Taylor Alarcon
I use data from the U.S. Census and MappingPoliceViolence.org to estimate the causal effect that local police violence has on mobility modality across neighborhoods in Illinois. I employ propensity score matching to match census tracts on a variety of demographic and economic covariates, assign exposure to fatal police violence as the treatment, and estimate the average treatment effect on the treated (ATT), where the outcome variable is percentage of citizens either driving, walking, taking public transit, etc. to work. My findings insist that exposure to local police violence causes a %2.5 decrease in citizens who drive to work and a %2 increase in citizen population who take public transit to work, building on theory that local violence – police violence in particular – have substantial impacts on collective fear, anxiety, and distrust in exposed populations.
Does exposure to police violence cause changes in urban mobility?
There exists a body of work that looks at the causal relationship between one’s exposure to violence and one’s social behavior. A 2016 study in the Journal of Research in Crime and Delinquency examined the association between witnessing community violence and criminal behavior in a representative sample of young adults, finding that recent exposure to violence in the community increased the risk for young adult criminal activity, while another 2012 study tangentially suggestsed the negative impact that local violence has on psychological distress, attention span, and impulse control. These past studies suggest that exposure to violence in general, let alone violence conducted by those meant to serve and protect, may have some sort of causal relationship with one’s likelihood to engage in different social behaviors.
While past research suggests that exposure to violence has a potential association with anti-social behavior, there is currently no research that looks specifically at how police violence affects how citizens move through their communities, to go to places like work, church, or other spaces. This project seeks to understand inequality based not on where people live, but where and how they move about the city. Building on past research on mobility disadvantage, it examines separation, isolation, and integration for communities exposed vs. not exposed to local acts of fatal police violence.
The two datasets I decided to use for this analysis are MappingPoliceViolence.org and U.S. Census Demographic Data (Kaggle). I am using the data from MPV to collect a list of census tracts in Illinois that have experienced at least one instance of local police violence, i.e., exposure to what I refer to as the treatment. Based on this list of exposed census tracts, I then added a column to the U.S. demographic dataset – where each observation is a census tract and its corresponding demographic and economic information - with a dummy variable, where 1 denotes having been exposed to police violence and 0 denotes no exposure.
This final combined data structure has variables such as gender and racial compositions, income dynamics, poverty and employment rates, and mobility measures for each census tract in the U.S. Because I am only performing this analysis on Illinois, I subset this final dataset to only include Illinois census tracts. There are over 3,000 census tracts in Illinois and, according to MPV, 141 instances of police violence through the year 2017.
My outcome variable(s) are any variables that have to do with mobility (driving, walking, or taking public transit to work). My treatment variable is whether or not that census tract has been exposed to local police violence.My explanatory variables (and confounders) are everything else: gender and racial compositions, income dynamics, poverty and employment rates, and all other pre-treatment variables included in the U.S. Census Demographic Dataset.
I include some data visualizations and descriptive statistics in Data Wrangling and Analysis section of the paper.
Propensity scores are a one-number summary of all your covariates. In ordinary terms, propensity score for an observation is generated by summarizing all the other data on that observation. Propensity score matching is the process of identifying incredibly similar observations in a way that artificially constructs treatment and control groups to estimate the treatment effect, getting as close as possible to a randomized experiment.
Most often, propensity score matching is implemented via a statistical package that can be loaded in software, but it can also be hard-coded by (1) defining the variables for each observation in the dataset that are confounders, (2) denoting which observations have been treated vs. those that haven’t, (3) using logistic regression to predict, for each observation, the propensity to receive the treatment. This prediction is the propensity score: the probability of having received the treatment.
Once these steps are taken, the researcher can match observations that received the treatment with similar observations (i.e., similar propensity scores, or similar summary of their covariates) that have not received the treatment, restructure the dataset to include the matched data only, check to ensure the major assumptions (overlap and balance, which are included below), and estimate – either via difference in means or via regression – the effect of the treatment on the outcome of interest.
A few benefits of propensity score matching include being able to adjust for one variable instead of many, requiring less strict parametric assumptions, reasonably addressable diagnostics, that it can be incorporated into a doubly-robust strategy for causal inference, and that it the researcher can choose the model specification without ever looking at an outcome.
Costs, on the other hand, include the time commitment, the need for a decent estimate of the propensity score, lack of clear guidelines for success, and vulnerability to p-hacking.
Given the nature of the research question, structure of the data, and method, I found it most appropriate to estimate average treatment for the treated (ATT) in this analysis. I am making inferences solely about census tracts that have been exposed to the treatment, i.e., census tracts that have experienced local instances of police violence.
The average treatment effect on the treated (ATT) is defined as the average treatment effect of observations that were defined as treated. In mathematical terms and using the Potential Outcomes Framework, this can be defined as E(Y(1)-Y(0)|Z=1). ATT generally makes the most sense for propensity score matching given that matching fundamentally restricts the dataset to observations that are matches. The researcher can only reasonably make inferences about the treated observations, since non-treated observations who are not similar to treated observations are thrown out from the sample.
The assumptions for propensity score matching include the two standard methods for all causal inference problems – ignorability and SUTVA – as well as sufficient balance and overlap.
Ignorability: the ignorability of the treatment assignment says that if you can’t control for confounders, your statistical model is showing correlation and not causation; if the factors that determine treatment assignment are observable, we can control for them in order to obtain causal effects.
Sufficient overlap: Enough observations with similarities between groups (enough matches).
Appropriate specification of the propensity score model/ balance achieved: Low standard mean differences and other statistics of difference, and enough observations that there is a sufficient number of both treated and control observations (low number of unmatched units).
SUTVA: that response of a particular unit depends only on the treatment to which the observation was assigned (no alternative explanations for affecting the outcome) and does not affect the outcome of any other observations.
The plausibility of the assumptions under the current research design is high. Ignorability is plausible given that we can control for all observable factors that determine treatment assignment, and SUTVA is plausible since census tract data points are fundamentally bounded. Overlap and balance are also plausible given the volume of census tracts in the entire sample, the diversity of tract demographics across the state of Illinois, and overlapping covariate characteristics between observations.