The team drew on an unusually rich data infrastructure. Using the Centers for Medicare and Medicaid Services Chronic Conditions Data Warehouse, they identified beneficiaries aged 65 and older who were alive and Alzheimer’s-free in 2011 and then tracked whether claims-coded Alzheimer’s disease appeared between 2012 and 2016. After excluding movers, people who died or could not be linked across the follow-up window, and those with prevalent disease at baseline, the analytic cohort comprised 181,614 residents of Miami-Dade County, of whom 7,606, or about 4.2 percent, developed claims-identified Alzheimer’s disease during follow-up. Each beneficiary’s residential Census block was linked to vegetation measurements derived from Landsat satellites, the workhorse of long-term environmental monitoring, and models adjusted for age, sex, race and ethnicity, neighborhood median household income, and neighborhood walkability measured by Walk Score.

The core exposure metric was the Normalized Difference Vegetation Index, or NDVI, a spectral index computed from the contrast between how vegetation absorbs red light for photosynthesis and scatters near-infrared radiation from its internal leaf structure. Higher NDVI values indicate denser, more photosynthetically active vegetation, and the index has become the default greenness measure in health studies because it is reproducible, spatially continuous, and available in archives stretching back decades. Yet NDVI has well-known limits: it measures photosynthetic activity, not tree canopy, park quality, accessibility, or vegetation type, and it says nothing about whether residents actually use the green spaces near them. The new study adds another layer of caution by showing that even the numerical value of NDVI assigned to a neighborhood depends heavily on choices researchers make before the analysis begins.

Those choices fall into two categories that geographers have long formalized as the modifiable areal unit problem and the modifiable temporal unit problem. The first, known as MAUP, recognizes that statistical relationships can change when data are aggregated at different spatial scales or zoning systems. The second, MTUP, is its temporal analogue: results can shift depending on how measurements are segmented and averaged across time. Both problems are textbook material in spatial epidemiology, but they had rarely been tested systematically for a chronic neurodegenerative outcome. The Miami-Dade team did exactly that, generating 32 distinct NDVI products by combining four raster resolutions (30, 120, 250, and 500 meters), two temporal acquisition strategies (single-date imagery versus seasonal composites averaging all cloud-free September-to-November scenes), two temporal specifications (2011 baseline versus a 2011-plus-2016 sum), and two metrics (mean NDVI and NDVI range within each block).

The headline result was reassuring in direction but revealing in detail. Across all four spatial resolutions, higher seasonal mean NDVI in 2011 was consistently associated with lower odds of incident Alzheimer’s disease. Each 0.1-unit increase in NDVI was linked to roughly 18 to 23 percent lower odds, with the strongest association at the 250-meter resolution, where the odds ratio was 0.77. Notably, the native 30-meter Landsat resolution, the finest available, did not produce the best-fitting models; moderate aggregation at 120 and 250 meters captured neighborhood-scale greenness more consistently, apparently by smoothing pixel-level noise while retaining meaningful local variation. The researchers used the Quasi-likelihood under the Independence Model Criterion, a generalization of the Akaike Information Criterion for their clustered regression framework, to compare model fit, and the 250-meter seasonal model scored best within its set, though neighboring resolutions performed similarly.

Temporal aggregation mattered just as much. When the team used a single cloud-free image from fall 2011 instead of the seasonal composite, the inverse association weakened considerably, with odds ratios closer to 0.92 to 0.96 per 0.1-unit NDVI increase. Seasonal compositing, the authors argue, reduces the influence of anomalous acquisition dates, residual cloud and atmospheric contamination, and short-term vegetation fluctuations, all of which are especially relevant in subtropical Miami where cloud-free imagery is scarce. In a striking contrast, the single-date models fit best at the finest 30-meter resolution, while the seasonal models favored moderate aggregation, suggesting that the apparent spatial scale of a greenness effect depends on how the exposure is captured in time. A two-time-point metric combining 2011 and 2016 NDVI confirmed that the spatial pattern of associations was robust, though it was interpreted strictly as a sensitivity analysis because the 2016 measurement postdates some incident cases.

The study also probed an exploratory metric with a very different personality: NDVI range, the difference between the maximum and minimum pixel values within each Census block. Whereas mean NDVI was consistently protective, higher NDVI range was generally associated with higher Alzheimer’s incidence, and the association was fragile, weakening or disappearing at coarser resolutions. The authors are careful to note that range is not a measure of greenspace fragmentation, connectivity, or landscape morphology; it is simply a crude indicator of within-block spectral variability that is highly sensitive to extreme pixels and to the number of pixels available. Still, its positive association hints that heterogeneous residential environments, where lush vegetation coexists with sparse cover, may tell a different story than uniformly green blocks, a question that will require proper landscape-configuration metrics to resolve.

Why should anyone outside the remote-sensing community care? Because the greenness–health literature is increasingly cited in urban planning debates, from tree-planting initiatives to heat mitigation programs, and because prior studies in this very field have reached conclusions using wildly different exposure definitions. Some used 300-meter buffers, others Census blocks or road-network buffers; some used a single satellite image, others annual averages or multi-year moving windows. Reviews have found that buffer sizes in published studies range from about 30 to 5,000 meters. The new analysis shows that holding the population, outcome, covariates, and statistical model constant while varying only the exposure definition changes both the magnitude of the effect and which model fits best. Scale selection, the authors conclude, is not a technical preprocessing detail but a substantive component of epidemiologic inference that should be tested and reported explicitly.

The authors are equally candid about the study’s limits. Alzheimer’s status came from insurance claims rather than clinical adjudication, so delayed diagnosis, underdiagnosis, and coding variation are possible, though such misclassification would likely attenuate rather than inflate the observed associations. Residential Census blocks capture where people live, not where they walk, socialize, or spend their days, a limitation known as the Uncertain Geographic Context Problem. The cohort excluded movers and people with incomplete address linkage, potentially producing a more residentially stable and healthier sample than the county at large. Walk Score data from 2013 were assigned to 2011 residences, and the analysis did not adjust for air pollution, heat, or noise, which correlate with greenness and may independently affect cognitive health. The observed associations therefore reflect residential greenness within a broader urban environmental context rather than isolated effects of vegetation alone.

Even so, the practical message is clear and, in its way, empowering. NDVI remains a valuable, scalable screening tool for identifying neighborhoods where low greenness coincides with older-adult vulnerability, and the protective signal itself survived every scale test the researchers threw at it. But translating a satellite index into a greening prescription requires more: information on tree canopy versus grass, vegetation type, park access, safety, land ownership, and community priorities. Future work, the authors suggest, should pair higher-frequency satellite products such as MODIS or harmonized Landsat-Sentinel data with mobility-aware exposure measures and policy-actionable greenspace indicators. In the meantime, the study stands as a reminder that in environmental epidemiology, the map you choose to draw helps determine the world you find.

Subject of Research: Spatial and temporal scale sensitivity of satellite-derived NDVI greenness exposure assessment in relation to Alzheimer's disease incidence among older Medicare beneficiaries

Article Title: Spatial and Temporal Scale Sensitivity in NDVI-Based Greenness Exposure Assessment for Alzheimer’s Disease Incidence: A Miami-Dade County Medicare Cohort Study

Article References: Southworth, J., Chaudhary, A., Dewald, J. R., Ma, R., Safaei, M., Szapocznik, J., & Brown, S. C. (2026). Spatial and Temporal Scale Sensitivity in NDVI-Based Greenness Exposure Assessment for Alzheimer’s Disease Incidence: A Miami-Dade County Medicare Cohort Study. Environmental Advances, Article 100765. https://doi.org/10.1016/j.envadv.2026.100765

Image Credits: AI Generated

DOI: 10.1016/j.envadv.2026.100765

Keywords: Alzheimer's disease, NDVI, greenness exposure, remote sensing, Medicare cohort, modifiable areal unit problem, modifiable temporal unit problem, Landsat, spatial epidemiology, urban green space, Miami-Dade County, dementia risk