Quantifying Equity and Equity Biases in Transportation and Data

Abstract

Equity is a critical field of study in transportation. The built transportation network does not serve the needs of the population to achieve equal levels of economic vitality and prosperity. Because of these concerns, there has been a recent effort to address these equity issues in the transportation network. This effort coincides with a massive growth in the data available for transportation practitioners, known as big data. The growth of data has led to data-driven decision-making to allow for more effective transportation policies and decisions than was afforded with classical methods. However, as the amount of data available has grown, there is a great concern for understanding the biases within this data. Though significant effort has been spent to mitigate the biases present in transportation datasets, there is little understanding of the equity implications of these biases. Since biases inherently misrepresent certain population segments, there is a possibility that critical populations will be underrepresented in our data sources and therefore our data-driven decision-making, which is directly counter to our goals of increasing equity in transportation research and practice. One of the key pillars in addressing equity in transportation is ensuring that the decision-making process includes rigorous representation of all impacted parties for ongoing transportation projects. Therefore, it is critical that representation also be maintained when using data in data-driven decision processes. How do we identify and quantify representation for transportation datasets that influence decision making? To understand this issue, we can reframe our idea of data biases as equity biases: an equity bias is any bias in a data source that produces a negative equity outcome by underrepresenting critical populations. Here, critical populations can mean any population of interest, including communities of low income, communities with high poverty levels, or historically disadvantaged communities such as black, indigenous, and peoples of color (BIPOC) communities. This definition focuses on the equity outcomes of data biases as opposed to the precipitating causes of bias.With this definition, this dissertation presents a methodological framework to address the key challenges relating to equity biases: how do we identify, quantify, and address equity biases and representation such that transportation datasets best serve in data-driven decision-making? This work addresses these issues by utilizing ecological regression to define the representation of datasets to accurately understand which populations are under- and over-represented in transportation datasets. Ecological regression is well suited for this task as, unlike other regression methods, it can address individual level characteristics (such as income, racial demographics, etc.) in aggregate datasets, allowing us to consider these critical equity demographics when assessing the representation of a transportation dataset. This allows us to define representation for different demographic strata, thus allowing for the comparison of representation between datasets. This framework is tested on datasets of tolling data collected from the five tolling facilities in Washington State.