How to Be a Well-Prepared Organizer: Studying the Causal Effects of City Events on Human Mobility

Jiyuan Chen, Hongjun Wang, zipei fan, Xuan SongCommunications 2024 · DOI · OpenAlex

Abstract

The analysis of how city events causally affect human mobility is of critical importance. The city government will be thrilled to know how an impending event will influence mobility beforehand, so that they can either decide specifically when and where the event will be held (or not), or be more prepared for some possible circumstances such as crowd collapses and crushes. Previous research on human mobility mainly focuses on simple future prediction based on data correlation, yet the study on the underlying causal effect is woefully inadequate. Motivated by the recent tragedy, the Itaewon Halloween disaster, in this paper we try to explore the causal effects of city events on human mobility using counterfactual prediction. The main technical challenge here lies in capturing and debiasing the time-varying unobservable confounders (e.g., people’s willingness to go outdoors) that affect both the event organization and the number of event participants. Fortunately, the increasing sources of time-varying data offer the possibility to refactor such confounding effects from observation. To this end, we utilize multiple sources of observation data in New York City to construct a neural network-based causal framework, which automatically learns and balances the time-varying unobservable confounders representations and provides estimations for the ITE problem.

Introduction

Urban planners and city governments often struggle to predict how a specific event—a parade, a street market, or a construction project—will actually change human mobility. While we have plenty of models that show correlations (e.g., "it's crowded when there's a festival"), these models fail to answer the "what if" questions essential for safety management.

Researchers have developed a neural network-based causal framework to move beyond simple traffic prediction and into the realm of causal inference. By accounting for unobservable "confounders"—like a population's sudden reluctance to go outside after a crime—the model estimates the "Individual Treatment Effect" (ITE) of city events, providing urban planners with a tool to predict and prevent crowd disasters like the 2022 Itaewon stampede.

What Problem Does This Paper Address?

The core challenge is the presence of unobserved confounders. For example, if a shooting occurs in a neighborhood, people's "willingness to go outdoors" drops. This hidden variable affects both the likelihood of an event being held and the eventual turnout. Without accounting for this, a model might incorrectly conclude that a specific farmers' market is unpopular, when in reality, people were simply staying home due to safety concerns. This paper aims to "debias" these hidden factors to reveal the true causal impact of events on city movement.

The increasing frequency of city events over time.

As shown in the data, the number of city events has grown significantly over the years. Parallel to this growth is a concerning rise in casualties associated with such gatherings, emphasizing the urgent need for proactive safety modeling.

The rising number of deaths and injuries associated with large gatherings.

Methods

The authors propose a causal framework that utilizes counterfactual prediction. Essentially, the model asks: "What would the mobility in this precinct look like with the event vs. without it?"

Key technical components include:

  • Proxy Variables: Since "willingness to go out" isn't a data point you can download, the authors use proxies: Google Trends for safety-related keywords (e.g., "gunshot"), NYC crime records, and weather data.
  • Time-Varying Representation: Confounders change over time. The model uses a Recurrent Neural Network (RNN) to capture how historical trends (like a recent tragedy) influence current behavior.
  • Shared Bottom Architecture: The framework uses a Graph Convolutional Network (GCN) as a "shared bottom" to learn the general features of the city's geography. It then branches into two auxiliary Multi-Layer Perceptrons (MLPs) to predict the two potential outcomes: $Y(1)$ (with event) and $Y(0)$ (without event).
  • Distribution Balancing: To ensure the model doesn't just memorize correlations, they employ a Wasserstein-1 distance constraint. This forces the model to learn representations that are balanced across different regions, reducing bias in the causal estimation.

Datasets and Experimental Setup

The study focuses on New York City in the first half of 2015, integrating several high-resolution sources:

  • Mobility: NYC Taxi & Limousine Commission (TLC) trip records (pick-ups and drop-offs).
  • Events: Data from the Office of Citywide Event Coordination and Management (CECM), categorized into Entertainment, Sales & Market, and Special Events.
  • Proxies: Google Trends (shooting-related keywords), NYC OpenData (crime records), and Dark Sky (weather data).
  • Geography: The city was divided into 77 disjoint regions based on police precincts.

Key Insights and Results

The proposed model significantly outperformed standard machine learning and causal inference baselines in terms of Mean Absolute Error (MAE). On workday pick-up predictions, the framework achieved an MAE of 23.61, compared to 28.50 for TARNet and 184.97 for a standard Causal Forest.

  • Causality > Correlation: Simple prediction models might suggest construction sites cause traffic, whereas a causal model shows that people actively avoid those areas, leading to a negative ITE (decreased mobility).
  • The "Fear Factor": By correlating Google Trends for "gunshots" with actual shooting incidents, the authors demonstrated that public awareness is a measurable variable.
  • Spatial Risk Mapping: The authors calculated a "risk" metric ($Risk = ITE / Area$). They found that certain precincts (like 7, 9, and 2 in Manhattan) have disproportionately high risks during large gatherings, suggesting these areas need more aggressive police presence.

Conclusion and Future Work

While the model shows great promise, the authors acknowledge hurdles such as the unsupervised nature of causal estimation and the reliance on taxi data as a proxy for all mobility. Future work could incorporate subway swipes or bike-share data for a more holistic view. By moving from correlation to causation, this research provides a vital blueprint for safer, better-prepared urban event management.

Summary updated

Abstract

Abstract The analysis of how city events causally affect human mobility is of critical importance. The city government will be thrilled to know how an impending event will influence mobility beforehand, so that they can either decide specifically when and where the event will be held (or not), or be more prepared for some possible circumstances such as crowd collapses and crushes. Previous research on human mobility mainly focuses on simple future prediction based on data correlation, yet the study on the underlying causal effect is woefully inadequate. Motivated by the recent tragedy, the Itaewon Halloween disaster, in this paper we try to explore the causal effects of city events on human mobility using counterfactual prediction. The main technical challenge here lies in capturing and debiasing the time-varying unobservable confounders (e.g., people’s willingness to go outdoors) that affect both the event organization and the number of event participants. Fortunately, the increasing sources of time-varying data offer the possibility to refactor such confounding effects from observation. To this end, we utilize multiple sources of observation data in New York City to construct a neural network-based causal framework, which automatically learns and balances the time-varying unobservable confounders representations and provides estimations for the ITE problem.
BibTeX
@inproceedings{chen2024how,
  author = {Jiyuan Chen and Hongjun Wang and zipei fan and Xuan Song},
  title = {{How to Be a Well-Prepared Organizer: Studying the Causal Effects of City Events on Human Mobility}},
  booktitle = {Communications},
  year = {2024},
  doi = {10.1007/978-3-031-53770-7\_4},
}
Figure