Datasets COVID-19 Time Series Combined
- Rows
- 231,744
- Columns
- 6
- Data range
- 2020-01-22 to 2022-04-16
- Auto-refresh
- Re-checks its source every 30 days
Combined global COVID-19 time-series cases, deaths, recovered.
AI analysis
COVID-19 Time Series Combined — Dataset Analysis
1. Dataset Overview & Value for Correlation Studies
This dataset represents a global longitudinal record of COVID-19 pandemic progression, tracking cumulative confirmed cases, deaths, and recoveries across 198 countries and regions over 816 distinct dates — spanning what appears to be the bulk of the pandemic's recorded history. Sourced from the [Johns Hopkins University CSSE data pipeline via the datasets organization on GitHub](https://raw.githubusercontent.com/datasets/covid-19/main/data/time-series-19-covid-combined.csv), this is a well-established, community-maintained dataset with strong provenance. Its time-series structure makes it exceptionally well-suited for temporal correlation analysis, including lag-effect studies (e.g., do case spikes precede death spikes by a predictable interval?), cross-country comparative modeling, and policy-impact analysis when joined with intervention timeline data. The granularity at both the country and sub-national (province/state) level adds analytical depth, though that granularity comes with caveats discussed below.
---
2. Data Quality Observations
The dataset's overall quality is moderate to good at the country level but degraded at the sub-national level. Of the 231,744 total rows, there are 172,176 null cells across 6 columns — a figure that sounds alarming but is heavily concentrated in two specific columns. Province/State accounts for 159,120 nulls (68.7%), which is structurally expected: most countries report only national-level data, making sub-national breakdowns sparse rather than missing in a problematic sense. However, analysts must treat this column with care — nulls here mean "national aggregate," not "unknown," and conflating the two in a join or groupBy operation could produce duplicate-counting errors. More concerning is Recovered with 13,056 nulls (5.6%), representing a meaningful gap; notably, many jurisdictions stopped reporting recoveries mid-pandemic as a formal metric, meaning this column becomes increasingly unreliable in later time windows and should not be used as a primary metric without date-range filtering. The Confirmed and Deaths columns report zero nulls, establishing them as the most analytically reliable fields. Duplicate row counts are flagged as pending recomputation, which should be resolved before any aggregation work — given the composite key structure (Date + Country/Region + Province/State), duplicates at the province level would silently inflate national rollups.
---
3. Key Column Distributions
The three numeric columns tell a vivid story of extreme skewness and outlier concentration — a hallmark of pandemic data where a small number of large nations dominate global totals:
- Confirmed: Mean of 513,236 vs. a median of only 5,725 reveals massive right skew (skewness = 14), confirmed by the quartile spread: 75% of records sit below 109,564 cases, yet the maximum reaches ~80.6 million. The 39,937 statistical outliers likely represent the US, India, Brazil, and a handful of other high-burden nations in later time periods. Any correlation model using raw confirmed counts should apply log-transformation. - Recovered: Even more skewed (skewness = 21.1), with a median of just 55 and Q1 = 0, meaning the bottom quartile of records shows no recoveries at all — consistent with early pandemic data and nations that stopped tracking. The 43,957 outliers are the highest outlier count of any column, further undermining this field's reliability. - Deaths: Skewness of 10.4, mean of 9,760 versus a median of 73, with Q1 = 2 and Q3 = 1,777. The max of 988,609 represents peak US cumulative deaths. With 41,039 outliers but zero nulls, this column is the most statistically complete while still requiring log-transformation for linear correlation work. The case fatality rate (Deaths/Confirmed) derivable from these two columns is one of the most analytically valuable computed fields possible here.
---
4. Recommended Join Key Columns
The natural composite join key is Date + Country/Region for country-level analysis — both columns carry zero nulls and together produce a highly distinct record identifier across the dataset's temporal span. For sub-national analysis, extend the key to include Province/State, but implement explicit null-handling (treat null as a sentinel value like "National" rather than excluding rows). When joining to external datasets:
| Join Scenario | Recommended Key(s) | Notes | |---|---|---| | Cross-dataset temporal joins | Date | Standardize to ISO 8601; 816 distinct values suggest daily granularity | | Country-level policy/demographic data | Country/Region | 198 distinct values — watch for naming inconsistencies (e.g., "US" vs "United States") | | Sub-national health data | Province/State + Country/Region | Only 89 distinct provinces; sparse coverage |
Country name normalization (e.g., mapping to ISO 3166-1 alpha-2/3 codes) is strongly recommended before any multi-dataset join to avoid silent mismatches on country name variants.
---
5. Recommended Companion Datasets for Correlation Discovery
Several dataset categories would unlock high-value correlations when joined to this time series:
1. Government Policy/Intervention Data (e.g., Oxford COVID-19 Government Response Tracker / OxCGRT): Joining on Date + Country/Region would enable analysis of whether lockdown stringency indices correlate with case growth rates — a foundational pandemic research question with direct lag-analysis potential. 2. Vaccination Rollout Data (e.g., Our World in Data vaccination timeseries): A Date + Country/Region join would support modeling the relationship between vaccination rates and the deceleration of confirmed cases and deaths — arguably the most policy-relevant correlation available. 3. Demographic & Health System Data (e.g., World Bank WDI — population, GDP per capita, hospital beds per 1,000): Joining on Country/Region would allow normalization of raw counts into per-capita rates and enable cross-sectional analysis of whether health infrastructure predicted mortality outcomes. 4. Mobility / Human Behavior Data (e.g., Google COVID-19 Community Mobility Reports): Temporal joins on Date + Country/Region could reveal whether reductions in retail/transit mobility preceded case declines, supporting causal inference studies. 5. Variant Surveillance Data (e.g., GISAID or Nextstrain phylogenetic timelines): Joining on approximate date ranges and country would allow analysis of whether specific variant emergence events (Alpha, Delta, Omicron) correspond to inflection points visible in the Confirmed and Deaths time series in this dataset.
Bottom line: This dataset is a strong analytical foundation — temporally rich, geographically broad, and sourced from a credible pipeline. Its primary weaknesses (Recovered reliability, country-name normalization needs, sub-national sparsity) are well-understood and manageable. The extreme skewness across all three numeric columns means log-transformation and per-capita normalization are prerequisites for any meaningful correlation or regression analysis.
Columns
- Date (date)
- Country/Region (string)
- Province/State (string)
- Confirmed (integer)
- Recovered (integer)
- Deaths (integer)