Due to Bayesian Network's ability to identify causation, we believe it is an effective tool to itially determine whether Uber & Lyft can directly or indirectly cause parking violations in NYC.

The key step for building a Bayesian Network is structure learning, which estimates a directed acyclic graph (DAG) that captures the dependencies between the variables given a set of data records. In this project, score-based structure learning and constraint-based structure learning algorithms were are used to learn the causal structure

Score-based Structure Learning construes model selection as an optimization task. It has two building blocks: first apply a 'scoring function' that maps models to a numerical score, based on how well they fit to a given data set; second perform 'search strategy' to traverse the search space of possible models and select a model with optimal score.

Constraint-based Structure Learning attempts to capture the directionality of causal relationships correctly. It has two building blocks: identifying independencies in the data set using hypothesis tests and constructing a DAG according to identified independencies.

Fixed effects model (FEM) is widely used to control for the effect of unobserved variables when performing causal inference with panel data, so it fits our project well. To develop a FEM, dummy variables of each taxi zone is added to a standard OLS model to evaluate the fixed effects of the number of Uber & Lyft trips across different taxi cab zones. Considering the model for taxi zones i = 1, ...... , N which is observed at certain time periods t = 1, ...... , T:

where Yit and Xit are, respectively, the number of parking tickets (dependent variable) and the number of Uber & Lyft trips (independent variables) in the i-th taxi zone at time t; Zi are unobserved, time-invariant dummy variables in different taxi zones; β and γ represent coefficients of the independent variables and dummy variables; α and uit refers to the intercept and the error term. The coefficient, β , of the independent variables would indicate the causal effect exerted by Uber & Lyft trips on parking violations.

Within the framework of natural science, two identical NYCs, one having TNC companies operating the other not, were needed to strictly prove the causal effect of Uber & Lyft on parking violations. However, such ideal 'treatment groups' and 'control groups' were impossible to be found or developed in social science research. Consequently, DID model was applied in this project to solve this problem. The Figure below shows the general idea of DID and how it works.

The 'treatment object' and 'control object', P and S, have status P1 and S1 on time period 1, and status P2 and S2 on time period 2. Based on the DID's most critical assumption, parallel trend assumption, that any existing factor would have same effect on P and S, shown by the parallel S1 to S2 line and P1 to Q dotted line DID, the difference between Q and S2 on the second time period equals the difference between P1 and S1 on the first time period, and the difference between P2 and Q represents the effect of the outside factor or treatment.

For this project, to fulfill DID's parallel trend assumption, cross-clustering was applied to find paris of multi-dimensionally homogeneous taxi zones, which can be used as 'treatment group' and 'control group'. Because the Uber & Lyft trip data demonstrated that the effect of ridesharing cars became noticeable since 2015, the parking violation data before 2015 and other related socio-economic, demographic, and facility data of taxi zones were taken into clustering to identify homogeneous taxi zones.

Uber & Lyft has a direct impact on parking tickets

According to the above figure, there are four variables that directly impact parking ticket: Meter Parking Capacity, Parking Lot Capacity, Misdemeanor Crime Rate and Number of Uber & Lyft trips. However, Bayesian Network cannot tell us whether the impact is positive or negative and how strong the impact is.

Uber & Lyft are reducing NYC parking violation but the effect is not significant

The fixed effects model (FEM) controls for the unobserved factors. According to Figure above, the coefficient of Uber & Lyft trips (pickup) is -0.0007, indicating Uber \& Lyft trips reduces NYC parking violation slightly. However, this modeled relationship is not statistically significant, because the p-value of 0.677, is much larger than the significance level, 0.05, making the result of FEM less reliable.

Uber & Lyft is able to reduce NYC parking violation significantly

The figure above shows Flatiron and Midtown Center were identified as a pair of highly homogenous taxi zones, and, in total, 900 pairs (2.7% of all possible pairs) are determined by the DID model. Based on these pairs, the DID model gave an average causal effect intensity of 224.2, which implies that an increase of 224.2 Uber & Lyft pickups is associated with a decrease of 1 parking ticket.