[Paper Review] Time Series Analysis and Correlation of Subway Turnstile Usage and COVID-19 Prevalence in New York City

Sina Fathi-Kazerooni, Roberto Rojas‐Cessa+2 more|arXiv (Cornell University)|Aug 1, 2020

COVID-19 epidemiological studies27 references7 citations

TL;DR

This study establishes a strong correlation between New York City subway turnstile usage and COVID-19 deaths/cases using LSTM and ARIMA models. It demonstrates that subway ridership, adjusted for incubation and symptomatic delays, predicts reported deaths, and estimates the first cases and deaths in NYC occurred earlier than officially reported—between January 28 and February 24, and March 7–21, 2020, respectively.

ABSTRACT

In this paper, we show a strong correlation between turnstile usage data of the New York City subway provided by the Metropolitan Transport Authority of New York City and COVID-19 deaths and cases reported by the New York City Department of Health. The turnstile usage data not only indicate the usage of the city's subway but also people's activity that promoted the large prevalence of COVID-19 city dwellers experienced from March to May of 2020. While this correlation is apparent, no proof has been provided before. Here we demonstrate this correlation through the application of a long short-term memory neural network. We show that the correlation of COVID-19 prevalence and deaths considers the incubation and symptomatic phases on reported deaths. Having established this correlation, we estimate the dates when the number of COVID-19 deaths and cases would approach zero after the reported number of deaths were decreasing by using the Auto-Regressive Integrated Moving Average model. We also estimate the dates when the first cases and deaths occurred by back-tracing the data sets and compare them to the reported dates.

Motivation & Objective

• To investigate the correlation between NYC subway turnstile usage and the prevalence of COVID-19 cases and deaths during the early pandemic.
• To address the limitations of reported case data due to testing shortages by using death counts as a more reliable metric for analysis.
• To estimate the true onset of the pandemic in NYC by back-tracing data using ARIMA models to identify the first possible cases and deaths.
• To forecast the dates when COVID-19 deaths and cases would reach zero based on reported trends.
• To validate model predictions against official reporting dates and assess discrepancies due to data limitations and reporting delays.

Proposed method

• Applied long short-term memory (LSTM) neural networks to model time-series relationships between subway turnstile entries and reported COVID-19 deaths and cases.
• Used day-shifted subway entry data (1 to 25 days prior) as features to account for incubation and symptomatic periods before death or case reporting.
• Employed Auto-Regressive Integrated Moving Average (ARIMA) models to forecast future trends in deaths and cases, including zero-crossing dates.
• Conducted reverse-time ARIMA forecasting to estimate the origin of the pandemic by projecting backward from the first reported cases and deaths.
• Used 20% of the NYC DOH dataset as a test set to evaluate model performance and reliability.
• Compared model-estimated dates for first cases and deaths with officially reported dates to assess accuracy and data limitations.

Experimental results

Research questions

• RQ1 Is there a statistically significant correlation between NYC subway turnstile usage and the number of reported COVID-19 deaths and cases?
• RQ2 How do incubation and symptomatic delays affect the predictive power of subway ridership data for tracking pandemic progression?
• RQ3 When did the first SARS-CoV-2 cases and deaths likely occur in NYC, based on back-tracing the ARIMA model of reported data?
• RQ4 By what date are the number of daily COVID-19 deaths and cases expected to reach zero, according to ARIMA forecasts?
• RQ5 How do the model-estimated dates for the first cases and deaths compare to the officially reported dates, and what explains the discrepancies?

Key findings

• A strong correlation was established between NYC subway turnstile usage and reported COVID-19 deaths and cases using LSTM models, with lagged entries (1–25 days prior) serving as key predictors.
• The ARIMA model forecasted that the number of daily deaths would reach zero between May 16 and July 28, 2020, with the lower confidence bound indicating May 16 as the zero-deaths date.
• The model estimated that the first COVID-19 case in NYC likely occurred between January 28 and February 24, 2020, earlier than the officially reported date of February 29, 2020.
• The first death was estimated to have occurred between March 7 and March 21, 2020, earlier than the official report of March 11, 2020.
• The ARIMA forecast for cases projected a zero-case trend by June 26, 2020, based on the upper confidence interval, though the upper band did not cross zero, suggesting possible resurgence.
• Discrepancies between estimated and reported dates are attributed to early pandemic data limitations, including low testing availability and reporting delays.