NCHS – Provisional Drug Overdose Death Counts
- Rows
- 88,182
- Columns
- 12
- Data range
- 2015-01-01 to 2026-01-01
- Auto-refresh
- Re-checks its source every 30 days
- Last refreshed with new data
- 2026-11-13
County-level provisional counts for drug overdose deaths from the National Vital Statistics System. Covers multiple drug types and demographic groups in a current-flow time series.
AI analysis
Dataset Analysis: NCHS Provisional Drug Overdose Death Counts
1. Dataset Overview and Research Value
This dataset, sourced from the CDC's National Center for Health Statistics via the public API at data.cdc.gov, represents provisional county-level counts of drug overdose deaths drawn from the National Vital Statistics System. With 83,160 rows spanning 11 years, 54 state/territory designations, and 12 distinct drug-type or demographic indicators, it constitutes a rich longitudinal surveillance dataset for one of the most pressing public health crises in the United States. The "provisional" and "current-flow" nature of the data means it reflects near-real-time mortality reporting, making it particularly valuable for trend detection and policy evaluation rather than purely retrospective analysis. Its breadth across drug types (likely including synthetic opioids, heroin, stimulants, and combined categories) and demographic breakdowns creates strong potential for multivariate correlation studies linking overdose patterns to socioeconomic, healthcare access, or law enforcement variables.
2. Data Quality Observations
The dataset carries notable quality concerns that require careful handling before analysis. Across 12 columns, there are 123,034 null cells — a figure that is disproportionately concentrated in three columns. Percent Pending Investigation is null in 72,666 rows (87.4%), rendering it nearly useless for analysis and suggesting it was likely populated only for a specific data vintage or reporting jurisdiction. Predicted Value is missing in 28,861 rows (34.7%), which limits its use as a smoothing or forecasting supplement. Most critically, Data Value — the primary outcome variable — is null in 21,507 rows (25.9%), likely reflecting suppressed counts where death totals fell below reporting thresholds for privacy reasons, a common CDC practice. Duplicate row counts are flagged as pending recomputation, so deduplication should be verified before any aggregation. The Percent Complete column shows zero variance (min=max=mean=100, σ=0), making it a constant and analytically inert — it can be safely dropped.
3. Key Column Distributions
The Data Value column is the analytical centerpiece and exhibits a dramatically right-skewed distribution (skew=20.7), with a mean of 14,261.7 but a median of only 377 — a roughly 38x gap that signals the presence of large aggregate national or cumulative records coexisting with granular county-level counts. The interquartile range (Q1=111, Q3=1,527) confirms that the bulk of records represent relatively small counts, while 9,489 outlier rows pull the maximum to an extreme 3,538,080 — almost certainly representing cumulative national totals or multi-year rollups rather than single-period death counts. The Predicted Value column is better behaved but still heavily skewed (skew=10.5, max=112,418, mean=1,435 vs. median=308), with 5,658 outliers, suggesting it too reflects tiered geographic aggregation. The Indicator column's 12 distinct values are critical for stratified analysis — filtering by a single drug type or demographic group before computing statistics will be essential to avoid mixing incompatible units.
4. Recommended Join Key Columns
The most reliable join keys for cross-dataset correlation are Year (11 distinct values, 0 nulls) and State or State Name (54 distinct values, 0 nulls — the duplication between these two columns should be resolved, keeping one). The combination of Year + State + Month + Indicator forms a natural composite key for granular time-series joins. It is worth noting that 54 state values exceeds the 50 U.S. states, implying inclusion of D.C., Puerto Rico, and possibly other territories — any joining dataset must account for this to avoid row loss. Month (12 distinct, 0 nulls) enables seasonal join alignment. The absence of a county-level FIPS code column is a limitation given the dataset's described county-level granularity; if FIPS codes are embedded elsewhere or available in the full extract, they would unlock far richer geographic joins.
5. Recommended Pairing Datasets for Correlation Analysis
Several dataset categories would pair powerfully with this data. Prescription drug monitoring data (e.g., CDC's opioid dispensing rate by county) could reveal correlations between prescribing patterns and overdose mortality. Socioeconomic indicators such as unemployment rates, poverty levels, and median household income from the U.S. Census Bureau's American Community Survey (ACS) would test whether economic distress predicts overdose surges at the state or county level. Healthcare access datasets — including HRSA data on treatment facility density, naloxone distribution programs, or uninsured rates — could illuminate harm-reduction correlations. Law enforcement seizure data (e.g., DEA drug seizure statistics) would allow analysis of supply-side shocks on overdose mortality. Finally, COVID-19 mortality or policy stringency datasets would support interrupted time-series analysis of how pandemic disruptions affected overdose trajectories, given the dataset's multi-year span almost certainly covers 2020–2022.
Columns
- Year (date)
- State (string)
- Month (string)
- Period (string)
- Indicator (string)
- Data Value (integer)
- Percent Complete (integer)
- Percent Pending Investigation (integer)
- State Name (string)
- Footnote (string)
- Footnote Symbol (string)
- Predicted Value (integer)