How Your Emergency Room Visit Becomes NEISS Data: A Researcher’s Guide to the NEISS Database

Every year, roughly 400,000 injury records leave American hospital emergency departments and land in a public data file anyone can download for free. That file is the NEISS database, the National Electronic Injury Surveillance System, run by the U.S. Consumer Product Safety Commission (CPSC) since 1972. If you were treated for a product-related injury at one of the sampled hospitals, a de-identified version of your visit is probably in it, reduced to about twenty coded fields and one sentence of free text.

This article follows a single injury the whole way through: from the chart a clinician writes at 2 a.m., to the coder who translates it, to the weight that scales it to a national number, to the row a researcher downloads a year later. It also corrects the most damaging misconception about this dataset, which is that it uses ICD-10 codes. It does not, and a study designed on that assumption fails before it starts.

Key Takeaways

• A stratified probability sample of 100 hospitals from a frame of 4,723 eligible U.S. emergency departments. Not all report each year: 96 did in 2023, 77 of 100 in 2024.
• It collects almost 400,000 product-related cases annually. That is a record count, not a national figure.
• NEISS does not use ICD-10-CM or ICD-10-PCS. It uses closed internal code sets: 26 body-part codes, 30 diagnosis codes, four-digit product codes, up to three per case.
• Every case carries a Weight; the file also ships Stratum and PSU. All three are required for a defensible estimate and a correct standard error.
• The files are free and public-use for treatment years 2006 through 2025. The system is being rebuilt, with CPSC’s modernised version expected fully effective by the start of 2027.

What the NEISS Database Actually Is

CPSC calls NEISS its “largest and most dynamic data collection and follow-back surveillance system.” Its purpose is to let CPSC analysts make timely national estimates of injuries associated with (but not necessarily caused by) specific consumer products. That parenthetical is CPSC’s own wording on the official NEISS FAQ, and it is the most important sentence on the page. The NEISS database establishes association, not causation, and no statistical sophistication applied afterwards converts one into the other.

The sample is 100 hospitals selected from a 2020 frame of 4,723 eligible emergency departments representing 155,707,175 annual ED visits. Frame eligibility requires a U.S. location, at least six inpatient beds, and a 24-hour emergency department. The five strata are four ED-size bands plus a children’s-hospital stratum.

Stratum	Annual ED visits	Hospitals sampled	Hospitals on frame
Small	1 to 33,280	43	2,984
Medium	33,281 to 56,995	26	860
Large	56,996 to 92,195	12	528
Very large	92,196 and above	11	286
Children’s hospitals	Children’s designation	8	65
Total		100	4,723

Source: CPSC Supporting Statement for Collection of Information, Part B (OMB submission). Prepared by Editverse.

Two things follow. The sample is exactly 100, not “over 100,” which is how most secondary sources describe it. And the number actually reporting moves: 96 hospitals in 2019, 96 in 2023, and 77 of 100 in 2024. Weights adjust for non-response, but a study comparing 2019 with 2024 is comparing periods with materially different reporting bases, and a reviewer who knows the system will ask. Our companion piece on whether NEISS injury statistics are reliable works through what that does to estimate stability.

Fix one more distinction early. The NEISS database has sub-systems drawing on different hospital subsets and case definitions: NEISS-AIP (all injuries, with CDC, from 2000), NEISS-Work (occupational, with NIOSH, roughly 67 EDs), NEISS-CADES (adverse drug events, 60 EDs in 2017 to 2019), and NEISS-FISS (firearm injuries). Citing a NEISS-Work rate as a “NEISS” figure is a factual error even when the underlying paper is impeccable.

Step 1: What Gets Captured at the Hospital

Someone falls off an electric scooter, arrives at a sampled emergency department, and is treated for a wrist fracture. The clinical record from that visit is a full chart: triage note, examination, imaging, discharge instructions. Almost none of it reaches the NEISS database.

What reaches the NEISS database is what a trained abstractor extracts from the chart into a fixed structure. Under the legacy design this is manual medical-record abstraction, not an automatic feed from the electronic health record, which is exactly what CPSC’s modernisation is meant to replace. The abstractor answers a closed set of questions, then writes one short sentence describing the event.

The complete public-use variable list is short enough to print: CPSC_Case_Number, Treatment_Date, Age, Sex, Race, Other_Race, Hispanic, Body_Part, Body_Part_2, Diagnosis, Diagnosis_2, Other_Diagnosis, Other_Diagnosis_2, Disposition, Location, Fire_Involvement, Alcohol, Drug, Product_1, Product_2, Product_3, Narrative, Stratum, PSU, Weight. That is the whole record. Anything needing comorbidity, cost, length of stay, imaging findings, or post-discharge outcome cannot be answered here at all.

One detail catches almost every new user: Age is not a plain integer. 0 means age not recorded, 2 to 120 are years, and 201 to 223 are ages of 1 to 23 months. Read the column naively and you create a cohort aged 201 to 223 years, which passes every syntax check and destroys every age-stratified estimate you produce.

Step 2: Coding, and Why the NEISS Database Does Not Use ICD-10

This is the section that matters most, and it is where the previous version of this article was wrong. It stated that NEISS coders classify injuries using ICD-10-CM/PCS. That is false. The consequence is worth being blunt about: a researcher who plans an ICD-10-based NEISS analysis discovers partway through the project that the variable does not exist.

The NEISS database uses closed internal code sets maintained by CPSC. Every permitted value is enumerated in the NEISS Data Dictionary, and the rules for applying them are in the NEISS Coding Manual:

• Body part: a two-digit code with 26 current values, for example 30 Shoulder, 34 Wrist, 35 Knee, 75 Head, 77 Eyeball, 82 Hand, 92 Finger, 94 Ear. Four legacy codes (78, 86, 90, 91) are flagged “(Old)” and need deliberate handling in long series.
• Diagnosis: a two-digit code with 30 values, for example 52 Concussion, 57 Fracture, 59 Laceration, 63 Puncture, 64 Strain/Sprain, 68 Poisoning, 71 Other/Not Stated. When diagnosis is 71, the free-text Other_Diagnosis field carries what the coder meant.
• Product: a four-digit CPSC code with no external equivalent, up to three per case, unused slots filled 0000.
• Disposition: seven values, including 1 treated and released, 4 treated and admitted, and 8 fatality. Location: nine values, from 1 Home to 9 Place of recreation or sports.

That is the entire clinical vocabulary. Thirty diagnosis codes is not a lot: no laterality, no severity scale, no external-cause hierarchy, no procedure coding. If your question needs any of that, NEISS is not your dataset, and choosing correctly at the outset is the whole point of our guide to which health database your study should use. The system that does use ICD-10-CM/PCS is NEDS, compared later in this article.

Code sets also move. CPSC maintains a Product Code Comparability Table documenting every product code merged, split, deleted or reactivated since 1972. Corkscrews (0416) were deleted in 1983, restored in 1994, deleted again in 1995, restored in May 2000. Stationary power saws (0825) were split into 0841 to 0845 from 1974. Query one code across two decades without that table and you will publish an artefactual trend. Several columns also have hard start dates: Body_Part_2, Diagnosis_2, Product_3, Hispanic, Alcohol and Drug exist only from 2019, and Sex = 3 only from 2021. Any series crossing 2018 to 2019 or 2020 to 2021 contains a structural break. For how a coder chooses between two plausible codes, see how NEISS product and diagnosis codes are assigned.

Step 3: The 400-Character Narrative That Makes the NEISS Database Unique

After the coded fields comes the reason many researchers choose the NEISS database over larger databases: Narrative, a free-text description of the event capped at 400 characters.

The coding manual is prescriptive. The narrative must open with an age-and-sex abbreviation (“25YOM”, “19MOF”), use sentence case and correct spelling, name the specific product, put the clinical diagnosis at the end preceded by “DX:”, record blood alcohol in three digits (“BAC 080”), and contain no identifying information. A real one reads: 28YOF riding an e-scooter struck a curb and fell onto outstretched hand. DX: wrist fracture.

Four hundred characters written under time pressure is a thin instrument. It is also the only place in the federal injury-surveillance apparatus where the circumstance of an injury is recorded at national scale. NEDS has nothing like it. That single field is why the following studies were possible:

• Shields and colleagues read pressure-cooker narratives from 2003 to 2019 to classify precipitating factors into four categories that exist in no coded field: burning agent, struck by product, injured while opening, other. From 759 actual cases they produced a national estimate of 28,337 injuries (95% CI 24,588 to 32,086), about 1,667 per year.⁹
• Dell and colleagues identified probable homelessness in NEISS-FISS firearm records from the de-identified narrative alone, in 0.10% of cases (n = 3,225), with assault the mechanism in 82.64%. NEISS has no housing-status variable; the narrative supplied one.¹⁷
• Parker applied natural language processing to 59,025 unweighted gunshot narratives from 1993 to 2015 and changed the national picture. Prior estimates put the home first at 23.4% ahead of street or highway at 22.9%; after NLP classification, street or highway was most frequent at 46.1% and the home 37.7%.⁵
• Omaki and colleagues hand-coded 4,546 narratives to train a model, applied it to 91,325 coded cases of falls in children under five, and found infants fell from the bed in 33% of fall injuries.⁶
• Chen and colleagues showed the labour cost is tractable: an interactive machine-learning tool cut classification of tens of thousands of TBI narratives from days to minutes after about 60 minutes of training.¹⁹

Narrative work is a design decision, not an afterthought. If your question needs circumstance detail you are committing to manual review or a classifier plus a validation subsample, and that belongs in the protocol. Our post on the free-text narrative field goes deeper into extraction strategy.

Step 4: The Weight That Turns One Record Into a National Estimate

This is where most NEISS papers succeed or fail.

Because the NEISS database is a probability sample, every case carries a statistical weight. CPSC’s own formulation is that national estimates “are derived by summing the statistical weights for the appropriate injury cases.” The weight is not a constant. It is the inverse of the hospital’s probability of selection within its stratum, adjusted for non-response and for hospital mergers. Schroeder and Ault give it as:

NEISSwt = (N_h x n'_h) / (n_h x r_h)

where, within stratum h: N is hospitals on the frame, n is hospitals selected, n’ is hospitals in scope, r is hospitals participating. So a record from a small-stratum hospital can carry a far larger weight than one from a very large emergency department. An unweighted NEISS row count is therefore not a smaller version of the national number. It is a different quantity with a different shape.

Two figures make the point, and the previous version of this article conflated them. The cases collected figure is CPSC’s own: “almost 400,000 product-related cases” per year. The national weighted estimate is a different order of magnitude; CPSC’s most recent published all-products figure is 10,994,000 ED-treated consumer product-related injuries for calendar year 2020, with two caveats: 2020 was a pandemic year with depressed ED volumes, and CPSC does not routinely publish a single de-duplicated all-products total.

The reason it avoids that total is instructive. Every Data Highlights release states that “the data system allows for reporting of up to three products for each person’s injury, so a person’s injury may be counted in three product groups.” Sum the published categories and you double- and triple-count people. Single-category 2024 estimates are safe to quote: stairs, ramps, landings and floors at 3,101,789; beds, mattresses and pillows at 1,114,811; chairs, sofas and sofa beds at 756,953; bathroom fixtures at 658,680; exercise equipment at 564,845.

The design variables ship in the public file as Stratum, PSU (the sampled hospital) and Weight, or Wt in the SAS file. Applying the weight without declaring the strata and clusters gives a defensible point estimate attached to a standard error that is usually far too small, so intervals come out too narrow and p-values too generous. Durand and colleagues used NEISS itself as their test case and reported that “in almost all cases, design-corrected approaches contrasted profoundly with standard statistical techniques.”⁴ If you have never seen how far a design can move a result, the clearest demonstration on this site is what happened when we ran the same NHANES analysis three ways. For the sampling theory behind the NEISS weights specifically, see how NEISS weights scale a sample to a national estimate.

CPSC also publishes reportability rules worth mirroring. Variance is reported as a coefficient of variation, CV equals the square root of the variance divided by the estimate; the 95% interval is estimate plus or minus (1.96 x estimate x CV); and CPSC treats an estimate as unstable, and will not report it, when the estimate is below 1,200, contributing cases number fewer than 20, or the CV exceeds 33%.

Step 5: Release, and How to Download the NEISS Database Files

Records are validated, weighted and released annually. CPSC states that the most recent calendar year of treatment dates generally becomes available in April of the following year. In practice the schedule slips: data for 2025 treatment dates became available on 7 May 2026. Plan around “roughly a year plus a few months,” not around the invented 72-hour or two-to-four-week processing windows that circulate on content-farm pages about this dataset.

Getting the data is genuinely easy. No application, no data use agreement, no fee:

• NEISS Estimates Query Builder and download portal, with annual files for treatment years 2006 through 2025 in Excel, tab-delimited text and SAS formats.
• CPSC’s explanation of how its own estimates are produced, including the CV and reportability rules above.
• CPSC’s programme description of NEISS, useful for the opening of a methods section.

One caveat CPSC states plainly, which belongs in your limitations paragraph: estimates a user generates through the website “may differ from CPSC publications due to analytical differences in data processing.” Say so rather than letting a reviewer find it. A walkthrough of the download and file structure is in how to get the NEISS data files.

A Correct Worked Example: NEISS Database Rows to a National Estimate

The previous version of this article shipped a Stata snippet using discwt and hosp_ed. Those are HCUP/NEDS variable names; they do not exist in a NEISS file, and the code could not run. Here is the corrected version.

* ---- Stata: NEISS public-use file, one treatment year ----
* Design variables ship in the file: Stratum, PSU, Weight

* 1. Declare the design ONCE. Everything downstream depends on this line.
svyset PSU [pweight=Weight], strata(Stratum) singleunit(certainty)

* 2. Age is not a plain integer.
*    0 = not recorded; 2-120 = years; 201-223 = 1-23 MONTHS.
gen byte infant  = inrange(Age, 201, 223)
gen double age_y = .
replace    age_y = Age             if inrange(Age, 2, 120)
replace    age_y = (Age - 200)/12  if infant

* 3. Case definition. Product codes are four digits, up to THREE per case,
*    so check all three slots. Replace 1234 with your product code.
gen byte casedef = inlist(1234, Product_1, Product_2, Product_3)

* 4. Unweighted case count. This is NOT a national number.
*    Report it anyway; reviewers expect both.
count if casedef == 1

* 5. Weighted national estimate with a design-based standard error.
svy: total casedef

* 6. Subgroup estimate, domain-restricted. Do NOT drop rows first;
*    that breaks the variance estimation.
svy, subpop(if age_y < 18): total casedef

* 7. Apply CPSC's reportability rules to whatever comes back:
*    suppress if estimate < 1200, unweighted cases < 20, or CV > 33%.
estat cv

The same design in R with the survey package:

library(survey)

neiss_design <- svydesign(
  id      = ~PSU,        # primary sampling unit = the sampled hospital
  strata  = ~Stratum,    # four ED-size strata + children's hospital stratum
  weights = ~Weight,     # CPSC statistical weight
  data    = neiss,
  nest    = TRUE         # PSU ids are nested within strata
)

# National estimate of fractures (NEISS diagnosis code 57):
svytotal(~I(Diagnosis == 57), neiss_design)

# Then CV = SE / estimate. Suppress if estimate < 1200,
# contributing cases < 20, or CV > 33%.

In SAS the equivalent is PROC SURVEYMEANS or PROC SURVEYFREQ with STRATA Stratum; CLUSTER PSU; WEIGHT Wt;.

Two habits separate publishable NEISS analyses from rejected ones. Always report the unweighted case count beside the weighted estimate: the cleanest recent example is the JAMA Ophthalmology pickleball paper, reporting an estimated 3,112 ocular injuries derived from 73 unweighted cases, with a survey-weighted regression trend of 405 additional cases per year (95% CI 288 to 522, P = 0.004).¹⁰ A reader sees immediately how thin the underlying data are. And subset with a subpopulation option rather than by deleting rows, because dropping rows removes the information the variance estimator needs about the full design. To sanity-check a rate you have derived, our free incidence rate calculator handles the arithmetic.

What Researchers Have Answered With the NEISS Database

A CDC-authored scoping review screened 6,944 citations and identified 167 peer-reviewed articles using NEISS-AIP as the primary data source between 2001 and 2021, averaging 8.0 articles per year across 72 journals.³ That is the NEISS-AIP subset alone; the broader NEISS literature is considerably larger. Recent work shows where the system is strongest.

Micromobility. Neuroth and colleagues estimated 60,554 e-scooter injuries treated in U.S. emergency departments from 2015 to 2019 (95% CI 37,525 to 84,594), roughly 19% involving a motor vehicle; those were admitted more often (14.6% versus 8.8%, p < 0.01).¹² Akomaning and colleagues, from 4,020 unweighted e-scooter and e-bike records, reported a weighted estimate of 279,990 e-scooter visits and a roughly three-fold rise from 22,835 in 2019 to 65,892 in 2022, with males at 2.61 times the odds of alcohol involvement (95% CI 1.84 to 3.69).¹³

Lithium-ion battery injury. The previous version of this article carried two contradictory figures, 37% and plus 142%, neither sourced. What is actually published: FDA authors analysing 2020 to 2024 NEISS data estimated 2,043 emergency department visits (95% CI 1,326 to 2,760) for electronic nicotine delivery system battery explosions and burns, derived from 52 unweighted cases, with 26.3% requiring admission.¹¹ Note the shape: few records, a wide interval, an unusually high admission rate. A legitimate analysis, and a good illustration of how thin NEISS can be for an emerging hazard.

Paediatric injury. Xiao and colleagues reviewed 20 years of paediatric oesophageal foreign bodies: 52,315 unweighted cases producing a national estimate of 1,589,325, with coins at 37.6% and batteries at 6.8%, and incidence rising from 7.3 to 14.2 per 10,000 children (p < 0.0001).¹⁴ The NEISS database answers this better than anything else available, because the product is the exposure and the product is coded.

Falls. Hammig and Haldeman estimated 687,902 emergency department visits for stair falls among people aged 15 to 64 in 2019 alone, with alcohol intoxication associated with higher prevalence of traumatic brain injury (prevalence ratio 2.7, 95% CI 2.3 to 3.1).¹⁵

Firearm injury, and the sub-system trap. Bhagavathula and colleagues, using NEISS-FISS from 2000 to 2021, estimated 2.36 million non-fatal firearm injuries among U.S. working-age adults, males accounting for 85.7%.¹⁶ Alongside it sits the best available summary of when to pick which database: Carpenito and colleagues state that HCUP-NEDS “provides accurate data for the number of nonfatal firearm injuries, [but] injury intent is not coded reliably,” while NEISS-FISS reliably codes intent but “is too small to produce stable estimates of the number of these events.”⁷

Medication harms. The NEISS-CADES analysis in JAMA is worth reading as a presentation template. From 96,925 cases across 60 emergency departments, the authors estimated 6.1 visits for medication harms per 1,000 population annually (95% CI 4.8 to 7.5), of which 38.6% (95% CI 35.2 to 41.9) resulted in hospitalisation.⁸ Every estimate carries an interval; the participating ED count and the case count are both named. That is the standard.

What the NEISS Database Cannot Do

A post that only sells a dataset is not useful. If any of the following is central to your question, choose something else.

1. ED-treated injuries only. Injuries handled in urgent care, a physician’s office, at home, or not treated at all are invisible, as are deaths that never reach an ED.
2. No longitudinal follow-up. Each record is a single visit. No patient identifier, no readmission, no post-discharge outcome, no cost field. Recovery, recurrence and long-term outcome are outside the system.
3. No state, county or metro estimates. In the CDC scoping review of 167 NEISS-AIP articles, the limitation most commonly identified by authors (n = 38) was that the system produces national estimates only.³ The previous version of this article carried a table of Arizona and Ohio state-level policy findings. No such state estimates exist, and the table has been removed.
4. Real geographic gaps. Under the legacy design 14 states had no participating hospital, with Alaska, Hawaii, the Mountain West and northern regions underrepresented.
5. Small-cell instability. With about 100 hospitals, rare exposures yield very few unweighted cases. If you are unsure whether a subgroup can support an estimate, work it out before extraction using our power and sample size guidance.
6. Association, never causation. A product code records that the product was present, not that it produced the injury.
7. Some records are not injuries. A record-linkage validation at a NEISS children’s hospital found 13.32% of sampled cases (385 of 2,890) were non-injuries by the ICD-9-CM diagnosis in the same record; among true injuries, agreement was kappa 0.87 (95% CI 0.85 to 0.88). Best practice, the authors conclude, “should include manual review of case narratives.”¹
8. Coding accuracy is not one number. Treat any page quoting a single NEISS accuracy percentage with suspicion; the earlier version of this article quoted seven and all seven have been deleted. Beyond the validation results above, the one narrow finding available is a CDC evaluation reporting that more than 99% of the cause and intent variables in NEISS-AIP self-directed violence data for 2018 were coded consistently with the chart narrative.¹⁸ That is internal consistency for two variables, one sub-system, one year. It is not a system-wide accuracy rate.
9. System changes break time series. The 2019 and 2021 variable additions, the product-code history, and the varying number of reporting hospitals mean a naive twenty-year trend line is a claim you must defend, not a fact you can plot. Our post on the biases baked into NEISS sampling covers what peer review is likeliest to raise.

Some of what NEISS omits belongs to other systems rather than to nobody. Implanted medical devices and infusion pumps are FDA MAUDE territory, not CPSC’s, which matters when choosing a database for a device failure analysis.

NEISS vs NEDS: Choosing Between Them

These two get confused constantly, including in the previous version of this article. They are complements rather than competitors.

	NEISS	NEDS
Sponsor	U.S. Consumer Product Safety Commission	AHRQ, via HCUP
First data year	1972 (modern design from 1978)	2006
Scope	Injuries associated with consumer products, plus all-injury, occupational, firearm and adverse-drug supplements	All hospital-owned ED visits, any cause
Sample size	100 hospitals (77 to 96 reporting in recent years)	995 hospital-owned EDs (2020)
Unweighted volume	About 400,000 product-related cases per year	28,037,034 ED visits (2020)
Coding	NEISS internal codes: body part, diagnosis, product, disposition, location	ICD-10-CM/PCS (ICD-9-CM before Oct 2015)
Narrative text	Yes, 400 characters	No
Product identification	Yes, up to three four-digit codes per case	No
Geographic base	National probability sample; 14 states had no hospital under the legacy design	41 states plus DC, covering 85.0% of the U.S. population
Cost	Free public-use files	Purchased through the HCUP Central Distributor
Best used for	Product and mechanism questions; circumstance detail from narratives; early detection of emerging hazards	Volume, cost, payer, comorbidity, disposition and clinical-severity questions at scale

Sources: CPSC NEISS documentation; AHRQ HCUP NEDS overview and 2020 introduction. Prepared by Editverse.

The short version: if the exposure is a product or a mechanism, use the NEISS database. If the outcome is cost, payer, comorbidity burden or clinical severity, budget for NEDS. If you need both, price it in the protocol rather than discovering it later.

The NEISS Database Is Being Rebuilt: What NEISS-R Means

If you are planning a NEISS study in 2026 or 2027, this is not optional background. CPSC has announced a modernisation, NEISS-Remodel or NEISS-R, moving from manual medical-record abstraction at roughly 70 hospitals to standardised electronic data exchange covering all 50 states. The modernised system is “expected to become fully effective by the beginning of 2027.”

Three implications. Name the vintage of the data you used, because “NEISS data” will shortly mean two different collection mechanisms depending on year. Expect a discontinuity, since a change from manual abstraction to electronic exchange combined with a changed hospital set is exactly what produces a step in a time series that has nothing to do with injuries. And if your study spans the transition, plan the sensitivity analysis now. CPSC’s guidance already warns that “data users should carefully consider the likely impact of system updates” on longitudinal comparisons; the full CPSC modernisation announcement is worth reading directly.

Five Mistakes That Sink a NEISS Database Study

1. Publishing an unweighted count as a national number. The most common error and the easiest for a reviewer to spot. Report both figures.
2. Weighting without declaring the design. The point estimate survives; the standard errors do not.
3. Reading Age as an integer. Values 201 to 223 are months. This silently corrupts every paediatric analysis it touches.
4. Querying one product code across two decades. Check the Comparability Table first. Codes merge, split, disappear and return.
5. Assuming a variable exists because it exists in NEDS. No ICD codes, no cost, no comorbidity, no readmission, no state identifier.

Reporting matters too. A NEISS analysis is an observational study, and reviewers expect it written up against the STROBE checklist for observational studies, with sampling design, case definition and weighting described well enough to reproduce from the public file.

Professional Services for a NEISS-Based Study

Most NEISS projects fail in one of three places: the question was already answered, the variables were not available across the years the design needed, or the analytic sample could not support the estimate once restricted. All three are cheap to detect before extraction and expensive to discover afterwards.

There is no NEISS-specific package, because a NEISS project lands in one of the three above depending on how far it reaches. Everything in this line is listed under Registry and Real-World Data Studies. If you already hold your extract and need only the survey design handled properly, our analysis-only service is the shorter route.

Conclusion

An emergency room visit becomes NEISS data through a chain shorter and blunter than most researchers expect. A trained abstractor reads the chart, assigns a handful of two-digit and four-digit codes from closed CPSC vocabularies, writes one sentence of up to 400 characters, and attaches nothing else. CPSC weights the record by its hospital’s selection probability within one of five strata, and about a year later the file is published free to anyone who wants it.

Two facts carry almost all the practical weight. NEISS does not use ICD-10, so your case definition must be built and defended from its own code sets and usually from the narratives. And the roughly 400,000 records collected each year are not a national number, so the weight and the design variables are the difference between an estimate and a meaningless count. Get those right, report the unweighted count beside the weighted estimate, apply CPSC’s reportability thresholds, and be honest about what the system cannot answer, and the NEISS database remains one of the most useful public datasets in injury research.

Frequently Asked Questions About the NEISS Database

Does the NEISS database use ICD-10 codes?

No. NEISS uses closed internal code sets maintained by CPSC: 26 body-part codes, 30 diagnosis codes, four-digit product codes with up to three per case, plus disposition and location codes and a 400-character narrative. Every permitted value is listed in the NEISS Data Dictionary. The database that uses ICD-10-CM and ICD-10-PCS is NEDS, the Nationwide Emergency Department Sample, a different system run by AHRQ that must be purchased.

How many cases does the NEISS database contain each year?

CPSC’s FY2023 Annual Report states that NEISS supplies information on almost 400,000 product-related cases annually, and its OMB supporting statement gives a planning figure of about 470,000 cases, of which roughly 2,375 are selected for follow-back investigation. Those are record counts. The weighted national estimate is a different quantity: CPSC’s most recent published all-products figure is 10,994,000 ED-treated consumer product-related injuries for calendar year 2020, a pandemic year with depressed ED volumes.

How many hospitals are in the NEISS sample?

Exactly 100, drawn from a 2020 frame of 4,723 eligible U.S. emergency departments, in five strata: 43 small, 26 medium, 12 large, 11 very large and 8 children’s hospitals. Not all 100 report in a given year. CPSC’s Data Highlights record 96 reporting hospitals in 2019, 96 in 2023, and 77 of 100 in 2024.

Do I need the sampling weights to publish a NEISS estimate?

Yes, and you need the design variables as well. The public file ships Stratum, PSU and Weight. Summing the weights of cases meeting your definition produces the national estimate; declaring Stratum and PSU as the survey design is what makes the standard errors correct. A weighted point estimate computed without the design variables usually has standard errors that are too small, which produces confidence intervals that are too narrow.

Can the NEISS database give me state-level or county-level estimates?

No. This is the most frequently reported limitation in the published NEISS literature. In a CDC scoping review of 167 NEISS-AIP articles, the limitation most commonly cited by authors was that the system produces national estimates only and cannot be used for state or county-level surveillance. Under the legacy design, 14 states had no participating hospital at all.

How accurate is NEISS coding?

There is no single accuracy figure, and any source quoting one should be treated with suspicion. A record-linkage study at a NEISS children’s hospital found 13.32% of sampled cases were non-injuries by ICD-9-CM standards, with agreement among true injury cases of kappa 0.87 (95% CI 0.85 to 0.88). Against the CDC ICD-9-CM definition, the recommended NEISS traumatic brain injury definition had sensitivity 91.0% and positive predictive value 76.5%. The authors of the first study recommend manual review of case narratives as best practice.

Is NEISS data free, and how far back does it go?

NEISS public-use files are free, with no application and no data use agreement. Annual files are downloadable through the CPSC query-builder portal for treatment years 2006 through 2025, in Excel, tab-delimited text and SAS formats. Each new year generally becomes available in April of the following year, though the release can slip: data for 2025 treatment dates became available on 7 May 2026.

Can NEISS tell me whether a product caused an injury?

No. CPSC’s own wording is that NEISS supports estimates of injuries “associated with (but not necessarily caused by) specific consumer products.” A product code records that the product was present in the event, not that it produced the injury. NEISS also has no patient identifier, no follow-up and no outcome data after the visit, so causal and outcome questions both fall outside what the system can support.

References

1. Thompson MC, Wheeler KK, Shi J, Smith GA, Xiang H. An evaluation of comparability between NEISS and ICD-9-CM injury coding. PLoS One. 2014;9(3):e92052. doi:10.1371/journal.pone.0092052. PMID: 24658100.
2. Thompson MC, Wheeler KK, Shi J, Smith GA, Groner JI, Haley KJ, Xiang H. Surveillance of paediatric traumatic brain injuries using the NEISS: choosing an appropriate case definition. Brain Inj. 2014;28(4):431-7. doi:10.3109/02699052.2014.887146. PMID: 24564802.
3. Navon L, Chen LH, Cowhig M, Wolkin AF. Two decades of nonfatal injury data: a scoping review of the National Electronic Injury Surveillance System-All Injury Program, 2001-2021. Inj Epidemiol. 2023;10(1):44. doi:10.1186/s40621-023-00455-4. PMID: 37679835.
4. Durand WM, Stey PC, Chen ES, Sarkar IN. Trend analysis of aggregate outcomes in complex health survey data. AMIA Jt Summits Transl Sci Proc. 2018;2017:349-58. PMID: 29888093.
5. Parker ST. Estimating nonfatal gunshot injury locations with natural language processing and machine learning models. JAMA Netw Open. 2020;3(10):e2020664. doi:10.1001/jamanetworkopen.2020.20664. PMID: 33052403.
6. Omaki E, Shields W, Rouhizadeh M, Delgado-Barroso P, Stefanos R, Gielen A. Understanding the circumstances of paediatric fall injuries: a machine learning analysis of NEISS narratives. Inj Prev. 2023;29(5):384-8. doi:10.1136/ip-2023-044858. PMID: 37399309.
7. Carpenito T, Miller M, Manjourides J, Azrael D. Using multiple imputation by super learning to assign intent to nonfatal firearm injuries. Prev Med. 2022;163:107183. doi:10.1016/j.ypmed.2022.107183. PMID: 35964778.
8. Budnitz DS, Shehab N, Lovegrove MC, Geller AI, Lind JN, Pollock DA. US emergency department visits attributed to medication harms, 2017-2019. JAMA. 2021;326(13):1299-1309. doi:10.1001/jama.2021.13844. PMID: 34609453.
9. Shields W, Dong Y, Jager L, Shiang E, Frattaroli S, Omaki E. Using the NEISS database to understand pressure cooker related injuries in the USA. Inj Prev. 2023;29(6):506-10. doi:10.1136/ip-2023-044910. PMID: 37666516.
10. Lacher CR, Koc I, Tsui JC. Pickleball-related ocular injuries among patients presenting to emergency departments. JAMA Ophthalmol. 2025;143(11):938-44. doi:10.1001/jamaophthalmol.2025.3577. PMID: 41100112.
11. Blakney AM, Griggs E, Farrell S, Lindsey K, Chang JT. Electronic nicotine delivery system (ENDS) battery-related explosions and burns presenting to United States emergency departments, 2020-2024. Prev Med. 2026;207:108563. doi:10.1016/j.ypmed.2026.108563. PMID: 41875938.
12. Neuroth LM, Humphries KD, Wing JJ, Smith GA, Zhu M. Motor vehicle-related electric scooter injuries in the US: a descriptive analysis of NEISS data. Am J Emerg Med. 2022;55:1-5. doi:10.1016/j.ajem.2022.02.004. PMID: 35228017.
13. Akomaning E, Osei SP, Bhagavathula AS. Alcohol-related injuries from e-scooter and e-bike use in the US (2019-2022): a retrospective study. Inj Prev. 2026;32(4):565-70. doi:10.1136/ip-2024-045461. PMID: 39653429.
14. Xiao JB, Siu JM, Propst EJ, Wolter NE. Consumer trends reflected in the contents of the pediatric esophagus: a 20-year review. Laryngoscope. 2024;135(1):438-44. doi:10.1002/lary.31728. PMID: 39177065.
15. Hammig B, Haldeman S. Stair-related falls in the USA: traumatic brain injury and the role of alcohol intoxication. J Epidemiol Community Health. 2023;77(1):44-8. doi:10.1136/jech-2022-219396. PMID: 36328421.
16. Bhagavathula A, Price JH, Khubchandani J. Trends and disparities in non-fatal firearm injuries among working-age adults in the United States, 2000-2021. J Community Health. 2025;50(3):454-63. doi:10.1007/s10900-024-01431-9. PMID: 39702661.
17. Dell NA, Vaughn MG, Salas-Wright CP. Firearm injury among people experiencing homelessness: cross-sectional evidence from a national survey of United States emergency departments. Public Health Pract (Oxf). 2023;6:100446. doi:10.1016/j.puhip.2023.100446. PMID: 37954558.
18. Ehlman DC, Haileyesus T, Lee R, Ballesteros MF, Yard E. Evaluation of the National Electronic Injury Surveillance System – All Injury Program’s self-directed violence data, United States, 2018. J Safety Res. 2021;76:327-31. doi:10.1016/j.jsr.2020.12.002. PMID: 33653565.
19. Chen W, Wheeler KK, Lin S, Huang Y, Xiang H. Computerized “Learn-As-You-Go” classification of traumatic brain injuries using NEISS narrative data. Accid Anal Prev. 2016;89:111-7. doi:10.1016/j.aap.2016.01.012. PMID: 26851618.
20. U.S. Consumer Product Safety Commission. NEISS Data Highlights 2024. Available from: cpsc.gov.
21. Schroeder T, Ault K. The NEISS Sample (Design and Implementation) From 1979 to 1996. Washington DC: U.S. Consumer Product Safety Commission; June 2001. Available from: cpsc.gov.
22. Agency for Healthcare Research and Quality. Nationwide Emergency Department Sample (NEDS) Overview. Healthcare Cost and Utilization Project. Available from: hcup-us.ahrq.gov.