The 15-Month Year in the Federal Case Database

Somewhere in a chart of federal civil filings there is usually a bump around 1992 that nobody comments on. It is not a surge in litigation. It is that the file for that year covers fifteen months instead of twelve, and a quarter more time produces roughly a quarter more cases.

The Federal Judicial Center says so plainly on the page the data is downloaded from. It is not hidden, it is simply upstream of where most people start reading.

Two different years wearing the same label

Before 1992 the reporting period was not the federal fiscal year. It ran July to June, and it was called a statistical year, so statistical year 1990 meant July 1989 through June 1990. In 1992 the period was moved to align with the government's October-to-September fiscal year.

The files were not retrofitted. Everything from 1970 to 1991 stays on the old July-to-June definition. Everything from 1993 onward is on the fiscal year. And 1992 is the seam, covering July 1991 through September 1992, which is fifteen months, because that is what it took to get from one calendar convention to the other without dropping a quarter on the floor.

Three consequences, in descending order of how often they bite. A count for 1992 is not comparable to a count for any other year. A series that crosses 1992 has a step in it that means nothing. And a year label before 1992 does not denote the same twelve months as the same label used by any source on the fiscal year, so joining this data to another series on the year column alone will silently misalign the older half.

The unit changes between files, and the guide says so

The second trap is that the database is not one table. The research guide is explicit that the criminal data is not counted the way the civil data is:

Criminal cases often include multiple defendants, so the information provided in the criminal IDB is at the defendant (not case) level. The defendant number needs to be part of the unique identifier created for each observation.

Federal Judicial Center, The Integrated Database: A Research Guide

So a row in the civil file is a case and a row in the criminal file is a person. Counting rows across both and calling the total cases overstates criminal litigation by however many defendants the average prosecution carried, and that ratio is not stable over time or across districts.

The appellate files are a third shape again, described in the guide as a hybrid carrying both lower-court and appellate record fields, including which side appealed. Four court types, at least three units of observation. Nothing is wrong with that, and nothing warns you at the point of use.

The outcome is not a field

People arrive at this database looking for a column that says who won. On the appellate side the guide describes something more awkward: for matters that began in the courts of appeals the outcome is recorded as granted or denied, but for appeals from the lower courts the answer has to be assembled from a combination of disposition together with outcome, procedural termination, or method.

That is three or four fields read jointly, with the correct combination depending on how the matter arrived. Any pipeline that maps a single column to a win or a loss has quietly chosen one reading and discarded the rest. The civil side has its own version of this problem, which the companion article in this section works through in detail.

What is outside the database entirely

A footnote in the guide draws a boundary worth knowing before designing any study: only cases before district, appellate, or bankruptcy judges are reported.

Proceedings handled by magistrate judges are therefore not represented as such. Given how much federal civil practice runs through magistrate judges, particularly on discovery and on consent for full disposition, this is a substantial exclusion and it is invisible in the data. Nothing in the file signals the absence.

Why the code books are per-year, and what a blank means

The FJC maintains code books that document how the case data has changed over the life of the database, including fields that were added, often because new legislation expanded federal jurisdiction, and fields that were deleted when the courts stopped reporting them.

That is the single most useful sentence on the landing page, because it means an empty cell has at least three possible meanings. The value was not applicable to this case. The value was applicable and not reported. Or the field did not exist in the year the record comes from.

Those are not interchangeable, and no amount of staring at the data distinguishes them. Only the code book for that year does. A researcher who treats blanks as missing-at-random has assumed the second meaning across a period where the third may dominate.

How to use it without being caught

• Read the code book for every year the analysis touches, not one representative year.
• Name the unit before counting anything, and never pool civil rows with criminal rows.
• Treat 1992 as its own category, or exclude it and say that you did.
• Check whether each field you rely on existed across the whole period, and when it changed.
• Where a published table from the Administrative Office already answers the question, cite that instead of rebuilding it.

None of this is a criticism of the database, which is unusually well documented for administrative data and is free. The documentation exists precisely because the seams exist. The failure mode is not using the data, it is using it as though the last fifty years were collected under one unchanging scheme.

The neighbouring section covers the datasets outside the courts where the same reading discipline applies. For the whole scope, our other work on court data lays it out.