AURORA

Methodology

Meta-Aggregation Across Reference Studies

Aurora is additive across studies. Each reference study is loaded as an independent bundle — its own records, its own pre-registered hypothesis, its own reported finding, its own caveats. The analysis engine never merges records across studies (which would mask the per-study evidence and inflate apparent power). Instead, each study is analyzed on its own terms and the p-values are combined.

• Combination method: Stouffer's Z-score method, weighted by √n. Each study's p-value is converted to a one-sided Z-score via the inverse standard normal CDF: Z_i = Φ⁻¹(1 − p_i). The weighted combined Z is Z_combined = Σ(w_i · Z_i) / √(Σ w_i²) with w_i = √n_i. The combined p-value is p_combined = 1 − Φ(Z_combined).
• Why √n weighting: Stouffer-style sample-size weighting gives larger studies proportionally more influence on the combined estimate, which is appropriate when studies share a common true effect and primarily differ in sampling precision.
• Metadata-only studies: Studies that ship only with a citation and reported p-value (no raw records) still contribute to meta-aggregation through their reported p and n. They do not contribute to the user's distribution chart.
• Re-analyzed studies: Studies that ship with raw records are re-analyzed by Aurora using the same chi-squared engine as the user cohort. The re-analyzed p replaces the originally-reported p in meta-aggregation. This is more conservative than trusting the original analysis as-published.
• Variable targeting: A study contributes to a variable's meta-aggregation only if it (a) has records that can be re-analyzed on that variable, or (b) lists the variable in its preRegisteredVariables or primaryFinding.variable. This prevents post-hoc fishing across variables a study didn't target.
• What meta-aggregation cannot do: It cannot rescue biased samples. If every input study draws from a flawed source database with the same selection effects, the combined p-value carries those biases forward. Replication across independent samples with different selection criteria is what builds evidence; combining p-values from samples with shared selection bias is not.

Reference Studies — Schema

Reference studies are JSON bundles conforming to schema version aurora-study-1.0. Each contains:

• id — unique short identifier
• citation — author, year, title, journal, volume, optional URL
• n, optional nTimedSubset, controlN — sample sizes
• hypothesis — plain statement of what was tested
• preRegisteredVariables — array of Aurora variable IDs the study targeted
• primaryFinding — variable, test description, observed/expected statistics, p-value, plain interpretation
• caveats — array of methodological caveats and limitations
• records — optional array of subject birth records (same schema as user records)

Use the "Study JSON Template" button on the Reference Studies tab to download a starter schema. Studies are loaded additively — importing a new study does not replace existing studies or user records.

Variables Computed from Birth Date Alone

• Sun sign — date-precise tropical zodiac mapping (12 categories). Western astrological boundary dates per IAU/conventional ranges.
• Sun modality (Cardinal/Fixed/Mutable) — 3-category collapse of sun sign by quality. Cardinal = Aries/Cancer/Libra/Capricorn; Fixed = Taurus/Leo/Scorpio/Aquarius; Mutable = Gemini/Virgo/Sagittarius/Pisces. The dimension where Ruis (2008) reported the strongest signal in his serial-killer sample.
• Sun element (Fire/Earth/Air/Water) — 4-category collapse by triplicity.
• Decan — 1st, 2nd, or 3rd decanic division of the sun sign (3 categories per sign × 12 signs = 36 possible, but tested as 3-category collapsed across signs).
• Life path number — Pythagorean numerology: digital reduction of full birth date with master-number preservation (11, 22, 33). 11 categories.
• Day of week — 7 categories. Included as a sanity check: a strong day-of-week effect would suggest selection bias in the dataset (not an astrological signal), since day-of-week is uncorrelated with birth astrology.
• Season — meteorological seasons (Northern Hemisphere). 4 categories.
• Chinese zodiac year — 12-year cycle, approximated with February 4 cutoff (true lunar new year boundary varies; this introduces ~1.5% misclassification at year boundaries).

Variables Computed from Birth Date + Time

• Moon sign — computed from simplified Meeus lunar longitude formula. Accuracy ≈ ±1–2° for sun-relative position; sign-accurate except within ~2° of sign cusps. When birth time is missing, noon UTC is used as a default (introduces up to ±6° error and may cross a sign boundary; flagged in record metadata).
• Moon modality / Moon element — same collapse logic applied to moon sign.
• Sun+Moon factors in Mutable — count of how many of {Sun, Moon} fall in Mutable signs for each subject. Values 0, 1, or 2. Partial replication of Ruis's 8-factor Mutable composite, which used Sun, Moon, Mercury, Venus, Mars, Jupiter, Saturn, and Ascendant. Full replication is deferred to v2 pending Swiss Ephemeris integration.

Statistical Procedure

• Test: Pearson chi-squared goodness-of-fit. χ² = Σ (O − E)² / E with df = k − 1. The expected distribution is either uniform (default) or empirically derived from a control cohort selected in the Analysis tab.
• Control cohort (empirical null): When a control category is selected, expected probabilities are computed from observed control proportions (with Laplace smoothing of 0.5 to avoid zero-cell issues). This matters because real birth-date distributions are not uniform — there is documented seasonal variation in human births (Roenneberg & Aschoff, 1990, cited in Ruis 2008). Testing against uniform when truth is non-uniform inflates significance.
• Bootstrap p-values (Ruis 2008 method): When enabled with a control cohort, the engine resamples n records (matching test cohort size) with replacement from the control, computes the chi-squared statistic for each resample, repeats 1,000 or 5,000 times, and reports the proportion of bootstrap statistics ≥ the observed cohort statistic as the bootstrap p-value. Smoothed by (extreme + 1) / (B + 1). Bootstrap is more robust than analytic chi-squared when expected cell counts are low or when the empirical null deviates from the chi-squared asymptotic distribution.
• Effect size: Cramér's V, computed as √(χ² / (n × (k−1))). Interpretation: ≥0.1 small, ≥0.3 medium, ≥0.5 large.
• Confidence intervals: Wilson score intervals at 95% on each category proportion.
• Multiple comparison: Bonferroni correction. Corrected α = α / (number of variables tested). A test is "significant" only if the effective p-value (bootstrap if available, else analytic) is below α-corrected.
• Minimum expected frequency: A warning is emitted if any expected cell count is below 5 (chi-squared assumption violated; the bootstrap p-value, if available, should be preferred).

Replicating Ruis (2008)

Aurora can be used to attempt partial replication of Jan Ruis's 2008 study, "Statistical analysis of the birth charts of serial killers" (Correlation 25(2)). To do so:

• Tag offender records with a category like homicide or serial, with verified birth data and source citations.
• Tag a control cohort (general population sample with no offender records) with category control. Aim for control n ≥ 5× cohort n.
• In Analysis: select test cohort, select control cohort, enable bootstrap (1000 resamples), check the variables Sun Modality, Moon Modality, and "Sun+Moon factors in Mutable."
• Run analysis. The Mutable signal, if present and large enough to detect with only 2 factors instead of Ruis's 8, will appear as elevated counts in the Mutable category and depressed counts in Cardinal/Fixed, with a low bootstrap p-value.

Note: Ruis's original finding required 8 astrological factors and a sample of 293. Detection power with only 2 factors (Sun and Moon) and a smaller sample will be lower. A null result with 2 factors does not falsify the Ruis finding; it indicates the signal, if real, is below the detection threshold of this partial replication.

What Aurora Cannot Tell You

• It cannot establish causation. Even a real correlation between an astrological variable and offender status would not establish that astrology causes anything. Selection effects, recording bias, era effects, and confounders all produce spurious correlations.
• It cannot validate astrology. A null result means the signal is absent in this dataset; it does not "disprove astrology." A positive result with a properly powered, properly controlled, pre-registered analysis would warrant further investigation, not belief.
• It cannot score individuals. Aurora reports group-level distributions only. There is deliberately no individual-prediction output. Birth-data-based scoring of individuals for criminality risk is not a valid use of any statistical instrument and is not what this tool is for. The Case Lookup tab walks through one subject's chart and shows how each computed characteristic appears across loaded data; characteristics are never summed into a per-subject score, and no "% match," "risk score," or prediction is produced.
• It cannot fix biased samples. AstroDatabank and similar registries have known selection effects — whose birth time gets recorded is non-random. Famous criminals are over-represented; less-publicized cases under-represented. Era effects (recording practice changes over decades) further confound. A positive result in a biased sample is not evidence of an astrological signal.

Known Limitations · Roadmap

• Mercury, Venus, Mars, Jupiter, Saturn positions and the Ascendant are not yet computed. These require Swiss Ephemeris integration plus birth time and location, planned for v2. Full Ruis 8-factor Mutable replication awaits this.
• Planetary aspects and house placements are not computed (also v2 pending ephemeris).
• No pre-registration enforcement. For genuine research use, analysis variables and α should be specified before data is collected; Aurora cannot enforce this discipline.
• localStorage persistence is not secure storage. Do not enter identifying information about living subjects without appropriate IRB / data-protection review.

Citation

If results from Aurora are reported, methodology and version should be cited. Statistical procedures follow Pearson (1900), Cramér (1946), Wilson (1927), and Efron (1979) for the bootstrap. Lunar position calculations follow Meeus, Astronomical Algorithms, 2nd ed. (Willmann-Bell, 1998), simplified main terms. The methodological design of this instrument draws from Ruis, J. (2008), "Statistical analysis of the birth charts of serial killers," Correlation 25(2), and inherits its limitations regarding source-database selection effects.