7 Conclusion

After conducting the experiments in Chapter 6, it appears that using all available data is worthwhile; the multizone method offers small benefits over the zone-by-zone method; and hierarchical margins offer a very small benefit for addressing random rounding issues.

The next stage in the ILUTE synthesis effort will link this population with other data sources to synthesize the vehicles owned by each household, the place-of-work of the household members who are active in the labour market, and the business establishments that provide employment.

In conclusion, several of the problems of existing population synthesis procedures were successfully resolved in this research. The first major contribution is a sparse list-based Iterative Proportional Fitting procedure that combines the advantages of IPF and reweighting: an entropy-maximizing procedure that preserves the association pattern in the PUMS while fitting a set of disparate marginal distributions, making possible a large set of agent attributes with fine categorization. This technique produces results identical to the IPF procedure, but makes more efficient use of memory and time when a large number of attributes are synthesized.

The second major contribution was a technique for synthesizing relationships between agents using IPF and conditional probabilities. This allows persons to be grouped into aggregations such as families and households while fitting known distributions at the person, family and household level, and enforcing a limited number of constraints between the members of an aggregation. The results show that these relationships can be synthesized with only a minimal impact on the fit at any single level.