Chapter 1 Overview

TransLink’s HandyDART ridership has declined since the COVID-19 pandemic arrived in early 2020. Some HandyDART trips, particularly medical trips, have returned to almost pre-pandemic levels. Others, such as group workshop trips, have yet to return. In preparation for a return to normal, HandyDART has engaged the services of Nelson\Nygaard to develop a model that would estimate various ridership return scenarios and the corresponding amount of service that would need to be provided. The results from this model will be used to help TransLink develop future budgets and estimate peak fleet needs, and adjust service provision/budget strategy as needed in 2021 based on observed demand. This effort builds upon the work that was done in Phase 1, where a simpler model was developed.

This website provides documentation of the underlying methodology of the model, and links to the Shiny dashboard where the model results are presented. Additionally, steps for adjusting and re-running the model in the future are documented.

1.1 Key Findings

The model results are detailed on the Shiny dashboard linked throughout this documentation. Key findings regarding the model results are identified here to guide the interpretation of the results:

  • The model results provide an understanding of the capacity limitations of the HandyDART system in each social distancing regime based on when excess trips must be loaded onto taxis and no more HandyDART trips can be provided at higher levels of ridership demand. This is based on a) limits to the number of total vehicles in the HandyDART system and b) limits to the number of vehicles that can originate from each dispatch garage. Beyond this, Translink must consider when/if the local Taxi system would meet its own capacity limitations.

    • 6-foot distancing: The HandyDART system appears to be capacity limited in this (the existing, as of this writing) social distancing regime at approximately 60% ridership return
    • 3-foot distancing: The HandyDART system appears to be capacity limited in this social distancing regime at approximately 85% ridership return
    • 0-foot distancing: The HandyDART system appears to be capacity limited when social distancing is no longer in effect at approximately 105% ridership return.

  • As described in the limitations section below, the model is intentionally designed to be conservative (i.e. provides estimates on the high end of demand and service needs).For example, a higher number of service hours estimated for the 0-foot social distancing, 100% ridership return scenario was predicted than were observed in 2019. The model predicted 11% more service hours for this scenario than were observed in 2019. In addition to its conservative design, training upon the dispatch allocation and service hour assignment behavior observed in 2020 (as well as 2019) would tend to make the system less efficient than it was prior to the COVID-19 pandemic. Translink also indicated that HandyDART’s service ran too tightly in 2019, and that in the future they would like to provide more slack to create a better customer experience.

  • Service Hour Estimates can be used to budget for future months and years, especially as more information comes in about ridership return. The chart shown below (duplicated in the results dashboard) shows the top-line service hour results needed given a ridership return scenario, and you could also interpolate between estimates given a ridership return rate between one of the 5% scenario points.

Figure 1.1: 2021 Projected Service Hour Needs for HandyDART by Social Distancing and Ridership Return Scenario

2021 Projected Service Hour Needs for HandyDART by Social Distancing and Ridership Return Scenario

1.2 Limitations

Key limitations of the results are documented here – it is important to consider these when interpreting the model results.

  • Trip arrival rates: A key design feature of the model that makes it conservative (i.e. provides estimates on the high end of demand and service needs) is it attempts to estimate the daily trip rate for every TAZ pair ever observed to have a trip booked. The trip rate estimates how many trips are requested per day, or what is the probability of a trip being requested on a given day if the trip rate is less than one. This feature makes it conservative because sometimes a TAZ pair is only observed once or a couple of times in the nearly two years of training data, making it hard to estimate what a representative trip rate is because of the small number of observations. The estimates produced end up being higher than what is observed because of this lack of data, and we would rather assume this than the alternative (e.g., a lower trip rate than what was observed). This feature is likely responsible for the higher number of service hours estimated for the 0-foot social distancing, 100% ridership return scenario than were observed in 2019. The model predicted 11% more service hours for this scenario than were observed in 2019.
  • Taxi capacity not considered: The feasibility of trips being assigned to taxis is not assessed – i.e. there is no consideration of whether there is an upper limit on the capacity of the taxi system. All excess trips that could not be accommodated on the HandyDART system were assigned to Taxis.
  • Weekend HandyDART/Taxi Assignment Over-Simplified: Because Taxi trips are only assigned when no HandyDART vehicles are available, there are essentially zero weekend trips being assigned to Taxis (because there is not enough total demand to exceed available HandyDART vehicles). In reality, taxis can more efficiently serve some trips than HandyDART, and so are still assigned on weekends. The model currently does not take this into account when assigning trips.
  • Ridership return rates are assumed constant across geography: The ridership return rates (by trip purpose) enumerated for each scenario in the methodology chapter are assumed constant across the entire HandyDART service area. This is likely an oversimplification – ridership is returning more quickly in the more densely populated parts of the region, and this would have impacts in downstream calculations, such as how efficiently (in terms of service hours) those trips can be served. In future versions of this model, it may make sense to further divide ridership return assumptions by geography.
  • Training data limitations: As with any machine learning/modeling project, predictions are only as good as the data they are based upon, and subject to the same limitations. We had data for January 2019 through October 2020. Having more data before COVID (e.g., 2018) would make the model more robust, and it would be good to revisit the modeling effort periodically going forward, as training the model on more post-COVID data would be helpful. As it is, we have only since mid-March 2020 to represent post-COVID behavior, and obviously there is a lot of variability in behavior within that period.