Estimating Predictors of Frequent Riders in the Puget Sound
For a course in Maximum Likelihood Estimation, I used a beta-binomial model along with data from TRAC, the EPA, and Washington State Open Data to estimate the biggest predictors of frequent riders in the Puget Sound. Frequent riders refers to transit users who make at least 4 trips per week over a 10-week period. Beta-binomial models were compared to an ordinary least squares model and cross-validated.
Where are Frequent Riders?
The map to the right shows the percentage of ORCA-eligible population (over 6 years old) in each block group where ORCA data was collected. TRAC uses a probabilistic model to estimate where riders live based on ridership patterns observed through ORCA card usage. This formed the basis of the modeling process.
The map shows some expected results, where block groups with the most frequent riders are in denser more urban neighborhoods. There are a few interesting and less-expected block groups such as the dark blue just North of Everett.
Predictors: Significance & Magnitude
Plot showing the percent change given an increase in 1 standard deviation, with all other variables held at their means. Filled dots are significant at a .05 level, open dots are not, and lines show 95% confidence intervals. Plot made using R packages tile and simcf.
The study tested a host of demographic, environmental, and economic variables before settling on these 11 variables and a control for municipality (or lack thereof). Expected results included the influence of access to jobs by transit – defined by the EPA as jobs accessible within a 45-minute transit ride – and transit frequency, the number of transit lines stopping within the block group every hour. Unexpectedly, all densities tested offered negative impacts, possibly suggesting that controlling for municipality explained some of the density variance across all block groups, but that those densities offered some specific predictive power within each municipality.
Transit Predicts Riders
Basically, this study suggested that access to transit and transit’s ability to take riders to work were instrumental in supporting improved ridership. The plots to the left show a stark difference in magnitude between the two, however. Using R to simulate these counterfactual hypotheses, it is clear that frequency was a much stronger predictor than jobs access.