-
Notifications
You must be signed in to change notification settings - Fork 99
Telecommute
This plan is for a person level telecommute model that will be developed in cooperation with SEMCOG. The general model is first described before the SEMCOG specific model is later described. See Phase 6a Task 3 for more information on the task scope.
Telecommuting is defined as workers who work from home instead of going to work. It only applies to workers with a regular workplace outside of home.
The purpose of the model is to represent effects of telecommute program participation on model outcomes, to test implications of changes in telecommute program availability and participation, to account for likelihood of different worker characteristics, to model COVID and post-COVID scenarios, and to test telecommute program participation versus telecommuting on simulation day.
The telecommute model consists of two submodels - a person work from home model and a person telecommute frequency model.
- The work from home model would be run after usual work location choice and before the coordinated daily activity pattern (CDAP) model. It predicts for all workers whether they usually work from home. A separate work from home model is proposed instead of adding a work from home alternative to the the usual work location choice model since someone can work out of home but in their home zone, it makes it easier to estimate, and so shadow pricing applies only to working outside of home. An additional reason for running work from home after usual work location choice is so the work accessibility term calculated by usual work location choice can be re-used by the work from home model. If work from home is choosen, then the work from home model overrides the previously calculated work_zone field since it is no longer valid.
- For all workers that work out of the home, the telecommute models predicts the level of telecommuting. The model alternatives are the frequency of telecommuting in days per week (0 days, 1 day, 2 to 3 days, 4+ days).
Explanatory variables include household and person level variables, including employment categories, sex, age, presence of children, industry/occupation, auto ownership and variables associated with accessibility to work, such as the mode choice logsum and parking costs. Person industry/occupation are key determinants of telecommuting but are often not included in regional population synthesis models and so they are optional attributes. The model software will be built to work in either case - with or without these additional input person attributes.
With a person telecommute frequency variable now available to the system, several of the downstream models can be updated to include sensitivity. The downstream models that will be updated are the CDAP model, the Individual Non-Mandatory Tour Frequency Model, and the Non-Mandatory Tour Stop Frequency Model. The results of these models are all improved with an awareness of person telecommute frequency, and other models can be updated as well if desired by simply referencing the new person variable in the expressions.
More information on the model design is in Joel's proposed design presentation.
The model design is largely inspired by the SANDAG model design, see the telework methodology starting on page 10 of the SANDAG ABM2+ Enhancements to support 2021 RTP (2020) and the telework testing starting on page 74 of the SANDAG ABM2+ Sensitivity Testing Report (2020). The SANDAG telecommute model was estimated from the 2017 household travel survey. The dependent variable in the telecommute model is a person-level variable collected during the recruitment phase indicating the telecommute frequency for persons with job type other than 'work at home'. The analysis involved cross tabulating the share of workers who actually worked on days surveyed by reported telecommute frequency. A multinomial logit model was estimated to predict telecommute frequency based on household and person variables. An ordered logit model was also attempted, however, the specification was discarded due to illogical coefficients, which was consistent with other studies. Occupation, household size and structure, income, work and student status, number of vehicles, and distance to work were significant. The outcome of the telecommute model was reflected in adjustments made to the CDAP model, the mandatory tour generation model, and the nonmandatory tour frequency model. As summarized in the SANDAG work, the revisions to these downstream models were made to reflect the findings from the survey that workers who telecommute one or more day per week are: 1) less likely to go to work; more likely to stay home or engage in non-mandatory travel (roughly equally), 2) somewhat less likely to engage in multiple individual non-mandatory tours, and 3) less likely to make intermediate stops on non-mandatory tours.
The model design was also inspired by the MAG telecommute model, which was specifically developed to address COVID scenarios and is sensitive to employment status, industry/occupation, distance to work, age, income, gender, and presence of young children in the household. For what-if analysis, industry-specific coefficients were calibrated to reflect assumptions regarding telecommute frequency for closed businesses (in addition to 'usual' telecommuters) during COVID.
Beyond the typical data required to estimate an ABM, person industry/occupation is also needed since these are key determinants of telecommuting. This data is not always available in a household travel survey, or not disaggregate enough, and so it is optional. It is preferred to have this information in the synthetic population, but if it is not available, then these person variables can also be omitted from the model expressions.
The two new submodels will support what-if type analysis by allowing the user to specify constants to override/assert work from home and telecommute frequency shares. User defined expressions for both submodels means users can add alternative specific constants for whatever dimensions of the outcomes they would like to override - for example, a 50% share of occupation group XXXXX will work from home. These constants will be manually added as needed for now. The automatic calculation of an asserted share could be done through a semi-automatic calibration add-on that runs the submodel, compares the outcomes to the desired target, adjusts the constant up or down, re-runs the model with the new constant, and repeats the process until the desired target is met.
The SEMCOG implementation follows the design discussed above, with one notable exception. The SEMCOG synthetic population does not include person occupation, but does include person work industry. As a result, the person work industry will be used as an explanatory variable in predicting work from home. The rest of the model is unchanged.