Abstract

Background to Problem

Uncertainty in travel demand model forecasts can significantly affect the ability of transportation agencies to achieve policy and investment objectives, particularly in projects involving toll roads and managed lanes. Traditional travel demand models are subject to uncertainty in key inputs, such as traveler value of time, land use assumptions, and future network conditions. These uncertainties can impact revenue projections, project feasibility, and long-term infrastructure planning. However, addressing uncertainty typically requires running travel models tens or even hundreds of thousands of times under varying assumptions, which is computationally prohibitive given current model runtimes.

Description of Application

This study develops a synthetic modeling approach to efficiently quantify forecast uncertainty in travel demand models. The approach integrates an experimental design framework to systematically vary key model inputs while significantly reducing computational costs. The application builds upon recent enhancements to a tour-based travel demand model, explicitly incorporating the effects of flexible work arrangements and visitor occupancy rates—two variables that have exhibited high variability in recent years, in addition to the impact of land-use, value of time, and levels of network investment in the implementation area.

A fractional factorial design is employed to capture uncertainty’s effects of the five key inputs on model outputs without requiring the exhaustive simulations of a full factorial approach. This enables the estimation of a global response surface model, which serves as the basis for Monte Carlo simulations to generate comprehensive risk profiles. By evaluating forecast variability for planning horizon years 2035 and 2055, this framework provides transportation agencies with a scalable methodology for assessing risk in travel demand forecasts. Uncertainty is measured graphically by calculating the ratio of the 50^th percentile forecast versus the 90^th percentile forecast.

Why This Application is Noteworthy

This application advances travel modeling by offering a practical, computationally feasible approach to quantify and manage forecast uncertainty. Unlike conventional sensitivity testing, which often evaluates a limited set of scenarios, this method enables a statistically robust and scalable means of assessing a full range of plausible future outcomes. The integration of experimental design principles ensures that uncertainty analysis can be conducted more efficiently, supporting more informed policy and infrastructure investment decisions. This methodology is particularly relevant for agencies evaluating revenue risk in tolled facilities, where uncertainty in travel demand can have significant financial implications.

Project Status

This project is completed.

The impacts of the pandemic and its induced changes to telework have been extensively documented in the literature. Travel demand modelers and forecasters are now more conscious of the telework assumptions used in the travel demand models and have introduced ways to account for the potential effects of various telework assumptions on travel demand and ridership forecasting. In this study, we summarized our recent findings related to the uncertainties of transit ridership forecasting stemming from telework assumptions.

Specifically, we have used an advanced trip-based regional model to conduct the ridership forecasting under various telework scenarios, in support of several planning studies in the Washington metropolitan region. Telework trends were analyzed using the biennial regional State of the Commute surveys conducted by the Metropolitan Washington Council of Governments (MWCOG). Based on these telework trends and the baseline assumption used in the model, we derived several telework scenarios. These scenarios were then operationalized in the model, incorporating work-from-home rates differentiated by four household income groups, which reflects the survey finding that these rates tend to increase with income.

We analyzed the impacts on transit ridership across three modes—commuter rail, Metrorail, and bus—both in the short term (2030) and the long term (2045). Additionally, we calculated the elasticities of transit ridership with respect to work-from-home rates for both peak and off-peak periods, demonstrating significant differences in the potential effects of telework assumptions on transit ridership in different time periods of a day.

Over the past few years, remote and hybrid work have grown, driven by COVID-19 and shifting workplace norms. While their impact on mobility has been vastly studied, productivity effects remain less explored. This presentation examines how socioeconomic and behavioral factors influence remote work patterns and productivity. The findings indicate that work-from-home adoption and perceived productivity are influenced by perceived health risks and socioeconomic factors. Furthermore, the results highlight non-conventional land use and behavioral characteristics that can enhance analyses of telecommuting rates across diverse socioeconomic groups and regions. Incorporating these factors allows planners and modelers to generate more accurate, data-driven estimates that reflect both regional and individual-level influences on telecommuting rates used in regional models, while also providing stakeholders with insights into the impacts of health, mobility, and work arrangements on perceived productivity.

This presentation approaches the influence of travel patterns, socioeconomic characteristics, and a range of latent constructs, such as perceived health risk, social propensity, and tech affinity, on telecommute frequency and perceived work productivity changes after the COVID-19 pandemic. The dataset used in this study is derived from the COVID Future Survey, a nationwide, multi-wave panel survey conducted in the United States. This rich dataset includes, but is not limited to, socioeconomic and demographic attributes, travel patterns, attitudes and perceptions, lifestyle preferences, work arrangements and preferences, and pandemic-related behaviors. The methodology uses ordered probit models with two endogenous variables, namely work-from-home frequency and perceived productivity change, both ordinal. The exogenous variables include person- and household-level variables, in addition to commute attributes. Finally, the unobserved latent attitudinal constructs work as intermediate factors in the model structure. Thus, the exogenous, explanatory variables may influence the two endogenous variables both directly and indirectly through their effects on the latent constructs. As the proposed framework contains two endogenous variables and needs to explicitly account for latent attitudinal constructs, this study employs a joint equations modeling framework, which allows error correlations across both latent constructs and endogenous variables. Preliminary results show that individuals with higher telecommuting frequencies and those who are more comfortable with engaging in virtual activities felt more productive after the new work arrangements, whereas those with COVID-19 concerns, under the age of 40, and lower income, perceived their productivity as decreasing. By considering these behavioral and attitudinal characteristics, models can be refined to better predict regional shifts in mobility patterns, inform policymakers regarding health and productivity perceptions, and guide future infrastructure needs.

This work presents results from a wide range of angles, it may improve regional models by refining generalized assumptions regarding telecommuting rates and trends, especially in areas with public health concerns or with diverse population characteristics. Insights may guide policy and planning by linking long-term remote work adoption to productivity, office demand, and infrastructure needs.

This presentation builds upon "A Deep Dive into Telecommuting Behavior" that was presented at the TRB Innovations and Applications conference in 2023.  In this new presentation, the analysis is extended to use data from 2022, 2023 and 2024, providing post-pandemic evidence. More importantly, in addition to using American Communities Survey (ACS) data from 2006 to 2023, this analysis also uses the annual American Time Use Survey (ATUS) data for 2006 to 2024. The two surveys can be used to measure telecommuting in different ways, with ACS simply asking whether or not a worker usually telecommutes to work, and the ATUS data containing detailed activity codes for each 15 minutes during a single day. . 

Somewhat surprisingly, the ACS and ATUS data show very similer trends in percentage changes in telecommuting across the years, although the absolute levels differ due to the way that telecommuting can be measured with the very different survey data.  These large, long-term national datasets allow investigation into many factors that are correlated with differing telecommute rates, including profession and employer type (levels of government, private, self, etc), as well as a wide range of socio-economic factors. 

Also presented is an analysis of relationships between telecommute behavior and the amount of time spent in physical activityies. walking and biking). While the ACS data is not useful for such an analysis, the ATUS data are well-designed for looking at this relationship to determine whether telecommuting and phyisical activity levels are more often substitutes or complimentary, and how the rise of telecommuting during and after the pandemic have affected this relationship. The data indicates that, particulary since the COVID pandemic led to a rise in teleworking, people who telework spend signiciantly more time in participating in physical activities on that day, more than compensating for the decrease in time that some would spend walking or cycling to work. 

Background

In 2019, the Metropolitan Council undertook a major model improvement by upgrading their model to ActivitySim. The model was initially calibrated in and validated to 2018 household survey data. In 2021, the Council collected another round of household survey data. These two datasets were compared to show differences on how these should be used to calibrate the model. 

In the process of calibrating the model with 2021 data, the survey datasets were compared to illustrate the differences that exist between the two datasets. The comparison was specific to data that an activity-based model would use for calibration. One of the resulting model features that was implemented in the Council’s model is a factor that adjusts the work-from-home, telecommute frequency, and daily activity pattern models to pre-pandemic conditions. The non-mandatory tour frequency, joint tours, and mode choice models do not adjust based on the factor but are affected by the models that can adjust.

Description of Application

This application seeks to test this factor and ask the hypothetical question of what happens if we return halfway and entirely to pre-pandemic conditions. The comparisons will be large-scale and compare the outputs of these two models to both the 2018 and 2021 survey datasets.

Statement on Why Abstract is Noteworthy

Uncertainty has become a major discussion point. Following the release of the Transportation Model Improvement Program publication “Transportation Planning for Uncertain Times” in July 2022, the Transportation Research Board (TRB) Transportation Modeling Committee launched a book club and subcommittee (RIP) to help quantify the impact of uncertainty. At the 2025 TRB Annual Meeting, multiple posters and a well-attended Thursday morning workshop were presented. The issue with uncertainty is not a flash-in-the-pan, it is a reality that we all must address. This project is a step in the direction of addressing one component of the issue in a situation where we were fortunate enough to have good before-and-after data.

In addition to the reality of uncertainty, the Metropolitan Council model is open-source, as is ActivitySim. The implementation of this adjustment is included in their source code and available for implementation by others, whether it is directly in ActivitySim or ported to another platform.