Abstract
This paper concerns the classic problem of contingency tables with given marginals, also known as the problem of origin-destination table estimation. While traditional methods provide nonnegative real number estimates. the method herein provides integer estimates. The algorithm comprises two stages: the first is the standard iterative scaling which yields real value estimates; the second trims the real estimates into integers by combinatorial optimization. A penalty scheme converts the constrained combinatorial optimization into an unconstrained one so that any standard combinatorial optimization method can be used, such as simulated annealing. Bootstrapping experiments and an analysis of 39 sample sets of 12- and 18-floor elevator traffic measurement indicate that the integer estimate provides a higher resolution than the standard real estimate with roughly the same accuracy.
