We utilize data from the China Family Panel Studies (CFPS) 2012–2020. The analysis of restricted city-level data in this article was conducted in the restricted data room of the Center for Chinese Social Science Survey at Peking University. Since large-scale robot implementation in China began after 2010, we include survey data from 2012 onwards, which ensures sufficient variability in the independent variable by lagging the robot data by one period in the regression model. The CFPS is a national longitudinal survey conducted by the Institute of Social Science Survey (ISSS) at Peking University. It collects data at individual, family, and community levels, using a stratified, multi-stage sampling strategy to ensure the sample represents 95% of the total population of China (Xie and Hu, 2014). The sample covers compulsory education stage children aged 10 to 15, and each child is matched with their parents. After removing missing values and outliers, the final sample comprises 6,423 observations of children residing in 116 cities.

Our dependent variable is whether the child is severely depressed. The CFPS 2012, 2016, 2018, and 2020 employ the 20-item questionnaire from the Center for Epidemiological Studies-Depression Scale (CES-D), while the 2014 data utilize the 6-item Kessler Psychological Distress Scale (K6). Both scales are internationally validated, conceptually similar, and widely recognized measures of depressive symptoms (Rosenquist et al., 2011; Mewton et al., 2016; Tran et al., 2019). Combining CES-D and K6 allows us to maximize the use of all available CFPS survey waves and to provide a more comprehensive longitudinal perspective, a strategy also adopted by several recent studies (Wu et al., 2023; Zhu et al., 2025). Following the methodology of Zhu et al. (2025) and the CFPS technical report, the scores of negatively worded questions are reversed to ensure that higher scores consistently indicate a greater severity of psychological depression. Then, the total score is calculated by summing all question scores. However, we acknowledge potential limitations of combining two different scales, as differences in length and item content may introduce measurement heterogeneity. To mitigate this concern, following classic works such as Radloff (1991) and Prochaska et al. (2012), we set the cutoff value for severe depression in the CES-D questionnaire at 28 and in the K6 questionnaire at 13. Severe depression means necessitating mental health treatment and causing impairments in functioning, which is important for informing clinical interventions and health policy (Prochaska et al., 2012).

In order to obtain the key explanatory variable robot adoption, we use data from the International Federation of Robots (IFR). The IFR reports industrial robot data for approximately 100 countries from 1993 to 2022, covering more than 90% of the world robotics market. Accordingly, it stands as the leading authoritative robot statistics worldwide (Wang et al., 2022). Unfortunately, the IFR data on robots are available only at the country-year-industry level, but we require robot data specific to the city where each child resides. To convert robot use intensity from the industry level to the city level, China’s industry classification is matched with the IFR’s robot stock industry classifications, and industrial robot operations are grouped into 19 industries (j = 1,…,19). Following Acemoglu and Restrepo (2020), the paper employs a shift-share design in constructing a Bartik measure of robot penetration rate on the city level, see Equation (1):

where c , j , and t denote city, industry, and year, respectively. Robo t j , t represents the stock of robots in China in industry j in year t , where j is one of the 19 industries, namely, automobile, other vehicles (shipbuilding and aerospace), plastic and chemicals, metal products, machinery, food and beverages, basic metals, electronic products, non-metallic mineral products, wood processing, textile, paper and printing, other manufacturing industries, mining and quarrying, education and research, agriculture, forestry and fishing, utilities (electricity, gas, water supply), construction, and service. Given China’s extensive adoption of robots since 2010, the base year is set at 2005. Emplo y j , 2005 is the country’s total employment of industry j in 2005, and L jc , 2005 represents the base year employment share of industry j in city c. Employment data are either sourced from the WORLDKLEMS database or computed based on the 2005 1% National Population Sample Survey. The mean robot penetration rate stands at 0.206 robots per thousand workers, with a standard deviation of 0.414.

In estimation, we include a set of child and family demographic features that may influence children’s depression as controls, namely the child’s gender (1 = male, 0 = female), age, hukou status (1 = rural hukou, 0 = urban hukou), and number of siblings; and the parents’ age, years of education (illiterate or semi-literate = 0, primary school = 6, junior high school = 9, senior high school = 12, 3-year college = 15, 4-year college = 16, master’s degree = 19, doctoral degree = 21), and hukou status (1 = rural hukou, 0 = urban hukou).

An array of city characteristics are also included as control variables, namely (1) economic development level (LnGDP), obtained by taking the natural logarithm of GDP, (2) income level, measured as the logarithm of the city’s average wage of employees (in yuan), (3) cost of living, proxied by the logarithm of average house prices (in yuan), (4) environmental quality, measured as the logarithm of the annual average PM2.5 concentration in the city, and (5) educational resources, represented by the ratio of the number of students to the number of teachers in the nine-year compulsory education stage. Regarding the data source, the house prices are sourced from the China Real Estate Information database, the data for PM2.5 concentrations are provided by the Center for International Earth Science Information Network (CIESIN), and other city characteristics are obtained from the “China City Statistical Yearbook.”

The summary statistics of the sample are presented in Table 1. The sample covers 6,423 matched child–parent observations residing in 116 cities. According to the CFPS definition of adolescents, the sample covers 10 to 15-year-old adolescents who had sufficient cognitive ability to complete self-reported questionnaires. The children are on average 12.5 years old and relatively balanced in both genders. In the data, 2.2% of children are severely depressed. The figure is comparable to research focusing on Chinese children and adolescents (Xu et al., 2018), warranting the validity of the data. The statistics for other individual characteristics in the sample are also consistent with those that use child data from the CFPS, such as Wang and Yang (2022).

To identify the impact of robots on children’s mental health, this paper estimates Equation (2):

where the subscripts i, c, and t represent the child, the located city, and the survey year. The dependent variable Severely _ depresse d ict is a dummy variable indicating whether the child is severely depressed in year t. Recognizing that the impact of robots on children’s mental health may take additional time, we use the one-period lagged natural logarithm of the robot penetration rate in city c as the independent variable, denoted as Robot s c , t − 1 . X ict stands for individual characteristics, and Z c , t − 1 represents city characteristics lagged by one period. Moreover, the regression controls for city fixed effects δ c and year fixed effects φ t to account for city-specific and time differences in child depression. The standard errors are clustered at the city level to account for potential correlations across individuals in the same city. β 1 is the parameter of interest, capturing the average effect of robots on the likelihood of a child being severely depressed.

The main challenge to our identification is the omitted variable bias. A city with more extreme climates may tend to use more robots, while having children with systematically worse mental health conditions. To address the potential endogeneity issue, we follow Acemoglu and Restrepo (2020) in using robot adoption of foreign countries to construct an instrumental variable for domestic robot adoption, as shown in Equation (3):

where Robo t j , t US and Emplo y j , 2005 US are, respectively, the stock of robots in industry j in year t , and the total employment in industry j in 2005 in the United States. L jc , 2005 indicates the proportion of employment in industry j relative to the total employment in city c in 2005 in China.

The US industrial robot data has been widely used in the literature for constructing IV for robot adoption in China (Giuntella et al., 2022; Liu et al., 2024). The validity of this practice is outlined as follows: Firstly, although the US outperforms China in robotics technology, the developmental trajectories between the two countries exhibit a certain degree of similarity (Zhu et al., 2023), thus affirming the relevance between the instrumental variable and robot exposure in Chinese cities. Secondly, automation in the US has a limited direct impact on the mental health of Chinese children, thereby ensuring the exogeneity of the instrumental variable. A potential concern is that global supply chains may transmit shocks from US automation to Chinese industries. For instance, increased robot adoption in the US might reduce demand for Chinese exports or shift the production strategies of multinational corporations, which could in turn influence Chinese labor markets and even children’s mental health. However, such channels are unlikely to be quantitatively important. While US robot adoption may marginally affect specific sectors in China through trade, the overall macroeconomic impact is limited due to China’s diversified export structure (Jiao et al., 2024). Additionally, the period under study coincides with China’s rapid domestic industrial automation and the expansion of the internal market. These domestic forces dominate employment and social outcomes, making indirect effects through global supply chains quantitatively minor.

The validity of the identification is further illustrated by testing for the exogeneity of the shift-share instrumental variable. According to Borusyak et al. (2022), it would suffice to prove the exogeneity of the shock from robot adoption. We follow Borusyak et al. (2022) and conduct a balance test on the city level to demonstrate that the foreign adoption of industrial robots is uncorrelated to the pre-trend demographic and socioeconomic conditions in China. The results presented in Table 2 indicate that the instrumental variable derived from foreign robot data is not correlated to city-level pre-trend characteristics such as the population share of women, average age, share of college-educated population, average income, average working hours, unemployment rate or undergraduate admission rates in each city in China. This suggests that the robot adoption shock can be considered as randomly assigned with respect to China’s regional demographic features, economic conditions or competition in higher education. The results provide support for the assumption that the instrumental variable satisfies the exogeneity condition.

Columns 1–2 of Table 3 present the results of OLS regressions. Column 1 controls for city and year fixed effects. Column 2 additionally controls for individual and city characteristics. The OLS estimates suggest a significantly positive relationship between robot adoption and children being severely depressed.

Columns 4 and 6 report the two-stage least squares estimates with specifications of controls identical to those of Columns 1 and 2, and Columns 3 and 5 present the first-stage results for the two regressions, respectively. Both Kleibergen-Paap F statistics exceed 10, confirming the relevance of the instrumental variable. Specifically, based on the 2SLS estimates in Column 6, a one standard deviation increase in robot adoption (0.414 robots per thousand workers) significantly increases the probability of children being severely depressed by 0.75 percentage points.

In understanding the salience of the result, we note that China’s robot stock has experienced a rapid growth in recent years. From 2011 to 2019, China’s overall stock of robots on average increased by 0.144 standard deviation per year. Note that 2.7% of children are severely depressed in the 2020 sample. Based on the regression estimates, we can illustrate the potential impact of robot exposure during this specific period: the increase in robot adoption raised children’s risk of severe depression by 0.144*0.75*8 = 0.86 percentage points, accounting for 0.0086/0.027 = 31.85% of the children’s severe depression rate in 2020. This should be understood as an illustrative calculation based on the specific period of 2011–2019 and the 2020 prevalence rate. Nevertheless, the magnitude signifies a stark impact.

Tables 4, 5 show the results of a battery of robustness checks. Firstly, our results are robust under a series of alternative specifications in Table 4. In Column 1 of Table 4, we use the robot data from the US, Netherlands, France, South Korea, and Germany—countries ahead of China in the advancement of automation, whose distribution of robots across industries is positively correlated with China—to construct an alternative instrumental variable. Column 2 changes the base year to 2010 for constructing the independent variable and instrumental variable. The industrial structures in 2010 are more up-to-date, and may hence better capture the shock from robots experienced by the city. Column 3 replaces the dependent variable with the child’s standardized depression score. A higher score indicates greater depression in the child. Furthermore, Column 4 uses children’s self-reported mental pressure ranging from 1 to 5, as an alternative measure of mental conditions, and the results show a consistent effect of robots in worsening the child’s mental conditions. Recognizing that industrial robots capture only one dimension of automation, we follow Nepal et al. (2025) and employ the number of AI-related patents as a proxy for urban AI development. AI-related patents are identified in the China Intellectual Property Office (CIPO) database based on specific International Patent Classification (IPC) codes. We substitute the independent variable with urban AI development and use the same instrumental variables as Nepal et al. (2025). The results are reported in Column 5. To mitigate potential errors caused by combining two different scales, CES-D and K6, to measure depression, we exclude the CFPS 2014 sample and rely solely on the CES-D questionnaire to determine whether a child is severely depressed. The corresponding results are reported in Column 6. Column 7 includes individual fixed effects in addition to the baseline specification to account for any individual-specific characteristics that might also be correlated with robot exposure, such as attitudes towards robot application. Besides, since we are using a shift-share identification strategy, the spatial correlation may occur across regions with similar sectoral composition. However, such correlations are not accounted for by the inference procedures (Adao et al., 2019). Column 8 applies the inference procedures proposed by Adao et al. (2019), which account for cross-region correlation in residuals within shift-share designs. The results are all largely similar.

Additionally, we control for the effect of a series of factors that were changing concurrently with the application of robots in the city in Table 5. Column 1 of Table 5 controls for the internet penetration rate to eliminate the impact of rising information technology development on the results, ensuring that the observed effects are specifically attributable to robot exposure rather than general advancements in digital infrastructure. Besides, the offshoring of industries away from China due to the rising labor cost may be correlated with robot adoption, and it could affect job opportunities and academic pressure on children, which would bias the estimates upward. This concern is addressed by controlling for the city’s foreign direct investment in Column 2. A linear trend for the share of the largest industry (by employment) in the city is added in Column 3 to account for the dynamics of employment structure due to robot adoption. Column 4 incorporates both the child and elderly dependency ratios to control for the effects of family structure. Column 5 includes the strength of clan culture to account for local cultural influences. Following Huang et al. (2022), clan culture is measured by the number of regional genealogies relative to the regional population. This measure is further interacted with a time dummy variable to capture temporal variations. In Column 6, to exclude the effect of the COVID-19 outbreak in 2020, the sample is restricted to the CFPS 2012–2018 data.

Moreover, the admission rate of the college entrance examination has gradually increased with the expansion of China’s higher education in recent years, and government investment in education has also changed over time. These factors may correlate with the increasing use of robots and affect children’s academic burden and mental health, thereby potentially biasing the regression results. To address this concern, we collect data on the undergraduate admission rate and the first-tier college admission rate in each province from the Education Examinations Authority, and include them as controls, respectively, in Columns 7 and 8. Column 9 adds the ratio of regional education expenditure to total fiscal expenditure to account for variation in local education investment. The results remain largely consistent with these additional controls. Overall, the robustness of our results to various checks verifies the validity of our causal interpretation.

This section explores the role of academic burden in mediating the impact of robots on child mental health. In light of this conjecture, we look into various indicators of children’s academic burden.

Columns 1 to 3 of Table 6 test the effect of robot adoption on the child’s studying time, whether or not the child attends tutorial classes, and the hours spent taking tutoring classes per week. The results indicate that robots increase both the amount of time children spend studying and the likelihood and duration of their attendance at tutoring classes. Columns 4 to 6 further study whether robots affect children’s studying attitude, using the degree to which children concentrate on studying, whether they check homework, or finish homework before playing. These questions were answered on a scale of 1 to 5, with higher numbers indicating greater agreement. As presented in Columns 4 to 6, robots have significantly improved children’s learning attitude. One possible side effect of intensified academic activities would be crowding out physical exercise. We examine the impact of robot adoption on children’s physical exercise frequency. The results indicate that robot adoption significantly reduces children’s weekly exercise frequency (Column 7).

Academic burden is proven to be critical to children’s mental health (Kandola et al., 2020; Jiang et al., 2021). Specifically, a higher level of academic burden is associated with a higher level of school burnout, which in return, leads to a higher level of depression (Jiang et al., 2021). Furthermore, increased sitting time and reduced exercise time are linked to a higher risk of depression (Kandola et al., 2020). Our results confirm that one of the important mechanisms linking robot adoption to children’s mental health is the increased academic burden.

Negative income shocks resulting from robot adoption may lower children’s living standards, thereby worsening their mental health (Lauritzen and Sivertsen, 2012; Baird et al., 2013). Firstly, we consider the living environment of children. The negative income shock caused by robots may worsen children’s living conditions, potentially leading to mental health issues (Lauritzen and Sivertsen, 2012). Column 1 of Table 7 uses a dummy variable to indicate whether the family’s living condition is poor, assigning a value of 1 if the family faces living difficulties and 0 otherwise. Living difficulties are characterized by several conditions, including: children over the age of 12 sharing a room with their parents; three generations of family members living in the same room; children of different genders over the age of 12 sharing a room; beds being laid out at night and folded up during the day; beds being placed in the living room; and other related issues. Column 2 also evaluates the impact of robot adoption on household-reported housing crowdedness, which is measured on a scale from 1 to 7, where higher values indicate greater crowding.

Moreover, consumption is a useful indicator of living standards. However, the CFPS only reports overall household consumption and lacks data on consumption by children. Therefore, we examine the impact of robots on whether children have pocket money and the amount of their monthly pocket money (log-transformed). The pocket money allowances of children depend largely on the parents’ budget, and have been proven to significantly affect children’s mental health (Baird et al., 2013).

Columns 1 to 4 of Table 7 indicate that robots exert no significant impact on children’s living conditions or pocket money. These results suggest that robots have little influence on children’s living standards. One possible explanation is the substantial savings and assets of Chinese households, which help maintain living standards during sudden financial shortfalls. According to CFPS 2020 data, the average annual household consumption is 49,575 yuan, while the average household cash and deposits are 76,519 yuan, and the average household net assets are 807,786 yuan. Without reducing family living quality, household cash and deposits can cover 1.5 years of household consumption, and household net assets can cover 16.3 years of household consumption. Another explanation lies in the role of social safety nets, which support individuals negatively affected by shocks or emergencies. Social safety nets are commonly categorized as formal or informal, depending on whether they are legally mandated (Paitoonpong et al., 2008). In China, formal social safety nets include Dibao (the minimum living standard guarantee program), unemployment insurance and other statutory schemes, while informal social safety nets involve transfers from family members, friends, neighbors, communities, and other sources such as NGOs. Both systems play a crucial role in sustaining living standards when household income is disrupted by automation. Consequently, the short-term labor market impact of robots has not significantly undermined children’s living standards.

The income difficulties caused by robots may cause parents to care less about their children’s lives and reduce positive parent–child interactions (Brody et al., 1994), which could exacerbate children’s stress responses and harm their mental health (McLoyd, 1990).

We investigate this mechanism using survey questions from the CFPS related to parent–child relationships in Table 8. In Column 1, the dependent variable is the frequency of heart-to-heart talks with parents over the past month answered by children. Columns 2 to 4 examine more detailed questions to assess the impact of robots on parent–child interactions. Three questions are used, answered by parents or children: parents take the initiative to actively communicate with the child, parents or guardians ask about what happened to the child at school, and parents or guardians praise the child. Responses to these questions are rated on a scale of 1 to 5, where 1 = Never, 2 = Occasionally, 3 = Sometimes, 4 = Often, and 5 = Always. The results from Columns 1 to 4 show that robot adoption reduces both the frequency of positive parent–child interactions and parents’ care for their children’s lives.

Additionally, Column 5 examines the impact of robot adoption on the negative interactions between parents and children. The frequency of quarrels with parents in the past month answered by children is used as the dependent variable. The results show that robot adoption has no significant effect on the frequency of quarrels between parents and children. These tests suggest that reduced positive interactions between parents and children also serves as a channel through which the introduction of robots contributes to children’s depression.

It has long been established that women and men respond differently to adverse shocks (Hua et al., 2023). Therefore, girls and boys may exhibit distinct psychological responses to the introduction of robots. Columns 1–2 of Table 9 show that robots indeed contribute to severe depression in girls but not boys. Investigating the source of this gender difference reveals that robots have increased the academic burden and reduced parent–child interactions for both genders (Table A1).¹ As girls generally display a higher vulnerability in terms of mental conditions (Kuehner, 2017), robot adoption would contribute to a greater deterioration of girls’ mental conditions.

Given the varying importance that families place on education in rural and urban areas (Li et al., 2017), we would expect a different effect of robots between those two areas. As Columns 3–4 of Table 9 imply the effects of robots on children’s mental health are indeed significant in urban areas but not in rural areas. This is corroborated by the fact that urban children’s studying time has increased significantly in response to robot adoption (as shown in Column 1 of Panel A, Table A2). In contrast, robots have only increased the amount of time rural children spend in tutoring classes (see Columns 2–3 of Panel B, Table A2), without affecting their overall learning time (see Column 1 of Panel B, Table A2). This result may arise because urban parents, in response to advancements in robotics, are more likely to enhance their children’s competitiveness by extending study time. In contrast, rural parents may have limited educational awareness, focusing primarily on supervising tutoring sessions and paying little attention to extending learning time after training. The increased study time for urban children has heightened academic burden, leading to a significant deterioration in their mental health.