Multilayer networks characterize human-mobility patterns by industry sector for the 2021 Texas winter storm

TL;DR

This study introduces time-varying multilayer mobility networks where layers correspond to industry sectors via NAICS codes to fingerprint human movements. Using SafeGraph GPS data for Harris County during Winter Storm Uri, the authors show pronounced reductions in movements to healthcare, education, and dining, while movements to grocery stores and gas stations were prioritized. They analyze in- and out-degrees of census tracts across layers, reveal that outward movements are more predictable and strongly tied to population, while inward movements depend more on local POI infrastructure and become harder to predict during the storm. The work demonstrates how sector-specific mobility framing can inform emergency planning and resilience strategies, and it outlines future directions for refining the model, including normalization, age structure, and cross-event comparisons. Overall, the paper advances a granular, industry-stratified perspective on mobility dynamics in extreme weather scenarios by combining multilayer networks, hierarchical NAICS stratification, and regression-based predictability analyses.

Abstract

Understanding human mobility during disastrous events is crucial for emergency planning and disaster management. We develop a methodology to construct time-varying, multilayer networks where edges encode observed movements between spatial regions (census tracts) and network layers encode movement categories by industry sectors (e.g., schools, hospitals). Using the 2021 Texas winter storm as a case study, we find that people markedly reduced movements to ambulatory healthcare services, restaurants, and schools, but prioritized movements to grocery stores and gas stations. Additionally, we study the predictability of nodes' in- and out-degrees in the multilayer networks, which encode movements into and out of census tracts. Inward movements prove harder to predict than outward movements, especially during the storm. Our findings on the reduction, prioritization, and predictability of sector-specific movements aim to support mobility-related decisions during future extreme weather events.

Figures (6)

• Figure 1: Networks summarize human movements between census tracts.(a) Visualization of observed movements from home neighborhoods to hospitals in Harris County, TX during the storm week beginning on February 15, 2021 (Monday). Home locations are recorded using U.S. census block groups, whereas destinations locations are Points of Interests (POIs) with known latitudes, longitudes, and other information such as industry category (e.g. hospitals). (b)For different industry categories, we construct networks that are each encoded by a time-varying adjacency matrix in which $A_{ij}(t)$ encodes movements from home census block groups spatially contained in a census tract, which we enumerate by $CBG_i$ and $CT_i$, to POIs in census tract$CT_j$ during week $t$. Much of our study focuses on studying the movements in and out of census tracts each week as defined by their node degrees: $d^{in}_i(t) =\sum_j A_{ji}(t)$ and $d^{out}_i (t) = \sum_j A_{ij}(t)$.
• Figure 2: Hierarchical stratification of movement categories by industry sector.(a) Toy illustration for the hierarchical stratification of a mobility network into network layers that encode different behavioral categories of movements, defined using the North American Industry Classification System (NAICS). The number of digits in a NAICS code determines the hierarchy depth (i.e., level of coarseness when refining movements categories into subcategories). (b) A map of census tracts in Harris County (bottom) overlaid by three example networks at three different coarseness levels: all movements (top), health care and social assistance (middle), and hospitals (lower). (c) Fraction of POIs in each NAICS category for Harris County (left) and fraction of total observed movements in each NAICS category (right). In both panels, for each category different coloration indicates finer subcategories.See Figure \ref{['fig:4']} and Supplementary Figure 1 for additional details about the stratification of categories into subcategories and their industry sector NAICS codes.
• Figure 3: Decreased movement during storm, quantified by z-scores.(a)We plot the total movements $m^{(n)}(t){\color{black}=\sum_{i,j} A_{ij}^{(n)}(t)}$ during each week $t$ for the network layer that encodes observed movements to locations associated with health care and social assistance (NAICS code 62). Red and gray shading highlight the storm week and the weeks used to construct a baseline for comparison. We quantify the change in movements during the storm week, $t=8$, using a z-score $Z^{(n)}(t) \approx -27$, which is visualized in the right-hand panel and is discussed in Section \ref{['sec:zscore']}. It's calculation uses a baseline mean, $\mu^{(n)}$, and standard deviation, $\sigma^{(n)}$. (b) We plot the z-scores, $Z^{(n)}(t)$, versus $t$ for the eight NAICS categories with largest total movements across the 25-week study duration.
• Figure 4: Comparing the storm's impact on movement categories and sub-categories. Z-scores quantify the storm's impact on movement categories defined using the NAICS hierarchical classification scheme. These are shown using both a coarse scale with 2-digit NAICS codes (left) and a finer scale using 3 or 4-digit NAICS codes (right). See Supplementary Figure 1 for the industry sector NAICS codes. Both sets of movement categories are ordered top-to-bottom based on their computed z-scores (shown in colored boxes) so the most-decreased movement categories are at the top. Curved lines depict how coarse movement categories separate into finer categories, and the line widths are proportional to the number of observed movements for each category.
• Figure 5: Node degrees reveal heterogeneous flows among census tracts.(a)We show distributions of in-degrees $d^{in}_i(t)$ and out-degrees $d^{out}_i(t)$ for the mobility network combining all movement categories (left) during the six baseline weeks (blue) and storm week (red). (b) Focusing on network layers associated with high-movement categories, we plot the distributions of in- and out-degrees for both the baseline weeks and the storm week. The probabilities decay linearly in a log-log scale suggesting a power-law relation. (c) Scatter plots reveal that (left) a census tract's out-degree is strongly correlated with the population residing in that census tract, and (right) a census tract's in-degrees is strongly correlated with its infrastructure (i.e., the number of POIs in the census tract). (d) For high-movement categories during the baseline weeks, Pearson correlation coefficients ($r$-values) measure correlation between census tracts' in- and out-degrees versus their populations and the number of POIs for each industry sector. We report how the r-values changed during the storm week (i.e., $r$ for storm week minus $s$ for the baseline weeks). (We note that all p-values were smaller than 0.05 except for one instance, which is outlined by a black box.)