Which parts of a near-death experience actually go together
The site already shows how often each element appears. This page asks a harder question: when two elements turn up in the same account, does that mean anything? The answer is sometimes yes and often no, and separating the two is more interesting than either.
What is being counted
One cohort, defined to match exactly what the public archive shows, so every figure here can be checked against the accounts.
published accounts, analysed, with an element assessment
What is left out, and why
- 1,556 — After-death communications, which /explore does not show in production
- 860 — Not yet analysed
- 1 — Withdrawn at the experiencer's request
Excluding after-death communications is the single largest choice made here. It nearly halves the apparent rate of meeting deceased relatives, because that is what an after-death communication usually is. The cohort matches the public archive so that clicking any number lands on the same set of accounts it was computed from; the cost is that this page is silent about a category where one of these elements dominates.
How common is each element?
The share of accounts carrying each element, with a 95% confidence interval. Click any row to read the accounts behind the number.
The denominator is not the cohort size. The analysis records absence explicitly, so an element’s denominator is the accounts where the extractor considered it at all, which varies from 6,135 to 6,299. Using the cohort size instead would understate every rate by a percent or two, consistently and invisibly.
Which elements travel together?
Every pair of elements, tested for association, then re-tested with a control for how much the person wrote.
First, the finding that shapes everything below: nothing repels anything.
Of the 105 possible element pairs, 1 occurs together less often than chance would predict, and even that one misses by a rounding error (Out of Body and Peace, Joy or Love, lift 0.997). There is no pair of elements that tends to exclude the other. Whatever these accounts are, they are not drawn from competing scripts where seeing a tunnel means you do not meet anyone.
That is a statement about co-occurrence, not about meaning. Two elements can appear together constantly because they genuinely belong to one experience, because the same person is simply describing more, or because the model that read the account was primed by one to look for the other. The next two panels separate the first from the second. Nothing here can separate out the third.
Second: longer accounts contain more of everything.
Element count and narrative length move together (r = 0.37). Accounts with one element average 415 words; accounts with 15 average 2,346. Some of that is real, a richer experience takes more words to tell. Some of it is not: an extractor reading two thousand words has more chances to find anything than one reading four hundred.
Elements present (left) against mean narrative length. Rows with fewer than 20 accounts are omitted.
Third: what survives the control.
Each pair below is measured twice. The crude odds ratio is the raw association. The adjusted one pools the same comparison within narrative-length quintiles, so a pairing that only looked linked because verbose accounts contain more of everything falls toward 1.0. These are the pairs that hold.
Most weakened by the control
These associations are real but noticeably smaller than they first appear. The gap is the part that was length, not linkage.
Effectively independent
An odds ratio near 1.0 means knowing one element tells you nothing about the other. Leaving the body, the most frequently reported element in the corpus, is close to independent of almost everything it could be paired with.
All 105 pairs were tested at once, so p-values are corrected with a Benjamini-Hochberg false discovery rate of 5%; 102 survive. Correcting matters here: at an uncorrected threshold, five pairs would look significant by chance alone. The control uses narrative length in five quintiles, which handles the verbosity confound but not the deeper one, that the same extraction pass produced both labels for a given account.
Given one element, what else follows?
Pick an element to see what accompanies it, measured against how common each companion is to begin with.
Bar is the conditional rate; the vertical tick is that element’s rate across the whole cohort. The number on the right is the difference in percentage points.
Read these 3,822 accountsWhat happens first, and what happens last?
Each element's average position through the account, on a scale from the opening moment to the return.
Ordering exists only for the 3,895 accounts with an extracted journey, and every one of them is a near-death experience. There is no ordering data for out-of-body, meditation, shared-death, or spontaneous accounts, so this panel describes NDEs alone. It also describes a selected half of them: sequenced accounts average 946 words and Greyson 17.9, against 832 words and Greyson 12.8 for the 847 NDEs without one. Elements appearing in fewer than 30 accounts are omitted.
How do these accounts open, and how do they end?
The first and last steps of the extracted journeys, as a share of all journeys that have one.
It begins with
The tunnel is not how these accounts start. It is not even close to the most common opening, despite being the single image most associated with the subject.
It ends with
Read the endings against each other rather than as rates: these are shares of the journeys that reach each terminal element, so they say what the last step tends to be, not what fraction of people chose to come back.
Does what you saw predict what you came back believing?
For each element, the share of accounts reporting a strengthened belief in an afterlife, and a reduced fear of death, against the cohort baseline.
The two columns behave completely differently. Afterlife belief spans 21 percentage points across elements, from the baseline of 46% up to 67% for accounts that describe a being of light. Fear of death spans only 9 points and barely moves off its 61% baseline no matter what happened. Whatever reduces the fear of dying appears to be common to the experience itself, not to any particular thing seen within it.
Both measures are the analysis model’s reading of what the account says changed, not a validated instrument administered to the experiencer, and both are self-reported retrospectively, often decades later. Elements also co-occur, so these columns are not independent of one another; an account describing a being of light usually describes several other things too.
How to read this page, and where it can mislead you
The limits are not footnotes. They determine what these numbers can and cannot support.
These are labels applied by a language model, not observations
Every element on this page was extracted from free text by a model reading an account. It can miss what is there and assert what is not, and it read both elements of every pair in the same pass, so a pairing may reflect the model’s own associations as much as the experiencer’s. The narrative-length control cannot touch that. Nothing here is a measurement of an experience; it is a measurement of how these accounts were written down and then read.
The people here selected themselves, twice
Everyone in this corpus chose to write their experience down and chose to submit it to a public archive, mostly NDERF and OBERF, mostly in English, mostly from the United States. Whatever makes someone do that is not random with respect to what they experienced. Rates on this page describe this archive and cannot be read as rates among people who have had these experiences.
Denominators differ per element
Absence is recorded explicitly rather than inferred from silence, so each element’s denominator is the accounts where it was assessed, which ranges from 6,135 to 6,299. Pair statistics use the accounts where both were assessed. Ordering statistics use a much smaller sub-cohort of 3,895, all of them near-death experiences.
Association is not sequence, and neither is cause
An odds ratio of 11 between choosing to return and reaching a border says those two get told together. It does not say the border caused the choice, or even that the border came first, though the ordering panel suggests it usually did. None of this is evidence about what a near-death experience is.
Discarded labels
11 element names appear in the data outside the standard vocabulary, each in at most 3 accounts (vibrations, darkness, voice, code_red, weightlessness, fear, mystical_realm, communication_with_light, buzzing_sound, eccentric_sensations, adc). They are excluded from every figure here.
Statistical choices
Proportions use Wilson score intervals rather than the normal approximation, which misbehaves at the rates the smaller elements sit at. Pair p-values come from a chi-square with Yates’ correction and are adjusted across all 105 tests with a Benjamini-Hochberg false discovery rate of 5%. The length-adjusted figures are Mantel-Haenszel odds ratios pooled over narrative word-count quintiles, dropping any stratum with fewer than 30 accounts.
Computed August 20, 2026 from a snapshot of 6,301 accounts. Recomputed by scripts/build-element-analytics.ts.