HCHS/SOL is a longitudinal cohort of 16,415 self-identifying Hispanic/Latino adults who were at least 18 years of age at enrollment and were recruited from the Bronx, Chicago, Miami, and San Diego [28]. More details on the HCHS/SOL sample and study design have been previously reported [29, 30]. Briefly, the first in-person visit and enrollment occurred between 2008 and 2011 in each of the four cities. Yearly follow-up phone calls were conducted, and a second round of in-person visits (visit 2) occurred between 2014 and 2017.

Questionnaires and clinical examinations were administered at both in-person visits. The questionnaires assessed demographics, current health and medical history, socioeconomic status, and residential information [30]. Clinical examinations included the collection of blood and urine for various assays and analyses [30]. The yearly follow-up phone calls contained questions on general health, potential updates from their doctors or health professionals, and details on any potential hospitalization or emergency room visits that occurred since the last follow-up, as well as any updates to their residential location [30]. The study population for this analysis included participants who completed the baseline and second visits, were free of diabetes at the baseline visit, and had geocoded residential address data at both visits. All participating institutions received approval from their respective institutional review boards, and written informed consent was received from all participants.

Ascertainment of diabetes occurred at baseline, during the yearly follow-up phone calls, and at visit 2. During calls and visits, participants were asked if a doctor or health care professional had told them they had diabetes or high sugar in their blood and if they were receiving any treatment [30]. However, only during the in-person visits was blood collected and assayed to measure fasting plasma glucose (FPG) and glycosylated hemoglobin (HbA1c) levels [30]. A 2-h oral glucose tolerance test (OGTT) was also conducted on participants who did not self-report a diabetes diagnosis and on participants with FPG < 150 mg/dL [31]. Ascertainment of diabetes in the HCHS/SOL cohort was not able to distinguish between type 1 and type 2 diabetes.

For this analysis, we use two definitions of diabetes. Each definition is based on the American Diabetes Association criteria [32] and on data collection practices. Our primary definition for diabetes was having ascertainment of diabetes at visit 2 and any of the following: an FPG ≥ 126 mg/dL; an HbA1c level ≥ 6.5%; a post-OGTT glucose ≥ 200 mg/dL; or use of antihyperglycemic drugs. Our secondary definition, while also limited to those with ascertainment of diabetes at visit 2, included those who fit the primary definition of T2D and those who self-reported diabetes or high sugar in their blood during the in-person visits or the annual follow-up phone calls.

Pre-diabetes was assessed at baseline and at visit 2. Based on the American Diabetes Association criteria, pre-diabetes was defined as having an FPG ≥ 100 mg/dL and < 126 mg/dL and an HbA1c level ≥ 5.7% and < 6.5% [32].

In order to generate an interpolated disinvestment measure for each residential address, we conducted a virtual street audit using an established, validated approach and Google Street View [25, 26]. For each of the four cities, we selected 1000 street locations to be virtually audited via Google Street View. Locations were selected within the 351 census tracts where participants lived at baseline using a 600-point grid. To increase precision in neighborhoods where most participants lived, we embedded an additional 600-point oversample in the census tracts with the top quintile of HCHS/SOL participants in each study center. All points were randomly jittered up to 250 m. We used the Computer-Assisted Neighborhood Visual Assessment System (CANVAS) to conduct the virtual street audit [33]. We uploaded our set of coordinates to CANVAS, which tried to match each coordinate to Google Street View imagery within 50 m. If there were no matches within 50 m, the location was not used. For each location, auditors answered a set of 53 questions related to neighborhood disinvestment based on the most recent image (range, 2007–2023; median, 2022) and the oldest image (range, 2007–2019; median, 2011) using the drop and spin method (Table S1). The drop and spin method refers to how CANVAS accessed Google Street View at each location, and auditors were only able to spin around to answer the questions. They were not able to “walk” up and down the street, as this could change the date of the images.

Audit questions were taken from previously validated audits designed to assess neighborhood disinvestment and pedestrian safety [26, 34, 35]. Examples of the questions auditors had to answer were “Is there garbage, litter, or broken glass in the street or on the sidewalk?,” “Are there abandoned cars?,” and “Do you see boarded up or abandoned buildings?” The full list of questions is available in the supplementary material (Table S1). All responses were coded categorically, and most were dichotomous. The virtual street audit was conducted by 12 undergraduate and graduate students at the University of Washington between July 2022 and May 2023. All auditors were trained on the same set of example locations. During training, all questions and examples were discussed to limit subjectivity among auditors. For example, multiple images of graffiti and murals were shown during training to highlight the differences between them. To assess agreement between the auditors, approximately 15% of the locations in each study center were randomly selected as a reliability subsample. Auditor training materials are available upon request from the corresponding author.

To generate a neighborhood disinvestment measure for each participant’s residential address, we used a validated approach involving an item response theory (IRT) model and ordinary kriging [25, 26]. Only data from the most recent or only image available at a location were used to fit the IRT model. We considered creating a spatiotemporal disinvestment model with data from the oldest and most recent images; however, we found there was too little change over time to warrant a model [36]. We used the most recent, or only, images instead of the oldest images because the oldest images were less spatially dense. An IRT model was fit to indicators (litter, graffiti, under-maintained buildings, bars on windows, and abandoned buildings) to form a scale measuring a latent level of disinvestment. The IRT model’s posterior probability of observing the indicators we actually observed was used to estimate a latent level of disinvestment at each audited location. We then used ordinary kriging to estimate levels for each residential address within the HCHS/SOL census tracts via spatial interpolation. Geocoded addresses with the disinvestment measure were then matched with participants’ geocoded addresses in a secure research workspace. Our neighborhood disinvestment measure is unitless, with higher values indicating more disinvestment. For this analysis, we have transformed the measure to its z-score, indicating a unit increase as a change in 1 standard deviation. More information on how the neighborhood disinvestment measure was developed is presented in the supplementary material.

Of the 16,415 HCHS/SOL study participants, ascertainment of diabetes was collected at visit 2 for 11,619 participants. Of the 11,619 participants, we had a neighborhood disinvestment measure for 11,503 at their baseline residential address. For our primary analysis, participants needed to be free of diabetes at baseline based on the primary diabetes definition (N = 9120). For our analysis with the secondary diabetes definition, 8989 were free of diabetes at baseline.

To compare the two definitions of diabetes, we computed demographic statistics for both samples and calculated an age-centered adjusted incidence rate. Range, median, mean, and standard deviation of neighborhood disinvestment were also calculated for each sample overall and stratified by study center. We ran Pearson’s correlations on disinvestment at baseline and a neighborhood socioeconomic status (NSES) index [37], population within a 1 km buffer, and percent Hispanic at the census tract level. We ran Spearman’s rank correlations between disinvestment and ordinal data on individual-level education, income, and years living in the US. The NSES index is a compositional measure of NSES at the census tract level that represents education, employment, housing, income/wealth, occupation, and residential stability [37]. Higher index scores indicate higher NSES disadvantage [37].

For our primary analysis, we investigated the association between neighborhood disinvestment at baseline residential addresses and incident diabetes at visit 2 using the primary diabetes definition. We ran Poisson regression models using a log link and years between baseline and visit 2 as the offset for follow-up time to generate an incidence rate ratio (IRR) comparing groups with a one-standard deviation difference in neighborhood disinvestment. All regression models used complete-case analyses and sampling weights to account for the HCHS/SOL complex sampling design, those living in the same household, and participant non-response at visit 2, using the survey package in R [29]. The sampling weights were also calibrated to age, sex, and Hispanic/Latino heritage from the 2010 US Census.

Minimally adjusted models were adjusted only for age and sex. Our fully adjusted model was adjusted for age, sex, years living in the USA (born in the 50 USA states, foreign born in the USA for more than 10 years, or foreign born in the USA for less than 10 years), a 5-level income variable, a 4-level educational attainment variable, family history of diabetes, an NSES index [37], and a 17-level variable combining study center and Hispanic/Latino heritage group. Study center-specific models were created using the same sets of covariates, except that the 17-level combination variable was replaced with the 8-level Hispanic/Latino heritage variable (Cuban, Dominican, Puerto Rican, Mexican, Central or South American, more than one heritage, or other).

We repeated the regression models to investigate the association between neighborhood disinvestment and three different progressions of diabetes: (1) progression from pre-diabetes to diabetes, where the sample only included those with pre-diabetes at baseline; (2) progression of those free of diabetes and pre-diabetes to diabetes; and (3) progression of those free of diabetes and pre-diabetes at baseline to pre-diabetes at visit 2. Results from all models were exponentiated to report the IRR for each unit increment in neighborhood disinvestment.

We repeated the Poisson regression models using the secondary definition of diabetes, which included a self-reported diagnosis. As 46.4% of participants moved residences during the follow-up period, we also conducted a sensitivity analysis restricted to those who did not move (N = 6164).

For the analytic sample, the only variable with greater than 5% missingness was income, with 7.6% of participants missing data. Due to concerns about missing data and the computational intensity of multiple imputations by chained equations (MICE), we conducted a MICE analysis as the sensitivity analysis in two model iterations. The MICE analysis consisted of 10 imputations and 50 iterations on the fully adjusted model for both the primary and secondary definitions of diabetes. Results from the MICE analysis were very similar and did not change the overall interpretation of results. Since completing MICE analyses on all our models would be computationally intense, and the MICE analysis tested in our two iterations yielded very similar results, we used complete-case analysis for our reported analyses.

All analyses were conducted in R Studio, with R version 4.1.3 (R Foundation for Statistical Computing, Vienna, Austria). The “ltm” package (version 1.2) was used for the IRT models. The “mice” package (version 3.15.0) was used for the MICE analysis [38], and the “survey” package (version 4.2–1) was used for incorporating the survey design and weights into the regression models [39].