Does Urban Mobility Have a Daily Routine? Learning from the Aggregate Data of Mobile Networks

Abstract

Abstract Does the distribution of Rome's population follow routine hourly, daily, or weekly patterns? And if it does, how do such patterns vary in different parts of the city? This paper reports on our investigation of the aggregate patterns of urban mobility in Rome, Italy for which we used novel data from a mobile phone operator. Unlike research that chartered urban mobility through individual travel surveys, our research determined the aggregate distribution of Rome's population over time by using the volume of call activity in mobile network cells as the unit of spatial analysis. In this paper, we first illustrate and confirm that there is significant regularity in urban mobility at different hours, days, and weeks. We then show how mobility between network cells differs at various times, and we account for the differences by using demographic, economic, and (built) environment indicators. Acknowledgments The research for this article was made possible by the generous support of the government of Portugal through the Portuguese Foundation for International Cooperation in Science, Technology, and Higher Education and was undertaken as part of the MIT-Portugal Program. The research was also generously supported by the MIT SENSEable City Consortium members. Notes The U.S. Department of Transportation estimates that only 17.7 percent of all trips are now work related. (U.S. Department of Transportation) A similar dataset has been previously used in Ratti et al. and Reades et al. We use the terms call-volume distribution and activity distribution interchangeably in the following. We made numerous attempts to obtain these indicators from TIM for the present research, but were unsuccessful due to the sensitivity and proprietary nature of the data. The distribution of raw Erlang shares were highly skewed towards lower values: there were many cells with low shares, and very few cells with extremely high shares. The natural log transformation helped normalize the distribution. The counting system for weeks follows the actual number of the week in the year. At level-1 we express the share of call volume of a cell at a particular time as the sum of the cell's intercept (π0i), and a random error (ϵij) associated with the ith cell in time jth measurement period: At level-2, we express the cell level intercepts as the sum of the city-wide mean (γ00) and a series of random deviations (ζ0j) that differentiate each cell's own mean from the city-wide mean: However, a detailed study of interaction effects in future research would allow one to explore whether the effect of a specific hour (e.g., 15—16pm) differs in an average cell on different weekdays (e.g., Monday versus Tuesday) and different weeks. M4 thus contains fixed effects for hours, weekdays and weeks, as well as random effects for hours and weekdays. The level-1 structure of M4 is given as: At level-2, we include random effects for hourly and daily time dummies, as well as the intercept, thus allowing the effects of these predictors to vary across cells. (We could not add random effect estimates for weekly dummies, since such a fully controlled model would result in only one estimate per case and no statistical comparison.) Since the parameter estimates of the fixed effects refer to a “city-wide average cell,” we should interpret the fixed effect coefficients for hours, weekdays, and weeks with caution. As we had already normalized the outcome by the present overall total call volume in the city (see Equation 1), then the fixed effects simply register additional noise in the data that occurs due to problems with our initial assumptions. If the total population in the city were constant at all measurement periods, and the likelihood and duration of phone calls uniform over time, then we would expect to find zero and insignificant fixed effects. This is because the city-wide average share of call volume would always remain constant. The fact that the hourly fixed effects are significant and non-zero in M4 signals either that the total population of Rome, or the likelihood and duration of calls, fluctuates from hour to hour. Capturing these fluctuations in the fixed effects thus offers a second degree of correction to the random effect estimates. Slight variation is observed on Wednesdays and weekends. In the United States, non-work commuting now constitutes the majority of trips. Unfortunately, we have no travel survey data on Rome at the present time, but suspect that the emergence of irregular trip patterns is also characteristic of Italian cities, as it is of U.S. cities. The historical center is an area of about 6 km2, which constitutes less than 1 percent of the municipal territory, but hosts 13 percent of the total workers, but only 2 percent of the total night-time population.