In this step, the model's mathematical form is specified (e.g., regression, cross classification, logit, lookup table) and the variables of interest are identified.
In model estimation, one or more mathematical procedures are used to determine the likely values of the model parameters and coefficients. For example, when estimating the likely coefficient values for a logit model, the method of maximum likelihood estimation (MLE) is generally used. Empirically estimated models rely upon data, which is derived from surveys (e.g., Census, household travel surveys, air passenger surveys), traffic counts, or transit counts. Most estimation work is done with software packages such as SAS, SPSS, R, Stata, Alogit, LIMDEP/NLOGIT, or Biogeme.
Once a model is estimated, it needs to be implemented so that it can be applied. Most travel models are implemented and applied using computer software. The TPB travel model makes use of software packages that are designed both specifically for travel demand forecasting (e.g., Cube Voyager and Cube Base) and more general software packages (e.g., Fortran, ArcGIS, Visual Basic).
Model calibration and validation generally occur in an iterative fashion. The model is validated in a "base year" against observed data to make sure that it is performing adequately and reasonably. Based on the performance of the model in model validation, small adjustments are made to the model ("model calibration") until the model accurately replicates observed patterns and behavior. Ideally, the model is validated to a different set of observed data than was used for model estimation. A "future year" validation can also be performed. Although there are no observed data for a future year, one can make sure that the model forecasts are reasonable and consistent with expectations. All TPB travel models are validated against observed data.
In the final step of the process, models are applied, generally using computer software, so that they may be used for developing forecasts.
The above process is usually iterative and each step can feed back to the previous step.