The performance of eight models for predicting ridership levels on new transit routes by using early performance data is summarized.Seven of the models are based on least-squares estimates of linear and nonlinear functions; the eighth model is a manual method based on quarterly ridership statistics. Coparisons are based on r-square statistics, leverage estimates, and ability to predict ridership levels for the second year of operation. The results of these coparisons indicate that (a) forecasts based on less than 6 months of data are unreliable for all conventional statistical models, (b) a simple manual method based on prior experience with other local routes is more effective than least-squares models if ridership forecasts must be produced on the basis of limited amounts of data, and (c) probit-, logit-, and power-function and linear-log models perform acceptably if more than 6 months of data are used. This paper appears in transportation research record no. 1209, Transit administration and planning research.
Abstract