Bayesian Negative Binomial Regression of Afrobeats Chart Persistence Conference: Make sure to enter the correct conference title from your rights confirmation emai; Sept 01, 2022; Atlanta, GA

Cabansag, Ian Jacob and Ntegeka, Paul · 2026

Read Bayesian Negative Binomial Regression of Afrobeats Chart Persistence Conference: Make sure to enter the correct conference title from your rights confirmation emai; Sept 01, 2022; Atlanta, GA by Cabansag, Ian Jacob and Ntegeka, Paul (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2601-01391.txt.

Introduction

Ian Jacob Cabansag and Paul Ntegeka ask whether teamwork helps an Afrobeats song stay on the charts. Using daily Nigeria Spotify Top 200 data from 2024, they count for each track how many days it charted and how many streams it earned in Nigeria that year. Collaboration status marks whether a track is solo or multi-artist.

A Bayesian negative binomial regression, suited to count data that vary more than simple models allow, weighs chart days against collaboration and total streams. Sampling methods produce the estimates, checked with rate ratios and predictive tests. After streams are accounted for, collaboration tracks tended to spend slightly fewer days on the chart than comparable solo tracks. The finding cuts against the easy assumption that features always extend a song's life. For artists and labels weighing the cost of a guest spot, total reach seems to matter more than the pairing itself.

Catalog record

CREATORCabansag, Ian Jacob and Ntegeka, Paul
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE18 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2601-01391

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Bayesian Negative Binomial Regression of Afrobeats Chart Persistence Conference: Make sure to enter the correct conference title from your rights confirmation emai; Sept 01, 2022; Atlanta, GA Cabansag, Ian Jacob and Ntegeka, Paul Abstract 1. Abstract Afrobeats tracks compete for listener attention on streaming platforms, where chart visibility can translate directly into revenue and cultural impact. This project studies the relationship between collaboration and chart persistence using the 2024 Spotify Top 200 daily charts for Nigeria. Each track is summarized by the number of days it appears on the Top 200 over the year and its total annual streams in the Nigerian market. A Bayesian Negative Binomial regression model is fitted in which the response is the count of days on chart and the predictors are an indicator for collaboration (solo versus multi–artist track) and the log of total annual streams. This model appropriately handles over–dispersed count data and allows a direct interpretation of collaboration effects after controlling for overall popularity. Posterior inference is carried out in PyMC using NUTS, and posterior rate ratios, the posterior probability that collaboration helps, and posterior predictive checks are reported. The results suggest that, after conditioning on total streams, collaboration tracks have slightly fewer days on the chart than comparable solo tracks, with an estimated rate ratio of approximately \exp(\beta_{\text{collab}})\approx 0.93 (95% interval [0.88,0.99] ) and posterior probability P(\beta_{\text{collab}}>0\mid\text{data})\approx 0.007 . 2. Introduction The global rise of Afrobeats has been driven in part by streaming platforms, where editorial playlists and algorithmic recommendations amplify the visibility of hit songs. For artists and labels, one key indicator of success is chart persistence : how long a track remains on a country’s Top 200 chart. Longer persistence implies sustained listener interest and more opportunities for discovery. A common belief among artists and industry observers is that collaboration for example, featuring another artist on a track helps a song stay relevant for longer. However, collaboration is also correlated with other factors such as marketing budgets and artist popularity. Simply comparing the average number of days on chart for solo and collaboration tracks confounds these effects. The objective of this project is to quantify the effect of collaboration on chart persistence in a principled statistical framework. Daily Spotify Top–200 data for the Nigeria market in 2024 are used, focusing on Afrobeats and related genres as reflected in the chart. The central question is: After controlling for overall popularity (total annual streams), do collaboration tracks stay longer on the Nigeria Top 200 chart than solo tracks? From a methodological standpoint, the project moves beyond a simple conjugate Bayesian A/B test and instead employs a Bayesian Generalized Linear Model (GLM) with a Negative Binomial likelihood. This choice is motivated by the count nature and strong over–dispersion of the response variable (days on chart). The model is estimated using Markov Chain Monte Carlo (MCMC) with the NUTS algorithm implemented in PyMC , providing full posterior distributions over parameters and derived quantities. The Data section describes the dataset and preprocessing. The Bayesian model section presents the Negative Binomial regression formulation and prior choices. The Computation section summarizes the computational setup and convergence diagnostics. The Results section reports posterior summaries and predictive checks, and the Discussion section concludes with a substantive interpretation of the findings and limitations. 3. Data 3.1. Source and structure The dataset consists of daily snapshots of the Spotify Top–200 for the Nigeria market over the period 1 January 2024 to 31 December 2024. Each row corresponds to one track on one day and includes, among others, the following variables: • date : calendar date of the chart entry. • uri : unique Spotify track identifier. • rank : chart position (1–200) on that date. • track_name : track title. • artist_names : comma–separated list of artist names. • streams : number of streams for that track in Nigeria on that date. The data were obtained from the official Spotify Charts website for the Nigeria regional daily chart ( 3 ) : https://charts.spotify.com/charts/view/regional-ng-daily/latest The raw dataset contains N=72{,}995 daily track–level observations, representing n_{\text{tracks}}=1{,}335 unique tracks. The date range covers the full year, from 2024–01–01 to 2024–12–31. 3.2. Track–level aggregation The regression analysis is conducted at the track level, not the day level. For each track i (identified by a unique URI), the following summary variables are constructed: • n_{i} : the number of distinct days in 2024 on which track i appears in the Nigeria Top–200. This is referred to as days on chart . • S_{i} : the total number of streams accumulated by track i across all days in which it appears on the chart, S_{i}=\sum_{t}\text{streams}_{it} . • x_{i,\text{collab}} : an indicator for whether track i is a collaboration ( 1 ) or a solo track ( 0 ). The collaboration indicator is derived from the artist_names field using a simple text heuristic. Tracks are classified as collaborations if the artist field contains a comma (multiple artists), an ampersand “&”, or tokens such as “feat”, “ft.” or “ x ” indicating a featured artist. This heuristic agrees well with manual inspection of a random subset of tracks. For numerical stability and interpretability in the regression, the log–transformed total streams, x_{i,\text{logStreams}}=\log S_{i}, are used as a predictor. Table 1 reports summary statistics for the main variables across the n_{\text{tracks}}=1{,}335 tracks. Table 1. Summary statistics for track-level variables (Nigeria Spotify Top 200, 2024). Variable Mean SD Min Max Days on chart n_{i} 54.7 81.9 1.0 365.0 Total streams S_{i} 1,908,473 3,532,364 10,132 29,251,917 \log S_{i} 12.735 2.100 9.223 17.191 Collaboration rate P(x_{i,\text{collab}}=1) 0.456 Exploratory plots (histogram of n_{i} , boxplot of n_{i} by collaboration status, and scatterplot of n_{i} against \log S_{i} ) show that the response variable is strongly right skewed and over dispersed relative to a Poisson distribution, and that both collaboration status and total streams are associated with days on chart. These features motivate the use of a Negative Binomial regression model in the next section. 4. Bayesian Model 4.1. Likelihood: Negative Binomial regression Let n_{i} denote the number of days track i appears on the Top 200 chart in 2024. These are non-negative counts with substantial over-dispersion: \operatorname{Var}(n_{i}) is considerably larger than \mathbb{E}[n_{i}] . The Negative Binomial distribution provides a flexible parametric family for such data. For each track i the model assumes (1) n_{i}\mid\mu_{i},\alpha\sim\text{NegBin}(\mu_{i},\alpha), where \mu_{i}>0 is the mean and \alpha>0 is a dispersion parameter. Under this parameterization, \mathbb{E}[n_{i}\mid\mu_{i},\alpha]=\mu_{i},\qquad\operatorname{Var}(n_{i}\mid\mu_{i},\alpha)=\mu_{i}+\frac{\mu_{i}^{2}}{\alpha}. To relate \mu_{i} to covariates, a log–link function with two predictors is used: collaboration status and log total streams. Let x_{i1}=x_{i,\text{collab}} and x_{i2}=x_{i,\text{logStreams}} . The linear predictor is (2) \eta_{i}=\beta_{0}+\beta_{1}x_{i1}+\beta_{2}x_{i2}, and the mean is linked via (3) \mu_{i}=\exp(\eta_{i}). The coefficient \beta_{1} captures the log–multiplicative effect of being a collaboration track, holding total streams fixed: \exp(\beta_{1})=\frac{\text{Expected days on chart for collab track}}{\text{Expected days on chart for solo track}}\quad\text{at fixed }S_{i}. Similarly, \beta_{2} describes how days on chart scale with overall streams. 4.2. Priors Weakly informative priors are placed on the regression coefficients and dispersion parameter. The coefficients (\beta_{0},\beta_{1},\beta_{2}) receive independent Normal priors with mean zero and standard deviation 2 : (4) \displaystyle\beta_{0} \displaystyle\sim\mathcal{N}(0,2^{2}), (5) \displaystyle\beta_{1} \displaystyle\sim\mathcal{N}(0,2^{2}), (6) \displaystyle\beta_{2} \displaystyle\sim\mathcal{N}(0,2^{2}). These priors allow for substantial variation (for example, a priori \exp(\beta_{j}) lies roughly between e^{-4} and e^{4} with high

Read the complete text