# How a registered clinical trial primary outcome moved

**10,099 trials** sampled from ClinicalTrials.gov, one row each, describing whether the
primary outcome the trial registered is still the primary outcome its record names, when
it last changed, and what kind of change it was.

Source: **ClinicalTrials.gov**. Processed **2026-09-04**. **This is a frozen snapshot and is
not maintained current.** Every row carries `nct_id`, which resolves at
`https://clinicaltrials.gov/study/<nct_id>` to the live record, which is.

## What this is for

Registering a trial fixes, in public and in advance, what it will measure. The registry
keeps every version of that promise and dates each one, so it holds the evidence about
whether the promise was kept. The published literature has measured how OFTEN registered
primary outcomes change; the question it names and does not settle at scale is what KIND
of change they were, because a measure added after the data are in is not the same event
as a typo corrected. `change_kind` is an attempt at that, computed by a stated rule.

## The sampling frame, so a rate computed here means something

The frame is every **interventional** study with a primary completion date between
**2008-01-01 and 2024-12-31**: 311,713 of the 601,524 studies registered at the time of the
snapshot. Their identifiers were shuffled with a seeded generator and their version
histories fetched in that order, so **any prefix of the fetch order is a uniform random
sample of the frame**, and this file is such a prefix.

- 10,099 rows in this file.
- 5,446 are `in_headline_set`: completed, prospectively registered, with a primary
  outcome field. Rates about "trials that promised something and then finished" belong to
  this subset.
- 3,820 have `primary_outcome_changed` true.
- 3,667 carry a `change_kind`.

## What is NOT here, and why

**No trial record text.** No titles, no outcome wording, no summaries, no sponsors. The
ClinicalTrials.gov terms contemplate redistribution and attach conditions to it, one of
which is that redistributed data be kept current, which a frozen snapshot cannot do; the
terms also note an international copyright and possible third-party copyright in record
content. So this file publishes **our reading of the record and points at the record**.
Everything needed to check a row is in it: the identifier, the dates, and the rule.

## The limits, named rather than buried

- **A changed primary outcome is not evidence of misconduct.** Registrations are corrected,
  clarified, restructured to fit the registry's own data-entry rules, and amended for
  reasons an ethics committee approved. The registry records that a change happened and
  when. It does not record why.
- **The primary completion date is a proxy for data availability**, not for unblinding or
  database lock, and it is itself an editable registry field.
- **The comparison is one-directional.** The registry returns the ORIGINAL outcomes as
  structured objects but the CURRENT ones only as flattened text, so every test asks "is
  this original measure still present", never the reverse.
- `change_kind` is a rule, not a judgement. Its self-test and a 60-record hand audit are
  published with the study.

## Files

- `trials.csv`: RFC 4180 CSV, one row per trial, UTF-8, LF. An empty cell means the field
  does not apply to that row (for example, no change date on a trial that never changed).
- `schema.json`, `sources.json`, `manifest.json`, `validate.mjs`.

## Method and code

Pre-registration, harvest, analysis, classifier and verifier:
`research/the-outcome-that-changed/` in this project's repository. The page that reads
this data is <https://artwaste.land/strata/added-dropped-or-just-reworded/>.

## Citation

> Artificial Wasteland (2026). *How a registered clinical trial primary outcome moved.*
> https://artwaste.land/data/outcome-change-history/ . Derived from ClinicalTrials.gov,
> processed 2026-09-04; derived fields CC0 1.0.