Skip to content
Likelier

About Likelier

Likelier compares what people think is dangerous with the estimated lifetime probability, using authoritative sources and a transparent methodology. The goal is to give readers — and the LLMs they ask — a better starting point than gut feelings and news cycles.

Content is available in 12 languages.

ARDEENESFRHIIDJAPLPTRUZH

What we publish

Native probability

A native-unit probability as the source reports it ("1 in 11,000,000 per flight").

Normalized lifetime risk

A normalized lifetime risk for a US adult so different topics are actually comparable.

Full claim ledger

A full claim ledger: source type, verbatim excerpt, calculation notes, uncertainty, archive snapshot.

Perceived-risk side

A perceived-risk side tied to a cited survey, or explicitly labeled as editorial intuition.

Open data

A suggestions page, corrections log, and the raw data as JSON.

What we don’t publish

When the evidence is weak, contradictory, or can’t be normalized to a comparable baseline, we say so explicitly using a no reliable estimate state rather than inventing a number.

Trust

Content governance is the point of the site. See methodology for the normalization assumptions and source-type policy, and suggest for the public log of fixes.

Who makes this

A research team backed by Claude agents for sourcing + verification, with human spot-check before publication.

Data freshness

Every entry tracks its last review date. The newest are within the last 30 days; the oldest are flagged for re-review when sources update.

Support

Likelier is free and ad-free. If it’s useful to you, consider buying us a coffee — it funds the research and the archiving that keeps citations durable.

If this helped you think more clearly about risk, consider supporting the project so we can keep the data trustworthy and ad-free.

Buy me a coffee
Methodology

This page is the most important one on the site. If you read nothing else, read this. Our numbers are only as trustworthy as the assumptions behind them, and every assumption we rely on is disclosed here.

The normalization problem

Different sources report risks in different units. Aviation fatalities are often reported "per flight"; car crashes "per mile" or "per year"; cancer "lifetime"; shark attacks "per beach visit". These units are not comparable directly. Saying "plane crashes are rarer than shark attacks" without picking a common denominator is meaningless. So we pick one.

Our canonical basis

Every entry that publishes a number publishes both:

  • Native: the stat as the source reports it ("1 in 11,000,000 per flight", "~6 deaths per year in the US", etc.). Full transparency on what the source said.
  • Normalized: a lifetime probability, with an explicit scope tag so you can see what population the number is for. The default and most common scope is lifetime probability of death for a US adult, assuming 59 years of remaining life starting from age 18 (rounded from SSA actuarial life tables) and an activity-exposure figure disclosed per entry. Other scopes are also supported: global_adult_lifetime for fears dominated by non-US populations (tsunami, malaria, tuberculosis), activity_specific_lifetime for per-event activities (per astronaut mission, per vaccine dose, per ski day), and subgroup_lifetime for fears defined on a narrower population (women for breast cancer, men for prostate cancer, soldiers in combat). The conversion uses 1 − (1 − p)n where p is the per-year probability and n is remaining years.

The detail page shows both the native number and the exact calculation steps used to convert between them, plus the scope badge.

An honest limitation on cross-scope comparison

The site’s ranking and pairwise comparison pages sort all entries on a single lifetime-probability axis, but the scopes mixed together on that axis are not strictly comparable. A per-trip dengue probability, a global lifetime-adult cancer probability, and a per-mission astronaut figure are three different things being plotted on one scale. We do this anyway because the alternative — splitting the site into disjoint sub-catalogs — defeats the point of letting people calibrate two fears against each other. The compromise is that every card and detail page carries a scope badge, and the per-entry assumptions field spells out exactly what the number represents. Read those before you use a number from this site to make a decision.

Source-type policy

Every source is classified as one of:

  1. primary_study — peer-reviewed research, original data collection.
  2. govt_report — CDC, NHTSA, FAA, NSC, WHO, ONS, Eurostat, etc.
  3. peer_reviewed — journal article (secondary analyses OK).
  4. reputable_reference — Our World in Data, NSC Injury Facts, official actuarial tables.
  5. news_article — allowed only as corroboration, never as the primary citation.
  6. encyclopedia — never as the primary citation.

Minimum bar: at least one source per entry must be primary_study, govt_report, peer_reviewed, or reputable_reference. Entries that fail this are either rewritten or routed to the no_reliable_estimate state.

Independence check

Two sources that both cite the same underlying dataset don’t count as independent verification. When this happens, the entry gets an independence_note explaining the shared upstream — that honesty is more valuable than a false sense of corroboration.

Uncertainty

Every published probability includes an uncertainty range [low, high]. Where the source reports a confidence interval, we use it. Where it doesn’t, we widen by a conservative multiplier and document the choice. Charts render the uncertainty band visibly, not as a cosmetic accent.

Perceived risk

The "perceived" side is labeled by how we sourced it:

  • survey — cited survey with methodology (Chapman Survey of American Fears, Gallup, Eurobarometer).
  • poll — linked polling aggregator or poll result.
  • intuition — editorial intuition, not polled. Entries marked this way appear with a visible demotion so you know we’re speculating on the perceived side.

Provenance durability

At publish time, every citation URL is submitted to the Internet Archive’s Wayback Machine via Save Page Now. The archived copy is stored alongside the live URL. When a CDC page gets reorganized 18 months from now, we still have the text we relied on — and so do you.

Quality review

Each entry is scored on eight dimensions before publication. Here is what each dimension checks and why it matters.

Risk entries

Source grounding. Every URL we cite must resolve, and every quoted excerpt must appear verbatim in the source. If a source goes dead between publication and re-review, we use the archived snapshot captured at publication time. A score of 1 here is an automatic reject.

Source authority. At least two independent sources per claim, and at least one must be a peer-reviewed paper, primary study, government report, or reputable reference work. News coverage and encyclopedia summaries on their own do not clear the bar.

Arithmetic. The chain from the raw number a source reports to the "1 in N" we publish to a lifetime probability for a US adult is shown step by step and re-checked independently. A wrong denominator or a missing exposure window fails this check. A score of 1 here is an automatic reject.

Uncertainty. The low and high band we publish must be defensible from the literature and must contain the point estimate. A suspiciously tight band on weak data is treated as a failure to calibrate, not a strength.

Scope. If the entry is not a US-adult lifetime probability — because it is activity-specific, global, or restricted to a sub-population — the scope field has to declare it and the underlying calculation has to match. Mislabeling an activity-specific rate as a population baseline is the textbook failure. A score of 1 here is an automatic reject.

Prose. Dry, skeptical, no "you should…", no false reassurance, follows the in-house style guide. Polished copy that sneaks advice back in gets returned for a rewrite.

Perception honesty. When we publish a perceived-risk number, we say whether it comes from a cited survey or from editorial intuition. We do not invent a survey to fill a gap.

Caveat completeness. Known limitations — population heterogeneity, surveillance gaps, the era the underlying data is from — are stated in the body of the entry, not buried in a footnote.

Threshold: every dimension scores at least 3, the average is at least 4.0 / 5, and any score of 1 on Source grounding, Arithmetic, or Scope is an automatic reject regardless of the average.

First audit data will be published with the next review batch.

Decision pairs

Source verification. Every URL we cite must resolve and every quoted excerpt must appear verbatim in the source. A score of 1 here is an automatic reject.

Source authority & independence. Each side of the decision — the action and the inaction — needs at least one authoritative source, and the two sources must be truly independent rather than the same study reframed two different ways.

Regret-rate accuracy. The regret percentage must come from a study that directly measures regret for this decision. A satisfaction proxy, a related-but-different question, or a regret rate borrowed from a sibling decision all fail this check. This is the dimension that fails most often in review. A score of 1 here is an automatic reject.

Source comparability. Both sides should use comparable methodology — same time horizon, same sample frame, same regret instrument — or we say so explicitly. A score of 1 here is an automatic reject.

Gilovich pattern. Thomas Gilovich and Victoria Medvec (1995) found that people regret actions more in the short term but inactions more over a lifetime. We classify each decision against that pattern; when our data inverts it, we publish the inversion and flag it rather than smoothing the framing.

Prose quality. Dry, skeptical, no "you should…", no false reassurance, follows the in-house style guide.

Caveat completeness. Known limitations — sample quality, generalizability, population assumptions — are stated in the body of the entry, not buried in a footnote.

Sample quality. Sample size and representativeness must be appropriate for a regret claim about US adults. Small, skewed, or non-US samples are called out in the body.

Threshold: every dimension scores at least 3, the average is at least 4.0 / 5, and any score of 1 on Source grounding, Regret-rate accuracy, or Source comparability is an automatic reject regardless of the average.

First audit data will be published with the next review batch.

Corrections & suggestions

Mistakes happen. When we find one — or you report one — we fix it, preserve the prior value in the git history, and log the change publicly on the suggest page. Every detail page has a "Report it" link.

Limits

We’re publishing population-level probabilities for a "typical" US adult. Your actual risk depends on your age, health, location, behavior, and luck. These numbers are a starting point for grounded thinking, not personalized advice. See the terms.

Suggest

Have an idea for a risk we should cover? Spotted something that needs fixing? We’d love to hear from you.

Open a suggestion: use the structured issue template (opens in new tab) .

Running an AI agent? File from there

If you reach Likelier through the MCP (Claude Desktop, Claude Code, Cursor, Continue…), ask your agent to call the suggest_data_gap tool. Suggestions filed that way land in the issue tracker automatically — no context switch required. MCP docs.

Corrections log

Per-entry change history lives in git — every fix bumps the source list and preserves the prior value. The aggregate log here will populate as the suggestion workflow scales.