Measuring the featured-artist effect with proper statistical controls — and catching two of my own overclaims along the way. An independent reconstruction and extension of a COGS 108 (UC San Diego) team project.
Featured tracks look more popular than solo tracks on Spotify (p ≈ 1e-26). This project shows the gap is real but tiny, and that almost all of it comes from genre — not from collaboration itself.
Do tracks featuring at least one other artist have higher Spotify popularity than solo tracks, controlling for duration, explicitness, and the audio features energy and danceability — and, in the fuller specification, valence, loudness, and genre?
A "featured" track lists another artist in its credits or title (e.g. "feat. Name"); a "solo" track does not. These aren't rows from the analysis sample — they're familiar catalog examples to make the definition concrete.
Spotify Tracks Dataset (maharshipandya, via Kaggle/Hugging Face) — 114,000 track–genre rows with Spotify's popularity score, audio features, and track_genre. Deduplicated deterministically by track_id.
Drop duplicates by track_id — deterministic, reproducible.
Convert duration_ms → duration_sec; drop tracks with duration = 0.
Cast the explicit flag to a clean 0/1 integer.
has_feature = 1 if multiple artists are credited (;-separated) or the title contains "feat.", "ft.", "featuring", or "with" as a whole word.
The original team's processed files were lost after the course ended. This dataset was re-identified by matching summary statistics from the final presentation: 89,740 unique tracks; explicit split 82,036/7,704; group means 32.81/34.26. This rebuild reproduces every one of those numbers to within one row.
A large share of tracks sit at exactly 0 popularity — low or no recent engagement. That shapes every test downstream: a rank-based test instead of a t-test, and a dedicated robustness check that excludes the zero spike.
Mann–Whitney U test (popularity is non-normal), plus rank-biserial r and Cohen's d, which the original presentation never reported.
The project proposal planned an OLS regression with HC3 robust standard errors, a 95% CI, a p-value, and a standardized effect size — with genre among the controls. The final presentation stopped at the bivariate comparison above. This section completes it.
| Specification | Coefficient | 95% CI | p | R² |
|---|---|---|---|---|
| Raw difference in means | +1.41 | [+1.10, +1.72] | ≈1e-26 (M–W) | — |
| Spec 1: duration + explicit + energy + danceability | +1.08 | [+0.76, +1.40] | 4e-11 | 0.007 |
| Spec 2: + valence + loudness + genre fixed effects | +0.28 | [−0.01, +0.57] | 0.062 | 0.329 |
Variance inflation factors across all Spec 2 predictors stay below 2.6 — well under the conventional threshold of 5. Multicollinearity is not driving the shrinkage.
Genre fixed effects take R² from ~0.01 to ~0.33 and cut the featured-artist coefficient by roughly three quarters. An exact shift-share decomposition splits the raw gap into genre composition and the within-genre gap — the two terms sum to the raw gap with no residual, by construction.
The first version of this decomposition compared each group's genre exposure to a blended genre-level mean, mixing featured and solo tracks together. That's a natural thing to write — but it's not algebraically correct: the two components didn't sum back to the raw gap, landing about 0.4 points short, because the blended mean already absorbs part of the within-genre effect it's supposed to be separate from. A regression test written for this rebuild (test_genre_decomposition_sums_to_raw_gap) catches exactly this. The fix, a proper shift-share decomposition, moves genre composition's share of the gap from an approximate 73% to an exact 99%. The within-genre conclusion — effectively zero — is unchanged either way.
Alternate detection rules, excluding the popularity = 0 spike, and excluding remix/live/acoustic/remaster versions — the adjusted effect stays between roughly zero and half a point every time.
The original computation was correct. The conclusion drawn from it was stronger than the evidence supported.
“Featured artists meaningfully boost visibility and popularity on Spotify… collaborations strategically elevate track success.”
Based on p ≈ 1e-29 alone — without ever reporting an effect size.
The raw gap is real but tiny (+1.4 points, d ≈ 0.07). 99% of it is genre composition. Controlling for genre and audio features, the adjusted effect is +0.28 points — between roughly zero and half a point across every robustness check.
At n ≈ 90,000, statistical significance is nearly guaranteed for any nonzero difference. The analytical question was always effect size.
Different outcome (Spotify's recency-weighted popularity score, not stream counts); different sample (a broad catalog snapshot, not new chart releases); and unobserved artist fame — which biases this estimate upward if anything, since famous artists both collaborate more and are more popular. That the adjusted effect is still near zero makes the null-ish result more credible, not less.
Analysis built on the Spotify Tracks Dataset (maharshipandya, via Kaggle). Popularity scores, audio features, and the Spotify name/logo are the property of Spotify AB, referenced here for academic, non-commercial commentary on publicly available data.
Jose Enrique Siono Gutierrez, Zulema Zermeno, Sahar Zahir, Sonali Singh, Teresa Jia.
This site is an independent reconstruction and extension by Sonali Singh. Original team roles: analysis, background research, visualization, writing (original draft).
Cover art is used solely to illustrate the has_feature definition and belongs to the respective label/artist. Not used commercially, not modified beyond thumbnail resizing.