Skip to content

Known limitations

No feedback from the assignment.
The land-use and activity models read fixed skims. Assigned morning-peak car times are a median 0.78 of the skim times on which locations were chosen.
Weak response to time and cost in mode choice.
All time and cost coefficients are negative and significant but small. A 20 % higher car operating cost lowers car-driver kilometres by about 0.2 % (elasticity about −0.01), below the published range; the transit-time elasticity is about −0.2. The models reproduce current mode shares through their constants and socio-economic terms and respond weakly to price and time.
Activity frequencies do not respond to travel conditions.
Frequencies, start times and durations come from survey distributions. They depend on the day's work or school episode, not on travel time, cost or the calendar year.
Activity levels of 2024.
The model keeps the workplace attendance of 2017/18. More residents have a work activity than in the 2021–2024 time use survey, and discretionary activities do not follow the fall after 2019.
Moves after 2020.
The calibrated move rate stays at its 2017–2020 level, while the observed rate fell after 2020.
Tight housing in some counties.
When a county has no vacant dwelling left, in-migrants that the population control would add are not placed. In a 2016–2024 run with one seed, the 2024 population is 1.6 to 4.0 % below the Census estimate in Carroll, Caroline, Wicomico, Queen Anne's, Calvert, Cecil and Frederick counties.
Racial composition drifts.
With the own-race share in the residential location model, the composition drifts toward stronger local majorities from 2016 to 2024 (for example, the white share in Baltimore County rises from 60.3 to 62.7 %, against 53.0 % in the ACS 2020–2024). Holding the shares at their 2016 values changes this by only about 0.7 points, so the drift comes from the own-race term itself. The term describes where households of each group move; it does not separate preference from discrimination, steering or history.
Residential location by PUMA.
Fairfax County (Virginia) remains over-predicted by about 2 percentage points of hold-out movers.
Residents only on the roads.
Trucks, commercial vehicles, visitors and through traffic are not modelled. Volumes fall short of counts outside the peaks and on roads with much freight.
Correlated distance terms in the stop models.
The stop models combine highly correlated distance terms (variance inflation factors up to 32), each kept because it lowers the BIC; their individual coefficients are less precise than the fit suggests.
Repeated use of the hold-out half.
The hold-out half was consulted more than once in the specification search of several models, so its fit is not a fully independent test.