→ Open the interactive dashboard — national estimates with confidence intervals and NCHS reliability flags, the variation decomposition, and the peak-hour null result. No install.
A design-correct analysis of 16,025 real emergency department visits from the CDC/NCHS 2022 National Hospital Ambulatory Medical Care Survey, weighted to the 155,397,747 ED visits made in the United States that year.
The question this project set out to answer was the one ED operations folklore always asks: when is the department worst? The data gave a more useful answer — it isn't "when", it's "where" — and one number that reframes the whole capacity conversation.
between-hospital spread in median door-to-provider 2.4 → 75 min (IQR 16 min)
between-hour spread in median door-to-provider 11 → 22 min (IQR 4.5 min)
────────────────────────────
between-hospital variation is 3.56× larger
| Estimate | 99% CI | ||
|---|---|---|---|
| National ED visits, 2022 | 155,397,747 | — | matches CDC's published total exactly |
| Median door-to-provider | 16 min | — | 90th percentile: 98 min |
| Left without being seen (LWBS) | 1.89% | 1.12 – 2.66% | |
| Admitted | 11.46% | 8.67 – 14.25% | |
| Median boarding time (admitted) | 62 min | — | 90th pct 460 min, 99th pct 2,470 min |
| ED bay-hours lost to boarding | 48.6 million / year | — | the single largest recoverable capacity item |
1. Which hospital you walk into matters ~4× more than what time you arrive. Across 45 EDs with ≥100 sampled visits, median door-to-provider ranges from 2 to 74 minutes. Across the 24 hours of the day it ranges from 11 to 22 minutes. The worst individual ED averages 139 minutes to provider with 6.2% LWBS; one runs 17.7% LWBS. Site-level operating practice, not the time on the clock, is where the variance is.
2. The peak-hour effect is real in direction and too small to act on — an honest null. Arrivals do surge 16:00–19:00 and waits are directionally worse (median 17 vs 13 min overnight; LWBS 2.18% vs 1.07%). But under design-based inference the difference does not clear significance at the α = 0.01 level CDC recommends for this file (wait p = 0.41, LWBS p = 0.082). A staffing case built on peak-hour degradation would not survive scrutiny. This is reported as a negative result rather than quietly dropped.
3. Boarding is the actual capacity crisis. An admitted patient holds a treatment bay for a median of 62 minutes after the decision to admit — but 460 minutes at the 90th percentile and 41 hours at the 99th. Nationally that is 48.6 million bay-hours a year of ED treatment capacity consumed by patients who have already been admitted and are waiting for an inpatient bed. No amount of ED-side staffing fixes a downstream bed problem.
4. Patients without coverage are the ones who give up and leave. LWBS runs 2.87% for self-pay and 2.45% for Medicaid/CHIP against 0.65% for Medicare — a 4.4× gap between self-pay and Medicare. Directionally unambiguous and consistent across every uninsured category, though the pairwise test lands at p = 0.014 and so does not clear the α = 0.01 bar either. Reported with that caveat attached.
5. Triage works, with one wrinkle. Mean wait climbs from 26.9 min (ESI 1 Immediate) to 40.5 min (ESI 4 Semi-urgent) as it should — then falls to 30.1 min for ESI 5 Nonurgent. That non-monotonicity is the signature of fast-track pathways pulling the simplest patients out of the main queue.
NHAMCS is a stratified multi-stage probability sample, not a table of patients. Three things have to be right or every number above is wrong:
-
Weighting. Each record carries
PATWT, an inflation factor to national totals. Unweighted, the "national LWBS rate" is a statement about which hospitals CDC happened to sample. The pipeline validates its own weighting by reproducing CDC's published national total to the visit — if the byte offsets were wrong, that check fails and the build refuses to write. -
Variance. Visits from the same ED are correlated, so treating 16,025 rows as 16,025 independent observations understates standard errors badly.
src/survey.pyimplements the ultimate-cluster (with-replacement) estimator overCSTRATM/CPSUMthat CDC prescribes, with Taylor linearisation for ratios and means. This is the difference between "peak hours are worse" and "peak hours are not distinguishable from noise" — the naive calculation would have confidently reported the former. -
Sentinel codes. NHAMCS encodes missingness as negative numbers with different meanings: −9 blank, −8 unknown, −7 not applicable. Reading them as numbers is the classic way to publish an ED with a mean wait of minus four minutes.
−7onWAITTIMEis kept as a distinct flag rather than dropped, because "never seen by a provider" is a finding, not a gap.
Estimates failing the NCHS presentation standard (n < 30 or RSE > 30%) are labelled unreliable everywhere they appear rather than silently printed.
The public use file is 2,382 bytes of unlabelled ASCII per record. The byte offsets for its ~900 variables exist only inside CDC's 268-page documentation PDF, which is why most published analyses hardcode a handful of positions copied by hand.
src/nhamcs_layout.py parses the codebook out of the PDF and
validates that the recovered fields tile the record, so the pipeline can be pointed at a
different survey year. The 16 disposition flags — including LWBS — defeat text extraction
because their columns are detached in the PDF, so they are transcribed explicitly and then
structurally verified: the parser asserts ADMIT still sits at bytes 503–504 immediately
after the block and that bytes 487–502 are claimed by nothing else. If CDC renumbers the
file, that fails loudly instead of silently reading the wrong bytes as LWBS.
Python (pandas, NumPy, SciPy, Matplotlib) · SQL (SQLite star schema with CTEs,
window functions, FILTER aggregates) · pypdf for codebook extraction · Parquet
interchange.
SQLite is deliberate: the entire warehouse is one file a reviewer opens with no install, and it still supports every SQL construct the KPI layer needs.
Grain: one row per ED visit. Conformed dimensions for hour-of-day, triage acuity, expected payer, disposition, and hospital.
The survey design columns live on the fact table, not behind a join — they describe how a row was sampled, not what it describes, and every correct aggregate must carry them. Hiding them behind a join makes it easy to write a query that silently drops the weighting and returns unweighted counts that look plausible and are wrong by four orders of magnitude.
python src/warehouse.py --list # available KPI queries
python src/warehouse.py --query hospital_variation
pip install -r requirements.txt
python src/download.py # fetch the PUF + documentation from CDC (~40 MB)
python src/build_dataset.py # parse fixed-width -> validated tidy parquet
python src/warehouse.py # build the SQLite star schema, run the KPI suite
python src/analyze.py # design-based estimates, figures, metrics.jsonOr python run.py for the whole chain. Everything is deterministic; there is no sampling
step and no seed to worry about.
The CDC survey file is not committed to this repository. It is a 38 MB third-party
dataset governed by NCHS data use terms, and vendoring other people's data into a git
history is poor practice regardless. src/download.py fetches it on demand and verifies
the byte count.
NCHS terms permit statistical reporting and analysis, prohibit any attempt to identify a person or establishment, and prohibit linkage to individually identifiable data from other sources. This project does aggregate throughput analysis only and links to nothing.
- Model LWBS directly rather than slicing it — a survey-weighted logistic regression would separate the payer effect from the hospital effect, which the marginal tables here cannot.
- Multi-year panel. 2022 still carries pandemic distortion; the 2019–2023 files share a layout and the parser already generalises across years.
- Join CMS Hospital Care Compare (measures OP-18 and OP-22) to attach facility characteristics to the hospital-level spread and ask why the worst EDs are the worst. Note this would be a facility-level join only — NCHS terms forbid record-level linkage.
CDC/NCHS, National Hospital Ambulatory Medical Care Survey, 2022 Emergency Department
Public Use File.
Data: ftp.cdc.gov/pub/Health_Statistics/NCHS/Datasets/NHAMCS/ed2022.zip ·
Documentation: doc22-ed-508.pdf
Analysis and interpretation are mine and are not endorsed by CDC or NCHS.