Applications of causal inference in environmental policy and transport studies
File(s)
Author(s)
Yang, Xiuleng
Type
Thesis
Abstract
Rapid developments in modern society have led to increased availability of observational data and an increased use of statistical analysis of this data in decision making. Traditionally, determining if a particular intervention or course of action has a causal effect on a particular outcome involves designing a trial to ensure that the results are not influenced by other factors, such as, potential confounding (either measured or unmeasured), or selection bias. This thesis introduces statistical techniques which allow us to account for some of these issues through causal inference methodologies for observational data. The focus of the applications of the techniques that we have developed are in transport studies, including those related to environmental policy.
One of the major observational data sets that we will be analysing in this thesis concerns physical activity through sustainable transport approaches. There is published evidence that people are more likely to be both less active and more stressed than several years ago, which may cause major health problems. One approach to reduce the risk of such health problems is to increase people’s physical activities, that is, to minimise people’s sedentary behaviour. A convenient and economically affordable way to achieve this is through the promotion of active travel, i.e. encourage people to substitute walking and cycling instead of motoring for essential travel, which demands less motivation and generates less financial costs compared to other physical activities (e.g. jogging in leisure time, and training in the gym). This has the potential to not only enhance people’s active mobility but also protect the environment given the link between most motorised transport and air pollution.
To understand whether active travel (i.e. walking and cycling) could reduce stress levels, we apply propensity score matching methods to analyse the causal relationships between different travel modes and stress. This method is useful in this context because it can provide balance across measured covariates and therefore the adapted data we use can be treated as approximately unbiased and as if it arose from a randomised controlled trial. In addition to investigating the causal effects of travel modes on stress, we will also investigate whether specific features of the surrounding environment (i.e. air pollution as a potential proxy of traffic volume and noise) could have a causal impact on people’s stress. This could lead to additional evidence to promote active travel as a means to improve the quality of people’s daily lives and also provide environmental benefits. Measures to reduce traffic flows through neighbourhoods have been expanded relatively rapidly during post-COVID times. We examine the effect of one such policy namely “low traffic neighbourhood” on NO2 and traffic using a difference-in-differences approach, which allows us to determine these measures are effective in terms of prevent ing increased air pollution and traffic.
In causal inference, an important outstanding issue is how to identify the existence of unmeasured confounding. Identifying if unobserved characteristics of confounders are present is important in longitudinal observational studies as it allows us to conduct more reliable assessment of conclusions drawn from the inference process. Recent work on mixed models provide some promising results for reducing the bias induced by unobserved confounders. Therefore, we plan to extend this work to the intervention analysis in the case where there is both observed and unobserved confounding, with developments of causal inference methodology that could potentially identify particular types of unmeasured confounding for observational studies. To illustrate our methodology, we collated a longitudinal dataset linking transport and the economy. It is believed that efficient design and management of large transportation networks in cities can have a major impact on the economy, and in order to maximise the economic returns of future schemes, it is useful to understand the causal effects of past interventions such as road expansions, especially in different localities due to their various economic conditions. We apply our developed methodology to identify and adjust for potential unmeasured confounding from unobserved characteristics of local authority classifications that contain many features that we expect to influence assignment of roads and productivity.
One of the major observational data sets that we will be analysing in this thesis concerns physical activity through sustainable transport approaches. There is published evidence that people are more likely to be both less active and more stressed than several years ago, which may cause major health problems. One approach to reduce the risk of such health problems is to increase people’s physical activities, that is, to minimise people’s sedentary behaviour. A convenient and economically affordable way to achieve this is through the promotion of active travel, i.e. encourage people to substitute walking and cycling instead of motoring for essential travel, which demands less motivation and generates less financial costs compared to other physical activities (e.g. jogging in leisure time, and training in the gym). This has the potential to not only enhance people’s active mobility but also protect the environment given the link between most motorised transport and air pollution.
To understand whether active travel (i.e. walking and cycling) could reduce stress levels, we apply propensity score matching methods to analyse the causal relationships between different travel modes and stress. This method is useful in this context because it can provide balance across measured covariates and therefore the adapted data we use can be treated as approximately unbiased and as if it arose from a randomised controlled trial. In addition to investigating the causal effects of travel modes on stress, we will also investigate whether specific features of the surrounding environment (i.e. air pollution as a potential proxy of traffic volume and noise) could have a causal impact on people’s stress. This could lead to additional evidence to promote active travel as a means to improve the quality of people’s daily lives and also provide environmental benefits. Measures to reduce traffic flows through neighbourhoods have been expanded relatively rapidly during post-COVID times. We examine the effect of one such policy namely “low traffic neighbourhood” on NO2 and traffic using a difference-in-differences approach, which allows us to determine these measures are effective in terms of prevent ing increased air pollution and traffic.
In causal inference, an important outstanding issue is how to identify the existence of unmeasured confounding. Identifying if unobserved characteristics of confounders are present is important in longitudinal observational studies as it allows us to conduct more reliable assessment of conclusions drawn from the inference process. Recent work on mixed models provide some promising results for reducing the bias induced by unobserved confounders. Therefore, we plan to extend this work to the intervention analysis in the case where there is both observed and unobserved confounding, with developments of causal inference methodology that could potentially identify particular types of unmeasured confounding for observational studies. To illustrate our methodology, we collated a longitudinal dataset linking transport and the economy. It is believed that efficient design and management of large transportation networks in cities can have a major impact on the economy, and in order to maximise the economic returns of future schemes, it is useful to understand the causal effects of past interventions such as road expansions, especially in different localities due to their various economic conditions. We apply our developed methodology to identify and adjust for potential unmeasured confounding from unobserved characteristics of local authority classifications that contain many features that we expect to influence assignment of roads and productivity.
Version
Open Access
Date Issued
2022-09-26
Date Awarded
2023-01-01
Copyright Statement
Attribution-NonCommercial 4.0 International Licence (CC BY-NC)
License URL
Advisor
McCoy, Emma
Publisher Department
Mathematics
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)