Both survey data and mobile phone network data are associated with benefits and limitations. Mobile phone network data is limited in its lack of ground truth for the bus and car split, in the lack of socio-economic information about the traveller, as well as in the lack of information about trip purpose. Survey data on the other hand nowadays often has too few observations to obtain significance in important parameters, and it is unlikely that they capture a representative sample of the population. In this paper we investigate to what extent combining the two data sources for mode choice demand model estimation can mitigate the challenges present when only using a single data source, as well as to what extent the estimation can be improved by adding variables unique to the survey data. We also provide recommendations for data collection and combination for future mode choice demand models and analyse policy implications of the estimated results. 

We show that in our case the business cost parameter becomes insignificant when basing the estimation on survey data only. The business cost parameter is important to compute values of travel times which are relevant for social cost benefit analyses. A benefit of combining the two data sources is that we can both identify latent class variables which capture the correct business group in the latent class structure of the mobile phone network data, and simultaneously obtain better estimates of those variables based on observations from the survey data. Most of the elasticity results in this paper are on par with previous studies; however our results suggest higher cross-elasticities in response to train travel time than what has previously been observed, which could have implications for the social valuation of investments in high speed rail. 

In this paper we have shown that the proposed method of data combination is indeed valid, and that by combining the two data sources we can mitigate the limitations of each separate data source and thus maintain good quality mode choice demand forecasting models, even in the face of declining survey response rates.