Examining the deterrence effective of DUI legislation

European countries and the USA were among the pioneers in DUI legislation [5]. For instance, in California, USA, the introduction of per se DUI laws in 1981 made it illegal to drive with a BAC of 0.10% or higher, irrespective of impairment [6]. To examine the effectiveness of DUI legislation, Rogers and Shoenig analyzed statewide monthly fatal and injury rates from 1979 to 1986 using an interrupted time-series analysis to compare the rates before and after the implementation of these laws [6]. Their findings indicated that the legislation served as a significant factor in reducing these rates.

In developing countries, the introduction of DUI legislation occurred later compared to the USA and European countries. For instance, Brazil enacted its first DUI-related law in 1997, making it illegal to drive with a BAC higher than 60 mg/100 ml [7]. In 2008, Brazil further amended the law to adopt a zero-tolerance policy for DUI, setting the legal BAC limit at 20 mg/100 ml [8]. However, while the mortality rates for pedestrians decreased, those for motorcyclists and cyclists increased [7]. Furthermore, Volpe et al. found no significant evidence of a reduction in traffic-related mortality following DUI legislation in three major Brazilian cities: Belo Horizonte, Rio de Janeiro, and São Paulo [8].

The findings suggest that while the implementation of DUI legislation is essential, its effectiveness in reducing traffic-related fatalities largely depends on the level of enforcement. Legislation alone may be insufficient to generate a meaningful deterrent effect without consistent and visible enforcement efforts. In other words, when enforcement intensity is low, the perceived certainty of punishment among potential offenders decreases, thereby undermining the law’s capacity to deter impaired driving and reduce traffic-related mortality.

DUI legislation and law enforcement in China

The DUI legislation in China was established significantly later compared to that in the United States and Brazil. Specifically, China enacted its DUI legislation in May 2011, categorizing driving with a BAC exceeding 20 mg/100 ml as a civil offense and BAC exceeding 80 mg/100 ml as a criminal offense, specifically categorized as the crime of dangerous driving (危险驾驶罪) [3]. Moreover, in the event of a fatal traffic accident, if a driver’s BAC level exceeds 20 mg/100 ml, the incident is classified as being caused by DUI and categorized as the crime of traffic violation (交通肇事罪) [9]. An interrupted time-series analysis was utilized to assess the deterrent effects of this DUI-related law. The results indicated that crash, mortality, and injury rates showed a gradual decline from 2011 to 2017, this suggests that the general deterrent effect of the legislation may have been effective [10].

Since the implementation of DUI legislation in China, the number of individuals convicted of dangerous driving has increased substantially, accounting for approximately one-quarter of all criminal cases nationwide between 2019 and 2020 [11]. However, according to traffic police data, annual traffic-related fatalities have remained relatively stable at around 60,000 deaths per year [9]. This suggests that, despite the sharp rise in criminal convictions driven by stricter DUI laws, these measures alone may have had limited impact on reducing overall traffic fatalities.

The gap in prior research regarding methodological considerations

DUI enforcement consists of two principal stages: police roadside apprehensions and subsequent judicial proceedings [12]. Roadside apprehensions can heighten the perceived probability of arrest, and the proceeding conviction of crime can reinforce the perception that driving under the influence is unlawful. Both factors contribute to enhancing the certainty of deterrence. However, relative to the substantial body of research on the deterrent impact of legislation, there is a notable paucity of studies specifically addressing DUI law enforcement [13,14].

In the USA, Yao et al. employed mixed-effects regression models and found that enforcement intensity, particularly annual DUI arrest rates, uniquely and significantly predicts reductions in alcohol-impaired crash fatalities. Even after controlling for factors such as traffic funding and demographic variables, increased DUI enforcement remained strongly associated with declines in DUI-related deaths. These findings highlight that legislation alone is insufficient, that effective and visible enforcement is crucial for deterring drunk driving and reducing fatal crashes, especially in urban areas [14]. However, Elvik argued that the relationship between enforcement intensity and crash reduction is non-linear. Specifically, he noted that diminishing returns occur as enforcement levels increase. That is, additional enforcement yields progressively smaller reductions in traffic crashes [15]. Little is known about how this non-linear relationship varies across different city sizes or enforcement contexts, particularly in countries like China where empirical evidence remains scarce.

In China, Zheng highlighted regional disparities in law enforcement intensity, noting that economically developed areas tend to allocate more resources to DUI enforcement, leading to higher level of deterrence. In contrast, enforcement efforts are relatively weaker in some remote regions due to limited law enforcement capacity [11]. Li et al. observed that in economically thriving first-tier cities in China, the mortality rates were lower compared to other tiers of cities [9]. Bhalla et al. studied the prevalence of speeding and drinking driving in two Chinese cities as part of a mid-project evaluation of ongoing road safety interventions. They found that targeted enforcement and public awareness campaigns were associated with a measurable reduction in these risky behaviors in the intervention city, while rates remained relatively unchanged in the control city [16]. Nevertheless, empirical studies examining the intensity of DUI law enforcement and its relationship with traffic mortality in China remain extremely limited. As a result, it is still unclear whether an increase in enforcement intensity could lead to a meaningful reduction in traffic-related deaths.

Previous studies have primarily relied on traditional time-series approaches, such as interrupted time series analysis, which can only compare changes in mortality trends between two time periods and are therefore typically applied to evaluate pre- and post-legislation effects. In contrast, mixed-effects regression models can accommodate hierarchical structures and control for observed confounders [14]; however, they generally impose a linearity assumption that may not adequately capture the complex relationship between enforcement intensity and traffic mortality. Moreover, traffic accident data often exhibit weak temporal dependence and non-stationary characteristics, making traditional time-series modeling challenging [17]. Therefore, it is necessary to complement conventional regression approaches with machine learning models that relax the linearity assumption, enabling a more flexible assessment of enforcement impacts while more effectively addressing confounding factors [18]. In addition, the use of SHapley Additive exPlanations (SHAP) values provides an interpretable framework to quantify the contribution of enforcement, among a set of potential confounders, to variations in mortality rates [19].

Therefore, to examine the temporal relationship between DUI law enforcement and traffic mortality, this study employs a combination of traditional time-series analysis and machine learning techniques to analyze the dynamic deterrent effects of DUI enforcement in China, aiming to provide actionable insights for refining enforcement strategies and enhancing traffic safety.

Data source

This study utilized voluntarily disclosed government documents, which are publicly accessible. Given the nature of these publicly available data sources, ethical review was not required. Data analysis was conducted using Python (version 3.12.4 on Windows). The data on roadside check and fatal traffic accidents used in this study were obtained from first-instance court judgment documents voluntarily disclosed by Chinese courts and available through the National Basic Science Data Center [20]. However, comprehensive disclosure only became possible after 2013, when a mandatory court file disclosure policy was implemented. Prior to that, very few traffic crime cases were publicly accessible. Furthermore, the implementation of the Data Security Law in 2021 significantly reduced the number of publicly disclosed court documents [21]. For these reasons, the study period is limited to traffic crime cases that occurred between 2014 and 2020, during which the availability of case data was most complete for epidemiological analysis.

In addition to traffic-related data, the study incorporates geographic, population, and socio-economic variables. Geographic information is based on the official 2019 administrative map of China (ID: GS(2019)1822). Population data were sourced from the sixth (2010) and seventh (2020) national censuses [22], while socio-economic indicators were obtained from municipal-level statistical bureaus.

Text mining

Chinese court documents follow a standardized structure, typically beginning with a summary of the criminal behavior, followed by a legal analysis linking the behavior to its consequences, and concluding with the sentencing outcome. To extract key information from these documents, we employed rule-based keyword matching using Python’s re package. The extracted dataset included information on the year and month of each traffic crime, the city in which the offense occurred, whether the offender was apprehended during a roadside DUI check, and the number of fatalities resulting from the incident. The code used for the data extraction process is available in the supplementary file S1 File (supple_code.txt).

After applying text-mining techniques, a total of 381,233 arrests resulting from roadside alcohol checks were identified during the study period. Additionally, 248,973 deaths caused by traffic violations were recorded, of which 23,111 were associated with DUI-related offenses.

Data preparation

Among the 360 cities in mainland China, 76 cities reported no roadside check arrests or traffic-related mortalities during any year of the study period. In addition, 75 cities had populations of fewer than 1 million residents, according to the Seventh National Census. To ensure data stability and reduce the impact of low case frequency or small population size, these cities were excluded from the analysis. After this screening process, 248 cities remained for inclusion in the study (Fig 1). This exclusion accounted for only 4.1% of roadside check arrests (leaving 365,753 cases), 8.5% of total traffic-related mortalities (227,896 cases remaining), and 9.0% of DUI-related mortalities (21,036 cases remaining). As a result, the final dataset retained the vast majority of roadside checks and traffic-related fatalities that occurred in China during the study period.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Study regions and city population.

Downloaded from the Standard Map Service System of the Ministry of Natural Resources of the People’s Republic of China (http://bzdt.ch.mnr.gov.cn/). The base map was freely provided for public use and modified by the authors to include population data from the national census.

https://doi.org/10.1371/journal.pone.0338886.g001

A linear interpolation method was used to estimate annual population figures for each city between the 2010 and 2020 census years. Based on these estimates, roadside check rates and mortality rates were calculated per 100,000 population for each of the 248 selected cities on a monthly basis from 2014 to 2020. These rates served as the independent variable (roadside check rate) and two dependent variables (overall traffic mortality rate and alcohol-related mortality rate) in the subsequent analysis. For months with zero reported cases, a minimal value (1e-6) was imputed to enable the application of classical time-series analytical methods and machine learning models.

A total of 4 socio-economic factors were acquired from city-level statistical bureaus and served as confounding variables, including GDP (100,000,000 RMB) per 100,000 people, road passenger volume (10,000) per 100,000 people, financial income of city government (10,000 RMB) per 100,000 people, and number of hospital beds per 100,000 people. Since these variables were originally available only as annual data, monthly values were obtained by linear interpolation between yearly observations to better match the monthly frequency of the dependent and independent variables. This interpolation assumes gradual changes within each year and is unlikely to introduce significant bias, as socio-economic indicators typically vary smoothly over time. Information on dependent, independent, and confounding variables can be found in the supplementary file S1 File (supple_df.csv).

Vector autoregression and impulse response analysis

To analyze the dynamic relationship between roadside alcohol checks and mortality outcomes, we employed a Vector Autoregression (VAR) model separately for small (1–5 million population in 2020) and large cities (> 5 million population in 2020). To ensure the VAR analysis, we first conducted unit root tests to examine the stationarity of time series variables. Specifically, the Augmented Dickey-Fuller (ADF) test was applied separately to the monthly panel data of small and large cities. The VAR model captures the linear interdependencies among multiple time series variables. Specifically, we modeled the response of total traffic mortality and alcohol-related traffic mortality to shocks in roadside check rates over time. The analysis was conducted on monthly panel data aggregated by city size. For each city group (small or large), we estimate the following VAR model:

Where is [] or [], and denotes a vector of innovations. The optimal lag order 𝑝 was selected using the Akaike Information Criterion (AIC) and applied consistently across model estimation and Impulse Response Functions (IRFs) analysis. To ensure robustness, additional estimations were conducted using alternative lag lengths ().

We computed the IRFs to trace the dynamic impact of a one-time shock to roadside alcohol check rates on mortality outcomes over the following 1–8 months. Formally, the IRF at horizon h is defined as:

Where is a one-unit innovation (shock) in the roadside check variable. We present IRFs with 95% confidence intervals computed using standard errors from the VAR estimates. Separate models were estimated for small and large cities to assess heterogeneous effects across different urban sizes.

Gradient Boosted Regression for marginal effect estimation

To account for confounding effects that traditional time-series methods cannot adequately address, we employed Gradient Boosted Regression Trees (GBRT) implemented via the XGBoost library. Before finalizing the model, we compared three candidate algorithms, namely Random Forest, Support Vector Regression, and XGBoost. A grid search procedure was used to identify the optimal set of hyperparameters for each model. Based on both predictive performance and interpretability considerations, XGBoost was selected as the final model because it provided an appropriate balance between coefficient of determination (R²) and root mean squared error (RMSE).

This machine learning approach is well suited to capturing complex, nonlinear relationships and interaction effects among predictors without requiring prior assumptions about their functional form. To further evaluate the contribution and impact pattern of each feature in the model, we utilized both SHAP values and marginal effect analysis. SHAP values offer a unified measure of feature importance and directionality at both global and local levels, while marginal effect curves illustrate the predicted change in the outcome variable as a function of a specific predictor, holding all other variables constant. This complementary analytical framework enables us to identify not only which variables are most influential but also how they shape predicted mortality outcomes.

Separate models were trained for small and large cities, using the same dependent variables as in the VAR models: mortality rate and DUI-related mortality rate. The predictors included the roadside check rate and its lagged values (Lag0, Lag4, Lag8), along with four confounding variables. To account for seasonal effects, monthly seasonality features were created using sine and cosine transformations of the month variable. An XGBoost Regressor (100 estimators, max depth = 4, learning rate = 0.1) was employed to model the relationships between predictors and each outcome. For each combination of city group and outcome, 3-fold cross-validation was conducted by randomly splitting the data into three subsets; in each fold, two subsets were used for training and one for testing. Model performance was assessed using the R² and RMSE, averaged across the three folds. Feature importance was interpreted through SHAP values, with the mean absolute SHAP value quantifying each feature’s overall impact, and the mean SHAP value indicating the directionality of its effect.

To visualize the marginal effect of roadside check intensity, we used partial dependence plots. For each lagged roadside check variable, we generated a sequence of hypothetical values across its observed range (5th to 95th percentile) while holding other variables constant at their mean values. The predicted outcomes were then plotted to depict how the mortality rate changes with varying roadside check intensity.

Formally, the partial dependence function for a single feature is defined as:

where is the trained model, are the fixed values of the other features (controls), and is the feature of interest (roadside check rate with different lags). Separate plots were generated for each outcome and city size group to illustrate potential heterogeneity in the effects. Lagged variables were used to explore delayed policy impacts.

Roadside check rates and mortality rates by city size

Fig 2 illustrates the trends of three variables over the entire study period. In both small and large cities, the roadside check rate shows a marked increase starting in September 2019, followed by a sharp surge. However, the rate declined significantly in January and February 2020 due to the nationwide COVID-19 lockdown. After reopening in March 2020, roadside checks rebounded and have remained at elevated levels since. The overall mortality rate follows a similar pattern to the roadside check rate, while the alcohol-related mortality rate remains relatively stable throughout the period.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Trends in roadside check rates, mortality rates, and DUI-related mortality rates by city size (2014-2020).

https://doi.org/10.1371/journal.pone.0338886.g002

IRF analysis

All time series passed the ADF test for stationarity, allowing for the direct application of the VAR model followed by IRF analysis. Fig 3 presents the IRF results. Regarding overall mortality rates, in small cities, the IRF reveals an initial positive response that peaks around the first month, indicating a short-term increase in mortality rates associated with roadside checks. This effect gradually diminishes and becomes negative by approximately the sixth month, suggesting a subsequent decline in mortality rates. The negative effect continues to intensify, reaching its lowest point by the eighth month. In large cities, a similar response pattern emerges, with an initial positive peak near the first month; however, the effect is less pronounced and more variable, showing fluctuations throughout the lag period. The response concludes with a negative effect by the eighth month, though this decline is less steep than in small cities.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. The IRFs for the effects of roadside checks on mortality outcomes.

https://doi.org/10.1371/journal.pone.0338886.g003

For alcohol-related mortality rates, the IRF for small cities shows a strong initial positive response peaking around the second month, indicating a significant short-term increase in alcohol-related deaths linked to roadside checks. This response fluctuates thereafter, with a notable dip near the sixth month, followed by a slight recovery toward the end of the period. By the eighth month, the effect turns negative, reflecting a decrease in alcohol-related mortality rates. In large cities, the response is more muted, with a less pronounced initial peak and greater variability throughout the lag period. Similar to small cities, the response ends negatively by the eighth month, indicating a reduction in alcohol-related mortality rates. To test robustness, the VAR models were re-estimated using lag orders of 4 and 8. The impulse response functions exhibit highly consistent patterns across specifications. The peak responses occur around the first month in all cases, and the magnitude and direction of the effects change only marginally. These results confirm that the estimated IRFs are robust to alternative lag selections.

Overall, these findings suggest that roadside checks have differential impacts on mortality and alcohol-related mortality rates depending on city size, characterized by initial increases followed by decreases or stabilization in both outcomes. The 95% confidence intervals indicate some uncertainty in the estimates, particularly for large cities.

SHAP feature importance with direction

Fig 4 presents the model performance and feature importance for predicting mortality rate and alcohol-related mortality rate using XGBoost regression, stratified by city size. For mortality rate, the model achieved an R² of 0.28 in small cities and 0.42 in large cities, indicating a moderate fit. In contrast, models predicting alcohol-related mortality rate showed lower predictive power, with R² values of 0.06 and 0.07 for small and large cities, respectively.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Contribution analysis for mortality rate prediction using SHAP values.

https://doi.org/10.1371/journal.pone.0338886.g004

Regarding overall mortality rates, roadside check rates and their lagged terms were consistently among the most important predictors across both city groups. In small cities, the current roadside check rate had the highest importance, substantially exceeding that of other factors. Conversely, in large cities, variables such as GDP per 100,000 people and hospital beds per 100,000 people demonstrated significant negative contributions. For both city groups, the 8-month lagged roadside check rate exhibited negative effects, suggesting a long-term reduction in mortality. Additionally, the cosine month term contributed positively, while the sine month term contributed negatively in both models, indicating that mortality rates tend to be higher during the winter months.

For alcohol-related mortality rates, roadside check rates (lag 0) were more important in small cities and exhibited a negative effect, implying a protective influence. In large cities, confounding variables such as financial income and passenger volume carried greater weight than in small cities, suggesting that the impact of roadside checks may be more pronounced in smaller urban cities. Furthermore, in both models of alcohol-related mortality, the cosine and sine terms for month demonstrated negative and positive associations, respectively, suggesting that alcohol-related mortality rates reach their peak during the summer months.

Marginal effects of roadside check rate on mortality outcomes

Fig 5 illustrates the marginal effects of roadside check rates (lags 0, 4, and 8) on both overall mortality and alcohol-related mortality in small and large cities. At lag 0, mortality rates increase stepwise with higher roadside check rates in both city types, indicating a positive association. However, by lag 8, the relationship turns negative for most outcomes, except for alcohol-related mortality in small cities. Notably, the patterns indicate a diminishing returns effect on overall mortality rates, in large cities, where the marginal benefit of increasing roadside check rates levels off around 0.002. In contrast, small cities show a higher threshold, with the marginal effect persisting up to approximately 0.006 before stabilizing. These findings suggest that higher levels of roadside enforcement may be more effective in smaller cities.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Estimated marginal effects of roadside check rates (lags 0, 4, and 8 months) on mortality rates.

https://doi.org/10.1371/journal.pone.0338886.g005

Limitations

First, this study is based on data from China, and the results may not be directly generalizable to other countries or regions. Significant differences in enforcement culture, social norms, and DUI laws across countries could impact the generalizability of policy effects [8]. Second, other traffic safety measures or policies, such as speed limit regulations and traffic safety education, may interact with DUI enforcement and influence the study’s findings. This study was unable to fully isolate the effects of these factors, which could lead to an overestimation or underestimation of the results. Furthermore, the impact of total vehicle miles traveled on traffic accidents and various traffic violations was not considered, as data was unavailable. Third, due to the non-disclosure of cases involving state secrets or juvenile offenders, this study was unable to include these subsets of data in the analysis. Specifically, traffic accident cases may involve juvenile offenders, but the lack of relevant data limits the generalizability and completeness of the findings, thereby affecting the accurate assessment of overall trends in traffic violations. Lastly, neither classical time series models nor machine learning methods were able to identify a causal pathway linking roadside checks to fatality rates. Therefore, although this study demonstrated a close temporal association between DUI law enforcement and traffic fatality rates, the underlying mechanisms of this interaction remain unestablished.