Music Data Research Methodology and Statistical Standards

What kind of study is this?

The song-title project is an exploratory observational corpus study. It describes eligible records available in LyricsWeb; it is not a random sample of every song ever released. Results therefore support precise statements about this corpus, but they must not be presented as universal estimates for all music.

Unit of analysis and eligibility

The unit of analysis is one eligible LyricsWeb song record with a non-empty title and a usable release year. Translation and romanization utility pages are excluded so alternate presentation pages are not counted as separate originals. This version treats each eligible database record as one unit. A documented identifier-based duplicate audit is still required before making claims about unique recordings.

How a release year is assigned

Structured song-level release year.
A four-digit year parsed from the song-level release date.
Embedded album release year or date only when song-level data is unavailable.

Every yearly row reports how many years came directly from the song and how many used the album fallback. Release-year analysis does not use advertising or audience attributes. Language labels use the primary catalog language value, with a separately validated catalog classification used only when the primary value is missing.

How titles are measured

Token matching is Unicode-aware: each contiguous sequence of letters or numbers is counted as one title token. This supports multiple scripts, but it is not linguistic word segmentation and should not be interpreted as if every language separates words in the same way. The analysis measures token count, Unicode character count, symbol-only titles, one- through four-token groups, five-or-more-token titles, numeric characters, question marks, and exclamation marks. A symbol-only title is non-empty but contains no Unicode letters or numbers. These titles remain eligible and are reported separately. Decade summaries are weighted by the number of eligible records in each year.

How title novelty and reuse are measured

The title-originality study normalizes each eligible title conservatively, then counts one artist–title association at its earliest observed year. A title is “new” only when no earlier association is observed in this corpus; “previously used” means the normalized title appeared in an earlier eligible year. Cross-artist reuse counts normalized titles associated with more than one artist. These are naming-pattern measurements, not judgments about artistic quality, intent, authorship, or creativity.

Because years contain very different numbers of records, the study publishes raw ratios alongside a fixed-size 1,000-association annual diversity measure. The current partial year is excluded from long-term trend tests. Results also report title concentration, collision probability, robust slopes, lag-one dependence, and false-discovery-rate-adjusted probabilities.

Descriptive statistics

Every result begins with sample size and the observed distribution. Counts, proportions, weighted averages, medians, quartiles, and dispersion are reported where the underlying snapshot supports them. Percentages are always shown alongside record counts so that a striking percentage from a small year is not mistaken for a high-volume pattern.

Uncertainty for proportions

Proportions use 95% Wilson score intervals. Wilson intervals remain within the possible 0–100% range and have more reliable coverage than the basic normal approximation in many sample-size settings. These intervals describe conditional uncertainty within the observed corpus; they do not remove selection bias in which music is represented.

Trend magnitude and statistical testing

Theil–Sen median slope estimates the direction and magnitude of a trend while reducing sensitivity to isolated extreme years. Mann–Kendall is used only as an exploratory test of monotonic trend. Before any inferential wording is published, lag-one serial dependence is evaluated because neighboring years are not guaranteed to be independent.

When several measurements or language groups are tested together, p-values are adjusted with the Benjamini–Hochberg false-discovery-rate procedure. We report effect magnitude and practical meaning rather than treating statistical significance as the finding itself.

Sensitivity checks

Exclude the incomplete current year.
Compare direct song years with the album-fallback-inclusive result.
Repeat highlights under different minimum yearly sample thresholds.
Separate sparse early periods from well-covered modern periods.
Compare explicit language metadata with inferred language only after validation.

How linguistic diversity is measured

The language-diversity study reports several complementary measurements because no single statistic captures both richness and balance. Effective-language count converts Shannon entropy into an intuitive equivalent number of equally common languages. Miller–Madow entropy reduces finite-sample bias. Simpson diversity emphasizes the probability that two records belong to different language labels, while largest-language share reports concentration directly.

Annual catalog size changes sharply over time, so raw active-language counts are accompanied by expected distinct languages in a fixed sample of 1,000 records. Trend tests cover 1950–2025 and exclude the partial 2026 year. Every headline result is checked in a known-label-only view and a stricter primary-label-only view. Unknown labels remain visible in the full descriptive totals.

Theil–Sen slopes, Mann–Kendall statistics, Benjamini–Hochberg-adjusted probabilities, and lag-one autocorrelation are reported together. Strong serial dependence limits inferential interpretation: the study describes patterns in the documented catalog and does not prove a causal cultural change or represent every recording worldwide.

Language labels and classification

Language filters use validated catalog language labels. The primary label covers 94.672% of the audited source collection; a secondary catalog classification is used only when the primary label is absent. Among records containing both values, exact agreement was 97.8803% after normalization. Missing values remain a visible “Unknown” category. Labels are normalized for case and whitespace, and the historical code “iw” is mapped to “he”. These are catalog categories rather than independently verified linguistic classifications, so cross-language interpretation remains cautious.

Research snapshots and reproducibility

Each approved run creates a versioned snapshot containing its generation time, schema version, source range, code-defined methodology, aggregate rows, statistical analysis, and SHA-256 data checksum. A new run creates a new document rather than silently rewriting an earlier publication. Given the same source state, code version, and configuration, the pipeline is designed to be repeatable; changing any of those inputs can legitimately change the result.

Known limitations

LyricsWeb coverage varies by era, language, artist, and source availability.
Release metadata can be absent, inconsistent, or inherited from an album.
Eligible records have not yet completed a documented identifier-based duplicate audit.
Unicode token sequences are not equivalent to linguistic words in every writing system.
Language labels are catalog classifications and may contain residual errors.
Current-year figures are incomplete and are labeled as partial.
A large corpus reduces random noise but does not automatically remove selection bias.
Observational patterns do not by themselves establish a cultural or causal explanation.
Normalization can merge typographically different titles or keep meaning-equivalent titles separate.
“First observed” means first within this corpus and is not proof of a title’s first use anywhere.

Plain-language glossary

Corpus: The collection of records included in the analysis.
Confidence interval: A range showing uncertainty under a stated statistical model.
Effect size: How large a difference or trend is, not merely whether a test detects it.
Percentage point: The direct difference between two percentages; 20% to 25% is five percentage points.
Selection bias: A systematic difference between records in the corpus and records not represented there.
Normalized title: A title transformed by documented rules so comparable forms can be counted consistently.
Artist–title association: One observed pairing of an artist identity and a normalized title, consolidated at its earliest eligible year.
Collision probability: The probability that two associations drawn from the same observed distribution share a normalized title.

Statistical references

Explore Are Song Titles Becoming Less Original? · Explore How Song Titles Changed Over Time. · Explore Is Music Becoming More Linguistically Diverse?