Executive Summary
Anyone counting severe weather from the official U.S. storm record faces a choice of denominator that the record itself does not make for them. NOAA's database holds events and episodes: an episode is an entire storm system and an event is an individual event type within it.
The qualification that governs every number below. 41.8% of the state-scoped episodes in this window carry more than one event type. An event-type figure in this paper therefore counts the records of that one type inside an episode, and is never a count of the whole storm system. No pooled total is used as a multiplier.
Within that bound, this paper measures how many county-coded records of a single event type appear within an episode. The window contains 82,121 county-coded event records and 10,236 state-scoped episodes, but the primary statistic never divides those pooled totals; it is computed separately by event type. Among qualifying cells, hail and thunderstorm wind run at a median of 3 records of the named type within an episode, and flash flood, flood and tornado at 2. The tails matter more than the medians — a 90th-percentile hail episode contains 16 hail records and the maximum is 160.
Of 428 state-year-event-type cells, 110 meet the 50-episode floor and 109 of those show material multiplicity under the prespecified rule. The 318 cells below the floor are reported separately in the suppressed-cell evidence table and excluded from every headline. This frequency does not identify why the observed multiplicity occurs.
The most important limitation is what multiplicity is not. It is not duplication, not error, and not evidence that anything was over-reported. For tornadoes NOAA documents the mechanism directly: a tornado crossing a county or state line is recorded as separate segments. For the other event types this paper reports observed record multiplicity and does not assert a mechanism.
The practical implication is a denominator rule: decide whether you are counting records of one event type or episodes, say which, and expect the two to differ by roughly threefold for hail and thunderstorm wind in the qualifying cells of these five states between 2015 and 2025.
Business Question
Question. Within a state, year and event type, how many county-coded records of the named type appear within each NOAA episode across the five represented states?
Audience. Analysts, practitioners, journalists and researchers who need governed public-data comparisons — the professional research reader.
Publication-value mode. Reference benchmark. NOAA documents the episode and event concepts and states that tornado county crossings are separate segments, but publishes no distribution of how many records of a single named event type appear within an episode.
Specific reader utility. A reader can state whether a figure counts records of a named type within episodes or counts episodes, and can quantify the former within the represented qualifying cells without treating it as a count of whole storm systems.
Positive takeaway. Where multiplicity is material, the number of records of the named type within an episode must be stated explicitly with its event-type and episode grain.
Credible null takeaway. A qualifying event-type cell whose records of the named type map near one-to-one onto episodes can be documented as having little record multiplicity at that grain.
Why it matters. Record multiplicity can change a type-specific count without being visible in a headline total, and its distribution has not been published at this grain.
Data & Scope
Source. NOAA National Centers for Environmental Information, Storm Events Database, as represented in the governed Eldhurst research data platform. NOAA's population nouns are the episode and the event; this paper's unit is the county-coded event record.
Universe. Colorado, Ohio, Oklahoma, Pennsylvania and Texas; complete calendar years 2015 through 2025. Calendar year 2026 is excluded as incomplete, a decision fixed before any outcome was computed.
Grain. State, calendar year, event type and state-scoped episode identifier.
Episode identity. An episode is identified by the pair of state and NOAA episode identifier, so state scoping is explicit rather than assumed. This was audited before any multiplicity statistic was computed: zero identifiers appear in more than one state, and zero state-scoped episodes span more than one calendar year.
Inclusion rule. A record is included when it carries a valid positive episode identifier. Every one of the 82,121 records in the window does; none was excluded for missing identity, and each carries a distinct event identifier.
Exclusions. The forecast-zone table is a separate governed table covering event types NOAA collects by forecast zone rather than by county, and it is excluded entirely. This paper describes county-coded events only.
What "events per episode" counts. Within a cell, only records of that cell's event type inside that episode are counted. 41.8% of state-scoped episodes carry more than one event type, so a cell describes part of a storm system and never all of it.
Source limitations disclosed by the publisher. NOAA receives event information from county, state and federal emergency management officials, law enforcement, spotters, damage surveys, newspaper clipping services, the insurance industry and the public; states that time and resource constraints leave some of it unverified; and does not guarantee accuracy or validity. Independent assessment reports broader reliability limits in the same database.
Method
For each state, calendar year, event type and state-scoped episode, the count of distinct county-coded event records of that type is computed. The distribution of those counts within a cell gives the cell's median, 75th percentile, 90th percentile and maximum.
A cell qualifies when it holds at least 50 valid episodes; 110 of 428 cells qualify, against a prespecified minimum of 5 qualifying cells. A cell shows material multiplicity when its median is at least 2 or its 90th percentile is at least 4. Both thresholds were fixed before any outcome was computed.
Quantiles are nearest-rank on the sorted counts, so every reported quantile is an observed value rather than an interpolation.
Two robustness views are reported. The identity audit described in Section 3 runs before computation and gates the analysis: cross-state and cross-year span shares above 0.01 would block the method. Annual record counts are also screened before the primary calculation; a year below 50% of the median annual count would be excluded as incomplete, and none was excluded. The year-stability check compares each state-and-event-type series between its first and last qualifying year; a series is unstable when its median moves by more than 2 records, and the benchmark fails if more than half of series are unstable.
Tornado cells are reported separately, because NOAA records a tornado crossing a county or state line as separate segments and a reader must not read tornado multiplicity as the same phenomenon as hail multiplicity.
There is no ranking rule and no decision rule. No storm, county or office is named.
Findings
The locked evidence artifact: multiplicity by event type
| Event type | Qualifying cells | Episodes | Event records | Median | p75 | p90 | Max |
|---|---|---|---|---|---|---|---|
| Thunderstorm Wind | 52 | 6,040 | 34,323 | 3 | 6 | 13 | 170 |
| Hail | 35 | 3,933 | 25,396 | 3 | 7 | 16 | 160 |
| Flash Flood | 15 | 1,378 | 6,321 | 2 | 5 | 11 | 72 |
| Flood | 4 | 227 | 888 | 2 | 4 | 9 | 71 |
| Tornado | 4 | 254 | 776 | 2 | 4 | 7 | 24 |
Figure 1 renders the median column.
The two most common event types also have the largest median multiplicity. The median is the least interesting number in the table: for hail, a 90th-percentile episode contains 16 records of the named type and the maximum is 160. In the qualifying five-state, 2015–2025 cells, the typical hail or thunderstorm-wind episode contains 3 records of the named type; this is not a multiplier for whole storm systems.
Record identity, audited before computation
| Check | Value |
|---|---|
| County-coded event records | 82,121 |
| Distinct event identifiers | 82,121 |
| Distinct state-scoped episodes | 10,236 |
| Identifiers appearing in more than one state | 0 |
| State-scoped episodes spanning more than one calendar year | 0 |
| Episodes carrying more than one event type | 4,275 (41.8%) |
Two of these lines carry more weight than their size suggests. Zero cross-state reuse means the episode identifier is observed to be contained within a state in this window — it does not prove the identifier is scoped to a forecast office, and the governed table carries no forecast-office field to test that with. And 41.8% multi-type episodes is the reason an event-type cell can never be presented as a whole storm system.
Coverage of the primary grain
Of 428 state-year-event-type cells, 110 meet the 50-episode floor and 109 of those show material multiplicity. The one qualifying cell that does not is reported in the full cell table rather than dropped.
Tornado segments, reported separately
| State | Year | Episodes | Event records | Median | p90 | Max |
|---|---|---|---|---|---|---|
| Texas | 2015 | 70 | 258 | 2 | 10 | 22 |
| Texas | 2017 | 60 | 185 | 1 | 7 | 24 |
| Texas | 2019 | 73 | 184 | 2 | 5 | 11 |
| Texas | 2024 | 51 | 149 | 1 | 3 | 21 |
All four qualifying tornado cells are Texas. The observed tornado multiplicity is consistent with NOAA's documented rule that a tornado crossing a county or state line is recorded as separate segments. The governed aggregate does not isolate how much of the multiplier that rule accounts for, so the result is not presented as proof that every additional record arises from a crossing.
Stability
Thirteen state-and-event-type series had two or more qualifying years. Zero moved their median by more than 2 records between first and last year, against a failure threshold of more than half. That stability applies only to qualifying cells in these five states during 2015–2025, counting records of the named type within an episode.
Interpretation, kept separate from the above
The direct findings are the distributions above. The interpretation is bounded to what the record can support: the number of county-coded records of one event type within an episode is large enough, and stable enough across qualifying years, to be stated as a five-state, 2015–2025 caution. The record does not identify the cause of that multiplicity. For tornadoes, the observed pattern is consistent with NOAA's documented county and state crossing rule, but the governed aggregate cannot attribute the full multiplier to that rule.
Association Analysis
No inferential association test is run, and none is justified by the locked method.
The locked method specifies a distributional benchmark at a fixed grain, a prespecified materiality rule and a year-stability check. It authorises no significance test, correlation or model.
The universe is every represented county-coded record in the window, not a sample of them, so a significance statistic would describe sampling variability that does not exist and would imply generalisation beyond the five states and eleven years, which Section 8 forbids.
The authorised check is the year-stability comparison, reported in Section 5: zero of thirteen series unstable. That is a stability result about the benchmark, not an association between variables.
Association is not causation. Nothing here claims that event type causes multiplicity; event type and the NWS collection rule for that type are not separable in this record, and the paper does not pretend otherwise.
Business Implications
The declared value mode is a reference benchmark, so the permitted implication is what a reader can now compare, declare and check.
An analyst or journalist using the storm record can:
- state whether a type-specific figure counts episodes or records of the named type within an episode;
- expect, in the qualifying cells of these five states between 2015 and 2025, roughly a threefold difference for hail and thunderstorm wind and roughly twofold for flash flood, flood and tornado, counting records of the named type within an episode; the 318 cells below the 50-episode floor are outside this guidance and Flood and Tornado each rest on only 4 qualifying cells;
- expect the tail to be much larger than the median, and check the 90th percentile rather than the median when a single large system could dominate a result;
- identify an episode by the pair of state and episode identifier rather than by the identifier alone.
What must not be inferred or automated. Event records are not independent physical storms. Episode identifiers are not universal or named storms. Segmentation does not inflate damage, and this paper reports no damage figure at all. Unreported weather is not absent weather. Observed containment of identifiers within a state is not proof of forecast-office scope. No event-type cell is a whole storm system.
Limitations
What the data exclude. Events NOAA collects by forecast zone rather than by county, which are a separate governed table excluded here; unreported weather; and any state or year outside the window.
What the method cannot determine. Why multiplicity differs by event type. The collection rule and the event type are confounded by construction — NOAA collects some types by county and others by zone, and the county-coded types are precisely those recorded per county — so the record cannot separate a reporting effect from a meteorological one.
Coverage and sensitivity weaknesses. The Flood and Tornado strata rest on 4 qualifying cells each against 52 for thunderstorm wind, so their figures are measured on a much narrower base. All four qualifying tornado cells come from a single state. 318 of the 428 cells fall below the 50-episode floor and are published separately in the suppressed-cell table but excluded from the headline.
Alternative explanations considered. The record does not identify why multiplicity occurs. For tornadoes, the observed pattern is consistent with NOAA's documented county and state crossing rule, but the aggregate does not measure that rule's contribution. For the other event types, the collection rule and event type are confounded in this record, so the paper reports observed multiplicity and adopts no mechanism. Every figure describes NOAA record multiplicity and never physical storm incidence.
The unanswered question that should guide follow-up. Whether the county-coded and forecast-zone halves of the database can be reconciled onto one denominator for a mixed-type analysis. That needs the zone table, a different grain and a different method.
References / Appendix
Authoritative sources
- NOAA National Centers for Environmental Information, Storm Events Database FAQ — defines an episode as an entire storm system containing multiple event types, defines an event as an individual event type, describes tornado county and state crossings as separate segments, lists the reporting sources, and states that some information is unverified and accuracy is not guaranteed.
- NOAA National Centers for Environmental Information, Storm Events Database — the official search surface and source of record.
- Renato P. dos Santos, Some comments on the reliability of NOAA's Storm Events Database, arXiv:1606.06973 — independent assessment reporting broader reliability limits in the same database.
All external sources were verified on 12 August 2026.
Field definitions
| Term | Definition |
|---|---|
| Event record | One county-coded storm event record carrying a distinct event identifier. |
| Episode | An entire storm system in NOAA's terms, identified here by the pair of state and NOAA episode identifier. |
| Event type | NOAA's individual event category, such as Hail or Thunderstorm Wind. |
| Named-type records within an episode | Within a cell, the count of distinct event records of that cell's event type inside one episode. |
| Qualifying cell | A state, calendar year and event type holding at least 50 valid episodes. |
| Material multiplicity | A cell whose median is at least 2 or whose 90th percentile is at least 4. |
| Segment | NOAA's separate record for a tornado crossing a county or state line. |
Permitted and prohibited claims
This article states: the distributions in Section 5, stated as properties of NOAA record multiplicity for county-coded events in five states across complete calendar years 2015–2025.
Outside what this article claims: describing event records as independent physical storms; describing episode identifiers as universal or named storms; saying segmentation inflates damage; treating unreported weather as absent; claiming observed state containment proves forecast-office scope; and presenting an event-type cell as a whole storm system.
Figures
Scroll chart horizontally to read all labels.
Full chart description
Horizontal bar chart titled "County Records of the Named Event Type Within an Episode". Each bar is an event type, showing the median number of county-coded records of that one event type found within a single episode, across Colorado, Ohio, Oklahoma, Pennsylvania and Texas in complete calendar years 2015 to 2025, for qualifying cells only. - Flash Flood: median 2 records of this event type per episode, 90th percentile 11, across 1,378 episodes and 6,321 records of this event type. - Flood: median 2 records of this event type per episode, 90th percentile 9, across 227 episodes and 888 records of this event type. - Hail: median 3 records of this event type per episode, 90th percentile 16, across 3,933 episodes and 25,396 records of this event type. - Thunderstorm Wind: median 3 records of this event type per episode, 90th percentile 13, across 6,040 episodes and 34,323 records of this event type. - Tornado: median 2 records of this event type per episode, 90th percentile 7, across 254 episodes and 776 records of this event type. An episode is identified by the pair of state and NOAA episode identifier. Each bar counts only records of that one event type within an episode; 41.8% of episodes carry more than one event type, so a bar is never a count of the whole episode. These are database records and never counts of physical storms. This figure reports observed record multiplicity and attributes no mechanism to it; for tornadoes the pattern is consistent with NOAA's documented county and state crossing rule, which does not explain the aggregate figures for other event types.