Resident Production Model

Introduction

The goal of the resident trip production models in TRMG2 is to determine how many trips each individual will make in a day. Unlike traditional trip generation models, attractions are not estimated in this step. Instead, the destination choice models estimate coefficients on employment as part of the size term, which function similar to trip attractions.

TRMG2 production models are person-level models. This means that detailed information about each person (e.g. age), as well as the characteristics of their household and geographic location, are used together in the prediction.

Using DT to guide segmentation

After attempting various forms of generalized linear models (GLM) and logistic regression models, Caliper opted to use Decision Trees (DT) from the machine learning stack. This model form had much stronger predictive power compared to traditional approaches while still being easy to understand and explain.

In the trees below, each node lists the average trip rate as well as the percent of the population that node represents. This lets you see the overall average trip rate (top of the tree) and how it changes as you segment the surveyed population.

To save room in the charts, many of the explanatory variables are abbreviated. Their meanings are listed below for reference:

• Person-level variables
  • is_worker: if the person is a worker
  • is_senior: if the person is >= 65
  • is_child: if the person is < 18
  • age: person’s age
  • gender: Male (1) and Female (2)
  • single_parent: if the person is the only adult in a household with children
• Household Variables
  • retired_hh: if the household contains only retirees
  • per_inc: per-capita income (household income / size)
  • oth_ppl: number of other people in the household
  • oth_kids: number of other children in the household
• Zonal variables (see accessibility page for more details)
  • g_access: general accessibility of the person’s home zone
  • n_access: nearby accessibility of the person’s home zone
  • e_access: employment accessibility of the person’s home zone

Work Tours

Non-Work Tours

Calibration

Calibrating the production rates is done to ensure that the final model is producing the same number of trips per person on average as the survey. One complicating factor in this comparison is that the survey and model socio-economic data have different total populations.

• Survey: 1.71 million
• Model: 1.83 million

As a consequence, the survey trip totals will be increased by 1.07 and then compared to model results. The table below shows the production model results compared back to the total trips in the survey (factored up). While close, the model consistently predicts fewer trips compared to the survey.

Another important consideration is which type of survey weight to use. RSG provides two weight fields on the survey:

• Household weights
• Trip weights

Importantly, trip weights have extra adjustment based on GPS traces to account for things like un-reported trips. As a result, the trip weight gives a higher number of total trips. For estimation, however, the decision trees worked best with household weights, which allowed each person to have an integer count of trips instead of 1.2 or 3.4. This allowed us to identify the features properly, but then calibration is required to scale up the trip totals as shown below.

Finally, the market segmentation created by the Auto Ownership model is not exactly the same as what is observed in the survey. While the share of total households owning 1, 2, 3, and 4+ autos is calibrated to match the survey, the AO model cannot guarantee that market segmentation will match. This is because auto sufficiency is defined by autos owned and number of adults. In other words, the production rate calibration factors are the first opportunity to correct market segmentation imbalanced introduced by the auto ownership model.

In the table below, the model results are compared to the survey trips. As expected (due to the weighting issue above), the model underpredicts trip making.

The calibration factors below are the ratio of the observed to modeled trips. These are applied by trip type and segment to ensure total trip making matches the survey (based on trip weight).