Every count on this site comes from a set of recorded conversations. This page is where the counting is explained — the whole chain from recording to denominator, the limitations stated before any finding rather than after, and the findings that were withdrawn.
This began as a podcast, not as a study. That is a limitation and it belongs in the first paragraph, not a footnote.
The conversations were recorded to be published as episodes. No research was contemplated when they were made, so nobody was sampled, nobody was randomised, no instrument was piloted, and no consent to research participation was sought at the time. The analysis was built three years later out of recordings that already existed.
What that buys, and what it costs. It buys length and candour — people talked for an hour without a research frame, about things a questionnaire would never have reached. It costs every property that makes a sample representative. Nothing here generalises to the profession, and no percentage on this site should be read as a rate in the population.
What was done to it afterwards. Every usable transcript was read against the same four questions, coded into a fixed set of variables, and counted. The coding rules, the exclusions and the failures are published. The design was retrospective; the coding was not.
Four different numbers are all correct and all count different things. Most confusion about this study comes from treating them as one number.
| Stage | Count | What it counts |
|---|---|---|
| Recordings identified | 336 | Distinct recordings the retrieval run attempted |
| Could not be downloaded | 3 | All three logged the same format error; whether they were interviews at all is unknown |
| Transcribed | 333 | Usable transcripts produced |
| Deliberately skipped | 2 | One host-authored segment with no guest; one “part 2” of an interview already coded |
| Coded | 331 | Transcripts read against the four fixed questions |
| Excluded from the denominator | 22 | Reasons below |
| THE DENOMINATOR | 309 | Every percentage on this site divides by this |
| Reason recorded in the data | Conversations |
|---|---|
| Not a chiropractor | 10 |
| Second episode with a person already coded | 5 |
| No reason recorded | 3 |
| Not yet enrolled | 2 |
| Host-authored training module, no respondent | 1 |
| Aborted recording, restarted | 1 |
| Total | 22 |
Three of the 22 are excluded with no reason recorded. That is a gap in the record and it is stated here rather than rounded away.
Conversations are not people. 331 conversations involve 327 distinct people — at least four sat for more than one recording. A count of conversations is not a count of individuals, and every figure on this site counts conversations unless it says otherwise.
138 of 331 conversations carry a verifiable publication date. 193 do not. Of the 138: two in 2019, 34 in 2020, 34 in 2021, 31 in 2022, 37 in 2023, the latest being November 2023. So “recorded between 2019 and 2023” is supportable, but it describes the 42% that carry a date, not the whole set.
Every transcript was read against the same four questions. Not asked identically in every conversation — these were interviews, not a survey — but the transcript was read looking for an answer to each:
That produced 1,324 coded records — exactly 331 × 4. Every conversation has a record for all four questions, including where the answer is “not discussed.”
Alongside the four answers, 40 further variables were coded — how the person first got inside a practice, who suggested the career and at what age, what they say formed them, whether they have hosted or employed a student. That is 27,729 filled cells across 2.39 million words of transcript.
Nothing was inferred to fill a blank. Where a conversation does not address something, the cell says so. “Not discussed” is a value, and it is never treated as a no.
A summary is not speech. Each record holds both a coder’s summary and, where one exists, the span of the person’s own words it rests on. The summaries total 105,485 words; the quoted spans 64,877. Counts are built on the spans, not on the summaries.
Being named is not being endorsed. Where a count records how often a school, a technique or an organization is named, some of those mentions are critical. The count measures presence in the story, and nothing else.
1. Everyone in this corpus stayed. These are people who entered the profession and remained in it long enough to be interviewed about it. Nobody who left is here. Nobody who tried to get in and could not is here — one person describes seeking access to a practice and never obtaining it, and she is the only trace of that population in 309 conversations, and she got in anyway or she would not be in the set. Any finding about barriers is bounded by this and it cannot be corrected for.
Nineteen claims were assembled that rested on more than one person independently saying the same thing. Each was put through a check for hidden connection between its members. Five did not survive.
| Outcome | Groups |
|---|---|
| Stands | 10 |
| Stands, with a caveat written into the claim | 2 |
| Rebuilt with different members | 2 |
| WITHDRAWN | 5 |
| Total | 19 |
Fourteen stand. Five are withdrawn and are not published anywhere on this site.
The pattern in the five failures is worth more than the five claims were. A group fails when its members turn out to be members of the thing the claim is about — people who agree because they were taught together, not because they found the same thing separately.
One failed differently, and it is the instructive one. Its members were not connected to each other; they were connected to the interviewer. That is a failure mode no scan of what respondents say could ever detect, and it is the reason the check is described below as a lower bound.
The check’s own limitation, which travels with every claim that passed it. Connections were found by matching names that respondents actually said. A person shaped by a teacher they never mention is invisible to it. So it is a lower bound on connection, never an upper one, and a group passing the check does not establish independence — it only means no connection was found.
Both are recorded as results rather than deleted. The first predicted a band of cases for a particular pattern; the actual count came in below it and the predicted failure condition occurred instead. The second named in advance the kind of case that would falsify it — and that case arrived. Two of its clauses did not survive, one survived weakened, and one survived strengthened and was restated.
A study that publishes only what worked is not reporting a method. These are here because they are the part that tests whether the rest can be trusted.
When a figure published here is found to be wrong, it is corrected in place, dated, and the superseded figure is left visible. Nothing on this site is quietly edited after publication.
If you find an error — in a number, a denominator, a source, or a date — tell us here. We will correct it, date the correction, and say that we did.
This applies to sources as well as figures. Where a document this site cites is superseded by a later edition, the citation is updated and the superseded reference stays visible, so a reader can see what changed and when.
This page is the short account. The complete write-up — every finding, every denominator, every caveat, and the material that did not make it into any published finding — is Volume One.
Published September 13, 2026 · Last verified September 13, 2026