Methods and input data

Trip generation

Trip generation is based on the National Household Travel Survey. The model uses four time periods: overnight (7:00 pm–5:59 am), AM Peak (6:00 am–9:59 am), midday (10:00 am–3:59 pm), and PM Peak (4:00 pm–6:59 pm). The model divides trips into three purposes: home-based work (HBW), home-based other (HBO), and non-home-based (NHB). While this is fewer trip types and time periods than might be included in production travel models, it is consistent with the general practice of dividing trips by time of day and type.

Household-level trip counts are estimated for each time period and trip type using a simple linear regression with the trip count as the dependent variable and independent variables for number of vehicles, household size, household income, Census tract residential density, and number of workers. This results in 12 regression equations, for each time period and trip purpose. Household income is represented by dummy variables for less than $35,000, $35,000–$74,999, $75,000–$99,999, and $100,000 or more.

These regression models are disaggregate, household-level models, whereas the four-step model is an aggregate model using marginal data at the TAZ level (for simplicity, TAZs correspond directly to Census tracts). Specifically, the model uses household size (topcoded at 4), number of workers (topcoded at 3), number of vehicles (topcoded at 3), and income (in the categories used in the regression). To apply the household-level model to this aggregate data, I disaggregate the data to household-level records (i.e. create a synthetic population) using iterative proportional fitting with a seed matrix derived from the Integrated Public Use Microdata Sample 2021 Five-Year American Community Survey data for the entire US (Ruggles et al. 2024). This seed matrix ships with the software. The regressions predict household-level tripmaking, which is re-aggregated to the tract level.

The NHTS does not provide sufficient spatial detail to estimate trip attractions. Instead, I use the Puget Sound Household Travel Survey, which includes origin and destination Census tract in the public-use dataset (Puget Sound Regional Council 2024). I calculate the number of HBW and HBO trips in each time period that have the non-home end in each Census tract in the Puget Sound region. For NHB trips, the production and attraction ends are not clearly defined. I therefore assume that NHB attractions and productions are half of the total number of NHB trips originating from or destined to each tract.

To extrapolate this data to tracts outside the region, I build linear regression models for each trip type and time period based on total employment and employment in retail, education, and accomodation/food services from the US Census Bureau Longitudinal Employer-Household Dynamics Origin-Destination Employment Statistics. I balance total attractions by trip type in each time period to match estimated productions.

Trip distribution

The trip distribution step uses a singly-constrained (at the production end) gravity model (Travel Forecasting Resource 2020). The exponent for the gravity model is calibrated for each trip type using the method introduced by Merlin (2020) based on median trip length. The method observes that half of the weighted destinations should be closer to the origin than the median trip, and half should be further away. I make two slight changes to the function presented in Merlin (2020); see appendix.

The median trip distance comes from the NHTS. I approximate the crow-flies distance for each NHTS trip by dividing the network distance by 1.3, a factor determined by Wang et al. (2024).

For intrazonal trips, I assume a travel distance of $0.52 \sqrt{s}$, where $s$ is the area of the TAZ. This is based on a Monte Carlo simulation of the average distance between random points in a square. There are two opposing factors that bias this. TAZ’s are not square, increasing average travel distance. However, development within a TAZ is concentrated in certain areas, decreasing average travel distance. I assume these roughly cancel out.

Mode choice

The mode choice model is a multinomial logit model based on the NHTS. Because there is not detailed information about each trip and the alternatives available in the NHTS, the model is based solely on trip type, time period, travel distance, and housing unit density in the home tract. For NHB trips, a separate model is estimated excluding density. Goodness of fit is poor, but the point of the model is to demonstrate a simple demand model, not to produce accurate forecasts.

Network assignment

The network assignment uses a Frank-Wolfe traffic assignment algorithm (Boyles, Lownes, and Unnikrishnan 2023). First, I convert production-attraction format home-based trips into origin-destination format, using directionality factors estimated from the NHTS. I calculate peak hourly vehicle flows during each period using average vehicle occupancy for each period and trip type, and an assumed “peaking factor” that accounts for the proportion of traffic during the time period that occurs in the busiest hour—for example, I assume 45% of the traffic in the three-hour PM Peak period occurs during the busiest hour. Then, I run the assignment algorithm, with impedances based on a Bureau of Public Roads-style function: $$t_{\mathrm{congested}} = \left(1 + 0.6 \left[\frac{f}{c}\right] ^ 5 \right) t_{\mathrm{freeflow}}$$

where $f$ is the predicted flow, $c$ is the link capacity, and $t_{\mathrm{congested}}$ and $t_\mathrm{freeflow}$ are congested and free-flow link travel times. The factors 0.6 and 5 are from the Southern California Association of Governments travel demand model (Southern California Association of Governments 2012).

I derive the network from OpenStreetMap. By default, I retain only the most major roads—motorways, trunk, and primary roads (and associated ramps). Furthermore, I run the assignment algorithm only until the “relative gap”—a measure of the error in the estimate—is 1%, rather than the typically recommended 0.01% (Boyce, Ralevic-Dekic, and Bar-Gera 2004), to improve performance.

This work © 2026 by Matt Bhagat-Conway is licensed under CC BY 4.0

References

Boyce, David, Biljana Ralevic-Dekic, and Hillel Bar-Gera. 2004. “Convergence of Traffic Assignments: How Much Is Enough?” Journal of Transportation Engineering 130 (1): 49–55. https://doi.org/10.1061/(ASCE)0733-947X(2004)130:1(49).

Boyles, Stephen D., Nicholas E. Lownes, and Avinash Unnikrishnan. 2023. Transportation Network Analysis. Vol. 1.

Merlin, Louis A. 2020. “A New Method Using Medians to Calibrate Single-Parameter Spatial Interaction Models.” Journal of Transport and Land Use 13 (1, 1): 49–70. https://doi.org/10.5198/jtlu.2020.1614.

Puget Sound Regional Council. 2024. “Household Travel Survey Trips.” https://psrc-psregcncl.hub.arcgis.com/datasets/household-travel-survey-trips/explore.

Ruggles, Steven, Sarah Flood, Matthew Sobek, Daniel Backman, Annie Chen, Grace Cooper, Stephanie Richards, Renae Rodgers, and Megan Schouweiler. 2024. “IPUMS USA: Version 15.0.” https://doi.org/10.18128/D010.V15.0.

Southern California Association of Governments. 2012. “SCAG Regional Travel Demand Model and 2012 Model Validation.”

Travel Forecasting Resource. 2020. “Destination Choice: Theoretical Foundations.” Travel Forecasting Resource. 2020. https://tfresource.org.

Wang, Shanshan, Henrik M. Bette, Michael Schreckenberg, and Thomas Guhr. 2024. “How Much Longer Do You Have to Drive Than the Crow Has to Fly?” June 10, 2024. https://doi.org/10.48550/arXiv.2406.06490.