According to data released by China’s two largest online delivery platforms, the total number of registered DRs has exceeded 5.7 million (9, 10). Official progressions estimate that the number of Chinese food DRs will surpass 10 million in the near future (11). These DRs broadly categorized into two types: Zhuansong DRs and Zhongbao DRs (12). Zhuansong DRs are formally employed under labor contracts with platform companies and are subject to direct daily management at designated work sites. In contrast, Zhongbao DRs voluntarily register with the platforms to provide delivery services. They operate more independently, accepting and fulfilling orders at their discretion (13).

Despite differences in employment status and management, both Zhuansong and Zhongbao DRs follow similar workflows, which generally include three phases: waiting, receiving/picking up, and delivering orders. First, riders must open the platform’s mobile app to log in and go online, entering a waiting phase during which the system assigns them orders. Zhuansong DRs are required to meet a minimum daily online time, while Zhongbao DRs are not subject to such a requirement.

Second, once a consumer places an order, the platform allocates it to a DR based on several factors, including the rider’s proximity to the customer and merchant, current workload, historical rejection rate, and external conditions such as weather and traffic. Zhuansong DRs are not permitted to refuse assigned orders, although they may transfer up to three orders per day to other riders. Zhongbao DRs, on the other hand, can reject assigned orders; however, each rejection is recorded by the system. As the refusal rate increases, the number of future assigned orders decreases. If an order is not accepted promptly, the system issues multiple reminders. Persistent inaction leads to reassignment of the order and a penalty for the DR.

Orders rejected by DRs are returned to a shared “order pool” accessible to all riders. DRs refer to this process of getting orders as “snatching orders.” Both riders waiting for orders and those currently delivering may participate in the “game of snatching,” where speed is the sole determinant of success (43). This competitive process contributes to the reasons foods DRs deliver get involved in accidents. Typically, most orders for Zhuansong DRs are system-assigned, whereas Zhongbao DRs rely more heavily on the snatching mechanism.

Third, the final step involves picking up and delivering the food. Most platforms allocate approximately 30 min for this process, from the time the order is accepted to when it is delivered to the customer. This time frame is often tight, especially considering riders must wait for merchants to prepare the food before pickup. The stereotypical image of a DR is someone speeding through traffic while a mobile phone repeatedly warns: “You are out of time.”

In addition to these standard procedures, DRs often handle multiple orders simultaneously. Their common strategy is to accept orders, align geographically, or even merge deliveries to increase efficiency and maximize income.

After an order is completed, platforms require customers to evaluate the DR’s performance based on service quality and whether the delivery exceeded the expected time. Ratings typically range from “very dissatisfied” to “very satisfied.” Delivery riders receive rewards based on customer evaluations and additional performance indicators. Platforms generally offer two types of rewards: cash and virtual points (14). Cash rewards commonly include attendance bonuses (usually for Zhuansong DRs who work for a minimum number of days per month), delivery volume bonuses (for completing certain orders within a specific timeframe), and subsidies for adverse weather conditions.

Virtual points, on the other hand, are linked to a DR’s “level,” which is directly related to their income. Higher-level DRs receive preferential treatment, such as priority access to high-value orders and higher commission rates. Table 1 presents DRs’ levels, commissions received from the platform, and corresponding points required for each level on a Chinese food delivery platform. Generally, points are credited to their virtual account each time a DR completes a delivery. Additional points can be earned through positive customer evaluations. Mirroring the reward system, there are also two categories of penalties. For instance, rejecting an order results in a fine, while receiving negative reviews or delivery orders late leads to point deductions.

This game-like, incentive-driven work system has led many DRs to become deeply engaged, sometimes to the point of overwork, unintentionally extending their working hours. Data shows that in 2018, the average daily working time of DRs in Beijing exceeded 10 h, while in Chengdu, average working hours ranged from 9 to 10 h per day in 2021 (15). As a result, the so-called flexibility associated with gig work may, in practice, constitute a form of “false freedom.” These workers are often described as being “stuck in algorithms.”

The well-established JD-R model is commonly employed to examine the effects of work-related pressures on DRs (13, 16). According to the JD-R model, all job characteristics can be classified into two categories: job demands and job resources. Job demands require individuals to exert sustained effort, often resulting in negative outcomes, whereas job resources can help mitigate these effects (17). The model suggests two primary pathways linking job characteristics to employee outcomes: (i) the “health impairment” or negative path assumption and (ii) the “buffer” assumption (17–19).

First, the “negative path” refers to the adverse health effects caused by excessive job demands, often resulting in burnout (20, 21). Several earlier studies on DRs support this assumption. For example, Zheng et al. (22) found that high workloads were significantly associated with traffic crashes. More recent studies have reinforced this view (23). For example, Chen (16) demonstrated that job overload and time pressure were positively associated with job stress among DRs, while self-efficacy helped reduce this impact. Drawing from the Conservation of Resources (COR) theory, Quy Nguyen-Phuoc et al. (24) found that a DR’s perceived control over their work environment played a significant role in the relationship between job demands and burnout. This led to the extension of the JD-R model to incorporate personal demands and resources.

Second, the “buffer” assumption proposes that job resources can attenuate the negative impact of job demands on employee well-being. Prior research has shown that when employees receive social support or timely feedback, job overload is less likely to lead to burnout (21). This buffering effect has also been observed among DRs. Studies using DRs as research subjects confirmed that job resources significantly moderated the relationship between job demands and job burnout (24).

Previous studies have established a link between job demands, job resources, and job outcomes among DRs, generally validating the JD-R model. However, most existing studies have focused on occurrence of injuries such as risky driving behavior or distraction. It remains unclear whether the JD-R model can also explain variations in physical health outcomes. Some recent studies have sought to extend the JD-R model to explore health-related issues targeting formal employed workers. For instance, Lehmann et al. (25) examined workers with chronic health conditions and discovered that job resources decrease the likelihood of illness, while job demands contribute to job burnout and functional limitations. Additionally, some researchers have utilized multiple health indicators, encompassing subjective health (26), fatigue (27), psychological health, and disease risks (28). However, whether this model is also applicable when explore gig workers’ health issues remain unclear. This study aims to address this gap by extending the JD-R model to explore the association between workloads and health status among DRs, thus contributing to the theoretical development of the model.

Specifically, the present study first tests the “negative path” assumption by investigating the correlation between workload and health outcomes among DRs. Following the extended JD-R framework (24), it then examines the underlying mechanisms of this association. As exposure to high-risk working environments may lead to physical fatigue and emotional stress, this study proposes that perceived risk serves as a mediator. Two potential pathways are hypothesized: (i) increased risk perception may prompt DRs to take more safety precautions, thereby reducing their likelihood of illness. (ii) Conversely, elevated perceived risk may trigger anxiety, which in turn contributes to physical health problems.

Finally, the study tests the “buffer” assumption by assessing whether organizational protection measures can alleviate the negative impact of high workload on the health of DRs.

Figure 1 illustrates the conceptual framework of this study. Based on this framework, the following hypotheses are proposed:

In summary, this study shares some similarities with previous research in that it is also based on the JD-R model to verify the negative impact of workloads on employee health and safety outcomes (13, 16, 17). For example, in consistent with Chen (16) and Zheng et al. (22), this study extends the JD-R model by emphasizing the mediating role of personal risk perceptions. Going beyond previous literature, this study yields some new results: first, existing literature has predominantly focused on occupational burnout (21) or accident risks (22, 29), whereas this research integrates the physiological-psychological association mechanism (H2). Second, this study employs a multidimensional perspective. Previous research has often been based on a single level. This study introduces “platform protection requirements” as a moderating variable (H3), advocating for the necessity of institutional interventions and comprehensively analyzing the influence mechanisms at both the individual and organizational levels.

Previous studies on DRs in China have commonly employed the road intercept sampling method due to its efficiency in collecting data over a short period. However, this non-randomized approach can introduce potential bias in the dataset. To address this limitation, the study adopted the respondent-driven sampling (RDS) method to improve the accuracy and representativeness of the sample and to better reflect the overall characteristics of the DR population (30).

To ensure sample quality, RDS analysis tool software was used to conduct an equilibrium sample distribution test on the survey data (31). Following established procedures, the weighted mean absolute discrepancies between the actual ( P s ) and equilibrium sample compositions ( P e ) were used to examine how the actual sample compositions approximated the theoretically computed equilibrium sample compositions ( { P s − P e } < 0.02 ).

The representativeness of the RDS sample was further evaluated by comparing the final RDS sample compositions with asymptotically unbiased estimates of the population composition (t-test for P s − P ¯ non-significant). The results indicated that the sample used in this survey showed good representativeness.

This study employed a structured questionnaire comprising six main sections: (1) socio-demographic information (e.g., age, education), (2) work characteristics (e.g., job type, platform), (3) insurance coverage, (4) occupational injury (e.g., type, number, reason, etc.), and (5) health status (self-rated health, illness, etc.).

A team of 18 trained graduate and undergraduate students recruited 0-round respondents (marked as “seeds”) at the DR’s gathering points. These seeds were selected through one-on-one interviews conducted by trained investigators. The seeds referred to 1^st-round respondents after finishing their questionnaire; then, 1st-round respondents referred the next round of respondents. The investigators reviewed each respondent to meet the requirements of the RDS. These steps are repeated to form long chains.

To comply with research ethics standards, investigators clearly explained the study’s purpose to all participants. Respondents were informed that participation was voluntary, responses would remain anonymous, and all data would be kept confidential. Before completing the questionnaire, each respondent was required to provide informed consent via a written consent form.

To compensate for their time, each received an incentive of CNY 20 (CNY 20 = USD 3) after their questionnaire was reviewed by the investigator. After the recommended respondents had completed their questionnaires and the completed questionnaires were reviewed, the recommender received an additional reward of CNY 5 (CNY 5 = USD 0.7) (for each successful referral). To control data bias, the investigators reviewed each questionnaire by “daily working hours” and “daily delivery” items and excluded invalid ones.

To ensure the integrity of the referral process and prevent manipulation, such as repeated submissions or referring unrelated individuals, each respondent was permitted to refer a maximum of five other participants. Respondents were assigned a unique identification code and chain number to track referral paths. The final effective sample size for this study was 1,092 participants.

The data analysis was conducted in three main steps. First, preliminary model assessments were performed to check for redundancy among variables, including a one-sample t-test, chi-square test, and multicollinearity test.

Second, benchmark regressions were run to examine the relationship between workload and DRs’ health. Given that the dependent variable “illness” is binary, this study used a binary Logit regression model constructed as follows: (1) Illnes s i ∗ = α + βworkload i + γZ i + ε i where Illness i is whether rider I has ever experienced illness as a result of his work, workload i (main independent variable) is rider i’s workload regarding daily delivery (DD) and daily working hours (DWH), Z i is a vector of covariates listed in the section below, and ε i is the error term.

Third, the study further investigated the mechanisms behind the main correlation. As stated above, this study investigated the mediation role of personal demands between workloads and DRs’ health by using risk perception as a mediator. The study further explored the moderation mechanism by adding organizational-level measures to the main regression. All analyses were performed using Stata version 16.0.

Table 2 presents descriptive statistics for the study sample. Among the 1,092 DRs, the vast majority were males (88.5%). The sample was relatively young, with an average age of 29.875 years (std = 7.072). Most riders were unmarried (53.3%) and employed full-time (76.1%). Additionally, a majority (68.9%) did not hold a college degree.

On average, riders delivered 39 orders per day and worked approximately 10 h daily. Nearly half of the respondents (around 50%) reported experiencing an illness related to their work as DRs.

Figure 2 provides further detail on the types of illnesses reported. Among respondents who experienced work-related illnesses, the most frequently cited conditions were gastroenteropathy (20.79%), COVID-19 (19.05%), and sunstroke (18.96%), followed by lumbar spondylosis (17.58%) and cervical spondylosis (16.21%).

The results of the one-sample t-test and chi-square test indicated that DD, DWH, age, and number of underaged children were significantly correlated with illness, at a significance of p < 0.05. Similarly, the chi-square tests for main supporter and marital status were significant (p < 0.05). Additionally, the results of sex (p = 0.366) and education (p = 0.494) were not significant.

For multicollinearity diagnostics, it was found that when moving the education variable from the regression model, the variance inflation factor (VIF) values of all the variables were lower than 10, and the average VIF value of the model was 1.61, meaning that it passed the test. While when the education variable was included in the regression, the VIF value reached as high as 35, with an average VIF value of the model at 9.44 (this study adopted a strict threshold, considering VIF > 5 as high risk). Common strategies for dealing with high VIF values include: removing variables (specifically, removing the variable with the highest VIF); merging variables (using PCA or factor analysis to transform related variables into new variables, although this approach is not applicable to the “education” variable in our study); and increasing the sample size (which was not feasible in this study). Comprehensively, since sex was a critical demographic variable, it was included in the final regression model while education was moved out.

As shown in Tables 3, 4, there was a significant positive correlation between DD and illness ( β = 0.010, SE = 0.004, p < 0.05 ). Regarding the marginal results, with the other variables held constant, the likelihood of suffering from illness increased by 0.2% when DD increased by one unit. DWH is also positive and significantly correlating with illness ( β = 0.051, SE = 0.025, p < 0.05 ). When controlling for other variables, the likelihood of suffering from an illness increased by 1.2% for each additional working hour.

These results indicate that increased workload, both in terms of delivery volume and working hours, is significantly linked to greater health risks among delivery riders. Thus, the findings provide empirical support for Hypothesis 1.

The bootstrap analysis was used to test the mediation effect of perceived risk in the relationship between workloads and illness, and results showed that the results of both DD and DWH were significant (DD: β = 0.001 , SE = 0.000 , p < 0.01 ; DWH: β = 0.008 , SE = 0.002 , p < 0.01 ). Results indicated that a full mediation effect exists between DRs’ workloads, level of risk perception, and illness.

The proportion of mediation effect when using DD and DWH as independent variables were 0.378 and 0.681, respectively. To test the robustness of the results, this study further used Karlson–Holm–Breen (KHB) analysis to test the mediation effect as the illness was treated as a binary dummy variable. As shown in Tables 5A,B, the mediation effects of both DD and DWH were significant when using perceived risk as a mediator. These results confirm that a delivery of a DR’s level of perceived risk mediates the relationship between workload and health. Therefore, Hypothesis 2 is supported.

As shown in Table 5D, extra platform requirements ( β = − 0.195 , SE = 0.067) limiting overwork tendency may significantly influence the path between DWH and illness. In the case of DD (Table 5C), the platform requirements did not significantly affect the likelihood of illness. Although not significant, the results for delivery also showed a negative pattern. Hence, both results suggest that a platform limiting riders’ overworking behaviors indirectly aids in alleviating the influence of workloads on illness. Therefore, Hypothesis 3 was partly verified.

Figure 3A decomposes the association between DD and illness based on whether the rider is the main financial supporter of the family. The first and second coefficients correspond to riders who are and are not the primary income earners, respectively. The figure reveals that riders, who are the main source of income for their families, are more susceptible to illness due to work overload.

Figure 3B highlights that workload has less of an effect on DRs who are working as full-time riders but leads to higher risks of illness among counter groups.