The study region is Franklin County, Ohio. Franklin contains Columbus—the second most populous city in the Midwestern United States—and has a large land area (1379 km²) and a mix of urban, suburban, and open space or rural-type land use (ie, some census tracts have Rural Urban Commuting Area codes other than “Metropolitan Area Core”) [22].

We randomly generated 3535 locations for auditors to assess GSV imagery. The minimum, average, and standard deviation distance in meters between each candidate location and its nearest neighbor was about 51, 259, and 168, respectively. At each location, auditors assessed imagery from up to 5 dates: earliest (limited to 2009 or later), most recent, middate between earliest and recent, 25th percentile between earliest and most recent, and 75th percentile between earliest and most recent. Date quantiles were selected with the goal of achieving a uniformly distributed sample of audits across time.

Items assessed were selected from previously validated neighborhood audit tools [23]. The audit protocol assessed 3 constructs: disinvestment from 7 items (garbage, graffiti, abandoned buildings, building condition, yard condition, dumpsters, and road verge condition), engagement from 4 items (team sport equipment, outdoor seating, yard decorations, potted plants), and sidewalk walkability from 2 items (walkway type and walkway condition).

Six paid undergraduate research assistants were trained in a 1-hour orientation session, during which they were introduced to the study’s goals, objectives, and key construct definitions, followed by a 4-hour session in which they were introduced to the CANVAS (Computer Assisted Neighborhood Visual Assessment System) audit platform and audit protocol [23]. Auditors were trained on 75 non-Franklin County locations in Ohio. A training manual that also served as a reference guide during auditing was provided. Auditing of the candidate study location dates began only after adequate agreement was achieved across the 75 test locations. Table S1 in Multimedia Appendix 1 summarizes each audit item by name, full wording, response options, and response frequency.

Auditing took place at The Ohio State University Comprehensive Cancer Center between August 2022 and April 2023 using 2 computer displays, one of which was 24″ with 1920 ×1080 resolution for viewing GSV scenes. A total of 8256 audits were completed, 5751 of which met eligibility for this spatiotemporal analysis across 2214 unique locations (Figure 1). Eligibility criteria were (1) appropriate streetscape (the imagery was not of highway, a backyard, or a nature trail, and was not within a building) (n=7784, 94.3% of audits), and (2) at least 2 different dates at the location (n=5751, 70% of audits, 62.8% of locations). Among the 2214 eligible audit locations, the percentage with 3 or more unique dates was 59.7%, and the percentage with 4 or more unique dates was 14.6%. Only 1.9% of locations had 5 unique dates.

CRC survival data used in external validation analyses were from the Ohio Cancer Incidence Surveillance System. Eligible cases were adults (aged 20 years and older), diagnosed from 2012 to 2019 with an invasive cancer of the colon or rectum (International Classification of Diseases codes C18.0-C18.9, C19.9, C20.9, and C26.0), vital status follow-up through December 31, 2020, and geocoded residential address at diagnosis in Franklin County, Ohio. Survival time was calculated as days from date of diagnosis to end of follow-up, where end of follow-up was defined as date of death, date of last contact, or December 31, 2020. Follow-up is passive in that cases are presumed alive unless a death is found in linkages with Ohio Vital Statistics, Social Security Death Index, or the National Death Index. Covariates also from Ohio Cancer Incidence Surveillance System included age at diagnosis and stage at diagnosis (localized, regional, and distant or metastatic), sex (male or female), race-ethnicity (non-Hispanic White, non-Hispanic Black, non-Hispanic Asian, Pacific Islander, American Indian or Alaskan Native, or other; Hispanic), marital status (married and marriage-like vs widowed or divorced or separated or never married), and type of insurance (private or commercial, Medicare, Medicaid, uninsured, other).

An item response theory analysis of responses to the 7 audit items was used to investigate item correlation structure (Figure 2), internal consistency reliability per area under the curve of the Item Response Theory model (0.955) [24], and generate an NDS [1,25].

To reduce physical disorder measurement error, we generated rater-adjusted scores by regressing physical disorder on the rater variable with 6 nominal levels (eg, 5 dummy variables treated as fixed effects). The residuals of the resulting model, which represents NDS adjusted for rater bias, were used in subsequent analysis. Throughout, we simply use NDS to refer to the rater-adjusted residuals of raw NDS.

Previous studies have reported that social stratification such as racial-ethnic segregation can be reinforced by such physical boundaries as major roads, railroads, and bodies of water [26,27]. To test this with NDS, we partitioned the study region into mutually exclusive areas with boundaries defined by major physical barriers and then aggregated those areas based on audit samples. We assessed highways, railways, and hydrography from 2010 TIGER (Topologically Integrated Geographic Encoding and Referencing) shapefiles for Franklin County, Ohio [28]. Buffers on either side of highway and railway line files were created as estimates of typical widths (30 ft) and easements (25 ft), respectively. Polygonal hydrography was processed to estimate those features acting as barriers with potential to separate communities from one another (eg, major rivers or large lakes). Through a series of operations, hydrography features with an average width of 10 ft and length of 1 mile were considered barriers and merged with the polygonal highway and railway files to yield the physical barrier file. Next, Voronoi polygons around unique audit locations (Figure 1) were generated and then split by the intersection of physical barriers. Through a series of operations, intact and split Voronoi polygons were aggregated based on proximity to one another and final audit sample frequency to achieve areas with a maximum of 15 audits and as few as 2 (median 10). We aggregated using the Geographic Aggregation Tool implemented in R (R Core Team) [29]. The resulting spatial clusters, together with physical barriers, are shown in Figure 3. Clusters are colored by sample frequency, with barriers shown in orange. In total, there are 530 clusters.

We were interested in accurately predicting the NDS, denoted as Y(s₀,t₀), at a given location s₀ and time t₀. To effectively capture the spatial and temporal dependencies among observations, we used a spatiotemporal model estimating NDS that accounts for correlations across both space and time:

where µ(s_i,t_j) is the population mean trend and ε(s_i,t_j) is a zero-mean residual process that captures spatiotemporal dependencies through a specified covariance structure. The specific choice of the covariance structure will be discussed in “Residual Process” section.

We used this spatiotemporal universal Kriging model to estimate averaged, monthly values of disorder at set time lags and durations from CRC diagnosis date. The spatial points used for Kriging are exact geocoded latitude and longitude of patients’ residential address at diagnosis. We then used these values to test associations with CRC accelerated failure time models using a Weibull distributional assumption on the survival time and adjusted for covariates. A priori times for estimating physical disorder were the exact date of diagnosis and prediagnosis monthly averages at 3-month intervals up to 2 years prediagnosis (0‐3 months, 0-6 months, 0-9 months,..., 0‐24 months).

Model covariates were age, sex, marital status, minoritized race-ethnicity, health insurance, stage at diagnosis, and the interaction between stage at diagnosis and physical disorder. Based on previous studies, we hypothesized a stage-neighborhood disinvestment interaction wherein survival time among those diagnosed at stage 2, but not stages 1 and 3, would be shorter for those living in disinvested neighborhoods [1].

The Ohio State University Institutional Review Board (IRB) approved this study (IRB number 2023C0088, most recent approval: September 3, 2025). The IRB approved a waiver of consent, as such no participant contact nor compensation for research participation was provided. Identifiable data were geocoded residential address and dates of diagnosis of cancer cases, which were analyzed only on encrypted, password-protected computers and by IRB-approved study personnel who completed all required responsible conduct of research training.

All analyses in “Results” section were implemented using the statistical computing software R (R Core Team) [36]. Specifically, the spatiotemporal model fitting and universal Kriging predictions were performed primarily using the gstat package [33,34]. The accelerated failure time model in exposure validation analysis is fitted using R package survival [37]. Reproducible code and the entire audit protocol training manual are provided in the supplementary materials.

Overall, 62.8% of locations had at least 2 available dates. Among these eligible locations, only 1.9% had 5 available dates that approximated a uniform temporal distribution important for assessing longitudinal trends (ie, dates available for each of: earliest 2009+, most recent, middate, and approximations of the 25th and 75th percentiles between earliest and most recent). However, 59.7% of locations had at least 3 unique dates available. The spatial distribution of audits representing the midpoint and 25th or 75th percentiles appear less dense toward the county periphery and away from the City of Columbus compared with distributions of the earliest or most recent images that cover all but the Southwest corner of the county (Figure S1 in Multimedia Appendix 1). Given the temporally uniform sampling protocol and audit dates (2022‐2023), the distribution of years by target date indicated complementary and some overlapping distributions of image years; 54% of “earliest” audits were 2009, 94% of “25th percentile” audits were between 2011 and 2015, 89% of “midpoint” audits were between 2014 and 2017, 59% of “75th percentile” audits were between 2018 and 2020, and 56% of “most recent” audits were 2021 or 2022. Wintertime images were rarer and 86% were taken between May and October.

Figure 4 shows the spatial and temporal distribution of NDS. Overall, the NDS ranges from −2.818 to 2.816, with mean −0.002, median −0.057, SD 0.809, skewness 0.225, and kurtosis 3.166. The distribution is roughly symmetric with no strong outliers. Each panel corresponds to an observation date (the first day of the month) and each point within a panel represents a location. Points are colored by values of NDS, with a gradient from blue through white to red. Notably, certain dates, such as 2017‐06 and 2019‐04, have only 1 data point. Also, for specific dates, such as 2009‐05 and 2014‐09, the spatial distribution of observations is unevenly dispersed, with central areas more extensively sampled and rural areas less frequently observed. These underscore the sparse and irregular structure of the observations of NDS and motivate the necessity for the modeling technique we chose, which can accommodate sparse and irregular spatiotemporal data. Additionally, spatial correlation patterns are noticeable within each panel, with nearby locations exhibiting similar NDS. In general, urban neighborhoods have higher NDS than rural areas. To quantitatively evaluate spatiotemporal autocorrelation, we compute a spatiotemporal Moran I [38], where we assign weight 1 to “spatiotemporal” neighbor defined based on the estimated spatial and temporal range in Table 1, which is generated based on the method described in the “Residual Process” section and will be discussed in the “Out-of-Sample Prediction Accuracy” section. The spatiotemporal Moran I for NDS is 0.320, with P value <.001 for a permutation test that randomly permutes NDS 1000 times and recalculates Moran I. This suggests that there exists positive spatiotemporal autocorrelation for NDS.

We compare and evaluate the 2 mean trend models, model 1 and model 2-R, from 2 perspectives: in-sample model fit and out-of-sample prediction accuracy. The corresponding results for model 2-F are shown in Multimedia Appendix 1.

Following methods described in “Residual Process” section, we fit the empirical and theoretical semivariogram based on the simplified sum-metric covariance model on the residuals obtained from the mean trend model fit. The residuals have a sample mean 0 (SD 0.745), skewness 0.054, and kurtosis 3.110. The 1-sample Kolmogorov-Smirnov test results in a test statistics D=0.009 (P value=.75). These suggest that there is no strong evidence of deviation from normality. Table 1 shows the 8 covariance parameter estimates. Recall from “Residual Process” section, these include spatiotemporal anisotropy stAni, nugget effect τ², partial sills σ_s², σ_t², σ_st² (spatial, temporal, and joint), and ranges φ_s, φ_t, φ_st (spatial, temporal, and joint). Both the spatial and temporal components contribute to the residual dependencies. The estimated spatial range ϕ^_s=2.769 suggests that spatial correlation becomes negligible beyond 2.8 kilometers. The temporal range estimate ϕ^_t=270 days suggests that temporal dependency exists in a moderate period of time, but the magnitude of the dependency, indicated by partial sill estimate σ^_t²=0.038, is very small. The joint range and partial sill estimates, ϕ^_st=0.005 and σ^_st²=0.052, imply that the joint effect is concentrated within a very small spatiotemporal neighborhood with small variance. The nugget effect estimate τ^²=0.268 suggests the measurement error and the amount of unexplained variance are relatively large. In the context of neighborhood disinvestment, such variability may arise from differences in auditing image quality or local features not captured by the broader spatial or temporal trends. Finally, a small spatiotemporal anisotropy estimate stAni^=0.056 indicates that temporal distance has much smaller contribution to the joint covariance than spatial distance.

Figure 5A shows the empirical versus theoretical semivariogram fit in 3D, and Figure 5B shows the residual of the theoretical semivariogram fit in 2D. The trends of the theoretical semivariogram in Figure 5A match the parameter estimates, for example, a moderate spatial range and a weak temporal dependency. From Figure 5B, the absolute residuals are larger when the spatiotemporal distance is small. This may be due to unmodeled fine-scale variation and measurement noise, which are more pronounced at short ranges and not fully captured by the smooth structure of the simplified sum-metric model.

This study addressed 3 questions related to the longitudinal assessment of built environment disinvestment; a potentially important exposure with growing evidence for a role in numerous health behaviors and outcomes [2,4]. First, historical GSV imagery was available to support repeated sampling across locations and dates, although uneven image availability and spatiotemporal clustering limited dense temporal sampling at some locations. Second, neighborhood disinvestment varied substantially across space but more modestly across time, with measurable correlation ranging approximately 2.8 kilometers spatially and 270 days temporally. Spring and summer have lower neighborhood disinvestment than fall. The traditional spatiotemporal regression-Kriging model using precise geographic locations achieved greater out-of-sample predictive accuracy than the physical boundary-based model, mean-trend-only prediction, and spatial-only Kriging alternatives. Third, greater prediagnosis neighborhood disinvestment was associated with shorter survival among individuals diagnosed with regional stage CRC but not among those diagnosed with localized or distant stage CRC; survival time ratios changed little across exposure windows. Together, these results support the feasibility and predictive value of modeling neighborhood disinvestment jointly across space and time, while indicating that its association with CRC survival differs by stage at diagnosis and is similar across selected prediagnosis exposure windows.

While over 60% of audit locations had 2 or more eligible dates to audit, only approximately 2% had 5 eligible dates. Moreover, trellis plots of included audits indicate that location-dates may be spatiotemporally patterned; most timepoint-specific audits appear to derive from a specific subregion within the study region as opposed to a spatially random or uniform sample throughout the larger region. Such patterns could create difficulties in empirically separating spatial from temporal trends—especially small-scale trends—in the data. Moreover, we found a lower density of eligible dates toward the region’s periphery, which together with spatiotemporal patterns in data availability is consistent with prior findings [39].

We expected the model incorporating physical boundaries into estimates to outperform the conventional spatiotemporal Kriging model owing to face validity and empirical research demonstrating social stratification changes (eg, racial-ethnic segregation) when traversing major roadways, waterways, or railways [26,27]. However, our results provide evidence that the conventional spatiotemporal regression Kriging model had higher accuracy than the model incorporating boundaries. While some abrupt changes in disinvestment could follow major physical boundaries in the study area, results suggest that smoother changes in disinvestment with respect to distance and time offer a better explanation of the data. Within the traditional Kriging model, neighborhood disinvestment exhibited both small- and large-scale spatiotemporal variability, with the majority varying across space compared with time. Across larger temporal scales, disinvestment increased slightly over the study period and both spring and summer yielded coefficients indicative of lower disinvestment compared with fall. In terms of smaller-scale variation, disinvestment values exhibit small but measurable temporal variation (semivariance=0.04) and a substantial amount of small-scale spatial variation (semivariance=0.24), such that disinvestment autocorrelation was measurable temporally within 270 days and spatially within 2.8 kilometers. These spatiotemporal model parameters should prompt future audit designs that sample spatially proximate locations across seasons of the year as well as dates within 270 days in order to generate more accurate disinvestment exposure estimates of participants in public health studies.

These spatial variability and autocorrelation magnitudes are comparable with other studies in various major cities and urban areas throughout the United States [1,15]. While few studies have explicitly tested neighborhood disinvestment temporal variation and autocorrelation [40], our findings of small to moderate temporal variation is in line with a recent study that reported κ coefficients for disinvestment audit items comparing earliest available and most recent images of the same location in back-to-back assessments conducted by the same rater [41]. The 5 disinvestment indicators and κ coefficients overlapping that of this study—buildings needing repairs (0.22), boarded or abandoned buildings (0.24), burned buildings (0.44), garbage or litter (0.43), and graffiti (0.64)—suggested “fair” to “substantial” agreement across a median image date separation of 10.5 years. For the typical public health study’s spatial and temporal scale (eg, city or larger spatially and 1‐10 years of exposure assessment temporally) and meaningful spatiotemporal measurement units (eg, tens of meters spatially and accumulations across years temporally), results of this study add to growing evidence indicating that neighborhood disinvestment varies more across space than time. On average, built environment change across time is incremental such that street-level visual transformations may not be apparent for many years. These findings can be useful in designing future spatiotemporal audits, indicating that lower-density temporal sampling and higher-density spatial sampling will yield more accurate neighborhood disinvestment exposure estimates than the opposite.

Our findings that neighborhood disinvestment as associated with survival only among those diagnosed at regional stage (stage 2) are consistent with previous studies of neighborhood disinvestment and breast cancer survival [1]. Paralleling our findings of small temporal variation in neighborhood disinvestment, there was little variation in the disinvestment—cancer survival associations as a function of exposure lags. Within each stage at diagnosis, the 95% CIs for survival time ratios are all heavily overlapped (comparing vertically within each subplot in Figure 7). The 95% CI lower bounds, however, for the longer duration exposure time lags were slightly further away from null, suggestive of better fitting associations when averaging disinvestment values 6-24 months prior to diagnosis. These small exposure lag variations could be spurious but, if true, could be due to better estimation of individuals’ disinvestment exposures over the course of multiple seasons. Central Ohio experiences 4 seasons with large variation in temperature, precipitation, and vegetation, which in turn could influence outdoor activity, perceptions of streetscapes, and experiences of stress.

A main limitation of this explicit spatiotemporal analysis is the lack of residential mobility patterns of CRC cases used in the external validation analysis. This potentially important temporal dimension prevented full assessment of the spatiotemporal models’ performance. As previously mentioned, emerging evidence suggests that between 20% and 40% of cancer cases change residence both pre- and postdiagnosis, depending on follow-up length, age, and several other factors [18,20,42]. We were also limited in our ability to adjust for other potential confounders, including CRC screening, socioeconomic factors, and social cohesion, that are not available in cancer registry data. Franklin County is likely not representative of US disinvestment or CRC, and similar to a previous study that found variation in spatial-only Kriging parameters across US cities of New York, Philadelphia, Detroit, and San Jose, results could be sensitive to spatiotemporal practices and policies that impact disinvestment (investment strategies, abandoned property handling, etc) [15,17,43]. Finally, variation in image availability may have affected measurement precision over space and time. However, a moderate to large sample of nearly 6000 audits across space and time, use of previously validated audit items and CANVAS audit platform, and testing of various neighborhood disinvestment spatiotemporal model parameters and frameworks, internal cross-validation, and exposure time lag testing within a single investigation are main strengths.

Auditing various built environment characteristics using GSV across a uniform distribution of US urban locations and dates (2009‐2022) is possible especially if a maximum of 3 dates per location are desired. A traditional spatiotemporal regression Kriging model was more accurate at predicting NDSs than a model accounting for sharp changes due to physical roadway, waterway, and railway boundaries. Despite relatively small temporal compared with larger spatial variation in neighborhood disinvestment, disinvestment was appreciably lower in spring and summer than in autumn and increased modestly over the 14-year period. Cancer stage–dependent associations between disinvestment and survival were slightly more precise when averaging 6‐24 months prior to diagnosis. In summary, built environment disinvestment exhibits appreciable variation over time, indicating a need for assessments across dates in addition to locations in order to capture important longitudinal trends.