CBECS microdata CSV download: the building file, delivered whole

Two real records from the 2018 file - one row per building, coded characteristics beside measured consumption:

PUBID   : 00001                REGION : 3      CENDIV : 5
PBA     : 2 (office)           SQFT   : 210000 NFLOOR: 994
YRCONC  : 2                    OWNTYPE: 8 (federal government)
MONUSE  : 12                   NWKER  : 350
HDD65   : 4463                 CDD65  : 1759
ELBTU   : 18708970             MFBTU  : 29727152
MFEXP   : 1072600              FINALWT: 2.17

PUBID   : 00002                REGION : 4      CENDIV : 9
PBA     : 2 (office)           SQFT   : 28000  NFLOOR: 5
YRCONC  : 6                    OWNTYPE: 2 (corporation/partnership/LLC)
MONUSE  : 12                   NWKER  : 12
HDD65   : 2424                 CDD65  : 189
ELBTU   : 1528667              NGBTU  : 201988
MFBTU   : 1730655              MFEXP  : 82030
FINALWT : 312.77

The anatomy matters more than any single value. Both rows carry PBA 2 - office - yet everything downstream diverges: a 210,000-square-foot federal building in a heavy-heating climate against a 28,000-square-foot corporate building in a mild one. And FINALWT spans two orders of magnitude between them, which is the survey design speaking: the rarer a building's shape in the stock, the bigger the weight each respondent of that kind carries. Treat the weight as part of the observation, not an appendix.

Get a sample cut to your building types and divisions - real rows come back with the complete field dictionary attached.

The nineteen-field load-bearing core stacks into three jobs. Identification: PUBID, REGION, CENDIV. Classification: principal building activity (PBA), gross square footage (SQFT) and its size band (SQFTC), floors (NFLOOR), construction era (YRCONC), owner type (OWNTYPE), months in use (MONUSE) and main-shift workers (NWKER). Measurement: degree days base 65 F (HDD65/CDD65) attached from weather data for climate normalization, per-fuel annual consumption in thousand Btu (ELBTU, NGBTU, FKBTU), the derived major-fuels totals MFBTU and dollar-side MFEXP, and FINALWT.

Beyond the core the file opens out into about 1,249 columns, and the honest way to read them is in three groups:

1. Roughly 620 questionnaire variables describing characteristics, activity, equipment inventories and the consumption-and-expenditure ledger.
2. Imputation flags marking every value the survey filled in statistically rather than observed - the difference between an honest model feature and a quietly biased one.
3. Replicate weights (the FINALWT1... series), existing precisely so standard errors reflect the survey design instead of a naive formula.

The classic mistake is averaging a flag column as though it were a measurement. The second classic mistake is ignoring FINALWT entirely and reporting unweighted averages - which answers a question nobody asked.

A fixed routine separates citable numbers from bar-chart filler, and none of it is optional:

1. Sum FINALWT within whatever segment you cut. That single habit scales results from 6,436 respondents to the 5.9-million-building universe - stock shares, market sizes, intensity baselines.
2. Take uncertainty from the replicate weights. They encode the survey's design; a naive variance formula will flatter your precision.
3. Respect the division floor. Nine census divisions are the finest geography ever released at building level - no state identifiers survive disclosure processing. Regional cuts are built on CENDIV or they are not built.
4. Expect small gaps against published tables. The building-level records are masked while the agency's pretabulated output is produced from unmasked data, so your tabulations will differ slightly - by design, not defect.
5. Date-stamp every figure with its cycle year. Waves are periodic rather than annual, so anything you publish is a snapshot claim, and reviewers will ask which snapshot.

Run honestly, the whole sequence is an afternoon of analysis rather than an afternoon of plumbing - which is the point of taking rows that arrive already shaped.

Three other delivered records circle the same built environment, and they answer different questions. Knowing which one you are asking saves a month of misfitted work.

Practical rule: CBECS when you need every variable side by side for custom segments nobody has published; the Building Performance Database when you need volume and can work in peer aggregates; Portfolio Manager for the buildings you already operate; GHGRP when the unit of analysis is emissions rather than real estate. Analysts usually end up needing two of the four - which is why they arrive with compatible keys rather than four incompatible formats.

• National benchmarking denominators. Weighted CBECS distributions define what "typical" consumption, expenditure or equipment looks like across 5.9 million buildings - the reference frame peer-group tools and rating scores sit on top of.
• Market sizing for building services. Summing weights by segment turns the file into a stock estimate, the denominator a janitorial or energy-services go-to-market plan needs before any directory is counted. Certified-provider registers then supply the demand side's counterpart: the ISSA CIMS certified cleaning organization directory lists about 390 certified organizations and the Green Seal certified products directory indexes roughly 360 certified companies.
• Model features and validation sets. Observed and imputed values ride side by side with flags distinguishing them - unusually honest training material for intensity, benchmarking and anomaly models. The fuller workflow lives on the data scientists use cases page.
• Program evaluation and policy claims. Before-and-after snapshots on consistent survey footing support efficiency-program impact statements without proprietary data.
• Cross-country context. When the question goes abroad, Eurostat waste statistics (municipal series from 1995) and the OECD municipal waste statistics (58 reference areas) extend the same discipline - weighted figures, cited vintages - to European and partner economies.

Market researchers and consultants get the definitive segmentation of the US commercial stock - activity, size, era, ownership and workforce on one survey footing - for bottom-up TAMs and category maps.

Data scientists and ML engineers get a labeled, weighted, weather-normalized building panel of 6,436 records by roughly 620 variables, ready for intensity and anomaly work.

Sales and growth teams at building-equipment and energy-services firms get territory sizing from the government's own denominators - buildings, floorspace and workers by segment and division.

Competitive-intel and product teams map how many targets exist, of what type, where, before committing to a wedge.

Journalists, academics and students get the number everyone else cites - sourced to a national probability sample rather than a vendor white paper.

Because the record is famous and still awkward. Coded categories arrive undecoded; flag columns masquerade as measurements; the weight apparatus intimidates anyone who has not read the codebook; and the file changes shape between cycles. Datadory ships the rows already shaped: categories decoded or preserved as your spec requires, imputation flags traveling beside the values they qualify, weights intact and documented, join keys stable across cycles.

Files, feeds, or straight into your warehouse. Daily, weekly, or hourly - your call. The sample comes first either way: name the building types and divisions, get real rows, then decide.

Start with the anchor record, EIA Commercial Buildings Energy Consumption Survey (CBECS), sampled to your building types and divisions before anything is committed. Product-level detail - sample rows, the full field dictionary, coverage chips - lives on its dataset page.

This page is one thread of a wider map. The Environmental & Facilities Services data guide walks the sixteen-record slice, the best environmental & facilities services datasets ranking scores all ten leaders side by side, and the environmental-facilities-services data hub indexes every record with its coverage statement.