Household travel surveys serve as the primary dataset for model estimation and are also used to support model calibration and validation. However, before household survey data can be used in model development, they must be expanded to meet regional control totals developed from national sources such as the Census or the American Community Survey.
Traditionally, household surveys have been expanded to match control totals for household-level variables such as household size, vehicle ownership, number of workers, residential location, and lifecycle variables. This methodology is effective and sufficient to support trip-based models where the household is the unit of analysis in determining the total amount of travel and the analyst strives to properly represent the total travel by market segment.
As activity-based models become more commonplace and reflect more disaggregate person-level choices of total travel, destinations, mode, and route, person-level socioeconomic variables such as age, gender, worker status, and student status are being increasingly used in travel demand modeling. However, since person-level variables are not explicitly used in survey expansion, they may be misrepresented in the expanded survey databases. This can lead to problems during estimation and validation. Therefore, we propose a framework that uses both household and person-level socioeconomic variables in survey expansion.
A multi-stage iterative proportion fitting (IPF) procedure that uses both household and person-level variables during survey expansion was developed for the Metropolitan Council region. Key steps in the process include the following:
• Use of four household-level variables in the IPF including household size, vehicle ownership, number of workers, and lifecycle;
• Perform the household-level adjustments first using the traditional IPF and assign the household weights to each member of the household;
• Use three person-level variables in the IPF including age, gender and worker status;
• Run the person-level adjustments and come up with separate weights for each member of the household;
• Calculate an average household weight using the person weights and begin the second round of household-level adjustments; and
• Identify an acceptable value for convergence.
This methodology has resulted in weights that better reflect both household and person-level distribution of demographics. The weights have helped model validation immensely since weighted data totals match very well when compared to socioeconomic results from the population synthesizer and Census data, thereby making a review of the model results faster.