Bicycle commuting in Melbourne during the 2000s energy crisis: A semiparametric analysis of intraday volumes

Cycling is attracting renewed attention as a mode of transport in western urban environments, yet the determinants of usage are poorly understood. In this paper we investigate some of these using intraday bicycle volumes collected via induction loops located at ten bike paths in the city of Melbourne, Australia, between December 2005 and June 2008. The data are hourly counts at each location, with temporal and spatial disaggregation allowing for the impact of meteorology to be measured accurately for the first time. Moreover, during this period petrol prices varied dramatically and the data also provide a unique opportunity to assess the cross-price elasticity of demand for cycling. Over-dispersed Poisson regression models are used to model volumes at each location and at each hour of the day. Seasonality and the impact of weather conditions are modelled as semiparametric and estimated using recently developed multivariate penalized spline methodology. Unlike previous studies that use aggregate data, the empirical results show a substantial meteorological and seasonal component to usage. They also suggest there was substitution into cycling as a mode of transport in response to increases in petrol prices, particularly during peak commuting periods and by commuters originating in wealthy and inner city neighbourhoods. Last, we extend the approach to a multivariate longitudinal count data model using a Gaussian copula estimated by Bayesian data augmentation. We find first order serial dependence in the hourly volumes and a ‘return trip’ effect in daily bicycle commutes.

► Model for hourly bicycle counts using matched weather and petrol price data.
► Penalized spline regression reveals strong nonlinear weather and seasonal effects.
► Commuters are forward-looking in the morning and backward-looking in the afternoon.
► Positive cross-price elasticity of demand with respect to petrol prices.
► Intraday serial dependence in counts captured using a Bayesian Gaussian copula model.