JMIR Publications

ABSTRACT

Background:

Neighborhood disinvestment, characterized by built environment disrepair and deterioration, has been linked to health behaviors and outcomes, including cancer survival. While prior studies have assessed spatial variation, temporal dynamics of disinvestment remain underexplored, despite potential relevance for long-latency outcomes such as colon and rectum cancer (CRC).

Objective:

We describe and validate a spatio-temporal neighborhood audit protocol using Google Street View (GSV) imagery, develop and compare predictive spatio-temporal models of neighborhood disinvestment and examine time-lagged associations between disinvestment and CRC survival.

Methods:

We conducted 8,256 virtual audits across Franklin County, Ohio (2009–2022) using seven disinvestment indicators (e.g., garbage, graffiti, abandoned buildings). After excluding ineligible images, 5,751 audits of 2,214 unique locations were included. A neighborhood disinvestment score (NDS) was derived using item response theory and adjusted for rater bias. We fit mean trend and spatio-temporal residual processes using universal kriging with a simplified sum-metric covariance structure, comparing two candidate mean trend models via out-of-sample root mean square prediction error (RMSPE). One model accounted for discrete physical barriers to define neighborhood clusters, the other was traditional kriging where NDS was allowed to vary continuously across space and time. For exposure time validation, predictions were linked to 2,727 CRC cases (2012–2019) from the Ohio Cancer Incidence Surveillance System (OCISS). We use accelerated failure time (AFT) models to estimate associations between NDS averaged over 0, 3, 6, 12, 18, and 24 months prior to diagnosis and CRC survival, adjusting for age and stage at diagnosis and demographic covariates. We test stage-specific associations between NDS and survival.

Results:

Spatio-temporal modeling indicated substantial spatial but modest temporal variation in disinvestment, with small-scale spatial correlation measurable within 2.8 km and temporal correlation up to 270 days. The traditional spatio-temporal kriging model (RMSPE = 0.651) outperformed alternate models including neighborhood cluster fixed (RMSPE = 0.671) and random (RMSPE = 0.662) effect models. Seasonal effects were significant, with lower disinvestment scores in spring and summer compared to fall. The AFT model fit shows higher pre-diagnosis NDS was significantly associated with shorter survival among patients with stage 2 cancer (acceleration factor < 1 across all lag windows, P value < .05), but not among patients with stage 1 or stage 3. Associations were consistent across time lags, with slightly stronger effects for averages over 6–24 months prior to diagnosis.

Conclusions:

Neighborhood disinvestment exhibits substantial spatial and moderate temporal variation. Traditional spatio-temporal universal kriging provides most accurate prediction compared to models accounting for discrete neighborhood boundaries. Neighborhood disinvestment is associated with reduced CRC survival time among cases diagnosed at regional stage, highlighting the utility of built environment spatio-temporal assessments within CRC research.