An alternative methodology for calibrating cross-classification models, namely multiple classification analysis (MCA), is described. This technique, which has been available in the social sciences for some time, does not appear to have been used in Transportation planning before, although it appears to be able to overcome most of the disadvantages normally associated with standard cross-classification calibration techniques. The MCA procedure is described briefly, and its merits - in terms of statistical assessment, ability to permit comparisons among alternative models, and lack of susceptibility to small samples in individual cells - are discussed in detail. In addition, the method is based on analysis of variance (ANOVA), which provides a structured procedure for choosing among alternative independent variables and alternative groupings of the values of each independent variable. These procedures are contrasted with standard procedures for cross-classification that estimate cell values by obtaining the average value of the dependent variable (e.g., a trip rate) for those samples that fall in the cell and are unable to use any information from any other cell. The process of selecting independent variables and selecting groupings of the chosen variables by ANOVA is illustrated with a case study. In this study the way in which this process works, and the degree to which there is statistical information provided to guide the analyst's judgment, is shown. In the case study the confirmation of intuitive selections of variables is noted, and also a more surprising result is produced that shows that the best household grouping is one that combines two- and three-person households. A second case study illustrates the use of MCA to calculate trip rates. A comparison of the conventional procedure of cell-by-cell averaging, a MCA design that corrects for interactions is given. It is shown that the MCA allows trip rates to be computed for some cells that are empty of data, and that MCA removes some possibly spurious rates that arise in the conventional method from small sample problems in some cells. It is concluded that MCA provides a strong methodology for cross-classification modeling and that the procedure is effective in surmounting most of the drawbacks of conventional estimation of such models. (A)
Samenvatting