Air quality impacts of new public transport provision: A causal analysis of the Jubilee Line Extension in London
Introduction
In 2016, 91% of the worldwide population lived in places where the air pollution level exceeded the World Health Organisation standards, and ambient air pollution is estimated to cause 4.2 million premature deaths annually (World Health Organization, 2018). As one of the main sources of air pollutant emissions, the transport sector has implemented various interventions aimed at mitigating air pollution, such as stricter vehicle emissions standards. Public transport is generally considered to be more sustainable than private road transport, with potential benefits including improving air quality (Slovic et al., 2016). An increase in public transport supply can cause complex reactions in various sectors in a city, consequently affecting activities and air pollutant emissions in a number of ways. Gonzalez-Navarro and Turner (2018) investigated the effects of a subway network extension on city population, transport mode substitution, and city configuration with data from 632 of the largest cities in the world. They found that subway expansion has little impact on population growth but can cause cities to decentralise, though the effect can be 10 times smaller than that resulting from highways. Their analysis also suggested that a 10% increase in subway extent causes a 6% increase in subway ridership, mostly due to mode substitution from other transport modes towards the subway, however with little effect on bus ridership. With a case study of the Regional Express Rail extension in the Paris metropolitan area, Mayer and Trevien (2017) found that the extension of urban rail transport had no impacts on population, but increased employment and firm location, and attracted highly-skilled households. Beaudoin et al. (2015) summarised the possible pathways by which public transport supply can affect air quality based on the traditional ‘four-step model’ in transport demand modelling. They highlighted the differences in mechanisms between the short-run equilibrium (such as trip redistribution) and the long-run equilibrium (such as changes in land use and car ownership). For ex-post assessment and to inform future investments, it is essential to understand whether and to what extent public transport provision has improved air quality in the past.
While statistical association describes a non-directed relationship between two variables, causation seeks to quantify the net influence of a treatment (e.g. public transport) on an outcome of interest (e.g. air quality) through all possible pathways directing from the treatment to the outcome (Altman and Krzywinski, 2015; Pearl, 2010, 2018). The term ‘treatment’ and ‘intervention’ are interchangeably used in this paper. The identification of a causal relationship can be biased by confounders and selection bias. A confounder is defined to be a common cause of both the putative cause and its outcome (Wunsch, 2007). Weather conditions and seasonality effects are considered to be important confounders when evaluating the causal effects on air quality time series and are generally required to be controlled (Hausman and Rapson, 2018). Selection bias arises when the treatment assignment is non-randomised (Bareinboim and Pearl, 2012). As a public transport intervention is generally a non-randomised and mostly non-experimental process, the main challenge is to identify and quantify the causal relationship between a public transport intervention and air quality in the presence of confounders and selection bias with non-experimental data.
Causal inference methods have been used to quantify the causal relationship between public transport intervention and air quality in previous studies, such as regression discontinuity design (RDD), difference-in-differences (DID), and instrumental variable (IV) regression. RDD assumes that the treatment assignment is determined by the value of a forcing variable being on either side of a threshold (Imbens and Wooldridge, 2009). DID involves two groups (treatment group and control group) and two periods (before and after the intervention), assuming only the treatment group in the after-intervention period is exposed to the treatment (Imbens and Wooldridge, 2009). IVs can be used to control for unobservable confounders. A valid IV needs to be highly correlated with the treatment yet only affect the outcome through its influence on the treatment (Wunsch, 2007). The choice among different causal inference methods needs to consider whether the problem setting satisfies the corresponding method assumptions.
By applying the RDD approach, Chen and Whalley (2012) found that a new urban transit system in Taipei reduced CO concentrations by 5–15% but had insignificant effects on NOx and O3 concentrations. Goel and Gupta (2015) found that the CO concentrations at a major traffic intersection were decreased by 34% by the opening of the Delhi metro system. Gendron-Carrier et al. (2018) focused on the impact of the subway opening in 39 cities around the world on aerosol optical depth. They found that aerosol optical depth in a 10 km radius disk surrounding a city centre decreased by 4% during 18 months after the subway opening.
Based on the DID method, Rivers et al. (2017) evaluated the effects of worker strike actions in the public transit system in 18 Canadian cities on air quality during 1974–2011. Their findings indicated a 3.5 ppb increase in NOx and no significant effects on CO or PM2.5. Bel and Holst (2018) found that a new bus rapid transit in Mexico City significantly reduced air pollutant concentrations (CO, NOx, and PM10) by 5–9%.
With IV regression models, Lalive et al. (2013) concluded that the railway service frequency growth between 1994 and 2004 reduced CO, NO, NO2 pollution, and infant mortality in Germany. Beaudoin and Lin Lawell (2016) evaluated the causal impact of public transit capacity on air quality (CO, Pb, NO2, O3, PM, and SO2) in 96 urban areas across the United States during 1991–2001. Elasticity estimates suggested that a 10% increase in transit supply resulted in a 2.29% increase in NO2 concentrations and a 2.87% increase in PM10 concentrations. According to their explanation, the deterioration in air quality after improving public transport supply can due to a relatively low cross-elasticity between road transport demand and public transport demand or due to induced demand in road transport following the increase in public transport supply.
Previous studies mainly focussed on the city-level or country-level evaluation. In the aforementioned studies, data from different monitoring sites were either averaged to represent a higher spatial level air quality condition or treated as panel data to estimate the treatment effect at a higher spatial level, except for Goel and Gupta (2015). Goel and Gupta (2015) evaluated the effects at different monitoring sites with separate models yet only two monitoring sites were considered, which cannot provide a full understanding of the spatial heterogeneity of treatment effects. Furthermore, different types of air quality monitoring sites (roadside/background) have not been distinguished in previous studies. While roadside concentrations are mainly affected by transport emissions, background concentrations can be affected by various sectors, such as domestic, transport, industry, and energy supply emissions sources. Public transport provision could therefore cause different effects on roadside and background concentrations. All but one aforementioned studies include weather and/or seasonality controls, but commonly by including variables in the causal inference model and assuming a parametric relationship. Due to complex atmospheric processes, collinearity and interactions can exist among meteorological variables and the air quality level, and the relationships among these variables can be non-linear (Grange and Carslaw, 2019). Therefore, representing the relationship between air quality and meteorological variables with parametric statistical models can be very challenging.
The reported causal effects of public transport provision on air quality in different cities are inconsistent, both in magnitude and direction (air pollution increase/decrease) of effect. This inconsistency is likely due to the interactions existing in the pathways from public transport provision to air quality, and to different choices of methodological approaches and estimation methods. The interactions in the impact pathways are dependent on various factors, including demographics and economic factors, transport network structures, and land use distributions (Beaudoin et al., 2015).
In this paper, we evaluate the short-run and long-run causal impacts on air quality of a major public transport intervention in London, the Jubilee Line Extension (JLE), with a sharp RDD model that captures the spatial heterogeneity of impacts for different air pollutants at different types of monitoring sites (roadside and background). The meteorological normalisation technique is applied with a non-parametric model, gradient boosting decision trees (GBDT), to control the weather and seasonality effects. Change point detection (CPD) is conducted to find mean-shifts in normalised concentration trends. Both the CPD and the meteorological normalisation are used to justify the assumptions of a valid RDD and to support the RDD model specification. The JLE in 1999 was a major addition to the London Underground network, aimed at increasing accessibility and supporting development along the route (Omega Centre of University College London, 2014). As a major investment and improvement in public transport provision, it is necessary to understand the effects caused by the JLE on air quality to inform future investments.
Section snippets
Materials and methods
To quantify the causal impacts of the JLE on air quality, a sharp RDD model is specified at individual air pollution monitoring sites, including roadside and background locations. Meteorological normalisation and CPD are conducted to support and justify the sharp RDD model before model specification. Five main air pollutants are considered: NO2, NOx, O3, CO, and PM10. NO2, O3, CO, and PM10 are regulated pollutants under the Directive 2008/50/EC within the European Union and the European
Results and discussion
In this section, we discuss the JLE's causal effects on air quality with respect to NOx, NO2, and O3 for the first phase and the latter phases of opening. The PM10 and CO concentrations generally had fewer significant identified responses and a discussion of these pollutants is included in the SI §S6. The robustness of causal effect estimates to the trend function specification is discussed in the SI §S8.
Conclusions
Our analysis shows that effect of the JLE on air pollution in some areas was greater than others, however across London the effect was small. Our results show that the JLE led to only small changes in air pollution at some specific locations; 63%, 43% and 29% of air pollution monitoring sites showed a response to the JLE in NOx, NO2 and O3 concentrations, respectively. For those sites that showed a response, the short-run changes were less than 3% in magnitude; −2% to +1% for NO2, -1% to 0% for
CRediT authorship contribution statement
Liang Ma: Conceptualization, Methodology, Software, Investigation, Writing - original draft, Visualization. Daniel J. Graham: Conceptualization, Methodology, Writing - review & editing, Supervision. Marc E.J. Stettler: Conceptualization, Methodology, Writing - review & editing, Supervision.
Declaration of competing interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgements
L.M. is funded by the Dixon and Skempton Scholarships from the Department of Civil and Environmental Engineering, Imperial College London.
References (54)
- et al.
Public transit investment and sustainable transportation: a review of studies of transit's impact on traffic congestion and air quality
Res. Transport. Econ.
(2015) - et al.
Evaluation of the impact of bus rapid transit on air pollution in Mexico city
Transport Pol.
(2018) - et al.
Detecting and characterising small changes in urban nitrogen dioxide concentrations
Atmos. Environ.
(2007) Stochastic gradient boosting
Comput. Stat. Data Anal.
(2002)- et al.
Subways and urban growth: evidence from earth
J. Urban Econ.
(2018) - et al.
Using meteorological normalisation to detect interventions in air quality time series
Sci. Total Environ.
(2019) Randomized experiments from non-random selection in U.S. House elections
J. Econom.
(2008)- et al.
The impact of urban public transportation evidence from the Paris region
J. Urban Econ.
(2017) Consistent cross-validatory model-selection for dependent data: hv-block cross-validation
J. Econom.
(2000)- et al.
Selective review of offline change point detection methods
Signal Process.
(2020)