Public Dataset Method Guides

A method guide is not a description of a dataset. It is an account of what the dataset was built to do, which determines what it can be asked, and it spends most of its length on the boundary rather than the contents.

What each guide covers

The same five questions, in the same order, because the order is what makes them useful.

• Who collects it, and under what obligation. A file assembled from mandatory reporting behaves differently from one assembled from voluntary submissions, and the difference shows up as holes in specific places rather than as noise spread evenly.
• What it was built to answer. Every field exists because somebody needed it for an operational purpose. That purpose is the best available guide to which fields are maintained carefully and which are filled in when there is time.
• How its categories are defined. The published code book, not an inference from the values.
• What it does not contain. Usually the most valuable section, and usually the one nobody writes.
• How stable its publication is. Both the update cadence and, separately, whether the files stay where they were.

The last question is not a formality

Assembling the material for this section produced a run of broken references that is worth stating plainly, because it changes how these guides are written.

A federal agency landing page for a major carrier data system could not be found at any of three documented addresses. A statistics index published under a widely cited path returns a not-found error. And two federal hosts turned out to be unreachable by script while working perfectly in an ordinary browser, which inverts the meaning of a failed check.

The general lesson is the one worth carrying: automated link verification is not evidence about whether a document exists, in either direction. A tool that reports a page as gone may have been refused, and a guide that trusts it will eventually tell a reader something live is dead. Confirmation here is manual or it is nothing.

None of that is exotic. It is the ordinary condition of public data infrastructure. The practical consequence is that a citation needs a date attached, and a guide needs to name the surviving route to a document rather than the canonical one, when those differ.

Why transport data sits alongside court data

The datasets covered here lean toward road and commercial vehicle safety, which may look like a departure from a publication otherwise concerned with legal technology. It is not. These are among the largest, best documented, and most heavily litigated public data collections available, and they are routinely cited in argument by people who have never read their documentation.

They also make unusually good teaching material. Federal and state collections of the same underlying events use different inclusion thresholds, different severity definitions, and different geographic units, so putting two of them side by side demonstrates the entire problem of counting from administrative records more quickly than any abstract discussion of it.

What does not get a guide

Three kinds of collection get turned down, and saying why is quicker than a criterion.

A file that does not publish its coding scheme cannot be written about responsibly. The whole value of a guide is telling a reader what a field means, and without the issuing body's own definitions that would mean inferring meaning from values, which is guessing with extra steps.

A file assembled by somebody other than the body that collected the records is a different object from the records, however faithful it claims to be. Republished extracts, tidied versions, and commercial repackagings all introduce decisions the reader cannot see. Where one of those is the only practical way to obtain the data, that is worth saying out loud, and it is a caveat rather than a source.

A file behind a licence that forbids describing its contents is not worth the trouble. This is rarer than it sounds in public data and it happens.

What is here

The first guide covers the federal picture for commercial vehicle crashes, where four separate sources feed into what looks like one number, and disagree by design rather than by error. The second covers the Texas state crash records system, which answers several questions the federal sources cannot and refuses several it appears able to answer.

Read either as reference rather than as an argument. Neither reaches a conclusion, because the point of a method guide is to let a reader reach their own and know what it rests on. For the analytical material that uses this kind of source, our other work on court data is the way through.