Open Conference Systems, ITACOSM 2019 - Survey and Data Science

Font Size: 
Multilevel time series modeling of mobility trends in the Netherlands for small domains
Sumonkanti Das, Harm Jan Boonstra, Jan van den Brakel

Building: Learning Center Morgagni
Room: Aula 210
Date: 2019-06-06 03:30 PM – 04:40 PM
Last modified: 2019-05-23


The purpose of the Dutch Travel Survey is to produce reliable estimates on mobility of the Dutch population. In this paper, multilevel time-series models have been developed to estimate reliable mobility trends at several aggregation levels, accounting for discontinuities induced by two different redesigns, and outliers due to less reliable outcomes in one particular year. The target variables in this paper are the average number of journey legs/parts per person per day (denoted by anjl-pppd),  the average distance of journey legs/parts per person per day (denoted by adjl-pppd), and the average distance of journey per person per day (denoted by adj-pppd), where journey legs/parts are characterized by journey motive and transportation modes for a particular journey. Predictions for higher aggregation levels are obtained by aggregation of the predictions at the most detailed breakdown into 504 domains defined by the combination of sex, age-class, motive and mode for the period 1999-2017. At first, two models for anjl-pppd and adjl-pppd are fitted using the annual input series of direct estimates and standard errors at the most detailed level. For anjl-pppd and adjl-pppd, the standard errors of the direct estimates are smoothed through Generalized Variance Function (GVF) method for obtaining reliable standard errors for some domains with few observations (or even zero observation). The models are fitted in an hierarchical Bayesian framework using Markov Chain Monte Carlo (MCMC) simulations by incorporating global-local priors for regularization purposes. The third target variable, adj-pppd, is estimated by combining the model predictions of anjl-pppd and adjl-pppd.The predictions at higher aggregation levels are obtained by aggregation of the most detailed domain predictions, resulting in numerically consistent set of trend estimates for all three target variables.

Keywords: Generalized variance function, Global-local priors, Hierarchical Bayesian approach, MCMC simulation, Small area estimation, Survey redesigns

Full Text: SLIDES