LARCH: A package for estimating multinomial, nested, and cross-nested logit models that account for semi-aggregate data

We present a summary of important computational issues and opportunities that arise from
the use of semi-aggregate data (where the explanatory data for choice scenarios are not necessarily unique for each decision-maker) in discrete choice models. This data feature is commonly encountered with large transactional databases that have limited consumer information, such as itinerary choice modeling. We developed a software package called Larch, written in Python and C++, to take advantage of this kind of data to greatly speed the estimation of discrete choice model parameters. Benchmarking experiments against Stata (a commonly used commercial package) and Biogeme (a commonly used freeware package) based on an industry dataset for airline itinerary choice modeling applications shows that the size of the input estimation files are 50 to 100 times larger in Stata and Biogeme, respectively. Estimation times are also much faster in Larch; e.g., for a small itinerary choice problem, a multinomial logit model estimated in Larch converged in less than one second whereas the same model took almost 15 seconds in Stata and more than three minutes in Biogeme. For more complex discrete choice models, such as the ordered generalized extreme value model, estimation times were two seconds in Larch and four to five days in Biogeme.