The clinical data was obtained from electronic health records of patients with diabetes receiving uniform access to care within a large, integrated, healthcare delivery system Kaiser Permanente Northern California (KPNC) from 2007 to 2011. Patients with type 1 or unknown type diabetes, histories of lower extremity amputation or pregnancy or cancer diagnosis within the study period and 1-year prior were excluded. Patients were retained for analysis if they had at a geocodable address record for at least two years over the period from 2006 to 2009 in one of nine Bay Area counties. An additional 5.2% percent had only one BMI measure over all four years and were excluded from the analysis.

Foreclosure data included address-level residential deed and housing price data compiled by DataQuick [40] from the nine Bay Area counties between 2006 and 2009. Data on area-level demographics were also collected from the American Community Survey (ACS) [41]. Data from the Local Area Unemployment Statistics database were used for county-level unemployment rates [42]. Additional individual-level data was based on a sub-set of patients (n = 8,923) within our population who had completed the Diabetes Study of Northern California (DISTANCE) survey at baseline, in 2006 [43]. The Institutional Review Boards of Kaiser Permanente Northern California and University of California, Berkeley approved this study. Clinical records were anonymized and de-identified prior to analysis.

The main outcome for this study, body mass index (BMI, kg/m²), was an annual average of all measurements during patient ambulatory visits from 2007 to 2010. The mean number of BMI measurements per patient was 1.3 per year and 8.1 over the four-year period.

We define foreclosure as the event in which the residential deed is transferred to the new owner. Our data did not differentiate between properties that were transferred back to the bank (real-estate owned) and those sold at auction. We used ArcGIS 10 and Data Scientist Toolkit in R to assign geographic coordinates to the foreclosures. Only 19 (<1%) of the foreclosure addresses not geocodable.

The primary exposure of interest was the annual foreclosure rate within the patient’s census block. The number of foreclosures was divided by the number of housing units in each block using data from the US Census 2000. In this region, the length of an average block is 100m, the distance at which a foreclosure is expected to have economic spillover effects on the values of neighboring properties [15].

Time-varying neighborhood-level covariates from 2006 to 2009 were used to control for area-level variation. We adjusted for mean housing prices at the zip-code level and county unemployment rate. Additional block-level variables were created using the American Community Survey (ACS) including population-density, proportion of the population under the federal poverty line, proportion of owner-occupied units, and proportion of the population identified as White.

The following individual-level time-varying covariates were collected from 2007 to 2010. Medicaid status (1/0) was included as a proxy for change in income or employment status. We included a Charlson comorbidity index score [44], which predicts the ten year mortality for a patient with a range of conditions, and annual indicators for use of medications clinically associated with weight change (insulin, oral diabetes medications, and psychiatric medications such as selective serotonin reuptake inhibitors (SSRIs) that can cause weight gain or weight loss.)

For the subsample with DISTANCE survey responses, we created measures of income (continuous), non-housing wealth (1 = $10k+, 0 = less than $10k), education (0 = graduate degree+, 1 = less than high school diploma, 2 = high school diploma, 3 = bachelors degree), employment status (1 = employed full-time, 0 = not employed), and partnership status (1 = married or co-habiting, 0 = non-married, divorced, separated, widowed). For the same subsample, we created homeownership status (1 = owner, 0 = renter) included individuals without missing data on housing status. Individual homeownership status was coded as missing for who reported living “rent-free” (i.e. with a relative) or moved at least once from 2007 to 2009 because we were unable to determine if they rented or owned for the entire period.

To examine the relationship between foreclosures and BMI, we fit a series of models of individual BMI at time t on the block foreclosure rate in the prior year. All statistical analyses were performed with Stata 13. The general form of the model was specified as:

Y_it = β_o + β₁F_it−1 + β₂Z_it−1 + β₃C_it + β₄X_i + β₅yearD_t + u_i + v_it, where Y_it is a measure of the mean BMI level of individual i in year t; F_it−1 is the block annual foreclosure rate for individual i in year t-1; Z_it−1 is a vector of lagged area-level controls for individual i in year t-1 (unemployment, housing prices); C_it is a vector of individual time-variant covariates including indicators for Medicaid status and medication use; X_i is a vector of individual time-invariant covariates such as sex and race; yearD are dummies for year t; u_i is an individual fixed-effect, and v_it is the time and individual specific error term.

Our primary parameter of interest, change in BMI for each one-unit change in the block foreclosure rate in the prior year, was estimated using an individual fixed effects approach by estimating a fixed-effect (u_i) for each individual. The inclusion of individual fixed effects allowed us to control for observed and unobserved factors that do not change over time that might be correlated with both individual body mass index and neighborhood foreclosure rate such as credit scores and propensity to be present-biased.

As foreclosures and health status are not randomly distributed, our estimates might differ based on health insurance or demographic characteristics. First, we restricted our sample to those under 65 who used Medicaid at least once during the period (n = 4,463) and for the entire period (n = 3,138). Next, we stratified our results based on the individual’s baseline age and created three groups of adults (20–49, 50–64, 65+). Then we stratified by groups potentially differentially affected by the recession using race/ethnicity [45].

In order to test one of the mechanisms by which the community foreclosure rate is hypothesized to influence BMI, we stratified by the main fixed effects models by homeownership status (owners or renters) among the DISTANCE respondents who lived in the same address for the period. If homeowners experience a nearby foreclosure as a stressful event because it signals a future potential loss of housing wealth and the stress response provokes an increase in BMI, we expect the relationship between foreclosure rate and BMI to be greater among owners than renters [46].

We conducted a series of sensitivity analyses to examine the consistency of the study results to different analysis specifications. First, we created two additional measures of foreclosure exposure and repeated the study analyses using these alternative definitions: a census block group foreclosure rate and a count of foreclosures within a 1-km Euclidean radius around the patient’s block centroid. Second, in order to facilitate comparisons to the single existing study on foreclosures and body mass index, we replicated the analytic approach used by Arcaya et al [20]. We used a linear mixed model fit with maximum likelihood containing both fixed and random effects with an independent covariance structure [47] to regress the lagged number of foreclosures in the patient’s census block on BMI, adjusting for education, income, age, race, sex, age, population density, foreclosures per 1km, and a quadratic time trend. Finally, we assessed the robustness of our main estimates to different lag structures, including two-year lag and contemporaneous specifications.

There were 105,919 individuals in the study population that met inclusion criteria and had at least two measures of BMI between 2007 and 2010. Over half (58%) of the sample was clinically observed in all four years and 17.3% of the individuals moved at least once during the study period.

The means and standard deviations for the time-varying variables are reported in Table 1. The average BMI (kg/m²) in this cohort was 31.15 (s.d = 1.2 units) over the period. For a person 5’6” and 193 pounds, the standard deviation is equivalent to 7 pounds. There were 0.3 foreclosures in the average block (20 homes) and the within-person standard deviation was 0.6.

Table 2 indicates substantive differences in baseline health, demographic and neighborhood characteristics between those whose blocks were least and most affected by the foreclosure crisis. Those who lived in blocks hit hardest by foreclosures during 2008 (i.e. top quintile) were more likely to be obese, Black, under 65, and living with a spouse or partner, while being far less likely to have a Bachelors degree, employment, and non-housing wealth of over $10,000 at baseline. These individuals also lived in neighborhoods at baseline with an average block foreclosure rate that was double that of the bottom quintile at baseline. The neighborhoods in the top quintile also had lower housing prices, more poverty, a higher rate of owner-occupied units, a lower proportion of Whites, and a lower population density.

The individual fixed effects models for the full cohort showed no statistically significant association between block foreclosure rate and BMI (Table 3). We observed a clinically irrelevant statistically significant relationship for the unadjusted model (Model 1), β = -0.007, 95% CI [-0.009, -0.005]. The addition of the year fixed effects attenuated the estimate (Model 2), β = 0.001, 95% CI [-0.001, 0.003], and coefficients on year indicators were negative, indicating that BMI decreased from 2007 to 2010. The model adjusting for unemployment rate and mean housing prices (Model 3) and the full set of individual time-varying covariates (Model 4) left the coefficient block foreclosure rate unchanged.

For each one standard deviation increase in the county unemployment rate, individual mean BMI increased by a half pound on average (p<0.05). Decreases in mean housing prices had a statistically significant relationship association (p<0.05) with increased BMI, however the magnitude of the effect was less than a tenth of a pound gain for one a one standard deviation change in prices, and was too small to be clinically meaningful. Starting insulin or oral medication during the study period was associated with an increase of approximately two pounds (p<0.001).

Table 4 contains results among Medicaid patients. Among those under 65 who were enrolled in Medicaid for at least one year during the period (Model 5, β = 0.004, 95% CI [-0.02, 0.03]) and for all four years (Model 6, β = 0.006, 95% CI [-0.01, 0.02]), there was no statistically significant effect of block foreclosure rate on BMI, contrary to our hypothesis.

S1 Table contains sub-group specific results stratified by age and race, while S2 Table provides contextual descriptive statistics by race. Estimates for patients over 50 years of age and for non-Hispanic Whites were similar to the full cohort, β = 0.001, 95% CI [-0.003, 0.004]. Among those ages 20 to 49 and among Asians, the coefficient on the foreclosure rate was negative and statistically insignificant. For Blacks, for a three percent (~ 1 standard deviation) increase in unemployment rate, there was a statistically significant (p<0.001) half unit increase in BMI or roughly three-pound gain for an average 5’6” and 193 pound person (β = 0.189, 95% CI [0.079, 0.300]). For those ages 20 to 49 and 65+, there was a statistically significant (p<0.001) change in BMI for every one standard deviation change in housing prices, however this effect was of a negligible magnitude. Among the DISTANCE cohort (not shown), there was no difference between the relationship between foreclosure rate and BMI among homeowners and renters.

The main models were fit using alternate exposures, the block group foreclosure rate and the number of foreclosures within 1-km from the patient’s block centroid, and yielded similar results to the block foreclosure rate, β = 0.001, 95% CI [-0.001, 0.003] (not shown).

For those who used Medicaid at least once during the period, there was a statistically significant (p<0.05) decrease of 0.26 kg/m² (~ 2 pounds) for every one percent increase in the block group foreclosure rate (Table 4, Model 7). We found no association between block group foreclosure rate and BMI among those who were enrolled in Medicaid for all four years (Model 8, β = -0.098, 95% CI [-0.287, 0.092]). There was no association between foreclosures per 1 km and BMI for those on Medicaid ever or for all four years (Models 9 and 10, 95% CI [-0.002, 0.003]).

The effect of unemployment on BMI among those who were observed to use Medicaid for all four years was robust across spatial measures of foreclosures; for every 3 percent (~ 1 standard deviation) increase in the unemployment rate, there was a statistically significant (p<0.05) increase of 1 kg/m² (~ 7 pounds).

We observed no association between body mass index and a 2-year lagged block foreclosure rate or contemporaneous block foreclosure rate. Finally, the mixed effects model that replicated the study by Arcaya et al [20] yielded no statistically significant association between block foreclosure rate and body mass index (n = 57,791, β = -0.0001, 95% CI [-0.002, 0.002] (data not shown).