Methodology

MoonAlibi fact-checks things blamed on the full moon using public statistics. The math is rigorous; the display is one word. This page discloses exactly how that word is decided.

What we measure is the correlation between the lunar phase and events — earthquakes, births, headaches, crime, and more. But "there were more on full-moon days" is not, by itself, a real correlation. We strip out the spurious correlation produced by confounders — season, day of week, long-term trends — and judge whether the gap between full-moon days and ordinary days exceeds chance. For only a few topics does a statistical link with the moon survive.

Defining a "full moon day"

• Moon phase is treated as a continuous value. For each event we measure the elapsed time from the nearest full moon instant (the syzygy moment, in UTC)
• A "full moon day" = within ±24 hours of the full moon instant (a 48-hour window). Same for new moons
• Phase instants are computed in-house using the algorithm from Jean Meeus, "Astronomical Algorithms", chapter 49 (no external API)
• All astronomy is computed in UTC; only the display is localized

Several famous "lunar effect" studies have been criticized for vague definitions of the full moon day — move the window arbitrarily and you can manufacture any conclusion. We therefore fix the definition and publish it here.

Verdict criteria

1. For the data in question, we derive the value expected if it were uniform with respect to moon phase (an expected count for counts; a baseline average for continuous quantities such as temperature) 2. We compute a 95% confidence interval for the gap between observed and expected (the observed/expected ratio for counts; the difference for continuous quantities) 3. If the interval contains "no difference" (1.00 for a ratio, 0 for a deviation), the verdict is "No difference" — unless the interval is too wide to detect even a meaningful effect, in which case the verdict is "Not enough data" rather than a claim of no difference 4. If the interval excludes "no difference", we grade by effect size using thresholds suited to the metric (±3% / ±10% for ratios, ±0.3°C / ±1.0°C for temperature, and so on): from smallest, "Practically none", "Slightly higher/lower (warmer/cooler)", "Higher/Lower (warmer/cooler)"

"No difference" reports an observation; it is not proof that no difference exists. Statistical tests cannot prove absence, which is why we avoid wording that asserts it.

Not printing the confidence intervals and effect sizes on every page is a deliberate design choice — the same reason a rain forecast shows "70%" rather than its internal model outputs. The criteria are fixed and published here, which is what keeps the verdicts honest.

Per-topic adjustments

We vary the baseline and preprocessing according to the nature of the data.

• Human-activity data (accidents, births, searches, news) carries strong weekday, seasonal, and holiday patterns, so we use a "same-condition (weekday or month)" average as the expected value and compare confounder-free indices
• Natural-phenomenon data (earthquakes, weather) does not follow the human calendar, so no weekday adjustment is needed; instead we apply phenomenon-specific preprocessing (for example, aftershock removal and a magnitude floor for earthquakes, or deviation from the monthly mean for temperature)
• Continuous quantities (temperature, wind speed, exchange-rate moves) are not counts, so we judge them with confidence intervals on differences or rates of change rather than ratios

Topic-specific processing is documented on each topic page.

Reproducibility

• The generator recomputes every verdict daily with the same procedure
• The same input data always produces the same output (idempotent builds)
• If a verdict label changes, the operator is notified and reviews the computation

Sources

• Moon phases and syzygy instants: computed in-house from Jean Meeus, "Astronomical Algorithms" (UTC, no external API). This is the shared foundation for every topic
• Each statistical dataset's source: differs by topic — USGS (earthquakes), the National Police Agency and NHTSA (accidents), ERA5 / Copernicus C3S (weather), GDELT (news), and so on. Specific sources, periods, and licenses are listed at the foot of each topic page