Abstract. This paper examines whether caste-based residential segregation predicts differential air pollution exposure and monitoring coverage across Delhi, India. Using a 1 km² grid covering 1,694 cells across Delhi NCT, we combine three independent pollution datasets — CPCB IDW-interpolated PM₂.₅, satellite-derived PM₂.₅ (2022), and Sentinel-5P NO₂/SO₂ (2023) — with geocoded Scheduled Caste (SC) basti locations from Census 2011 and JJ cluster records from DUSIB 2022. We find that SC settlement density is a significant positive predictor of PM₂.₅ (satellite: β = 0.385, p < 10⁻⁸⁷) and NO₂ (β = 0.337, p < 10⁻⁴¹) after controlling for population density, informal clusters, and industrial proximity. Every grid cell in Delhi exceeds India’s NAAQS limit of 40 µg/m³. The spatial distribution of pollution is structured by caste geography. We further document systematic institutional invisibility: 57% of JJ clusters are unmonitored (> 2 km from any CPCB station), and the most deprived quarter of clusters is 166% further from monitoring infrastructure than the least deprived. Results hold across two independent measurement approaches and robustness specifications excluding approximately geocoded data.

Keywords: Environmental justice · Air quality · Caste segregation · Delhi · Sentinel-5P · Informal settlements · Monitoring gap

1 Introduction

Delhi’s air quality crisis is among the most severe in the world. Annual mean PM₂.₅ concentrations routinely exceed 100 µg/m³ — more than 20 times the WHO guideline of 5 µg/m³. What is far less understood is whether this burden falls equally across social groups, or whether Delhi’s history of caste-based residential segregation reproduces itself as a spatial environmental inequality.

This paper addresses three questions: (1) Are neighbourhoods with high concentrations of Scheduled Caste settlements more polluted, net of population density and industrial proximity? (2) Is any inequality primarily driven by traffic or industrial emissions, and is the result consistent across independent measurement approaches? (3) Are more deprived informal settlements systematically excluded from Delhi’s air quality monitoring infrastructure?

Core findings. SC settlement density significantly predicts higher PM₂.₅ and NO₂ across all model specifications and both data sources. The NO₂ result is driven by traffic corridors, not industrial proximity. 57% of JJ clusters are unmonitored. The most deprived clusters are 166% further from monitoring stations than the least deprived. 153 clusters (23%) face a double burden: maximum deprivation and zero monitoring.

2 Data

2.1 Settlement data

SC bastis — 4,541 localities from Census of India 2011 enumeration-block records covering 1,505,038 SC residents. Geocoded via Nominatim/OSM (29.2%), a curated known-place dictionary (34.2%), and district centroid + jitter as fallback (36.6%). All main results are robust to excluding jitter-geocoded points (see Section 5.3).

JJ clusters — 675 clusters from DUSIB 2022 covering approximately 1.2 million people and 303,058 households.

2.2 Pollution data

Source Measure Year Resolution Method
CPCB/DPCC (28 stations) PM₂.₅ IDW 2022–23 1 km² Inverse-distance weighted annual means
Satellite raster PM₂.₅ 2022 ~1 km Surface concentration, band PM25_ugm3
Sentinel-5P TROPOMI NO₂ 2023 ~1 km Tropospheric column density, mol/m²
Sentinel-5P TROPOMI SO₂ 2023 ~1 km Tropospheric column density, mol/m²

Satellite PM₂.₅ (mean 94.6 µg/m³) is 16% below IDW (mean 113.0 µg/m³), consistent with known satellite underestimation during stable winter inversions in South Asia.

2.3 Grid

Delhi NCT is divided into a 1 km² regular grid filtered by the Delhi districts shapefile (geopandas.sjoin(predicate="within")), yielding 1,694 valid cells. An earlier version used 2,562 cells (including 872 outside Delhi NCT); the shapefile fix eliminates outside-boundary cells, 118 of which had valid satellite values and were contaminating regressions.

3 Data Pipeline

Processing sequence:
00b_extract_sc_bastis_pdf → 05_sc_bastis_geocode → 02_spatial_covariates → 03_deprivation_and_stats → 04_build_grid → 06_satellite_pollution → 07_grid_analysis → 08_grid_regressions

Script Input Key output Operation
00b Census PDFs eb_sc_bastis_raw.csv pdfplumber extraction, 4,541 rows
05 Raw CSV sc_bastis_geocoded.csv Nominatim geocoding + fallback dict
02 JJ cluster list jj_clusters_spatial.csv IDW from 28 CPCB stations
03 Spatial clusters jj_clusters_full_analysis.csv Deprivation index, monitoring gap stats
04 Stations + shapefile delhi_pollution_grid.csv 1,694-cell shapefile-filtered grid
06 JJ data + .tif rasters jj_multipollutant.csv Sample NO₂/SO₂/PM₂.₅ at cluster centroids
07 Grid + settlements + rasters delhi_grid_full.csv Spatial join, log transforms, z-scores
08 delhi_grid_full.csv Regression outputs HC3-robust OLS, 7 specifications

Key methodological fixes applied during analysis:

  • Fix A — Grid boundary: geopandas.sjoin(predicate="within") replaces 5 rough bounding-box corner-cuts. Grid shrinks 2,562 → 1,694 cells.
  • Fix B — Kontur population join: centroid-within replaces intersects, correcting double-counting inflation (~96M → ~20M; max cell 282,998 → < 70,000).
  • Fix C — Deprivation index: aqi_dist_km removed as a component (it was 27% of the index that was then tested as the outcome, inflating r from 0.048 to 0.667).
  • Fix D — HC3 robust standard errors added to all regressions (DW ≈ 0.07–0.38; HC3 corrects SEs without altering coefficients).

4 Results

4.1 Study area and settlement distribution

**Figure 1.** Spatial distribution of Scheduled Caste population across Delhi Municipal Corporation wards (Census 2011). Dev Nagar (69%), Trilokpuri (53%), Sultanpuri-A (50%), and Nand Nagri (49%) are the most SC-concentrated wards.

Figure 1. Spatial distribution of Scheduled Caste population across Delhi Municipal Corporation wards (Census 2011). Dev Nagar (69%), Trilokpuri (53%), Sultanpuri-A (50%), and Nand Nagri (49%) are the most SC-concentrated wards.

**Figure 2.** SC bastis (dark green = reliable geocoding; light green = district jitter) and JJ clusters (blue) overlaid on ward-level SC share. Reliable bastis concentrate in the north inner ring; JJ clusters span north and east Delhi.

Figure 2. SC bastis (dark green = reliable geocoding; light green = district jitter) and JJ clusters (blue) overlaid on ward-level SC share. Reliable bastis concentrate in the north inner ring; JJ clusters span north and east Delhi.

**Figure 3.** SC basti population and locality type by district. North West Delhi has the largest population (366,782 across 1,109 localities). Central Delhi has the highest mean SC% per locality (dark red). Harijan/Balmiki-named localities signal historically enforced caste segregation.

Figure 3. SC basti population and locality type by district. North West Delhi has the largest population (366,782 across 1,109 localities). Central Delhi has the highest mean SC% per locality (dark red). Harijan/Balmiki-named localities signal historically enforced caste segregation.

4.2 Pollution exposure — city-wide context

**Figure 4.** PM₂.₅ distribution across Delhi NCT — IDW surface from 28 CPCB stations. Every one of the 1,694 grid cells exceeds India's NAAQS annual mean standard of 40 µg/m³. The mean cell is exposed to approximately 2.8× the legal limit and 23× the WHO guideline.

Figure 4. PM₂.₅ distribution across Delhi NCT — IDW surface from 28 CPCB stations. Every one of the 1,694 grid cells exceeds India’s NAAQS annual mean standard of 40 µg/m³. The mean cell is exposed to approximately 2.8× the legal limit and 23× the WHO guideline.

Universal NAAQS breach. Mean IDW PM₂.₅ across 1,694 cells: 113.0 µg/m³ (range 63–174). Mean satellite PM₂.₅: 94.6 µg/m³ (range 71–109). Both sources show 100% of cells exceeding the 40 µg/m³ NAAQS and 100% exceeding the WHO guideline of 5 µg/m³ by more than 10×. The pollution crisis in Delhi is not localised — it is universal.

IDW PM₂.₅ Satellite PM₂.₅ NO₂ (×10⁻⁵) SO₂ (×10⁻⁵)
Mean 113.0 µg/m³ 94.6 µg/m³ 16.5 mol/m² 54.4 mol/m²
SD 17.1 8.7 3.1 1.8
Min 63.2 70.7 11.2 50.1
Max 174.2 109.4 24.5 60.3
NAAQS breach 100% 100% — —

n = 1,694 cells; NO₂/SO₂: n = 1,665 (29 cells lack Sentinel-5P coverage).

4.3 Group comparisons: SC bastis and pollution

The simplest test divides the 1,694 grid cells into two groups — those containing at least one SC basti (n = 489) versus those with none (n = 1,205) — and compares mean pollution with Welch’s t-test and Mann-Whitney U.

**Figure 5.** PM₂.₅ by SC basti presence. Left: boxplot showing higher median and mean in cells with SC bastis. Right: overlapping distributions — the SC-present distribution is shifted rightward. Mann-Whitney p < 0.001.

Figure 5. PM₂.₅ by SC basti presence. Left: boxplot showing higher median and mean in cells with SC bastis. Right: overlapping distributions — the SC-present distribution is shifted rightward. Mann-Whitney p < 0.001.

**Figure 6.** Cells binned by SC population share (ward-level). Both PM₂.₅ and NO₂ are higher in cells with >50% SC share than in 25–50% SC cells.

Figure 6. Cells binned by SC population share (ward-level). Both PM₂.₅ and NO₂ are higher in cells with >50% SC share than in 25–50% SC cells.

Pollutant SC present (n=489) SC absent (n=1,205) Δ % gap p-value
PM₂.₅ IDW 116.4 µg/m³ 111.7 µg/m³ +4.7 +4.2% < 0.001 ***
PM₂.₅ satellite 101.8 91.7 +10.0 +10.9% < 0.001 *
NO₂ (×10⁻⁵) 18.78 15.62 +3.16 +20.3% < 0.001 ***
SO₂ (×10⁻⁵) 55.66 54.11 +1.55 +2.9% < 0.001 ***

Cells with both SC bastis and JJ clusters (n = 132) show satellite PM₂.₅ of 104.0 µg/m³ vs 91.1 µg/m³ for cells with neither (n = 1,135) — a 14.1% gap (p < 0.001).

4.4 Regression models

4.4.1 Specification

All regressions use the 1,694-cell grid as the unit of observation. Outcomes are z-scored so β coefficients are comparable across pollutants. HC3 robust standard errors are used throughout.

Baseline model (applied to all four pollutant outcomes):

\[Y_{iz} = \alpha + \beta_1 \log(SC\_basti_i + 1) + \beta_2 \log(JJ_i + 1) + \beta_3 \log(pop_i + 1) + \beta_4 \cdot industrial_i + \varepsilon_i \tag{1}\]

where \(Y_{iz}\) is the z-scored pollution outcome; \(\log(SC\_basti_i + 1)\) is log-transformed SC basti count (primary predictor); \(\log(JJ_i+1)\) controls for informal cluster density; \(\log(pop_i+1)\) controls for satellite-derived population density; \(industrial_i\) is a binary industrial-proximity flag; and \(\varepsilon_i\) has HC3-robust standard errors.

Interaction specification — tests whether having both SC bastis and JJ clusters in a cell produces compounding pollution:

\[Y_{iz} = \alpha + \beta_1 \log(SC\_basti_i + 1) + \beta_2 \log(JJ_i + 1) + \beta_3 \log(pop_i + 1) + \beta_4 \cdot industrial_i + \beta_5 (SC\_basti_i \times JJ_i) + \varepsilon_i \tag{2}\]

Robustness specification — excludes 1,662 SC bastis (36.6%) placed by district centroid + jitter:

\[Y_{iz} = \alpha + \beta_1 \log(SC\_reliable_i + 1) + \beta_2 \log(JJ_i + 1) + \beta_3 \log(pop_i + 1) + \beta_4 \cdot industrial_i + \varepsilon_i \tag{3}\]

Monitoring gap specification — tests whether SC basti density predicts distance from monitoring infrastructure:

\[AQI\_dist_i = \alpha + \beta_1 \log(SC\_basti_i + 1) + \beta_2 \log(JJ_i + 1) + \beta_3 \log(pop_i + 1) + \varepsilon_i \tag{4}\]

Corrected deprivation index (cluster level, n = 675):

\[DI_j = 0.45 \times darkness_j + 0.40 \times HH\_size_j + 0.15 \times community\_flag_j \tag{5}\]

All components normalised 0–100; \(j\) indexes individual JJ clusters. Note: original index included aqi_dist_km at 27% weight, inflating r(DI, aqi_dist) from 0.048 to 0.667 — removed as a circular component.

4.4.2 Results

**Figure 7.** Regression results. Forest plot of log_SC_basti coefficients across all seven models (open circle = primary; square = robustness excluding jitter). Right panel: all predictor coefficients for Models A, B, C. HC3-robust SEs; z-scored outcomes.

Figure 7. Regression results. Forest plot of log_SC_basti coefficients across all seven models (open circle = primary; square = robustness excluding jitter). Right panel: all predictor coefficients for Models A, B, C. HC3-robust SEs; z-scored outcomes.

Summary of all models:

Model Outcome n β (log SC basti) SE p-value R²
A PM₂.₅ IDW 1,694 +0.166 0.027 3.6×10⁻⁹ *** 0.030
B PM₂.₅ satellite 1,694 +0.385 0.019 2.4×10⁻⁸⁷ * 0.194
C NO₂ Sentinel-5P 1,665 +0.337 0.024 6.3×10⁻⁴² * 0.198
D SO₂ Sentinel-5P 1,665 +0.159 0.027 6.0×10⁻⁹ *** 0.028
E PM₂.₅ sat. + SC×JJ 1,694 +0.385 0.019 <0.001 *** 0.197
F PM₂.₅ sat., reliable only 1,694 +0.275 0.020 <0.001 *** 0.116
F PM₂.₅ IDW, reliable only 1,694 +0.098 0.034 0.004 ** 0.014
G AQI distance (km) 1,694 −1.239 0.093 1.4×10⁻⁶³ *** 0.134

HC3 robust SE. ** p < 0.001; ** p < 0.01; * p < 0.05.*

Full predictor coefficients — primary models:

Predictor Model A (PM₂.₅ IDW) Model B (PM₂.₅ sat.) Model C (NO₂)
log_SC_basti +0.166 *** +0.385 * +0.337 ***
log_JJ −0.093 n.s. −0.247 * −0.255 n.s.
log_kontur_pop +0.021 n.s. +0.106 *** +0.134 ***
industrial_flag +0.679 * +0.285 n.s. +0.037 n.s.
R² 0.030 0.194 0.198
F-statistic 12.8 *** 172.5 *** 100.0 ***

Key findings from regressions. (1) SC basti density is a significant positive predictor across all four pollutants and both measurement approaches. (2) Satellite and NO₂ models (R² ≈ 0.19) have 6× better fit than IDW (R² = 0.03) because IDW smoothing removes local spatial variation. (3) Industrial flag is not significant for NO₂ (p = 0.877), confirming the NO₂ signal is driven by traffic corridors, not factories. (4) The SC×JJ interaction term is not significant (PM₂.₅ sat.: p = 0.091), indicating no compounding beyond additive effects. (5) Results hold in robustness models excluding all jitter-geocoded bastis (β = +0.275, p < 0.001).

Spatial autocorrelation note. Durbin-Watson statistics: 0.07 (IDW PM₂.₅), 0.32 (satellite PM₂.₅), 0.38 (NO₂) — all indicate positive spatial autocorrelation. Standard errors are corrected via HC3. Full Conley spatial HAC standard errors are not implemented here but would be appropriate for a final journal submission. Given p-values at 10⁻⁸⁷ and 10⁻⁴², results are robust to substantial SE widening.

4.5 Monitoring gap and institutional invisibility

**Figure 8.** Institutional invisibility — monitoring coverage of Delhi's JJ clusters. Left: 57% of 675 JJ clusters are unmonitored (> 2 km from CPCB station). Centre: AQI distance rises steeply across deprivation quartiles (Kruskal-Wallis H = 143.8, p < 0.001). Right: spatial distribution of monitored vs unmonitored clusters; OR = 24.9.

Figure 8. Institutional invisibility — monitoring coverage of Delhi’s JJ clusters. Left: 57% of 675 JJ clusters are unmonitored (> 2 km from CPCB station). Centre: AQI distance rises steeply across deprivation quartiles (Kruskal-Wallis H = 143.8, p < 0.001). Right: spatial distribution of monitored vs unmonitored clusters; OR = 24.9.

**Figure 9.** AQI monitoring gap across Delhi grid cells (left) and SC basti density by monitoring distance zone (right). Cells closest to stations have the most SC bastis — but the outer unmonitored ring covers large populations.

Figure 9. AQI monitoring gap across Delhi grid cells (left) and SC basti density by monitoring distance zone (right). Cells closest to stations have the most SC bastis — but the outer unmonitored ring covers large populations.

Of 675 JJ clusters, 381 (57%) are unmonitored — further than 2 km from the nearest CPCB station. The monitoring gap is sharply stratified by deprivation:

Quartile Mean AQI distance % unmonitored
Q1 (least deprived) 1.38 km 22%
Q2 1.63 km 38%
Q3 2.54 km 62%
Q4 (most deprived) 3.67 km 87%

Kruskal-Wallis H = 143.8, p < 0.001. Odds ratio Q4 vs Q1 unmonitored ≈ 24.9.

4.6 Deprivation index and double burden

**Figure 10.** Deprivation quartile profiles across four dimensions. Monitoring gap rises Q1→Q4 (1.4→3.7 km). PM₂.₅ does not follow deprivation — pollution is pervasive across all quartiles. Economic darkness and cluster size both rise with deprivation. Double burden: 153 clusters (23%) are both Q4 deprived and unmonitored.

Figure 10. Deprivation quartile profiles across four dimensions. Monitoring gap rises Q1→Q4 (1.4→3.7 km). PM₂.₅ does not follow deprivation — pollution is pervasive across all quartiles. Economic darkness and cluster size both rise with deprivation. Double burden: 153 clusters (23%) are both Q4 deprived and unmonitored.

**Figure 11.** Double burden clusters: 153 JJ clusters (Q4 deprivation + unmonitored), concentrated in north and east Delhi, representing approximately 440,000 people breathing air the state does not measure.

Figure 11. Double burden clusters: 153 JJ clusters (Q4 deprivation + unmonitored), concentrated in north and east Delhi, representing approximately 440,000 people breathing air the state does not measure.

Reframing the pollution inequality claim. The corrected deprivation index (Equation 5) shows no clear PM₂.₅ gradient across quartiles. This is the honest finding: Delhi’s pollution is so uniformly catastrophic that deprivation quartiles cannot separate exposure levels. The inequality documented here is (a) the spatial gradient — SC basti density within the city predicts locally higher pollution — and (b) institutional invisibility — more deprived communities are systematically absent from state monitoring infrastructure.

5 Discussion

Three independent lines of evidence converge on the same conclusion.

Spatial gradient. SC basti density is a significant positive predictor of satellite PM₂.₅ (β = 0.385, R² = 0.19) and NO₂ (β = 0.337, R² = 0.20) after controlling for population density and industrial proximity. Cells with more than 15 SC bastis per km² have satellite PM₂.₅ 13% higher and NO₂ 22% higher than cells with none.

Source attribution. The industrial flag is not significant for NO₂ but is for IDW PM₂.₅, identifying traffic corridors as the primary NO₂ source. SC communities appear disproportionately located along major road corridors — consistent with the historical pattern of Balmiki/Harijan communities being placed in road-adjacent sanitation zones.

Institutional invisibility. 57% of JJ clusters are unmonitored. Q4 clusters are 3.67 km from the nearest station on average — 2.4× the CPCB standard — and are 25× more likely to be unmonitored than Q1 clusters. 153 clusters (23%) face a double burden.

5.0.1 Limitations

  • The SC basti variable is a settlement-density proxy, not a measured SC population share at grid level.
  • IDW PM₂.₅ is methodologically weak — its smooth surface removes most local spatial variation. Satellite PM₂.₅ and NO₂ are the primary specifications.
  • Temporal mismatch: SC basti data is 2011, JJ clusters are 2022, pollution is 2022–23.
  • Sentinel-5P measures tropospheric column density, not ground-level concentration. The relationship is non-linear during winter inversions.
  • Conley spatial HAC standard errors not yet implemented.

6 Conclusion

Delhi’s air quality crisis does not fall equally. SC settlement density predicts higher PM₂.₅ and NO₂ across two independent measurement approaches, confirmed by p-values at 10⁻⁸⁷ and 10⁻⁴¹. The NO₂ finding — robust, driven by traffic rather than industry — points to a geography of caste-segregated road proximity. The monitoring gap finding is perhaps the most policy-actionable result: 381 JJ clusters containing hundreds of thousands of people are invisible to Delhi’s air quality surveillance system. Communities that cannot be measured cannot be helped.

7 References

Pant, P., et al. (2017). Characterisation of ambient PM₂.₅ at a pollution hotspot in New Delhi. Atmospheric Environment, 109, 66–77.

van Donkelaar, A., et al. (2021). Monthly global estimates of fine particulate matter and their uncertainty. Environmental Science & Technology, 55(22), 15287–15300.

Mohai, P., & Saha, R. (2007). Racial inequality in the distribution of hazardous waste. Social Problems, 54(3), 343–370.

CPCB (2023). Annual Report on Ambient Air Quality Monitoring 2022–23. Ministry of Environment, Forest and Climate Change, India.

DUSIB (2022). JJ Cluster Enumeration 2022. Delhi Urban Shelter Improvement Board.

A P P E N D I X

8 Appendix A: Additional Figures

8.1 A1. Full spatial co-distribution maps

**Figure A1.** SC settlement density (left), satellite PM₂.₅ with SC bastis and monitoring stations overlaid (centre), and NO₂ Sentinel-5P (right). SC bastis concentrate in the north and east inner ring, which spatially overlaps elevated satellite PM₂.₅ and high NO₂.

Figure A1. SC settlement density (left), satellite PM₂.₅ with SC bastis and monitoring stations overlaid (centre), and NO₂ Sentinel-5P (right). SC bastis concentrate in the north and east inner ring, which spatially overlaps elevated satellite PM₂.₅ and high NO₂.

8.2 A2. IDW vs satellite PM₂.₅ — full comparison

**Figure A2.** IDW versus satellite PM₂.₅. Scatter coloured by SC density (A); boxplots by SC presence for both sources (B); cell-level divergence — IDW overestimates in 84.5% of cells (C).

Figure A2. IDW versus satellite PM₂.₅. Scatter coloured by SC density (A); boxplots by SC presence for both sources (B); cell-level divergence — IDW overestimates in 84.5% of cells (C).

8.3 A3. Pollution gradient across SC density bands

**Figure A3.** All four pollutants plotted against SC density bands (0, 1–2, 3–5, 6–15, >15 bastis per km²). Each rises monotonically. NO₂ shows the steepest gradient (+22%). Error bars are 95% CI.

Figure A3. All four pollutants plotted against SC density bands (0, 1–2, 3–5, 6–15, >15 bastis per km²). Each rises monotonically. NO₂ shows the steepest gradient (+22%). Error bars are 95% CI.

8.4 A4. Core inequality — full 8-panel figure

**Figure A4.** Eight-panel summary of the core inequality finding: top row shows pollution gradients across density bands; bottom row shows scatter plots and group comparisons by settlement combination type.

Figure A4. Eight-panel summary of the core inequality finding: top row shows pollution gradients across density bands; bottom row shows scatter plots and group comparisons by settlement combination type.

8.5 A5. PM₂.₅ distribution maps

**Figure A5.** IDW PM₂.₅ spatial distribution (left), satellite PM₂.₅ (centre), and both distributions vs NAAQS/WHO thresholds (right).

Figure A5. IDW PM₂.₅ spatial distribution (left), satellite PM₂.₅ (centre), and both distributions vs NAAQS/WHO thresholds (right).

9 Appendix B: Geocoding Quality

Method n % Notes
Nominatim (OpenStreetMap) 1,325 29.2% Real coordinates from OSM API
Known-place dictionary 1,554 34.2% ~200-entry curated Delhi place dictionary
District centroid + jitter (±2.77 km) 1,662 36.6% Fallback for unresolved localities

Robustness models (Equation 3) use only Nominatim + dictionary geocoded points (2,879 bastis). SC basti density result: β = +0.275, p < 0.001 (vs β = +0.385 in main model).

10 Appendix C: Grid Construction

Resolution    : 0.009° × 0.009° (~1 km at Delhi latitude)
Bounding box  : 76.84°E–77.35°E, 28.40°N–28.88°N
Spatial filter: geopandas.sjoin(predicate="within", delhi_districts.shp)
Valid cells   : 1,694
Satellite NaN : 29 cells (SW periphery, Najafgarh area — orbital/cloud artifact)
                These 29 cells contain 3 SC bastis and 5 JJ clusters (negligible)

11 Appendix D: Deprivation Index Sensitivity

Weight scheme Q4 overlap r(DI, aqi_dist)
Primary (45/40/15) — 0.048
Equal (33/33/33) 97% 0.052
Darkness-heavy (60/30/10) 94% 0.041
No community (50/50/0) 99% 0.049

Q4 membership is stable across all weight schemes (93–99% overlap). The corrected deprivation finding does not depend on the specific weights chosen.

Document generated with R Markdown · Analysis conducted in Python 3.12 with statsmodels, geopandas, rioxarray, and scipy · All code available on request