Claims that poorer and non-white neighborhoods breathe dirtier air increasingly rest on dense networks of cheap air-quality sensors. Those networks are thinnest in exactly the neighborhoods being studied. Fresno has 47 sensors where Santa Clara has over 600 for a comparable area (GraPhy). Commercial sensors cluster in affluent white neighborhoods (mrDMDcc). The sensors that do exist drop out in contiguous blocks. Analysts fill these holes with a single "best guess" pollution map and then compute a disparity, such as the gap in average exposure between groups, as if the filled values had been measured.
We propose to treat the disparity itself as the quantity that imputation must get right. The method is multiple imputation: draw many plausible completed pollution maps rather than one, run the disparity analysis on each, and pool the results so the final uncertainty reflects what was never observed. We pair this with a sensitivity analysis that asks how much more polluted the unmonitored disadvantaged tracts would have to be, beyond what the model predicts, for the disparity conclusion to change. The test is a controlled one. We take dense networks where the true disparity is known, thin them out in the same demographically skewed way that real deployments are skewed, and check which pipelines still recover the true gap with honest error bars.
Core Claim
The estimand that matters for environmental justice is a population-weighted exposure difference. For two groups $A$ and $B$, this is
$$D = \frac{\sum_i P_i^{A}\,\bar{x}_i}{\sum_i P_i^{A}} - \frac{\sum_i P_i^{B}\,\bar{x}_i}{\sum_i P_i^{B}},$$
where $P_i^{A}$ is the number of group-$A$ residents in census tract $i$ and $\bar{x}_i$ is that tract's mean PM2.5 concentration over the study window. A second, allocative estimand is the set of tracts that cross a funding threshold.
Our claim has two parts. The first is that under demographically skewed sensor coverage, single deterministic imputation produces disparity estimates with confidence intervals that miss the true $D$ far more often than their nominal rate, however accurate the imputer is on average. The second is that a spatially structured multiple imputation repairs most of this, provided its imputation model includes the same demographic variables used in the disparity analysis. Statisticians call this requirement "congeniality": an imputation model that omits a variable the downstream analysis uses will bias that variable's association toward zero.
What remains after that repair is missingness that depends on the unseen values themselves, called missing-not-at-random. Examples are sensors that foul during smoke events, and unmonitored tracts where national models are known to under-predict. We handle that remainder with a calibrated "tipping point" analysis rather than pretending a model can identify it.
The prior work shows why each piece matters. Imputation error is systematically larger for marginalized groups in tabular data (Fairness in Missing Data Imputation), but nobody has measured this over continuous pollution fields. National land-use models under-predict street-level pollution more in less-white neighborhoods (Local exposure misclassification in national models), so using them as the imputation prior pushes the error in a known direction. Threshold-based screening tools flip 16.1% of tracts under plausible design changes (Mitigating allocative tradeoffs), so the size of the uncertainty has dollar consequences.
The closest prior work is GraPhy (Graph-Based Physics-Guided Urban PM2.5 Imputation), a physics-guided graph network that benchmarks Fresno against Santa Clara. It produces one deterministic map, evaluates pointwise error, and never asks whether the downstream disparity estimate is biased or correctly uncertain. Our contribution is not a better point imputer. It is an inference pipeline and an evaluation protocol aimed at the disparity.
Research Program
The imputer is deliberately conventional. The air-quality literature does not support the claim that heavier models win on real sensor outages. On a 4,400-sensor environmental network with realistic block outages, low-rank matrix completion reached MAE 0.341, while a deep denoising autoencoder reached 1.019 and splines 2.520 (Missing Value Imputation of Wireless Sensor Data). On mobile data in Delhi, a random forest using only target-day data matched models trained on long histories (AirDelhi).
We therefore use a Bayesian spatio-temporal low-rank model. This is matrix completion with a prior, so it returns posterior draws instead of a single fill. Its mean structure includes four inputs: a national land-use regression prediction, city-level wind interacted with the direction between neighboring sensors (borrowing GraPhy's convection idea as a covariate rather than as an architecture), tract race and income entered separately, and hour-of-week effects. Race and income enter separately because their geographic patterns decouple across cities (mrDMDcc). We draw $M=50$ completed fields, compute $D$ and threshold membership on each, and pool the results with Rubin's rules, the standard formulas that combine within-draw and between-draw variance.
The sensitivity analysis shifts imputed values in unmonitored or dropped-out locations by $\delta$ times a disadvantage indicator. It then reports the $\delta^*$ at which the pooled interval for $D$ first includes zero, or at which the threshold set changes materially. To make $\delta$ interpretable, we calibrate its plausible range empirically. The Bay Area mobile-monitoring data in Local exposure misclassification in national models give the observed relationship between local model error and tract demographics. A $\delta^*$ well outside that range means the conclusion is robust. A $\delta^*$ inside it means the network cannot settle the question, and that answer is itself useful because it identifies where new sensors are needed.
Baselines
The naive reference is complete-case analysis: compute $D$ using only tracts that contain a working sensor. This is what an analyst does when unwilling to impute, and it should be biased under skewed coverage. A second naive reference fills every gap with the national land-use regression. This represents the "no imputation needed, the national model is complete" position. It is the strongest version of that position because the model has no missing pixels.
The strongest point-prediction competitor is a physics-guided graph network in the style of GraPhy, used as a single imputation. GraPhy reports a 24%–56% MSE reduction over IGNNK and Geo-LSTM in Fresno and Santa Clara. There is no reported result found on Lacuna for its absolute error, or for any disparity or interval-coverage metric. If its code is unavailable, we use IGNNK, the baseline GraPhy compared against, with the same wind inputs. We also run single-draw matrix completion, since it was the most accurate method on realistic block outages (MAE 0.341 in Missing Value Imputation of Wireless Sensor Data).
The strongest multiple-imputation competitor is off-the-shelf MICE with predictive mean matching. MICE fills each variable in turn by regressing it on the others and drawing from that regression. We run it with $M=50$ draws, with the demographic variables included, and pooled with Rubin's rules. On Guro-gu PM2.5, MICE used as a point imputer had MAE 4.825 versus 1.170 for a deep decomposition model (Missing Value Imputation of Time-Series Air-Quality Data). That result concerns pointwise accuracy, not disparity coverage, which is why MICE stays in as a serious competitor.
Two ablations isolate our ingredients. The first removes the demographic covariates from our imputer's mean, testing congeniality. The second collapses our imputer to its posterior mean, testing multiplicity. For a fair comparison, every method sees the same retained sensors, covariates, and train/validation split, with hyperparameters tuned by the same 30-trial random search on held-out sensors from the retained set only.
Beating the baselines means two things in the thinning experiment below. Our pooled 90% intervals must cover the full-network $D$ at a rate within a few points of 90%. The single-imputation baselines must fall well short of that rate, with absolute bias in $D$ no larger than theirs. We also report the gap in tract-level imputation error between disadvantaged and other tracts, the spatial analogue of the imputation accuracy parity difference in Fairness in Missing Data Imputation.
Why not just…?
Why not just use the most accurate point imputer, such as GraPhy, and bootstrap the disparity? Bootstrapping tracts captures sampling variation but treats imputed concentrations as known. The uncertainty in unmonitored tracts is exactly the component it drops. Lower pointwise MSE also does not imply unbiased $D$. Error that is small on average can still be concentrated in the sparse tracts, which is the pattern Fairness in Missing Data Imputation found in tabular data. The thinning experiment measures coverage directly and will show whether this objection holds.
Why not just run standard MICE with demographics included? This is the real threat. If congenial MICE achieves nominal coverage, the honest conclusion is that the existing tool suffices once it is used correctly, and we will report that. Our expectation that it falls short comes from its pointwise weakness on sensor streams (MAE 4.825 vs 1.170 in Missing Value Imputation of Time-Series Air-Quality Data). It also comes from the fact that MICE has no spatial borrowing, so contiguous outages starve it. Under block outages we expect wide or biased intervals. Running MICE against our imputer with and without demographics separates the contribution of spatial structure from the contribution of congeniality.
Why not just use the national model, which has no gaps? That model's error correlates with race: mobile UFP observations were more than twice the national predictions, and the shortfall grew with non-white share (Local exposure misclassification in national models). It therefore biases $D$ in a predictable direction. It is still included as a baseline, and as the prior inside our model, so the experiment quantifies how much sensor data corrects it.
Why not just deploy more sensors? That is the right long-term remedy, and our tipping-point output tells planners where sensors would actually change a conclusion. But even dense deployments lose data in blocks (Missing Value Imputation of Wireless Sensor Data), and existing disparity claims must be judged with the networks that exist now.
Proposed Experiments
The decisive experiment is demographically skewed thinning. We start from the dense networks: Santa Clara (about 600 sensors), the 294-sensor calibrated Boston network from Mobility-driven estimate reveals elevated air pollution exposure, and Fresno (47 sensors) as a real sparse case, rebuilt from public archives where the original data are not released. On each dense network we compute the "true" $D$ for race and for income separately.
We then remove sensors with a retention probability that rises with tract income and white share. We tune the skew so that the retained network's demographic imbalance and density span the range from Santa Clara-like down to Fresno-like. On top of the thinning we overlay block outages drawn from the network's own observed dropout patterns, and a high-concentration dropout mode in which sensors fail during plume events of the kind documented in Characterizing the effects of structural fires. With 200 thinning replicates per setting, we measure bias in $D$, 90% interval coverage and width, the tract error-parity gap, and the rate at which threshold membership is wrong compared with the full network.
A second check validates the sensitivity analysis. In the Bay Area mobile data we hold out street-level observations in sensor-poor tracts. We then ask whether the realized discrepancy between the imputed and observed values falls inside the calibrated $\delta$ range, and whether conclusions flagged as fragile are the ones the held-out truth contradicts.
A third, applied, check runs the full pipeline on the actual Fresno network. We report whether the PM2.5 disparity between groups there survives pooled uncertainty, and at what $\delta^*$ it would vanish.
The proposal fails if any of three things happens. Single imputation might already achieve near-nominal coverage, which would mean uncertainty from missing sensors is negligible for disparity estimates; that would be worth knowing, and it would end this line of work. Congenial MICE might match our imputer. Or calibrated $\delta^*$ values might always fall far outside the empirical range, which would make the sensitivity analysis decorative.
References
- Graph-Based Physics-Guided Urban PM2.5 Air Quality Imputation with Constrained Monitoring Data
- Data‐Driven Placement of PM2.5 Air Quality Sensors in the United States: An Approach to Target Urban Environmental Injustice
- Fairness in Missing Data Imputation
- Local exposure misclassification in national models: relationships with urban infrastructure and demographics
- Mitigating allocative tradeoffs and harms in an environmental justice data tool
- Missing Value Imputation of Wireless Sensor Data for Environmental Monitoring
- AirDelhi: Fine-Grained Spatio-Temporal Particulate Matter Dataset From Delhi For ML based Modeling
- Missing Value Imputation of Time-Series Air-Quality Data via Deep Neural Networks
- Mobility-driven estimate reveals elevated air pollution exposure and socioeconomic disparities beyond residence-based approaches in Boston
- Characterizing the effects of structural fires on fine particulate matter with a dense sensing network