TL;DR
Large cities are more culturally diverse, but why has been debated: is it due to diverse people living there (demographic mixing) or people exploring more (cultural breadth)? This study uses 250 million music listening events from 2.5 million users to show that both factors play a role, but urban environments specifically foster individual cultural exploration, not just by bringing diverse people together.
What Problem Does This Paper Address?
The paper tackles a long-standing question in urban studies and cultural evolution: what drives the observed increase in cultural diversity in larger cities? Two main hypotheses exist:
- Demographic Mixing: Larger cities attract a wider variety of people from different backgrounds, leading to a more diverse cultural landscape simply by aggregation.
- Cultural Breadth: Living in larger, more stimulating urban environments exposes individuals to more diverse cultural inputs, broadening their personal tastes and engagement.
Previous research struggled to disentangle these hypotheses due to a lack of granular, individual-level behavioral data. This study aims to bridge that gap by analyzing large-scale music listening data to understand how both population-level diversity and individual engagement patterns are shaped by urban environments.
Methods
The study employs a multi-faceted approach:
- Between-Individual Diversity (BID): This metric quantifies how similar or dissimilar the music listening patterns are among individuals within a given geographical area. It uses Hill's number (specifically, the Shannon entropy for order 1) to measure the diversity of shared musical repertoires.
- Within-Individual Diversity (WID): This metric measures an individual's own breadth of music engagement by analyzing their listening history. It uses a "Generalist-Specialist Score" (GS-Score) derived from high-dimensional vector representations of songs (song embeddings). A larger "coverage" or radius in this vector space indicates broader musical taste.
- Song and User Embeddings: Deezer's platform data is used to create high-dimensional vector representations (embeddings) for songs, capturing relationships based on co-listening patterns. These song embeddings are then used to create user embeddings, representing each user's overall music taste.
- Causal Inference (Directed Acyclic Graph - DAG): To move beyond correlation and explore causal relationships, the researchers construct a DAG. This model explicitly defines hypothesized causal pathways between demographic factors, population size, and the diversity metrics (BID and WID). Inverse Propensity Weighting (IPW) is used to control for confounding demographic variables.
- Control Analyses: The study performs numerous robustness checks, including analyzing diversity at the artist and genre levels, accounting for algorithmic recommendations, using alternative population size metrics, and testing for spatial autocorrelation.
Datasets
- Deezer Music Listening Data: The primary dataset consists of approximately 250 million music listening events from 2.5 million users across France, Brazil, and Germany, collected over a four-week period in March 2023. This data includes information on song playback, listening duration, user demographics (age, gender), and stream type (organic vs. algorithmic).
- Geographical Data: NUTS3 unit areas (for France and Germany) and municipalities (for Brazil) are used as geographical units for analysis.
- Socio-Demographic Data: Data on median income, proportion of immigrants, and proportion of university degree holders were sourced from Eurostat and INSEE for France.
- Music Venue Data: The locations of 6618 music venues across France were collected from the SongKick database.
- Social Connection Data: International social connections were approximated using Facebook's Social Connectedness Index (SCI).
- Census Data: Population size, median age, and gender data from census sources were used for comparison and validation.
Key Insights and Ideas
- Urban Environments Foster Individual Exploration: The study provides strong evidence that larger cities don't just aggregate diverse people; they actively encourage individuals to explore a wider range of music. This supports the "cultural breadth" hypothesis.
- Age is a Key Driver of Individual Diversity: An inverted U-shaped pattern emerges, showing that individuals' music exploration broadens through their teens and early twenties, peaks in their late twenties, and then narrows with age. This is a robust cross-cultural finding.
- Demographic Mixing is Important, But Not the Whole Story: While demographic factors (like immigration and education) explain a significant portion of diversity, especially in large cities, they don't fully account for the observed trends. This suggests other urban-specific mechanisms are at play.
- A Saturation Point for Individual Exploration: Beyond a certain population threshold (around 500,000 inhabitants), the increase in individual cultural breadth appears to plateau, suggesting that sheer size alone might not indefinitely drive greater personal exploration.
- Country-Specific Repertoire Differences: While individual breadth is similar across countries, the shared musical repertoires (BID) show notable country-specific variations, indicating different national patterns of cultural convergence.
Key Results
- Population Size and Diversity: Both between-individual diversity (BID) and within-individual diversity (WID) show a significant positive correlation with population size across France, Brazil, and Germany.
- Age and WID: The inverted U-shaped relationship between age and WID is consistent across all three countries, with a peak in diversity around the late 20s.
- Demographic Control: In large urban areas, demographic variables explain a substantial portion of diversity (up to 84.1% for WID and 80.1% for BID). However, after controlling for these factors, population size still has a significant direct effect on BID.
- WID Plateau: The effect of population size on WID levels off in larger cities, suggesting a saturation point for individual cultural exploration.
- BID vs. WID Differences: While demographic factors strongly influence both BID and WID in large cities, their impact is more pronounced on WID, indicating that urban environments are particularly effective at broadening individual tastes.
Limitations and Future Work
- Music-Centric Study: The findings are limited to music consumption. Future work could extend this analysis to other cultural domains like language, literature, art, and societal values to test the generalizability of these mechanisms.
- Temporal Window: The study uses a narrow 4-week window. While chosen to avoid holiday travel biases, it might miss longer-term seasonal patterns. Future research could explore longitudinal data.
- Data Granularity and Proxies: While extensive, some demographic data (like income and immigration) are approximated at the municipality level, which might not fully capture individual nuances. The study also relies on self-reported age and gender.
- Platform Bias: The Deezer user base might not perfectly represent the general population, although demographic comparisons suggest reasonable alignment.
- Causal Inference Assumptions: The causal inference relies on assumptions about relationships between variables, which are laid out transparently via a DAG. Future work could further test these assumptions.
- Cohort Effects: The WID analysis reflects both age and birth cohort. Disentangling these would require longitudinal data tracking individuals over many years.
- Vertical Transmission: The study cannot capture vertical transmission mechanisms (e.g., parent-child cultural inheritance), which are important for cultural evolution. Future research could investigate these through longitudinal studies.
Summary updated