Inbound commuters and tourists linked to more theft in Shanghai communities with more non-hukou residents

In Shanghai, the link between the daily flow of commuters and tourists into a neighborhood and the amount of recorded theft there depends on who actually lives in that neighborhood, according to a new study published in the journal Crime Science. In communities where a larger share of residents held no local Shanghai hukou — the household-registration status that shapes access to housing, schooling and social benefits in China — the association between inbound mobility and theft intensity grew steadily stronger. Where the non-hukou share was small, the same inflows showed essentially no positive association with theft counts.

The study, led by Yuxuan Hu, is among the first to examine how an institutionally defined form of residential differentiation conditions the crime relevance of dynamic population exposure. Rather than treating residential composition and daily mobility as competing explanations of neighborhood crime, the analysis treats them as two distinct population processes: one relatively enduring and structured by the hukou system, the other dynamic and activity-space based, concerning who enters and uses a community during everyday activities.

To build the picture, the researcher combined three data sources for 4,067 Shanghai communities nested within 177 street-level clusters. Measures of residential composition came from the 2010 Population Census, in which the non-hukou resident share was calculated as the number of usual residents without Shanghai hukou divided by the total usual-resident population. Inbound mobility was constructed from China Telecom mobile-phone signaling records for 2016, capturing both commuter inflow — people whose inferred home and workplace lay in different communities — and tourist inflow, defined as users who were not Shanghai permanent residents and remained continuously within an area for more than three hours. Each inflow was expressed relative to the community’s residential population and compared with the corresponding ratio for the rest of Shanghai, then combined into a single score using principal component analysis, which explained 72 percent of the variance in the two indicators.

The crime outcome came from a distinctive source: theft records derived from court judgment documents publicly available through China Judgements Online, originally compiled by other researchers. The judgment-derived records were cleaned, geocoded and linked to community boundaries. The 2017 records supplied the two outcomes — whether a community recorded at least one theft, and how many thefts it recorded — while 2015 and 2016 histories served as prior-theft controls. In total, the analytical outcome contained 7,266 recorded theft incidents, with 1,722 communities, or 42.3 percent, recording at least one incident in 2017. The mean community count was 1.79 incidents, and the mean non-hukou resident share was 0.414.

Two modeling strategies separated the questions. Logistic regression estimated the probability that a community recorded any theft in 2017, while negative binomial regression modeled the count of incidents, with the 2010 census usual-resident population included as an exposure so that expected counts could be compared conditional on population size. Standard errors were clustered at the street level, and all models controlled for population density, a census-derived composite measure of community socioeconomic status, broad urban location, spatially lagged characteristics of neighboring communities, and prior recorded theft.

When entered separately, both focal variables were positively associated with theft occurrence. Non-hukou resident share carried a positive and statistically significant coefficient, and inbound mobility did as well, and both remained positive when entered simultaneously. For theft intensity, however, the additive pattern differed: inbound mobility was positively associated with the expected theft count, but non-hukou share alone was not statistically distinguishable from zero once mobility and the controls were in the model.

The central result emerged from the interaction between the two measures in the count models. That interaction was positive and statistically significant, indicating that the association between inbound routine activity mobility and recorded theft intensity became more positive as the non-hukou share increased. On the adjusted theft-rate scale, the average marginal effect of a one-unit increase in inbound mobility was −0.227 recorded thefts per 10,000 residents at a non-hukou share of 0.10 — a value not statistically distinguishable from zero — but rose to 0.312 at a share of 0.40, 0.973 at 0.70, and 1.554 at 0.90. Adjusted predicted rates showed the same gradient: at a non-hukou share of 0.10, predicted rates were 4.88, 4.81 and 4.65 incidents per 10,000 residents at the 25th, 75th and 90th percentiles of inbound mobility, while at a share of 0.90 the corresponding rates were 4.81, 5.16 and 6.13.

The occurrence models told a narrower story. The interaction between non-hukou share and inbound mobility in the logistic model was positive but not statistically distinguishable from zero on the log-odds scale, and the study’s authors caution that pointwise significance of marginal effects at particular moderator values does not itself establish that the effects differ across values. The probability-scale estimates did trend in the same direction, with the marginal association of inbound mobility rising from 0.002 at a non-hukou share of 0.10 to 0.033 at 0.90, but the analysis supports positive additive associations for occurrence rather than strong evidence of systematic moderation. The study also stresses that the logistic and negative binomial models estimate different outcomes and do not constitute a formal two-part model, so the difference between the two sets of results should not be read as evidence that occurrence and accumulation arise from distinct causal processes.

Robustness checks left the count-model conclusion intact. An intercept-only zero-inflated negative binomial specification reproduced the baseline estimates without improving fit, and across alternative neighborhood definitions for the lagged prior-theft control — including several k-nearest-neighbor specifications, a 2,000-meter distance specification and a trimmed contiguity specification — the interaction ranged only from 0.362 to 0.366 and remained significant throughout. A Moran’s I test detected a small but statistically significant degree of residual spatial autocorrelation, but after eleven Moran eigenvectors were added to the model the focal interaction remained positive and significant at 0.370, closely matching the baseline estimate. An exploratory test of whether the interaction differed across the Inner City, Inner Suburbs and Outer Suburbs found no systematic heterogeneity, and the authors explicitly warn against interpreting the significant estimate observed within the Outer Suburbs as evidence that the conditional association is stronger there.

The study is observational, ecological and confined to a single Chinese megacity, and its authors are careful about what the findings do not show. The theft outcome consists of recorded incidents derived from court judgments, not a complete census of theft, since reporting, investigation, prosecution, adjudication and geocoding all shape which incidents enter the record. The mobility measure is an aggregate of commuter and tourist inflows among users represented in telecommunications data, and cannot identify which types of incoming population, activities or time periods matter most. The authors also emphasize that the results concern community-level associations and do not establish causal mechanisms, nor do they imply criminogenic characteristics of non-hukou residents or mobile populations. The mechanism underlying the conditional pattern remains unresolved: the study does not measure collective efficacy, informal social control, or the actual guardians, offenders and targets that routine activity theory points to, and the authors note that non-hukou share is not a direct measure of those regulatory conditions. What the analysis does establish is a specific contingency — that the crime relevance of dynamic activity-space exposure varies with the residential context receiving those flows — and the authors suggest that mobility measures may be most informative when interpreted together with the characteristics of the places receiving them rather than as stand-alone indicators of place-based risk.

Subject of Research: How hukou-based residential composition conditions the association between inbound daily mobility and recorded theft in Shanghai communities

Article Title: Hukou-based residential composition, inbound routine activity mobility, and recorded theft in Shanghai

Article References: Hu, Y. (2026). Hukou-based residential composition, inbound routine activity mobility, and recorded theft in Shanghai. Crime Science, 15(1), Article 33. https://doi.org/10.1186/s40163-026-00299-5

Image Credits: AI Generated

DOI: 10.1186/s40163-026-00299-5

Keywords: Shanghai, hukou, recorded theft, commuter inflow, tourist inflow, routine activity theory, crime and place, mobile phone data, neighborhood crime, negative binomial regression, residential composition, China Judgements Online