Gender Differences in DUI Crash Injury Severity: A Partially Constrained Random-Parameter Logit Model Analysis

Abstract

Driving under the influence (DUI) has long been recognized as a major contributor to traffic accidents. However, the factors influencing the severity of crashes in DUI situations may vary significantly between genders due to physiological and psychological differences. This study analyzes DUI single-vehicle crash data from Texas to construct a random-parameter logit model that captures gender-specific differences in crash severity. A partially constrained method is employed to better identify these gender-specific factors, emphasizing the importance of separately assessing DUI behavior for males and females in traffic safety analysis. The results reveal notable gender differences in the severity of injuries from DUI crashes. A comprehensive evaluation was conducted from four perspectives: driver characteristics, vehicle features, roadway conditions, and environmental factors. Out-of-sample simulations provided additional insights, showing that even at lower blood alcohol concentration (BAC) levels, the probability of severe injury increases significantly. In conclusion, this study not only uncovers the gender-specific mechanisms behind DUI crash severity but also offers valuable empirical evidence for integrating gender considerations into future traffic safety policies and interventions.

1. Introduction

Alcohol consumption significantly impairs drivers’ cognitive abilities and reaction times, thereby increasing the likelihood of operational errors and substantially elevating both the risk and severity of traffic accidents [1,2]. Driving under the influence (DUI) has long been recognized as one of the major threats to road traffic safety. According to statistics, approximately 10,598 people in the United States die annually in DUI-related traffic accidents, accounting for about 30% of all traffic fatalities [3]. Thus, an in-depth exploration of the causal mechanisms underlying DUI-related crashes and the identification of their key contributing factors hold considerable theoretical significance and practical value for formulating effective intervention strategies and reducing road traffic casualties.

A substantial body of research has focused on identifying the key factors influencing the severity of DUI crashes. Behnood et al. [4] employed a latent class multinomial logit model to analyze single-vehicle DUI crash data, incorporating variables such as gender and blood alcohol concentration to classify drivers into subgroups. Their findings revealed significant heterogeneity in crash severity outcomes across latent classes. Romano et al. [5] utilized a binomial logit model to examine the contributing factors of fatal DUI crashes, and the results indicated that sociodemographic characteristics such as gender, race, and income level were significantly associated with fatality risk. In addition, Yang et al. [6] proposed a multivariate hierarchical random-parameter correlated-outcome logit model to investigate the key determinants of crash severity in alcohol-related two-vehicle collisions. By introducing a random slope structure, this model further captured the temporal instability in the effects of explanatory variables on crash severity, thereby providing a more flexible framework for modeling DUI crashes under complex scenarios.

Previous studies have also shown that women generally have a higher blood alcohol concentration (BAC) than men, primarily due to physiological differences such as lower body water content and alcohol dehydrogenase levels [7]. Yadav et al. [8] further reported that alcohol has a stronger negative impact on driving performance in women. As BAC levels increase, female drivers exhibit larger increases in speed variability and acceleration variability compared to male drivers. Therefore, examining DUI crash severity across genders is essential. Such analysis can support the development of targeted intervention measures and gender-specific traffic safety strategies, which carry significant public health benefits.

A large body of research has shown significant gender differences in traffic accidents, highlighting the importance of examining the interaction between gender factors and traffic safety [9,10,11]. Although both male and female drivers may be affected by DUI accidents, there are likely significant differences between the sexes in terms of accident occurrence, severity, and underlying causes. A deeper understanding of these gender differences is crucial for developing more targeted intervention strategies and optimizing road safety measures. For instance, Wang et al. [12] investigated the determinants of injury severity among motorcyclists without helmets and emphasized that the influencing factors for male and female drivers are non-substitutable. Moreover, previous studies have modeled gender differences in various crash contexts, including highway collisions [13], speeding-related crashes [14], bicycle crashes [15], and motorcycle crashes [16,17]. However, research on gender differences in DUI crashes remains relatively scarce, and a systematic theoretical framework has yet to be established.

Given the differences between men and women in vehicle control ability after alcohol consumption, it is necessary to develop separate models for each gender. This approach enables more accurate analysis and supports the development of targeted intervention strategies. In this study, a partially constrained random-parameter logit model is used to examine how gender influences the severity of DUI crashes. The model captures gender-related heterogeneity, and out-of-sample prediction is applied to further test differences in crash injury outcomes between male and female drivers. The findings clarify the mechanisms through which gender differences lead to different crash outcomes and provide a solid basis for more refined traffic safety measures. These results can help policymakers consider gender factors more carefully when designing DUI prevention and crash mitigation strategies, thereby improving their effectiveness. In addition, this study highlights the importance of integrating gender perspectives into traffic safety research and contributes to the fields of equity and social science.

2. Methods

2.1. Data

The DUI crash data used in this study were obtained from the Crash Records Information System (CRIS) of the Texas Department of Transportation. Previous studies have shown that the contributing factors to crash severity differ significantly between single-vehicle and multi-vehicle crashes [18,19]. Moreover, compared to multi-vehicle crashes, single-vehicle crashes are often associated with a higher fatality rate among drivers [20]. All variance inflation factor (VIF) values were below the commonly accepted threshold of 10, indicating no serious multicollinearity issues in the model. Therefore, this study focuses on DUI single-vehicle crash data from the year 2021, with the data description provided in

Table 1. Descriptive statistics of key explanatory variables that were significant in the models estimated.

Variables	Female	Male
	Mean (SD)	Mean (SD)
SUV (1 if yes; 0 otherwise)	0.302 (0.459)	0.160 (0.367)
Driver’s blood alcohol content below 0.08% (1 if yes; 0 otherwise)	0.037 (0.189)	0.051 (0.219)
Driver’s blood alcohol content above 0.15% (1 if yes; 0 otherwise)	0.306 (0.461)	0.294 (0.455)
The average daily flow below 20,000 (1 if yes; 0 otherwise)	0.174 (0.379)	0.206 (0.405)
Unsafe speed indicator (1 if driver violation driving at unsafe speed; 0 otherwise)	0.037 (0.189)	0.050 (0.218)
Rollover indicator (1 if vehicle rollover; 0 otherwise)	0.043 (0.203)	0.063 (0.243)
Weekend (1 if yes; 0 otherwise)	0.431 (0.495)	0.485 (0.499)
Night indicator (1 if crash occurred on night (18:01–24:00); 0 otherwise)	0.345 (0.475)	0.363 (0.481)
Darkness with streetlights (1 if yes; 0 otherwise)	0.538 (0.499)	0.479 (0.499)
Vehicle reversed before the crash (1 if yes; 0 otherwise)	0.024 (0.154)	0.023 (0.151)
Vehicle turned before the crash (1 if yes; 0 otherwise)	0.087 (0.281)	0.075 (0.264)
Crash occurred in rural area (1 if yes; 0 otherwise)	0.110 (0.313)	0.145 (0.351)
Clear weather (1 if yes; 0 otherwise)	0.810 (0.393)	0.798 (0.402)
Vehicle model year within 10 years (1 if yes; 0 otherwise)	0.319 (0.466)	0.209 (0.407)
Older driver indicator (1 if driver is over 55 years old; 0 otherwise)	0.066 (0.248)	0.094 (0.292)
Black driver indicator (1 if driver is black;0 otherwise)	0.140 (0.347)	0.129 (0.336)
Driver use restraint (1 if yes; 0 otherwise)	0.773 (0.419)	0.719 (0.449)
Airbag for driver was deployed (1 if yes; 0 otherwise)	0.496 (0.499)	0.428 (0.495)
Hit and run crash indicator (1 if the driver left the crash place, 0 otherwise)	0.050 (0.218)	0.059 (0.236)
High design speed indicator (1 if crash occurred where design speed was ≥60 mph; 0 else)	0.211 (0.408)	0.229 (0.420)
Low design speed indicator (1 if crash occurred where design speed was ≤30 mph; 0 otherwise)	0.326 (0.469)	0.306 (0.461)
Surface conditions indicator (1 if the road surface was non-dry; 0 otherwise)	0.124 (0.329)	0.132 (0.339)
Straight grade indicator (1 if the road grade was straight; 0 otherwise)	0.084 (0.277)	0.085 (0.279)
Curve leave indicator (1 if the road leave was curve; 0 otherwise)	0.134 (0.340)	0.143 (0.350)
Off-road crash (1 if yes; 0 otherwise)	0.649 (0.477)	0.684 (0.464)
Shoulder or median crash (1 if yes; 0 otherwise)	0.120 (0.324)	0.104 (0.305)

(Where mean denotes the average, and SD denotes the standard deviation).

2.2. Random Parameter Logit Model

The random-parameter logit model is an extension of the multinomial logit model, which is capable of capturing the heterogeneity of factors influencing crash severity. Its utility function is expressed as [21]:

$$U_{jn}={\beta }_j{X}_{jn}+{\epsilon }_{jn}$$

(1)

Here, $U_{jn}$ denotes the utility function that determines the injury severity level $j$ of driver $n$ in a crash; $X_{jn}$ represents the explanatory variables (factors influencing crash severity); and ${\beta }_j$ denotes the estimated parameters, and ${\epsilon }_{jn}$ is a disturbance term. To account for potential heterogeneity, If the error term is assumed to follow a generalized extreme value distribution, the result is a standardized polynomial Logit:

$$P_n(j)={\displaystyle \frac{EXP({\beta }_j{X}_{jn})}{{\displaystyle \sum _{\forall J}EXP({\beta }_j{X}_{jn})}}}$$

(2)

Here, $P_n(j)$ denotes the probability that collision $n$ will result in collision severity $j$. To accommodate the possibility that parameters are random and vary between observations in vector ${\beta }_j$, Equation (2) is modified as follows:

$$P_n(j)={\displaystyle \int \frac{EXP({\beta }_{jn}{x}_{jn})}{{\displaystyle \sum _j^{I}EXP({\beta }_{jn}{x}_{jn})}}}f(\beta |\phi )\mathrm{d}\beta$$

(3)

Here, $f(\beta |\phi )$ denotes the probability density function of ${\beta }_j$, and $\phi$ represents the parameter vector describing the density function and other related terms.

To illustrate the potential variation in the mean and variance of random parameters. ${\beta }_n$ is defined as a vector of parameters that vary across crash $n$ [22,23]:

$${\beta }_n={\beta }_{jn}+{\gamma }_{jm}{Z}_{jm}+{\sigma }_{jm}\exp ({\omega }_{jm}{W}_{jm}){\delta }_m$$

(4)

In the equation, ${\beta }_j$ represents the mean of the estimated parameters across crashes, $Z_{jm}$ denotes the mean heterogeneity captured for injury severity level $j$, and $W_{jm}$ reflects the variance heterogeneity in the influence on crash severity. ${\delta }_m$ is the disturbance term, and $m$ represents all crash-related factors included in crash $n$.

In addition, the average marginal effects of crash injury severity were calculated to better assess the impact of significant variables in the model on the probability of driver injury outcomes [24].

$$P_{X_k}^{P(i)}={\displaystyle \frac{1}{N}}{\displaystyle \sum _{n=1}^N(P_n(i|X_{nk}=1)-P_n(i|X_{nk}=0))}$$

(5)

In the equation: $P_{X_k}^{P(i)}$ denotes the probability of crash severity level $i$ when the $k_{th}$ explanatory variable changes while holding all other variables constant; $X_{nk}$ represents the value of the $k_{th}$ explanatory variable in crash $n$; $P_n(i|X_{nk}=1)$ indicates the probability of crash severity level $i$ for crash $n$ when $X_{nk}=1$ with all other variables held constant; similarly, $P_n(i|X_{nk}=0)$ represents the probability of crash severity level $i$ for crash $n$ when $X_{nk}=0$ with all other variables held constant.

Lastly, the estimation approach accounts for the potential gender-specific differences in DUI crash severity by applying both constrained and unconstrained modeling frameworks. To effectively identify and control for variations in parameter effects between male and female drivers, a partially constrained modeling approach was employed. In this approach, the full DUI crash dataset—combining data from both genders—was used, and statistical tests were conducted to examine whether specific parameters significantly differ across genders. First, separate unconstrained models were estimated for male and female drivers, and the results are presented in the Appendix A. Next, a series of likelihood ratio tests was performed to compare the constrained and unconstrained parameters for each variable, thereby identifying parameters that exhibit significant gender differences.

$$X^2=-2[LL({\beta }_C)-LL({\beta }_U)]$$

(6)

Here, $LL({\beta }_U)$ denotes the log-likelihood when the model converges with unconstrained parameters, and $LL({\beta }_C)$ denotes the log-likelihood when the model converges with constrained parameters (when specific variables are allowed to be constrained and share identical parameter values among females). Where $LL({\beta }_U)$ is the log-likelihood when the model converges with unconstrained parameters (meaning all parameters for specific variables are allowed to vary by gender), and $LL({\beta }_C)$ is the log-likelihood when the model converges with constrained parameters. The obtained $X^2$ value follows a chi-squared distribution. This statistic is used to decide whether to reject the null hypothesis that the constrained and unconstrained parameter values are equal. If the null hypothesis is rejected, the unconstrained model is considered statistically superior, and the separate parameters for each gender are retained. If the null hypothesis cannot be rejected, the constrained parameter values remain in the model. In the final model estimation, a likelihood ratio test is applied to each statistically significant variable. The resulting $X^2$ values show that, for all included variables, the null hypothesis can be rejected with more than 90% confidence.

3. Results

The study used a likelihood ratio test to validate the modeled differences in DUI crash severity between males and females with the following test formula:

$$X^2=-2[LL({\beta }_{FM})-LL({\beta }_F)-LL({\beta }_M)]$$

(7)

where $LL({\beta }_{FM})$, $LL({\beta }_F)$, $LL({\beta }_M)$ denote the log likelihood at convergence for the full model, female model and male model, respectively. The results of the test showed $X^2$ = 26.6 with a degree of freedom of 13. Therefore separate estimation was performed for the different gender models (Appendix A).

In addition, the analysis indicates that no definitive conclusion can be drawn regarding whether there is a significant difference between the unconstrained model and the partially constrained model. This conclusion is supported by the likelihood ratio test results, which suggest that the null hypothesis, assuming the equality of the partially constrained and unconstrained models, cannot be convincingly rejected. The test statistic for this hypothesis is as follows:

$$X^2=-2[LL({\beta }_{PC})-LL({\beta }_{UF})-LL({\beta }_{UM})]$$

(8)

Here, $LL({\beta }_{UF})$ and $LL({\beta }_{UM})$ represent the log-likelihood values at convergence for the unconstrained parameter model, and $LL({\beta }_{PC})$ is the log-likelihood at convergence for the partially constrained parameter model. The value of this statistic is 5.74 with 3 degrees of freedom, which means that the null hypothesis assuming the equality of the unconstrained and partially constrained models can only be rejected with 88% confidence. Therefore, the partially constrained model will be retained and used to interpret the results and provide insights. The results of the partially constrained model and the marginal effects are presented in

Table 2. Partially Constrained Random-Parameter Logit Model for DUI Crash Injury Severity for Female [F] and Male [M] Drivers.

Variable Description	Parameter	z-Value
Defined for minor injury
Constant [F]	−1.1132	−17.15
Constant [M]	−1.3800	−10.76
SUV [M]	0.1312	1.88
Driver’s blood alcohol level below 0.08% [M]	0.3365	2.93
Weekend [M]	−0.2029	−3.88
Vehicle rollover [F]	0.7449	6.98
Vehicle rollover [M]	0.6988	2.77
Vehicle reversed before the crash [M]	−1.4557	−4.63
Vehicle turned before the crash [F] [M]	−0.3706	−3.94
Rural area [F] [M]	0.5129	7.71
Vehicle model year within 10 years [F] [M]	−0.2718	−4.80
Drivers over 55 years of age [M]	0.4033	4.62
Black driver indicator	−0.2202	−2.75
Driver use restraint [F]	−0.8939	−2.77
Standard deviation of driver use restraint indicator	0.9365	1.74
Driver use restraint [M]	−0.6011	−10.19
Airbag for driver was not deployed [F]	1.3603	8.59
Airbag for driver was not deployed [M]	1.1045	20.43
Driver left the crash place [F]	−0.8033	−2.93
Road grade was straight [F]	0.3863	2.31
Defined for severe injury
Constant [F] [M]	−3.1261	−11.91
Pickup [M]	0.4074	3.33
Driver’s blood alcohol content below 0.08% [M]	0.8169	3.37
Driver’s blood alcohol content above 0.15% [F] [M]	0.5963	5.71
The average daily flow below 20,000 [F] [M]	0.5341	4.60
Driver violation driving at unsafe speed [F]	1.5562	4.92
Vehicle rollover [F]	1.3675	4.00
Vehicle rollover [M]	1.8052	4.62
Standard deviation of vehicle rollover indicator	4.0085	3.05
Darkness with streetlights [F] [M]	−0.5935	−5.43
Night [M]	0.2796	2.26
Vehicle turned before the crash [F] [M]	−1.3274	−4.63
Clear weather [F]	−0.5879	−2.95
Clear weather [M]	−0.4268	−2.88
Vehicle model year within 10 years [F] [M]	−0.8196	−5.85
Drivers over 55 years of age [M]	0.4969	2.69
Driver use restraint [M]	−2.1047	−12.78
Airbag for driver was deployed [F]	1.6694	8.50
Airbag for driver was deployed [M]	0.9036	3.65
Standard deviation of airbag for driver was not deployed indicator	2.1034	5.23
Road design speed was ≤30 mph [F] [M]	−1.0100	−6.61
Road design speed was ≤60 mph [M]	0.4937	3.77
Road surface was non-dry [F] [M]	0.4822	2.78
Road leave was curve [F]	−0.5880	−2.95
Road leave was curve [M]	−0.4268	−2.88
Off-road crash [F] [M]	1.1723	5.75
Shoulder or median crash [F]	1.2985	4.06
Shoulder or median crash [M]	1.7146	5.81
Heterogeneity in the mean of the random parameters
Driver use restraint: Vehicle rollover [M] [SI]	−3.7954	−2.44
Shoulder or median crash: Airbag for driver was not deployed [M] [SI]	0.6375	1.95
Driver use restraint: Airbag for driver was not deployed [M] [SI]	−0.7006	−2.13
Road surface was non-dry: Driver use restraint [F] [MI]	0.5656	2.38
Clear weather: Driver use restraint [F] [MI]	−0.4320	−2.30
Off-road crash: Driver use restraint [F] [MI]	0.3539	2.54
Model Statistics
Log-likelihood at zero LL(0)	−13,427.24
Log-likelihood at convergence LL(β)	−8515.90
McFadden Pseudo R-squared 1-LL(β)/LL(0)	0.3558
Number of observations	12,222

and

Table 3. Marginal effects results.

Description	Female			Male
	NI	SI	MI	NI	SI	MI
SUV				−0.0025	0.0027	−0.0002
Pickup				−0.0034	−0.0017	0.0051
Driver’s blood alcohol level below 0.08				−0.0030	0.0016	0.0014
Driver’s blood alcohol content above 0.15	−0.0061	−0.0031	0.0093	−0.0061	−0.0031	0.0093
The average daily flow below 20,000	−0.0047	−0.0031	0.0078	−0.0047	−0.0031	0.0078
Driver violation driving at unsafe speed	−0.0012	−0.0007	0.0019
Vehicle rollover	−0.0018	0.0006	0.0012	−0.0085	0.0040	0.0045
Weekend				0.0104	−0.0115	0.0011
Night				0.0020	0.0010	−0.0030
Darkness with streetlights	0.0063	0.0032	−0.0095	0.0063	0.0032	−0.0095
Vehicle reversed before the crash				0.0012	−0.0013	0.0001
Vehicle turned before the crash	0.0051	−0.0034	−0.0017	0.0051	−0.0034	−0.0017
Rural area	−0.0107	0.0132	−0.0025	−0.0107	0.0132	−0.0025
Clear weather	0.0073	0.0038	−0.0111	0.003	0.0017	−0.0047
Vehicle model year within 10 years	0.0135	−0.0082	−0.0053	0.0135	−0.0082	−0.0053
Older driver				0.0057	0.0045	0.0012
Black driver				0.003	−0.0033	0.0003
Driver use restraint	0.0144	−0.003	−0.0114	0.0712	−0.0421	−0.0291
Airbag for driver was deployed	−0.0366	0.0279	0.0087	−0.0873	0.0570	0.0303
Hit and run crash	0.0009	−0.0010	0.0001
High design speed				−0.0038	−0.0022	0.0060
Low design speed	0.0040	0.0020	−0.0060	0.0040	0.0020	−0.0060
The road surface was non-dry	−0.0124	−0.0068	0.0192	−0.0124	−0.0068	0.0192
The road grade was straight	−0.0013	0.0015	−0.0002
The road leave was curve	−0.0011	−0.0006	0.0017	−0.0019	−0.0011	0.0030
Off-road crash	−0.0267	−0.0149	0.0415	−0.0267	−0.0149	0.0415
Shoulder or median crash	−0.0016	−0.0008	0.0025	−0.0051	−0.0025	0.0077

Notes: NI = No Injury (no physical harm to the driver); SI = Severe Injury (life-threatening or serious injuries); MI = Minor Injury (non-serious injuries). The column order is NI/SI/MI.

.

3.1. Variables Producing the Same Parameter Values Across All Genders

Partial Constraint Model Results are shown in Table 2. Table 3 presents the marginal effects. The calculation of marginal effects in this study follows the approach of Song et al. [25], in which the marginal effects of partially constrained parameters and unconstrained parameters were estimated separately based on different samples. The results indicate that vehicle turning prior to a crash is statistically significant in both the male and female models, with consistent parameter estimates. The marginal effect analysis shows that this factor increases the probability of no injury (0.0051). This may be attributed to the fact that vehicles generally travel at lower speeds while turning, thereby reducing collision intensity and lowering crash severity [26]. In addition, a driver’s BAC greater than 0.15% increases the probability of severe injury (0.0093), which can be explained by alcohol’s impairment of cognitive processing and the prolongation of driver reaction times [27]. Furthermore, newer vehicle models increase the probability of no injury (0.0134) across both gender-specific models. This finding may be attributed to continuous technological advancements that have effectively enhanced overall vehicle safety performance [28].

Speed plays a critical role in determining crash severity [29,30]. The results show that when crashes occur on roads with speed limits below 30 mph, the probability of severe injury decreases by 0.006. This can be attributed to the reduction in kinetic energy at lower speeds, which mitigates crash consequences. Under dark conditions, the presence of lighting significantly reduces the likelihood of severe injury crashes, likely because street lighting improves visibility and thereby lowers crash severity [31]. At the same time, when the average traffic volume is below 2000 vehicles, the probability of severe injury increases by 0.0078, possibly because drivers under the influence of alcohol find it more difficult to maintain appropriate speeds on such roads, thereby elevating the risk of severe injury [32]. Finally, crashes occurring off-road are more likely to result in severe injuries. This is primarily due to the lack of buffering facilities and effective safety protection designs in such areas, which amplifies collision forces and structural damage, thus significantly increasing crash severity [33].

3.2. Variables Producing Random Parameters

In the female driver DUI crash model, the results indicate that driver use restraint is a random parameter. According to the parameter distribution, driver use restraint has an 83.01% probability of reducing the occurrence of minor injuries when a crash occurs. This finding is consistent with previous research, which has shown that driver use restraint significantly reduces the risk of injury in traffic accidents, further underscoring its critical role in traffic safety [34]. Meanwhile, the mean heterogeneity results reveal that this probability decreases to 63.7% and 71.79% when crashes occur on non-dry road surfaces and off-road locations, respectively. This may be because such environments are more complex, making it more difficult for impaired drivers to maintain effective vehicle control, thereby increasing the likelihood of minor crashes [35]. In addition, the mean heterogeneity results show that under clear weather conditions, the probability of minor injury decreases further. This can be attributed to improved visibility, which allows drivers to better identify potential hazards, thereby mitigating crash consequences to some extent [36].

In the male driver DUI crash model, two random parameters were identified. First, the results show that vehicle rollover significantly increases the probability of severe injury, while the marginal effects indicate that it notably decreases the probability of no injury (−0.0085). This phenomenon can be attributed to the fact that rollovers often cause severe deformation of the vehicle structure, subjecting occupants to greater impact forces and thereby aggravating injury severity [37,38]. Meanwhile, the mean heterogeneity results demonstrate that driver use restraint effectively reduces injury severity.

In addition, the deployment of airbags during a crash is associated with a 66.62% probability of severe injury. This finding is intuitive, as airbag activation typically indicates a high-impact collision [39]. The mean heterogeneity result further reveals that when driver use restraint is present, the severity of injuries is significantly reduced. Finally, crashes occurring on the shoulder or median are associated with greater injury severity. This may be explained by the prevalence of rigid roadside objects (e.g., guardrails, utility poles), which increase the likelihood of severe consequences when struck [40].

3.3. Variables Producing Fixed Parameters

With respect to driver characteristics, the model reveals that male drivers with a BAC below 0.08% still face an increased probability of severe injury. This finding further highlights that even low BAC levels can pose significant threats to traffic safety. Driver age is also identified as an important factor influencing injury severity. The marginal effects indicate that when male drivers are older than 55 years, the probability of severe injury increases by 0.0012. This result may be explained by age-related declines in reaction ability and increased physical vulnerability, which make older drivers more susceptible to severe injury in crashes [41]. In addition, the model shows that hit-and-run behavior is significant for female drivers, reducing the likelihood of minor injury. This suggests that when only minor crashes occur, female drivers are less likely to leave the crash scene. Furthermore, crashes involving male Black drivers are associated with a lower probability of minor injury.

With regard to vehicle characteristics, the model results indicate that for male drivers, SUVs increase the likelihood of minor injury crashes, while pickup trucks increase the probability of severe injury crashes. Vehicle rollovers increase the probability of severe injury for both male and female drivers. In contrast, crashes occurring during reversing reduce the probability of minor injuries. Finally, for female drivers, operating a vehicle at unsafe speeds increases the probability of severe injury by 0.0019, a finding consistent with previous research conclusions [42].

Regarding roadway characteristics, it was observed that female drivers are more likely to experience minor injuries when crashes occur on sloped road segments. Crashes occurring on curves are statistically significant for both genders and increase the probability of severe injury; however, model results indicate that males are at greater risk than females in curve-related crashes. This may be due to the increased difficulty of vehicle operation on complex road segments, especially under the influence of alcohol, which further reduces vehicle control and elevates both crash risk and severity [43]. Moreover, men’s more aggressive driving style amplifies this risk [44]. In addition, non-dry roadway surfaces significantly increase the probability of severe injury. For impaired drivers, slippery or poor road conditions make it more difficult to maintain control, thereby aggravating crash severity. Male drivers are also more likely to sustain severe injuries in crashes occurring on roads with speed limits exceeding 60 mph. Finally, crashes occurring on the shoulder or median are associated with higher injury severity for both genders.

As for environmental characteristics, crashes occurring at night increase the severity of crashes for male drivers, which may be related to reduced visibility and more complex traffic conditions during nighttime driving. Male drivers are also more likely to be involved in minor injury crashes during weekends, possibly due to increased alcohol consumption on those days [45]. Furthermore, crashes in rural areas are statistically significant in both gender-specific models and increase the probability of minor injuries. Favorable weather conditions, on the other hand, improve visibility and thus reduce the probability of severe injuries for drivers of both genders.

3.4. Out-of-Sample Prediction

For the out-of-sample prediction, we applied the estimated parameters from the male DUI crash model to the observed female crash data in order to predict the resulting injury severities. We then compared these predictions with those generated from the female crash model using the same parameters and data. This out-of-sample prediction approach is necessary as it involves estimating injury probabilities by applying parameters obtained from a different sample. To provide a deeper understanding of this technique, Hou et al. [46] offered a comprehensive explanation, discussion, and empirical evaluation. The relevant results are presented in

Table 4. Average Difference in DUI Collision Injury Probability Estimated Using Female Data and Male Model Parameters.

No Injury	Minor Injury	Severe Injury
−0.0363	0.0247	0.0116

.

According to the out-of-sample prediction results, when the estimated parameters from male drivers are applied to female crash data, the probability of no injury decreases by 0.0362, while the probabilities of minor injury and severe injury increase by 0.0247 and 0.0116, respectively. Male drivers are generally characterized by more aggressive driving styles and are more prone to high-risk behaviors such as driving under the influence of alcohol. However, systematic empirical evidence is still lacking regarding the physiological and behavioral differences between men and women and how these differences specifically affect crash injury severity. Further research is therefore warranted.

4. Discussion and Conclusions

Due to physiological and psychological differences, men and women show distinct driving performance under the influence of alcohol. This study investigates how these differences affect DUI crash injury outcomes and provides practical insights for policymakers to better account for gender factors. Using 2021 DUI crash data from Texas, a random-parameter logit model is estimated, and a partially constrained modeling approach is applied to capture gender-related variations in crash severity. Out-of-sample prediction is then used to more accurately identify and evaluate these gender differences.

The model results show that some variables are consistently significant in both female and male crash injury models. These include vehicle turning behavior, vehicle model year, and driver BAC levels above 0.15%. However, the impact of alcohol involvement differs between genders, and the underlying causes of DUI crashes emphasize different factors for male and female drivers. Therefore, these gender differences should be carefully considered when designing future traffic safety interventions and developing related policies [47].

The partially constrained model identified three random parameters. Among these, driver restraint use was found to significantly reduce crash severity. This result aligns with previous studies highlighting the crucial role of restraints in improving driver safety and further confirms their effectiveness in alcohol-impaired driving situations. Therefore, future traffic safety policies and interventions should place greater emphasis on promoting restraint use, especially in DUI contexts, to reduce crash-related injuries [48].

Vehicle type plays an important role in determining crash severity. This study finds that DUI crashes involving large vehicles, such as SUVs and pickup trucks, are more likely to result in severe injuries. This is because alcohol consumption impairs drivers’ ability to control the vehicle, while larger vehicles have greater inertia and more complex handling characteristics than regular passenger cars, thereby amplifying crash risks. Therefore, future research and policy initiatives should prioritize large-vehicle drivers as a key group for monitoring and intervention to effectively reduce the severity of DUI crashes.

Curved and sloped road segments significantly increase the likelihood of severe crash injuries. For alcohol-impaired drivers, these complex roadway environments further reduce vehicle control, leading to higher crash risk and severity. To address this, enhanced monitoring and inspection should be implemented on high-risk segments to lower crash probabilities. Additionally, adequate roadway lighting under dark conditions can improve drivers’ visibility and perception, thereby mitigating crash severity. Therefore, optimizing and expanding lighting facilities is recommended to enhance driving safety and reduce the severity of alcohol-related crashes [49].

In addition, the study finds that in the male crash model, even when BAC is below the legal threshold of 0.08%, the probability of severe crashes still increases significantly. This suggests that the current BAC limit may underestimate the risks posed by low-level alcohol impairment. Therefore, strengthening the regulation of alcohol-impaired driving and, where appropriate, considering a reduction in the legal BAC threshold could help mitigate severe traffic crashes attributable to alcohol consumption.

It should be noted that this study has certain limitations. First, the analysis is based solely on crash data from the year 2021, without accounting for the dynamic variations over time. In recent years, temporal instability has become a critical issue in crash research, and ignoring temporal evolution may introduce bias and weaken the effectiveness of policies aimed at mitigating crash severity [50]. To address this, future studies should employ partially constrained modeling approaches to capture parameter variations over time and further investigate gender-related differences across different temporal contexts.

Second, this study only considers single-vehicle DUI crashes and does not include multi-vehicle crashes. This limitation may restrict the external applicability of the findings. Future research should broaden the data scope to incorporate different crash types, thereby enhancing the generalizability of the conclusions and their relevance for policy development.

Finally, as this study relies on historical crash data, it may not fully capture the real-time risks associated with driving [51]. Future research should therefore integrate more diverse data sources, including traffic conflict indicators and dynamic monitoring information, to improve the identification and assessment of DUI-related risks and enhance the timeliness and effectiveness of intervention strategies.