Residential location choices (RLCs) significantly impact sustainable urban development. Existing RLC models either integrate RLCs into complex urban models, focusing on interdependencies between sub-modules, or explore factors affecting RLCs using Multinomial Logit Models, examining the impact of individual and location characteristics. However, these models lack integration of individual preferences, particularly travel behavior. This study addresses this gap by constructing an RLC model based on travel behavior, utilizing big data from mobile phone trajectory data. The use of revealed preferences from mass mobile phone signaling data offers advantages over stated preferences by reducing sampling bias and reflecting actual behavior. The novelty of this research lies in its focus on revealed preferences, extension of existing models to include individual residential preferences proxied by home-based travel behaviors, and ability to analyze RLC at both group and individual levels. The model's application is demonstrated through assessments of dynamic changes in RLC behavior and predictions based on past travel patterns.

The paper categorizes existing RLC models into two types: those integrated into complex urban models (like MUSSA II, RELU-TRAN, and UrbanSim) focusing on interactions between land, labor, industry, and transportation; and those exploring individual and location characteristics using Multinomial Logit Models. While travel behavior significantly impacts RLCs, with studies suggesting preferences for neighborhoods facilitating satisfying trips, no existing RLC models incorporate individual travel behavior. This study bridges this gap by focusing on analyzing residents' revealed preferences, using mass mobile phone signaling data, instead of stated preferences, thus reducing small sample issues.

The study develops an RLC model based on home-based travel behavior, combining population-level gravity model and individual-level utility maximization approaches. The gravity model (Equation 1) weighs the attraction (mj) of a location against its deterrence (f(rij)), with commuting time as a form of deterrence and housing prices as a financial constraint. The study uses HBNC time to represent a location's attractiveness, reflecting revealed preferences and the consumption of built environment. Equations (2)-(5) define the attractiveness (mj) and deterrence (f(rij)) terms using commuting time and HBNC time. Equation (6) presents the final RLC model, incorporating housing expenditure, HBNC time, and commuting time. The individual-level model (Equation 7-10) is based on utility maximization, with the utility function incorporating commuting time as an iceberg cost and HBNC time (with a consumption and travel component) along with an idiosyncratic shock. Equation (11) shows the probability that a resident chooses a specific location, showing the same structure as the population-based model but with different interpretations. The study uses spatiotemporal travel trajectory data from over 4 million users in Shenzhen and over 12 million in Beijing from 2018-2020. Housing expenditure data, POI data, and precipitation data were also utilized. The methodology involves fitting analysis using GEV distribution to verify the EVT hypothesis for commuting and HBNC time, followed by GLM fitting of the RLC model. Robustness checks involved adding control variables (amenities), using instrumental variables (precipitation, gender, age) to address endogeneity, testing scale effects (using different tile sizes), and incorporating time-lagged terms. The model's application is demonstrated through analysis of pre- and post-COVID-19 changes and predictive power assessment.

The Generalized Extreme Value (GEV) distribution fitting analysis supports the hypothesis that residents minimize commuting time and maximize HBNC time. The GLM results show that the probability of choosing a residential location decreases with increasing commuting time and increases with increasing HBNC time and decreases with housing expenditure, aligning with expectations. The Wald test shows commuting time has a significantly larger impact on RLC than HBNC time. The inclusion of control variables (amenities) improved model fit. Instrumental variable analysis confirms model robustness despite potential endogeneity. The model remains robust across different spatial scales (tile sizes). The inclusion of time-lagged terms further improves the model's goodness of fit. Analysis of pre- and post-COVID-19 data shows that while the sign and significance of commuting and HBNC time remain consistent, the relative importance of commuting time increased post-pandemic, likely due to reduced optional travel. Predictive power assessment, using 2019 data to predict 2020 RLC, shows positive correlation between predicted and actual values at various scales, both in terms of probability and rank.

The findings support the hypothesis that individual travel behaviors significantly influence residential location choices. The model successfully integrates individual preferences, offering a more nuanced understanding of RLC compared to traditional approaches. The robustness tests demonstrate the model's reliability and applicability across different contexts and scales. The impact of external shocks (COVID-19) highlights the model's ability to capture dynamic changes in RLC patterns. The strong predictive power suggests the model's practical value for forecasting housing demand and evaluating policy interventions. The use of revealed preferences from massive datasets contributes valuable insights.

This study presents a novel RLC model that integrates individual travel preferences, improving upon existing models. Its robustness across various tests and its predictive power highlight its usefulness for urban planning and policy analysis. Future research could explore additional variables (noise, air quality), utilize more granular data, refine the categorization of travel types, and account for co-occurrences of non-work site visits to further enhance the model's accuracy and explanatory power.

The study acknowledges limitations such as potential omitted variables (noise, air quality), use of secondary travel trajectory data limiting verification of data quality, data security concerns restricting the use of individual-level data, and a binary distinction between mandatory and optional travel that may oversimplify travel behavior. The underestimation of HBNC time due to the ignorance of co-occurrences of non-work site visits is another limitation.