Data source and study sample

Data we employed in this study were derived from the China Family Panel Study (CFPS) for the waves of 2012, 2014, 2016, 2018 and 2020. CFPS is a nationally representative survey launched in 2010. This biennial follow-up survey covered 25 provinces, cities and autonomous regions in mainland China, aiming to collect information at individual, household and community levels, including a wide range of data on demographic and socioeconomic characteristics, health conditions and health service utilization. CFPS adopted a multi-stage stratified sampling method to reduce operation costs and implicit stratification to increase efficiency. At the first two stages in the sampling process, official administrative entities were used, and the final onsite sampling stage used frames constructed from field-drawn maps of dwelling units, not official household registers. The survey sampled all members of selected households, therefore allowing us to identify the information of the individual’s spouse, children and parents. The baseline survey was conducted in 2010 and covered 14 960 households, 33 600 adults and 8 990 children. The detailed information of the survey design has been reported elsewhere (Xie and Hu, Reference Xie and Hu2014). Since China implemented the UTP in January 2016, we linked the 2012, 2014, 2016, 2018 and 2020 waves of CFPS data as a longitudinal dataset of an unbalanced panel to examine the impact of UTP. Since our study is focused on married females of childbearing age, we included 7481 currently married females (3 060, 3 563, 3 859, 3 782 and 2 814 women from 2012, 2014, 2016, 2018 and 2020 waves, and total 17 709 observations for the pooled sample) aged 20–40 years during the survey window from 2012 to 2020 with no missing value on socioeconomic information, CESD score and other health status (Fig. 1).

Figure 1.

Flowchart of sample selection.

Study design

We employ a quasi-experimental research design which is an effective approach for investigating the causal relationships between the implementation of the Universal two-child policy (UTP) and women’s depressive symptoms. The difference-in-difference (DID) approach is key to evaluating interventions to inform health policymakers and future strategies (Godard-Sebillotte et al., Reference Godard-Sebillotte, Karunananthan and Vedel2019). This approach can effectively disentangle the impact of intervention from the permanent disparities between the intervention and control groups, as well as the temporal trends in the outcome that are unrelated to the intervention of UTP.

Measures

We regard the survey year before 2016 (2012 and 2014) as the pre-UTP period and the survey year after 2016 (2016, 2018 and 2020) as the post-UTP period. Before the implementation of UTP, China implemented a selective two-child policy in 2014. The implementation of this policy allows couples to have two children as long as one of the couples is an only child. Therefore, UTP removed the restrictions on couples where neither spouse was from a one-child family, while it did not affect families where at least one of the spouses was an only child or families who had already had two or more children. We categorized women who were neither only children themselves nor had spouses who were only children, or who had either one or no children before 2016, as the UTP-exposed group. Women who were either only children themselves or whose spouses were only children, or who already had two or more children, were categorized as the control group. The impact of UTP is estimated by comparing the differences between two changes in outcomes: (i) changes in depressive symptoms between the pre- and post-UTP periods within the UTP-exposed group and (ii) the pre- and post-UTP periods in the control group.

Depressive symptoms were employed as the outcome variable. The questionnaire modules regarding depressive symptoms in the CFPS are different across waves. In 2010 and 2014, depressive symptoms were assessed based on the Kessler 6 Rating Scale (K6), which is a six-item scale asking how often the respondents experienced depression, nervousness, agitation, hopelessness, psychological fatigue, and worthlessness during the last month. The responses for the K6 items are on a 5-point scale: 0 (never), 1 (once a month), 2 (2–3 times a month), 3 (2–3 times a week), or 4 (almost every day). The total scores of these six items range from 0 to 24, with a higher score indicating more frequent or more severe depressive symptoms. In the 2012, 2016 and 2018 waves, items of depressive symptoms were adapted from the Center for Epidemiologic Studies Depression Scale (CES-D). The CES-D scale comprises 20 items, with 16 items measuring negative feelings and 4 items measuring positive emotions. The respondents are asked about the number of days on which they experienced depressive symptoms during the last 7 days. Responses are on a 4-point scale: 0 ‘rarely’ (less than 1 day), 1 ‘some days’ (1–2 days), 2 ‘occasionally’ (3–4 days), and 3 ‘most of the time’ (5–7 days). The positive questions were further reversed to calculate the final score, which ranges from 0 to 60, with higher values indicating more severe or more frequent depressive symptoms. To reduce respondent burden and refusal rate, the assessment switched to a shorter version of the scale – the CES-D scale to a subset of eight items (hereafter CES-D8) in 2016, 2018 and 2020.

It is noteworthy that both the K6 and CES-D8 scales have been widely validated and demonstrate good reliability and validity in general population studies. The K6 has shown high internal consistency (Cronbach’s α > 0.80) and strong discriminant validity in distinguishing clinical cases of depression, while the CES-D8 also exhibits acceptable reliability (α typically between 0.70 and 0.80) and correlates highly with the full CES-D scale. Therefore, to ensure continuity, we employed CES-D8 to assess the depressive symptoms in the 2012, 2016, 2018 and 2020 waves, and further standardized the total scores (CES-D8 and K6 in 2014) of each wave. The standardized score of depressive symptoms was the main outcome variable. Additionally, we used a binary outcome of depression for sensitivity analyses, which was categorized into ‘having moderate or severe depressive symptoms’ (5 or higher in K6 score, or 7 or higher in CES-D8, coded 1) or ‘mentally healthy’ (coded 0).

Control variables included the following: age (modelled as a continuous variable in years), residence (categorized as urban or rural based on official household registration status), education (categorized into three levels: primary school or below, junior high or high school, and junior college or above), financial status (operationalized as a categorical variable defined by tertiles of annual family income per capita: low, medium, or high), the gender composition of children, if any (categorized as: no child, only girl(s), both boy(s) and girl(s), or only boy(s)), self-rated health (‘In general, how would you rate your health?’) scored on a scale from 0 to 4 (where 0 = very poor and 4 = very good), and chronic conditions (defined as a binary variable indicating the presence or absence of any physician-diagnosed chronic disease within the past six months).

Statistical analysis

We employed a linear regression model for the continuous outcome variable (standardized score of depressive symptoms to estimate the impact of UTP on mental health among married women of childbearing age. To integrate multiple waves of the longitudinal survey, we structured the data in a panel format and included survey wave fixed effects in all regression models to control for time-specific shocks and systematic differences across waves. We adjusted results for several potential confounders including age, residence, education, financial status, self-rated health score, chronic conditions and gender of children. The regression model was:

\begin{align*}{Y_{it}} = &\,\alpha + \beta \times {\text{Perio}}{{\text{d}}_{it}} + \gamma \times {\text{UT}}{{\text{P}}_{it}} + \delta \times {\text{Perio}}{{\text{d}}_{it}} \nonumber\\ & \times {\text{UTP}} + \mu \times {X_{it}} + {A_i} + {B_t} + {\varepsilon _{it}} \end{align*}

Here, ${Y_{it}}$ is the standardized score of depressive symptoms of an individual i in the year t. ${\text{Perio}}{{\text{d}}_{it}}$ is a variable indicating whether the observation was during the UTP implemented period. ${\text{UT}}{{\text{P}}_{it}}$ is a variable indicating whether the observation was in the UTP-exposed group or control group. ${A_i}$ is the individual fixed effect, ${B_t}$ is a vector of dummy variables for a year-fixed effect, ${X_{it}}$ is a set of varying time, individual-level covariates and ${\varepsilon _{it}}{\text{ }}$is the error term. The key parameter of interest is $\delta $, the DID estimate, which measures the pre-post change in depressive symptoms between the UTP-exposed group and control group, thereby indicating the impact of UTP on women’s depressive symptoms. We performed stratified analyses by education, financial status and gender of children with the same model mentioned above to examine how socioeconomic and cultural factors moderate the policy’s mental health effects. This approach is grounded in the rationale that these variables capture critical dimensions of vulnerability – including resource constraints and son preference-related family pressures – that may lead to differential impacts of fertility policies on psychological well-being.

The DID analysis was based on the ‘common trend assumption’, stating that the intervention and control group would have the same variation trend in the pre-intervention periods. Thus, a parallel trend analysis was employed to assess whether the impact measured can be attributed to the implementation of UTP in the sensitivity analyses. Moreover, we used the binary outcome (having moderate or severe depressive symptoms or not) with logistic regression to reanalyse and check the robustness of the results. Furthermore, considering the variation in depression measurement methods across some survey years, we conducted additional analysis using only the 2012, 2016 and 2018 surveys, all of which utilized the CES-D8 for depression measurement, to further examine the robustness of the results. Additionally, we accounted for potential within-individual correlation across waves by clustering standard errors at the individual level in all regression models. All analyses were performed using the STATA version 17.0 for Mac (Stata Corp, College Station, TX, USA).