3.1. Exploratory Data Analysis

Histograms reveal leftward shift in heat week call distribution relative to normal weeks, evidenced by both mean and median decreases. This is a slight call reduction during extreme heat periods, conflicting with the main hypothesis. This may likely be due to behavioral shifts during extremes versus normals. While the 311 calls were selected based on their connections to heat and heightened aggravation with high temperatures, certain complaints like "banging" or "loud music" could decrease because individuals may want to insulate themselves indoors with AC, reducing audible and visible perceptions to outdoor or even sanitation issues.

Both distributions have pronounced positive skewness with long right tails extending beyond 15 calls per 1,000 population. However, heat week distribution per capita has more extreme outliers and higher variability, suggesting certain tracts experience disproportionate surges in QoL complaints during heat waves (about 852 tracts, or 38.3% of tracts have higher calls during heat weeks). These outlier tracts warrant individual investigation as they may be commercial districts with transient populations inflating per-capita rates, neighborhoods with vulnerable populations, areas with heightened civic engagement and 311 awareness, or tracts with potential data quality issues.

The great overlap between distributions suggests most tracts maintain relatively stable complaint rates regardless of heat conditions. However, some tracts do show meaningful elevation during heat weeks, highlighting importance of identifying characteristics that predict heat sensitivity rather than assuming universal heat response.

PCT_TREE_CANOPY: Tree canopy may be a powerful discriminator among neighborhood types as most tracts have very little canopy as opposed to others given the low mean and median yet high maximum.

NDVI: Seemingly shows a similar distribution to PCT_TREE_CANOPY.

WCR: Expectedly very low due to the urbanity of NYC unless tracts are bordering water bodies.

PCT_IMPERVIOUS: High percentage is expected due to the built density of metropolitan cities.

PCT_BACHELORS_PLUS: Over 30% of individuals attained high education, which also affects awareness and use of city services like 311. Again, high standard deviation indicates unequal distribution of education.

PCT_RENTERS: Expected high values due to NYC being a renter-dominant city. This means a lot of renants may experience infrastructural issues outside of their immediate control, as safety and building regulation responsibilities fall on the landlord.

PCT_LIMITED_ENGLISH: Very high skew shows that individuals who speak limited English are geographically stratified in different parts of NYC. This reflects a significant barrier to accessing 311 services.

MEDIAN_INCOME:: After removing placeholder values, income distribution reveals NYC's stark economic geography. Standard deviation exceeds the mean, indicating heavily skewed distribution, which unsurprisingly reflects extreme wealth inequality across census tracts.

POVERTY_RATE: High right skew and a 100% max indicate concentrated poverty in certain neighborhoods.

PCT_NON_WHITE: Racial composition reflects NYC's diversity, but high variability indicates persistent residential segregation patterns that could interact with target variables.

BD: Most tracts have low-to-moderate building coverage with some ultra-dense commercial/residential cores.

AH: Building heights have great skew and range, so there's significant vertical variability in NYC, likely reflecting skyscraper districts in the central core and mid- to low-rise as the geography eases out into the periphery.

POI_500M_DENSITY: Amenities are expectedly highly clustered, likely in denser commercial / business and residential areas.

KNN_SUBWAY_dist_mean: Average walking distance to nearest subway stations shows most tracts generally have reasonable transit access, and this is unsurprising as NYC is the only city in the US with public transport as its main commute / modality for citizens. Although the variation suggests transit deserts exist.

Extreme Heat and Normal Weeks: Heat week and normal week call rates show strong positive correlation, indicating neighborhoods with high baseline complaint rates also show elevated rates during heat periods. Again, imperfect correlation suggests heat amplifies complaints differently, with some tracts having proportionally larger increases than baseline activity.

Tree Canopy and NDVI Correlation: Tree canopy coverage and NDVI measure related environmental characteristics. But it also suggests that they both capture unique aspects, especially since NDVI reflects all vegetation including grass / shrubs while tree canopy specifically quantifies woody vegetation providing shade.

Impervious Surface Relationships: Negative correlations show expected trade-off between built environment and green space. But there are some tracts with high vegetation despite substantial impervious coverage, potentially through street trees, small parks, or green infrastructure.

Income and Education Correlation: Strong positive association between higher education and median income is expected.

Income and Poverty Correlation: Extremely strong negative correlation creates redundancy as these variables measure opposite ends of same economic spectrum.

Race and Poverty Correlation: Moderate positive correlation between percent non-white and poverty rate reflects persistent structural racism and residential segregation in NYC.

Building Density and Height Correlation: Moderate positive correlation suggests taller buildings tend to cluster in denser areas.

POI Density and Subway Access Correlation: Negative correlation indicates that commercial centers concentrate near transit. However, stats indicate deserts with commercial activity and transit-rich residential zones both exist.

Green Space and Heat Calls Correlation: Weak negative correlations support the hypothesis on buffering effects, though relationships are weaker than anticipated. However, this may require nonlinear modeling or interaction terms to capture threshold effects.

Income and Heat Calls Correlation: Near-zero or weakly positive correlation challenges simple vulnerability narrative, supporting one of the hypotheses of nuanced prediction that affluent areas may show increased reporting despite lower physiological vulnerability.

Critical VIF Violations (10 < VIF)

PCT_RENTERS, BD, and PCT_IMPERVIOUS are likely very correlated with one another as imperviousness and building density have overlapping measures of the physical city, whereas renters as opposed to owners will be concentrated in denser areas.

PCT_BACHELORS_PLUS and MEDIAN_INCOME are expectedly correlated with one another as higher educational attainment often meets that higher earning gradient.

NDVI's high VIF indicates a correlation with its inverse of impervious surface as well as correlation with PCT_TREE_CANOPY. It may have a higher VIF due to its broader measurements that capture more city characteristics than tree canopy.

PCT_NON_WHITE is likely associated with POVERTY_RATE, but the former may capture more broad sociodemographics, hence the higher VIF.

Moderate Multicollinearity Concerns (5 < VIF < 10)

POVERTY_RATE and KNN_SUBWAY_dist_mean are moderately multicollinear, the former with MEDIAN_INCOME and PCT_NON_WHITE and the latter with the urban form variables. However, these metrics do not have as drastic of VIF values as their related counterparts.

Acceptable VIF Range (VIF < 5)

Low VIF for WCR, PCT_LIMITED_ENGLISH, POI_DENSITY, AH, and PCT_TREE_CANOPY indicate that these are the most distinct and independent variables from 311 QoL reports in heat.

With this, it seems the variables cluster strongly and reveal that socioeconomic and environmental inequality is not composed of independent factors, but rather tightly bundled mechanisms that drive 311 reporting behavior. This means that OLS, while providing an interpretable baseline, is not ideal for capturing all these variables, reinforcing the need for models robust against multicollinearity like Random Forest.

The outliers significantly wash out the other tracts in NYC, many 311 super users are in the Queens Borough in the Long Island City area right around LaGuardia Community College. This general area looks like it has a higher non-white and higher poverty rate.

Most notably, the percent change map looks like it lacks spatial autocorrelation, but visually the impact must be influenced by some scattered tracts with low baseline activity that makes it look fragmented. This must indicate some kind of stress that's unique to the tracts and their mechanism shifts.

Despite the visual scattered perception, it does look like that the changes are more skewed toward peripheral edges of the boroughs.

Also, MEDIAN_INCOME and PCT_BACHELORS_PLUS show an almost inverse spatial pattern when compared to PCT_NON_WHITE.

3.2. OLS Model Results

3.2.1 Normal Heat Model

In the OLS model for normal heat week QoF 311 report density, the overall F-statistic is strongly significant, indicating that the set of urban features jointly contributes meaningfully to explaining variation in reporting behavior. The R-squared of 0.084 shows that these predictors account for roughly 8% of the spatial variation, which presents a modest and low level of explanatory power that is typical for urban planning research on 311-based analyses. Such 311 outcomes reflect that service-request behavior is shaped by complex human responses, institutional factors, and informal social dynamics that are only partially captured by observable indicators. The relatively low R-squared also suggests the non-linear relationships might exist between urban features and 311 reporting behavior, indicating that machine-learning models may be more suitable for uncovering these non-linearities.

Regarding individual predictors, under the 0.05 significance threshold and within the linear OLS framework, seven features show statistically significant associations with QoF 311 report density during normal heat weeks: PCT_TREE_CANOPY, PCT_IMPERVIOUS, NDVI, POVERTY_RATE, PCT_NON_WHITE, BD, and AH. All remaining variables exhibit no significant linear effect.

3.2.2 Extreme Heat Model

In the OLS model for extreme heat–week QoF 311 report density, the overall F-statistic is highly significant, indicating that the set of urban features jointly provides meaningful explanatory power. The model’s R-squared of 0.088 shows that these variables account for roughly 8–9% of the spatial variation, which also shows a limited explanatory power. The relatively low R-squared also suggests the non-linear relationships might exist between urban features and 311 reporting behavior in extreme heat weeks.

Regarding individual features, under the 0.05 significance threshold and the linearity assumption inherent to OLS, six features exhibit statistically significant associations with QoF 311 report density during extreme heat weeks: PCT_IMPERVIOUS, NDVI, POVERTY_RATE, PCT_NON_WHITE, BD, and AH. All other variables show no significant linear relationship after controlling for the rest of the model, implying that their effects may be weak, or more accurately represented through non-linear structures.

3.2.3 OLS Comparison

Across both the extreme-heat-week and normal-heat-week OLS models, only about half of the urban features exhibit statistically significant linear associations with QoF 311 report density, and both models yield low explanatory power (R² < 0.10). This consistency indicates that the linear models capture only a small portion of the urban environmental, socioeconomic, and built indicators driving 311 reporting. Taken together, these results reinforce the limitations of linear OLS for this problem. They highlight the need for machine-learning approaches that can better capture non-linear effects, enabling a more precise comparison of how urban features influence 311 reporting differently under extreme heat versus normal heat conditions.

3.3. ML and SHAP Results

3.3.1 ML Model Result

Across both models, Random Forest substantially outperforms the OLS baseline, demonstrating the importance of non-linear and complex effects in explaining QoF 311 report density. For the regular heat week model, the test R2 reaches 0.274, over three times higher than the OLS R2 of approximately 0.08. Similarly, the extreme heat week model attains a test R^2of 0.246, again far exceeding the linear model but slightly lower than the regular heat scenario. This gap suggests that reporting behaviors during extreme heat is more variable and influenced by additional unobserved or volatile mechanisms. The clear improvements in model performance in both cases further confirm that machine-learning approaches capture substantial structure in the data that linear models fail to represent.

SHAP results reveal a consistent set of dominant predictors across both heat conditions. In each model, AH, PCT_NON_WHITE, and NDVI emerge as the three strongest contributors, reinforcing the central role of high-density residential morphology, socio-demographic composition, and environmental greenness in shaping QoF 311 reporting patterns. Several other predictors, such as subway accessibility, median income, WCR, poverty rate, and PCT_RENTERS, also exhibit meaningful non-linear contributions that were not visible in OLS, underscoring the value of ML models for uncovering complex behavioral responses.

3.3.2 Extreme Heat vs Normal Heat

Comparing the two heat conditions reveals both stability and notable shifts in feature influence. The hierarchy of the top four features, AH, PCT_NON_WHITE, NDVI, and KNN_SUBWAY_dist_mean remains consistent across models, suggesting that built form intensity, demographic composition, and urban greenery systematically shape reporting behaviors in both normal and extreme heat contexts. These stable high-importance variables reflect structural neighborhood characteristics whose effects persist regardless of temperature severity.

However, several features display meaningful changes in importance under extreme heat. WCR increases in influence in the extreme heat model, indicating that proximity to water bodies may play a more substantial role when heat stress intensifies, possibly reflecting shifts in human activity patterns or uneven access to cooling amenities. In contrast, some variables such as PCT_IMPERVIOUS, BD, and PCT_TREE_CANOPY show moderate decreases in relative importance, suggesting that some physical morphology may matter slightly less once temperatures exceed critical thresholds. Overall, these patterns indicate that while the core drivers of 311 reporting remain stable, the marginal influence of secondary features is sensitive to heat severity.

3.3.3 Non-Linear Relationship for Features

Across the SHAP scatter plots for both the extreme heat and normal heat models, clear non-linear relationships emerge for most features. Only NDVI and POVERTY_RATE exhibit relatively consistent linear patterns, while all other features display visibly complex, non-monotonic trends. A prominent example is AH, the most influential features in both models: although AH appears significant in OLS, its relationship revealed by SHAP approach is distinctly U-shaped in both heat conditions. This means that both low and high AH values are associated with higher predicted 311 report density, whereas medium AH values correspond to the lowest predicted levels.

Similarly, PCT_NON_WHITE and PCT_RENTERS show a characteristic inverted-U pattern in both models, where extreme low or high values of these features correspond to lower 311 reporting density, and intermediate values correspond to higher levels. An interesting shift between the two heat conditions appears in the behavior of BD: in the regular heat model, BD follows an inverted-U shape, but in the extreme heat model, this pattern transitions into an approximately linear negative relationship.