RECS 2020 microdata: the household file, delivered as rows

One record per surveyed household, exactly as a delivery lands - this row is the research capture's observed first respondent:

DOEID        : 100001             REGIONC      : WEST
DIVISION     : Mountain South     state_postal : NM
TYPEHUQ      : 2                  BEDROOMS     : 4
TOTROOMS     : 8                  KOWNRENT     : <tenure code>
TOTSQFT      : <total sq ft>      TOTUSQFT     : <used sq ft>
HDD65/CDD65  : <degree days base 65F>
KWH          : 12521.48           DOLLAREL     : 1955.06

The load-bearing core stacks into four jobs. Identification: DOEID, the anonymized key joining every RECS file, table and cycle. Geography: REGIONC (example: WEST), the finer DIVISION (Mountain South) and state_postal, populated only where identification rules permit (example: NM). Structure and occupancy: housing unit type (TYPEHUQ), bedroom and room counts (BEDROOMS and TOTROOMS read 4 and 8 on the sample row), tenure (KOWNRENT) and square footage (TOTSQFT, TOTUSQFT). Measurement: total reference-year electricity (KWH), its bill (DOLLAREL) and the heating and cooling degree days base 65F (HDD65, CDD65) that keep a Minnesota-versus-Arizona comparison honest.

Beyond the spine the 799-variable layout opens out, and the honest way to read it is in groups:

1. End-use splits - space heating, water heating, cooling, refrigeration and lighting consumption and expenditures, resolved beside the household row.
2. Equipment and appliance inventories - heating, cooling and water-heating systems plus major-appliance holdings, keyed to the same DOEID. The appliance tables travel far enough to anchor their own cataloged record: EIA Residential Energy Consumption Survey (RECS) 2020 - Appliances Tables (HC3.x), scored 9/10, ships the HC3.1-HC3.9 ownership-and-usage series beside the same 18,496-household microdata.
3. Imputation and weight machinery - the flags marking filled-in values and the analysis weights scaling the sample, without which unweighted averages answer a question nobody asked.

Published household cycles run 1993, 1997, 2001, 2005, 2009, 2015 and 2020, with the survey first fielded in 1978 and an initial 2024-wave release underway. Each wave is a full-stock snapshot rather than a point on a continuous series, which shapes sensible planning: trend work stitches cycles on consistent definitions instead of reading a ticker, and any figure you publish is stamped with its vintage, because there is no such thing as last month's RECS.

Most teams take the full historical depth once as a backfill and hold it steady in the warehouse, then let new vintages land as fresh rows beside the old ones instead of overwriting them. That is also how the revision history stays honest - the 2020 file's own post-launch variable additions show why vintage discipline matters even inside a single cycle.

A fixed routine separates citable numbers from bar-chart filler:

1. Apply the analysis weight before summing anything. Weighted, the respondents become the 123.5-million-unit occupied stock; unweighted, they remain 18,496 households answering a question nobody asked.
2. Normalize before comparing households. Divide KWH by TOTSQFT and control for HDD65/CDD65 so equal-efficiency homes in different climates stop looking like outliers.
3. Respect the geography floor. Census division is the standard reporting grain; state detail exists only where disclosure rules allow it, and no addresses survive anywhere. Model within the boundary rather than through it.
4. Reconcile against published aggregates before modeling. Sum weighted KWH and DOLLAREL against the estimate tables first - validation rows ship with every delivery precisely so this step is a check, not a research project.
5. Date-stamp every figure with its cycle year. Waves are periodic, reviewers will ask which snapshot a claim came from, and the 2024 initial release makes the question newly live.

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

Partially - the release's best-known boundary, and a structural one rather than an oversight. State-level consumption and expenditure estimates publish for select states; elsewhere results roll up to census division, which is exactly why the household row carries both a DIVISION field and a state_postal field populated only where masking permits. CBECS and MECS stop earlier still, reporting national and regional tables, so no survey in the bundle answers an all-fifty-states question alone.

For complete state-by-state accounting, the pooled slice adds EIA State Energy Data System (SEDS): annual production, consumption, prices, expenditures and CO2 emissions for all 50 states, the District of Columbia and US territories, with series reaching back to 1960 (prices and expenditures from 1970) and a consolidated long-format extract exceeding 2.5 million records keyed on documented five-character MSN codes.

The standard pairing holds: RECS 2020 microdata for household behavior inside covered geographies, SEDS for complete state totals, joined on state postal code. Our residential electricity use by state cluster walks that join in detail.

One cataloged record bundles three national surveys aimed at different populations, so the first decision in any project is picking the right grain. The table below lines them up.

All five formats in the bundle - CSV, XLSX workbooks, SAS transport files, PDF documentation and ZIP archives - surface across these cycles, and each survey ships its questionnaire forms, codebooks and methodology reports beside the data. None moves more often than annually, and the cycles do not align: the newest household year is 2020, the newest building year is 2018, the newest manufacturing year is 2022, so any cross-survey model has to tolerate a two-to-four-year stagger between grains. Where a continuous series matters more than unit-level depth, EIA Electric Power Monthly bridges the gaps month by month - the pairing our Electric Power Monthly archive cluster documents.

• Data scientists and ML engineers get the strongest labeled panel in the slice: weighted, weather-normalized households with hundreds of candidate features, ready for intensity, appliance-adoption and synthetic load-curve work. The fuller workflow lives at data scientists use cases.

• Market researchers and consultants segment with official denominators already counted - households by structure type, tenure, region and income band - instead of commissioning a survey to relearn them. See market researchers use cases.

• Utility strategy and electrification teams size the switch market directly: which homes hold which heating equipment, at what efficiency, paying what bill - the demand model behind heat-pump and time-of-rate planning (demand forecasting).

• Efficiency program evaluators read before-and-after waves on consistent survey footing, grounding impact claims in a probability sample rather than a vendor study (market sizing).

• Investors and quants read residential electrification exposure out of the same tables regulators publish, ahead of sizing positions in utilities and equipment makers (ESG emissions analysis).

Because the record is famous and still awkward. Coded categories arrive undecoded; the weight apparatus intimidates anyone who has not read the codebook; state fields sit half-empty for reasons buried in disclosure rules; and the file changed shape between its 2023 finalization and its 2024 variable additions, so stale copies circulate silently. Assembling all of that by hand is a project; receiving it is a query.

Datadory normalizes before delivery. Enums like TYPEHUQ and REGIONC arrive with value-label references attached, imputation flags traveling beside the values they qualify, weights intact and documented, join keys stable across cycles - and end-use, equipment or estimate-table slices cut to specification when the full 799-column width is more than a given pipeline needs.

Files, feeds, or straight into your warehouse. Daily, weekly, or hourly - your call. Start with a sample: name the field families, divisions and cycles, and the extract arrives cut to exactly that scope with the field dictionary attached. The schema in the sample is the schema you ship against.

Start with the anchor record, EIA Consumption & Efficiency - Electricity Surveys (RECS/CBECS/MECS), sampled to your sectors before anything is committed - product-level detail, sample rows, the full field dictionary and coverage chips live on its dataset page.

And when you're ready to build, request a sample scoped to the variables, divisions and vintages your models need - household 100001 loads from the same table as whatever the 2024 wave brings next.