Worldwide COVID-19 Air Quality Dataset 2026

Rows: 1,564,571
Columns: 9
Data range: 2025-12-29 to 2026-10-30
Auto-refresh: Re-checks its source daily
Last refreshed with new data: 2026-10-30

Global air quality time-series (PM2.5, PM10, etc.) from hundreds of cities/stations, daily aggregates 2015-present. Open scientific source for environmental IoT/sensor data, zeitgeist on pollution trends.

AI analysis

Worldwide COVID-19 Air Quality Dataset 2026 — Analytical Overview

1. Dataset Representation & Research Value

This dataset, sourced from the World Air Quality Index project at [aqicn.org](https://aqicn.org/data-platform/covid19/report/14724-6c14d6e7/2026), represents a longitudinal, globally distributed time-series of air quality measurements aggregated at the daily level across 529 cities in 71 countries, spanning from 2015 to the present (2026). Each row captures a specific pollutant species (e.g., PM2.5, PM10, ozone, NO₂) measured at a city-level station on a given date, yielding per-day statistical summaries — count of readings, min, max, median, and variance — derived from underlying IoT sensor networks. With nearly 789,000 rows and coverage across multiple continents and pollutant types, this dataset is exceptionally well-suited for correlation studies examining how air quality shifted during the COVID-19 pandemic lockdown periods, how pollution levels trend seasonally or across years, and how industrial or traffic activity influences atmospheric composition across geographies. The AQICN platform is a recognized open scientific aggregator of real-time and historical sensor data, lending credibility and relative freshness to the observations. Notably, the time coverage join-key range appears as unresolved template placeholders ({{firstValue}} to {{lastValue}}), which should be resolved before any temporal scoping is performed in production pipelines.

---

2. Data Quality Observations

On the surface, data completeness is remarkable: total null cells across all 9 columns is zero, and no column carries a disproportionate null burden — a strong indicator that the ingestion pipeline from the API/CSV source is well-structured. However, the absence of nulls does not imply absence of data quality issues; several columns show deeply concerning distributional anomalies. The min column reports values as low as -3,233.9, and the median column dips to -42, both of which are physically impossible for air quality concentration measurements and strongly suggest sensor faults, calibration errors, or placeholder sentinel values that were not filtered upstream. Similarly, the max column reaches 3,152.3 — far exceeding any credible real-world pollution reading for most species — pointing to additional erroneous sensor spikes. The variance column's maximum of ~27 million is consistent with these pathological extremes inflating spread. The duplicate-row count is noted as pending Phase D recomputation against the Parquet store, which is a critical gap: given that the dataset joins on Date × Country × City × Specie (a four-part natural key), duplicate combinations across those dimensions could silently distort any aggregation or join operation. This must be resolved before the dataset is used analytically. The outlier counts are substantial across every numeric column — 119,727 outliers in min, 108,359 in variance, and 85,045 in median — collectively representing roughly 10–15% of total rows flagged in at least one column, which is a non-trivial contamination rate requiring cleaning before modeling.

---

3. Key Column Distributions

The count column (readings per city-day-species aggregate) has a mean of 123 but a median of only 72, with a right skew of 6.6 and a Q1/Q3 range of 43–144 — indicating that most station-days have moderate sensor throughput, but a long tail of high-density monitoring stations (max: 2,640 readings/day) pulls the mean upward. The 61,337 outliers in count suggest a subset of cities with dramatically higher sensor density, which could introduce heteroskedasticity in any cross-city comparison. The median pollution column is the most analytically central: its overall median is 12.5 µg/m³ (roughly in the "moderate" range for PM2.5 under WHO guidelines), but the mean of 116.1 is orders of magnitude higher due to extreme outliers, rendering the mean unreliable as a central tendency measure — the median or trimmed mean should be used for analysis. The variance column is the most extreme: a skewness of 216 and a max of ~27M against a median of just 103 makes this column nearly unusable without log-transformation or robust scaling. The 17 distinct pollutant species in the Specie column implies the data is wide in measurement types, so species-stratified analysis (e.g., analyzing only PM2.5 rows) will be essential to avoid mixing incommensurable units and scales in any correlation exercise.

---

4. Recommended Join Key Columns

The natural composite join key for this dataset is Date × Country × City × Specie, and all four columns carry zero nulls with meaningful cardinality: 153 distinct dates, 71 countries, 529 cities, and 17 species. For cross-dataset joins, Country + Date is the most broadly compatible key pairing, offering the widest match surface against macroeconomic, health, or policy datasets that typically report at the country-day or country-month granularity. City + Date is the preferred key for joins against urban mobility, traffic, or hospital admission data, though city-name normalization (e.g., standardizing to ISO city codes or UN LOCODE) will likely be required since raw city strings across datasets are notoriously inconsistent. The Date column alone (153 distinct values) appears suspiciously low for a dataset spanning 2015–2026 (~4,000 days), suggesting dates may represent aggregated reporting windows rather than individual calendar days — this discrepancy should be investigated before using Date as a high-resolution join key.

---

5. Suggested Complementary Datasets for Correlation Discovery

Several dataset categories would pair powerfully with this air quality data:

- COVID-19 Case & Mortality Data (e.g., Our World in Data, WHO COVID-19 dashboard): The most direct pairing — correlating lockdown stringency periods with sudden drops in PM2.5/NO₂ at city level is a well-validated research pattern, and this dataset's 2015-present coverage provides a clean pre-pandemic baseline. - Oxford COVID-19 Government Response Tracker (OxCGRT): Stringency index scores by country and date would allow regression analysis between policy interventions (school closures, mobility restrictions) and pollution metric changes. - Urban Mobility / Traffic Data (Google Mobility Reports, Apple Mobility Trends, Uber Movement): City-level traffic volume is a strong causal driver of PM2.5 and NO₂; joining on City + Date could reveal lagged correlations between reduced movement and improving air quality. - Global Weather / Climate Data (NOAA, ERA5 reanalysis): Temperature, wind speed, humidity, and precipitation are confounders for pollution dispersion; without controlling for meteorology, any COVID-lockdown air quality signal risks spurious attribution. - Industrial Production & Energy Consumption Indices (IEA, World Bank): Country-level manufacturing output and electricity generation (especially coal-based) are structural drivers of PM10 and SO₂, making this a natural pairing for longer-term trend analysis beyond the pandemic window.

Columns

Date (date)
Country (string)
City (string)
Specie (string)
count (integer)
min (decimal)
max (decimal)
median (decimal)
variance (decimal)