A category is often treated as a permanent fact about a record: this company is in that industry, this customer belongs to that segment, this job has that occupation code.
Usually, it is not that simple. A category is a definition applied at a particular time. The definition may be revised while its familiar label remains. A numeric code may persist while the records it covers change. And an apparent trend may turn out to be a change in the filing cabinet rather than a change in the world.
That is why a category needs a version number.
This matters whenever a report compares years, joins older files to newer ones, or tries to turn source data into a long-running business metric. A category value should be read as a fuller statement: *this record met this definition, in this classification edition, at this time.*
A matching code is not proof of a match
The U.S. North American Industry Classification System, or NAICS, offers a useful example. NAICS is the standard used by U.S. federal statistical agencies to classify business establishments for collecting, analyzing, and publishing information about the business economy.
Because it is widely used, NAICS codes often appear in vendor files, public datasets, market reports, and internal business tables. That familiarity can encourage a risky shortcut: treating the code as a stable identity across time.
The official 2022 NAICS materials explicitly warn against that assumption. They mark some entries with an “R,” indicating that a 2017 industry code was reused with different content.
Consider code 425120. In the 2022 edition, it is labeled “Wholesale Trade Agents and Brokers.” But the official concordance shows that its newer content includes material from both the earlier 425120 category and the former 425110 category, “Business to Business Electronic Markets.”
A table containing 425120 in two extracts may look easy to join. It is not automatically a like-for-like comparison. The number matches, but the definition does not fully match.
The same problem can occur without code reuse. Labels may be retained while their boundaries shift. Local categories and vendor-defined segments can change even more quietly, with no public manual or concordance to consult.
When the boundary moves, the trend can move with it
Retail classification changes between the 2017 and 2022 NAICS editions make the problem more visible. The 2017 system included “Electronic Shopping and Mail-Order Houses,” a category organized largely around how retail occurred.
For the 2022 revision, the recommended direction was to place less weight on whether goods were sold online, in a store, or through print delivery. The prior category’s content was distributed across numerous merchandise-based retail categories, including electronics and appliances, clothing, books, health and personal care, and general merchandise.
That does not mean online retail vanished. It means the classification changed its organizing principle.
Imagine a file with 2019 records coded as 454110, “Electronic Shopping and Mail-Order Houses,” and a 2024 file with product-based retail codes. A quick pivot table might show “e-commerce” falling and other retail categories rising. That could reflect changing business behavior. It could also reflect the changed classification boundary, or both.
Before telling a story about customers, companies, sales, or employment, ask whether the definition changed first.
Preserve the original, then map for a stated purpose
The answer is not to avoid mapping categories. Mapping is often necessary. The answer is to stop treating mapping as replacement.
Keep the category as received or historically assigned. Then add a separate harmonized category for the specific analysis at hand. A current portfolio view may reasonably use a current taxonomy. A long-run trend may require a stable higher-level rollup. Those are different analytical decisions and should remain visible in the data.
A compact metadata pattern can prevent later confusion:
| Field | What it preserves |
|---|---|
raw_category_label | The original words supplied with the record |
raw_category_code | The original code, if one exists |
category_system | NAICS, SOC, a vendor taxonomy, or an internal scheme |
category_edition | The applicable revision or version |
assigned_or_observed_date | When the value was assigned or observed |
harmonized_category | A separate value created for a defined analysis |
mapping_reference | The concordance, rule, or documentation used |
mapping_note | Important limits, assumptions, or exceptions |
The source label matters even when it is messy. A source may use an official standard, a local business label, or an informal spreadsheet term. Those are not interchangeable, and forcing them into one field can erase useful context.
Treat a crosswalk as an analytical choice
A concordance describes how one edition relates to another. It does not automatically solve every historical comparison.
At minimum, label each relationship as one-to-one, many-to-one, one-to-many, or partial or ambiguous. One-to-one mappings are often manageable after review. A one-to-many split deserves more caution: an old total cannot be assigned among several newer categories unless record-level evidence exists or an allocation rule is chosen and disclosed.
For trend reporting, there are usually three honest options:
- Compare periods only within a stable edition.
- Roll both periods up to a higher-level category that can be shown to remain comparable.
- Mark a break in the series and avoid a false comparison.
The third option can feel unsatisfying, especially when a chart needs a clean line. But an interrupted line is more useful than a smooth one built from incompatible definitions.
Industry classifications are not unusual in this respect. Occupational classifications also change as work changes. The 2018 Standard Occupational Classification revision added a net 27 detailed occupations, with meaningful updates in areas including information technology and health care. A broad job family may still sound familiar across editions, but detailed occupational trends may not be continuous by default.
Categories are useful because they compress a complicated world into manageable groups. The compression is necessary. It also has consequences.
Before comparing category totals across years, ask five questions: What system assigned the category? Which edition was used? What did the definition mean then? Did the code or label change meaning? Is the mapping truly one-to-one?
A category is not just a label on a row. It is a decision about what belongs together—and every such decision has a date.
Morrow Creek Data Services handles paid research, cleanup, verification, and structured delivery projects.
Start a data project