimates for several environmental exposures.

Materials and methods

Study population

The IWHS is a population-based prospective cohort of 41,836 postmenopausal women enrolled in 1986 and followed for cancer incidence and mortality. Study enrollment and data collection have been described.²⁸ Briefly, women aged 55–69 years were randomly selected from valid Iowa driver’s licenses and mailed an enrollment survey to collect information about dietary intake, demographics, family history of cancer, and medical and reproductive history. The participation rate at enrollment was 42%; response rates among living subjects for five follow-up surveys mailed in 1987, 1989, 1992, 1997, and 2004 were 91%, 89%, 83%, 79%, and 68%, respectively. The Institutional Review Boards of the University of Minnesota and University of Iowa approved the IWHS.

Residential mobility

We excluded 186 participants whose date of death was before the first follow-up (1987), because no prospective mobility information was available for these individuals. Mobility was assessed for the remaining 41,650 (99.6%) IWHS participants. Participant address information was collected over a 19-year period from enrollment (1986) and at the five follow-up surveys, with the last information collected in 2004. Addresses were also updated through routine tracing efforts and linkages with cancer and death registries for each follow-up period. We did not trace additional moves between survey periods. Addresses were geocoded as latitude and longitude coordinates with ArcGIS (v. 9.3). Geocoded addresses were linked to US Census 2000 Incorporated Places boundaries to describe town size for addresses within an incorporated place (population <10,000, 10,000 to <100,000, or ≥100,000) or otherwise categorized as an unincorporated place. We defined “well-geocoded” addresses as geocodes placed at the complete street address or nearest intersection.²⁹ Overall, 78% of enrollment addresses were well-geocoded; poorer geocode rates were observed for unincorporated places (eTable 1; http://links.lww.com/EE/A105). We censored addresses for follow-up surveys following death (n = 10,823 total deaths before 2004). We assessed residential mobility for each follow-up period by comparing address pairs between each consecutive survey. Among well-geocoded addresses, a move was defined when the distance between addresses was greater than 250 m, a distance selected to account for positional error, a frequently observed phenomenon especially among rural addresses.^30,31 If one or both addresses in a pair were not well-geocoded, we instead compared the combination of city/town and ZIP code and recorded a move if both attributes were mismatched.

Environmental exposures

We evaluated several environmental exposures previously investigated in relation to cancer risk in the IWHS^{4–6,25,26,29,32} and in other studies in Iowa.^33,34 Exposure to average annual nitrate–nitrogen (NO₃-N mg/L) in drinking water was estimated based on the city of the participant’s address using public monitoring data (1986–2004) for finished water samples from Iowa public water supplies (PWS)³⁵ (eFigure 1; http://links.lww.com/EE/A105). Exposure to agricultural land was estimated using land cover classification data in 1992 (the mid-point of follow-up) from the US National Land Cover Database³⁶ (eFigure 2; http://links.lww.com/EE/A105). We calculated the proportion of agricultural land (row crops and pasture or hay, respectively) within a 750 m circular buffer around the home, a distance previously associated with higher pesticide levels in homes.³³ Finally, we evaluated exposure to concentrated animal feeding operations (CAFOs) using publicly available information on permitted CAFOs (2003–2011) from the Iowa Department of Natural Resources³⁷ (eFigure 3; http://links.lww.com/EE/A105) and enumerating CAFOs within 5 km of the address. The percent of IWHS participants by county of residence at enrollment is shown in eFigure 4; http://links.lww.com/EE/A105.

Bias analysis

Study participants with well-geocoded addresses at enrollment and all follow-up periods (n = 29,226; 70% of the cohort) were assigned exposures for each address and included in the bias analysis. Participants excluded from the bias analyses were similar to those included, with the exception of town size at enrollment, drinking water source, and occupation (a greater proportion of excluded participants lived in an unincorporated place, used a private well, and/or were homemakers); eTable 2; http://links.lww.com/EE/A105. Exposure was censored at follow-ups where addresses were outside of Iowa (n = 2,002 participants [7%]). We assigned exposures to participants in two ways: (1) based on the enrollment address, therefore assuming the participant remained at the enrollment address for up to 19 years (“stability method”) and (2) based on calculating a duration-based average across all addresses during follow-up (“mobility method”). Exposures were estimated at each new address starting from the year of the follow-up survey when the move was observed. Exposures for agricultural land and CAFOs near homes were assigned for all 29,226 participants. Our analysis of NO₃-N included 20,057 participants who reported using a PWS in 1989 for which nitrate monitoring data were available.

We classified participants as exposed or unexposed using distribution- and regulatory-based thresholds that have been evaluated in previous IWHS analyses. We classified continuous exposures (NO₃-N mg/L and agricultural land percentage within 750 m of the home) based on the highest quartile (<75th and ≥75th percentile). We also evaluated exposure to NO₃-N at half the US Environmental Protection Agency maximum contaminant level (MCL) of 10 mg/L (1/2 MCL, <5 and ≥5 mg/L).³⁸ For CAFOs, we classified a participant as exposed if there was one or more CAFO within 5 km of the address.

Statistical analyses

We calculated the mobility rate (probability of moving at least once) and the distribution of Euclidean distances (km) between moves with well-geocoded address pairs overall and for each follow-up. We computed the proportion of moves that occurred within the same city and, for moves to a different city, the proportion of moves to a larger city (based on town population). We grouped participants into three categories according to their mobility over the 19-year period: “nonmovers,” those never observed to move; “ever movers,” those who moved at least once; and “frequent movers,” a subset of ever movers who moved two or more times. We examined differences in characteristics between movers and nonmovers, including age, race, body mass index, smoking status, occupation, education, town size at enrollment, follow-up participation, drinking water source (reported in the 1989 follow-up survey), historical duration at the enrollment residence (reported in the 1987 follow-up survey), and prospective residence duration (years after enrollment until a participant moved or was censored due to loss to follow-up or death). We estimated the odds ratios (OR) and 95% confidence intervals for the likelihood of moving between subsequent follow-ups using a generalized estimating equation (GEE) model. To account for the varying number of years between follow-ups, we used the natural logarithm of years between follow-ups as the offset. Categories of race, body mass index, smoking status, occupation, education, and drinking water source were based on information collected at a single time point (enrollment), while age and town size of the address varied by follow-up. Missing-level indicators were included for variables if the percentage of participants with missing values was greater than 2%. Participants with missing values for education level, race, smoking status, and/or town size were excluded from the GEE models (n = 1,114).

Bias analysis

Using the mobility method as the gold standard in comparison to the enrollment address, we computed the sensitivity and specificity of correctly assigning exposure at 3, 11, and 19 years of follow-up (survey periods beginning [1989], mid-point [1997], and end [2004]). Sensitivity of exposure assignment was defined as the percentage of truly exposed participants that were correctly classified as exposed (e.g., ≥75th percentile and 1/2 MCL NO₃-N; ≥75th percentile agricultural land; or ≥1 CAFOs within 5 km). Specificity was defined as the percentage of truly unexposed participants (e.g., <75th percentile and 1/2 MCL NO₃-N; <75th percentile agricultural land; or 0 CAFOs within 5 km) that were correctly classified as unexposed. We estimated the attenuation of the observed OR (OR_obs) due to misclassified exposures in a hypothetical nested case–control study of cancer assuming a true OR (OR_true) of 2.0, using an approach described by Blair et al.³⁹ We note that our estimated attenuation assumes there is nondifferential misclassification of exposure, in that the sensitivity and specificity are independent of the outcome.

For sensitivity analyses, we explored changes in misclassification with decreasing exposure prevalence. For NO₃-N and agricultural land exposures, we used cut points at or above the 50th, 75th, and 95th percentile. For CAFOs, we used increasing counts of ≥1 (40% exposure), ≥2 (25% exposure), and ≥8 operations (5% exposure) near the home.

RESULTS

The cumulative mobility rate of IWHS participants over 19 years of follow-up was 4% at the first follow-up (1987), 9% at the second follow-up (1989), 12% at the third follow-up (1992), 21% at the fourth follow-up (1997), and 32% at the fifth follow-up (2004). Almost half of all moves occurred within the same city (41%). For moves to a different city, 54% of destination cities were larger than previous cities. The median distance between addresses was 6.5 km (Table 1). Participants who moved from unincorporated places traveled farther (median, 10.8 km). The distance between addresses for follow-ups one through four were similar (median, 5.3–5.7 km) but greater between follow-up four to five (median, 8.7 km) (eTable 3; http://links.lww.com/EE/A105).

Table 1.

Characteristics of the IWHS participants who did not move, ever moved, and moved frequently during 19 years of follow-up

We observed differences between nonmovers (n = 28,381; 68%), ever movers (n = 13,269; 32%), and frequent movers (n = 3,090; 23% of ever movers) across demographic, health, and lifestyle characteristics (Table 1). A greater proportion of ever movers used to smoke, worked outside the home, had more than a high school education, and participated in the fifth follow-up compared with nonmovers. Similar and elevated trends were observed for the subset of frequent movers. Additionally, a greater proportion of frequent movers lived in larger towns compared with nonmovers. Most participants who did not move were at their enrollment address for 20 years prior (61%).

GEE model results indicated that when accounting for all other variables, mobility was significantly predicted by smoking status, occupation, education, town size, and drinking water source (Table 2). The odds of moving were greater for former or current smokers, those who worked outside the home, and those who had more than a high school education. Compared to living in a town size of <10,000, the odds of moving were 15% and 17% greater for those living in a town of 10,000–<100,000 and ≥100,000, respectively, and 18% greater for those in an unincorporated place. Compared to using a PWS as a primary water source, mobility was greater for those using bottled water or other source and less likely for those using a private well or unknown/missing water source.

Table 2.

GEE model^a,b for predicting mobility

Exposures to NO₃-N in drinking water, pasture/hay, row crop, and CAFOs assigned at the enrollment address were similar to the duration-based average across all addresses (Table 3). Sensitivity and specificity of exposure assignment at the enrollment address decreased for all environmental exposures with increasing years of follow-up as a greater proportion of women moved (Table 4). Sensitivity was lowest for NO₃-N exposure ≥75th percentile (93.5%) and pasture/hay exposure (94.1%). Minor differences were observed for pasture/hay and row crops. Sensitivity of NO₃-N exposure ≥5 mg/L (99.1%) was greater than exposure at ≥75th percentile even though exposure prevalence was lower (17%). The lowest specificity was observed for exposure to one or more CAFOs (40% prevalence) at all years of follow-up (96.9%). Attenuation of the OR_true for the study period ranged from 5% to 8% for NO₃-N, from 9% to 10% for agricultural land, and 6% for CAFOs.

Table 3.

Exposure distributions for participants with well-geocoded addresses (n = 29,226) assigned at the enrollment address and averaged across all addresses for 19 years of follow-up

Table 4.

Accuracy of exposure assignment when using the enrollment address^a for 3, 11, and 19 follow-up years and attenuation in an odds ratio (OR) of 2.0 for a hypothetical cancer analysis for 19 years of follow-up

Attenuation of the OR associated with NO₃-N exposure defined at or above the 50th and 95th percentiles (5%) was somewhat less than observed at the 75th percentile (8%); eTable 4; http://links.lww.com/EE/A105. Accuracy of CAFO exposures was similar for different cut points. The accuracy of agricultural land was influenced by exposure prevalence; at 5% exposure prevalence, sensitivities were lowest (range 86.0%–91.0%), and specificities were highest (range 99.3%–99.5%). Attenuation of the OR_true ranged from 5% to 8% for NO₃-N, from 9% to 16% for agricultural land, and from 6% to 7% for CAFOs.

DISCUSSION

Our study found the majority (68%) of the IWHS cohort remained at their enrollment address during the 19 years of follow-up. We found that mobility was predicted by occupation, education, smoking status, town size, and drinking water source. Overall, we observed low misclassification of environmental exposures and resultant bias of hypothetical risk associations. We found that mobility-related misclassification was influenced by years of follow-up, the spatial variability of the environmental measure, and the prevalence of exposure.

The proportion of participants in the IWHS who moved during 19 years of follow-up (32%) was lower than that in other more urban and younger populations, including the Canadian National Enhanced Cancer Surveillance System case–control study (60% mobility; 20-year residential history),¹⁹ the European Study of Cohorts for Air Pollution Effects subcohort in Sweden (70% mobility; 14.6-year follow-up),²² and the Cincinnati Childhood Allergy and Air Pollution Study prospective birth cohort (54% mobility; follow-up birth to 7 years).¹¹ The rate of mobility in our population was comparable to that in a study of Canadians 55 years and older, of which 32% moved over a 12-year period and 30% lived in a small city or rural area.¹⁵ We observed a median movement distance of 6.5 km (4 miles) among movers, which is similar to studies of younger and more urban populations in the United States that report median distances of 3.2¹⁰ and 4 miles.^11,40

We found that mobility varied by several demographic and residential characteristics, including age, education, occupation, smoking status, town size, and drinking water source. Mobility was less likely for homemakers and farmers compared to those who worked outside the home. Though there are few existing studies to directly compare this result, a study of older Canadians found participants who returned to work were more likely to move.¹⁵ The odds of mobility were greater for current and former smokers in the IWHS. These results agree with a study of pregnant women, which found greater mobility for those who had reported ever smoking.¹² Similar to our findings, older residents (60+ years old) in Rhode Island who were more educated were more likely to move.¹⁴ The greater likelihood of moving among women living in an unincorporated place reported in our study is comparable to another study of older adults who were more likely to move if they lived in a rural area.⁴¹ Increased mobility in rural or unincorporated places among older adults may be explained by a perceived inability to remain independent in their homes and communities.^41,42 However, we also found mobility was greater for women living in larger towns, similar to patterns of higher mobility in urban areas observed in one population of younger women.⁴³ Although not evaluated in our study, other predictors of mobility that may be relevant to older populations include marital status, home ownership, number of children, and use of community services.^13–15,42

We did not find appreciable differences in exposure distributions when exposures were estimated only at the enrollment address compared to an average across all addresses, suggesting mobility in the cohort was not likely motivated by exposure levels. Other environmental exposures, such as traffic-related air pollution, may decrease as people relocate away from the exposure source (e.g., major roadways).^11,20 We demonstrated that exposure misclassification from mobility increased over time in the IWHS by reevaluating exposure at the beginning, middle, and end of follow-up, indicating that the length of the exposure period will impact the extent of mobility-related misclassification. Even though misclassification increased over time, sensitivity and specificity remained >90% for all exposures at the end of the 19-year follow-up period. The low levels of misclassification we observed can be in part attributed to the stability of the population at the enrollment address because only 32% of the population moved during follow-up. Similarly, a study of older Canadians also with a 32% mobility rate observed no systematic biases in the associations between place-dependent variables (e.g., housing quality, material and social deprivation, urbanicity) when mobility was ignored.¹⁵ In our analysis, the sensitivity of NO₃-N exposure in drinking water at the end of follow-up was lowest at ≥75th percentile (93.5%) compared to both ≥5 mg/L (99.1%) and ≥95th percentile (97.5%). The lower sensitivity at the ≥75th percentile despite higher exposure prevalence may be attributed to the crude spatial scale of NO₃-N (city-level), which results in multiple people in the same city having identical exposure values and, consequently, percentile cutpoints inexact due to having many tied values. The accuracy of agricultural land exposure assignment was typically lower than NO₃-N at various cut points and can be in part explained by differences in the spatial variability of these two exposures.^17,40 Mobility may result in misclassification of exposure to agricultural land within 750 m buffers, regardless of a change in city of residence, while NO₃-N exposure would not be affected by intra-city moves (41% of moves in our data). The risk estimates of CAFOs were also less attenuated than agricultural land and could be in part explained by the larger buffer size (5 km) and limited range of possible exposures to CAFOs (i.e., a maximum of 29 facilities within 5 km of the home).

In an effort to evaluate relevant misclassification and epidemiologic bias, we used environmental exposures that have been previously assessed and analyzed for cancer risks in the IWHS^4–6,26,29 and have been examined in studies of other populations in Iowa.^33,34 Since our current study assessed prospective exposure, we cannot directly estimate attenuation of cancer risks in previous investigations, which evaluated historical exposures. Historical duration of >20 years at the drinking water source has been used previously to define subgroup analyses of cancer risks to control for potential influences of mobility.⁵ We noted similarities between historical and prospective mobility rates in this cohort, as the proportion of participants remaining at the enrollment address for 19 years of follow-up (68%) was similar to the proportion of participants reporting a historical duration of >20 years at the enrollment address (69%). Thus, we expect limited mobility-related misclassification in previous analyses due to exposure estimation using only the enrollment address.

Our study has several strengths. By using prospectively collected residential information, we did not have to rely on participants’ recall of previous addresses or retrospective tracing of addresses using commercial software. Our analyses provide context for studies of environmental exposures over long time periods, including insight about expected exposure misclassification and attenuation of the exposure-disease outcome due to unaccounted changes in residential addresses. We acknowledge some limitations to this work. Because residential addresses were collected at each follow-up period, we were not able to capture moves as they occurred between surveys. However, our rate of frequent mobility (23% among movers) is comparable to the rate of frequent mobility (more than one move; 23%) in a study of older Canadians that collected address information biennially.¹⁵ While we used tracing efforts to update address information for nonrespondents in follow-ups, participation rates were higher for those who moved, suggesting there was a diminished ability to detect mobility for nonrespondents. Because we did not have information about drinking water source at each survey, we assumed participants who reported a PWS in 1989 continued to use a PWS at their other residences. In our hypothetical case–control analysis, we assumed exposure misclassification was nondifferential with respect to outcome. Although exposure misclassification is often nondifferential when exposure is assessed through secondary data linkages,²² it is possible that mobility can result in differential misclassification if characteristics of the study population (e.g., education, smoking) are associated with both mobility and the outcome under study.^11,12 We note that the overall rates of exposure misclassification due to mobility in our study were extremely low. Therefore, even if the mobility-related exposure misclassification was differential, the effects on the estimated odds ratio should remain relatively minimal. Further exploration of the association between mobility and its correlates with specific health outcomes of interest is warranted. Finally, addresses within city or town boundaries had better geocoding accuracy, which led us to exclude more participants in rural areas in our bias analysis.

CONCLUSIONS

We evaluated residential mobility in a prospective cohort and found that the likelihood of participants changing residences differs by their demographics and other characteristics. We demonstrated how not incorporating residential changes in environmental exposure assessment can lead to misclassification and attenuation in the relative measure of association. Although we found low mobility-related exposure misclassification in the IWHS, we note the potential impacts on epidemiologic inference can vary by the characteristics of the study population (mobility rate), choice of exposure metrics (duration of exposure, thresholds used to define exposure), and characteristics of the exposure itself (spatial variability). Although limited to the exposures we evaluated in this study, our findings suggest that misclassification and attenuation bias in epidemiologic studies of long-term exposure to environmental factors can be mitigated by incorporating residential moves. Future studies should consider characteristics specific to their participants that could impact mobility and resultant bias.

Conflicts of interest statement

The authors declare that they have no conflicts of interest with regard to the content of this report.

Acknowledgments

We thank Abigail Flory of Westat Inc. for GIS support on this project.