This study was a longitudinal study of pregnant women who underwent pregnancy tests and delivered at Beijing Friendship Hospital in China. Participants were recruited from November 2019 to January 2022. Data were collected at early pregnancy within 12 weeks of the last menstrual period (T1) and at five separate time points during the perinatal period; specifically, mid-pregnancy with 22–24 weeks of gestation (T2), late pregnancy with 34–36 weeks of gestation (T3), 1 month postpartum with 2–4 weeks after delivery (T4), 3 months postpartum with 10–12 weeks after delivery (T5), and 6 months postpartum with 24–26 weeks after delivery (T6). All participants provided written informed consent after receiving both oral and written explanations of the study’s objectives and procedures. The study was approved by Ethics Committee of Beijing Friendship Hospital, Capital Medical University.

Data were collected in the hospital, including information on participants’ psychiatric conditions, stress, anxiety, and depression. All participants completed the Edinburgh Postnatal Depression Scale (EPDS) questionnaires. Additional data were gathered on participants’ characteristics, including age, educational levels, employment statuses, monthly household incomes, planned pregnancy, gravidity, parity, adverse pregnancy history, and perinatal mortality experience. The results of Harman’s single-factor test suggested that common method bias was not a significant concern in this study.

Edinburgh Postnatal Depression Scale (EPDS) is the most frequently used self-report tools for detecting perinatal depression based on the DSM-5 criteria (Cox et al., 1987). It includes 10 items of respondents’ experience of symptoms such as laughter, enjoyment, guilt, anxiety, panic attacks, sadness, sleep disturbances, feelings of being overwhelmed, crying, and suicidal thoughts over the past 7 days (Smith-Nielsen et al., 2018). Although EPDS was originally created to evaluate depressive symptoms in postpartum women, it has also been applied to screen for antepartum depression.

Several studies have confirmed its cultural applicability, good internal consistency and adaptability in screening perinatal depression among Chinese population (Lee et al., 1998; Lau et al., 2010; Song et al., 2024). Chinese version of EPDS demonstrated high internal consistency, with a split-half reliability coefficient of 0.74 and a standardized Cronbach’s α of 0.78 (Lau et al., 2010).

For categorical variables, frequencies and percentage were calculated. Prior to factor analysis, specimen validity was assessed using the Kaiser-Meyer-Olkin (KMO) test and Bartletts test of sphericity to confirm the data’s suitability for factor analysis. Factor structures were examined using exploratory factor analysis (EFA) with the R packages EFA tools v0.4.4 and lavaan v0.6–17. Parallel analysis was conducted on a polychoric correlation matrix to determine the number of factors, comparing actual eigenvalues to the 95th percentile eigenvalues of 5,000 simulated random datasets. Scree plots were also consulted to determine the number of factors. EFA models with maximum likelihood extraction and oblique rotation were applied, considering factors with loadings ≥0.30 (Škodová et al., 2021; Lautarescu et al., 2022). These analyses were conducted for each of the six groups. The factor models were iteratively evaluated with two and three factors, as no existing research suggested factor structures with more than three factors for the EPDS. The best-fitting model was then subjected to multigroup confirmatory factor analysis (CFA) to assess measurement invariance across the different stages of antepartum and postpartum periods.

Model fit was evaluated using standard indices such as the root mean square error of approximation (RMSEA), the standardized root mean square residual (SRMR), the comparative fit index (CFI) and the Tucker–Lewis index (TLI). The RMSEA was considered acceptable if below 0.08, while values under 0.05 signified a good fit (Škodová et al., 2021). Similarly, an SRMR under 0.08 suggested an acceptable fit (Matsumura et al., 2020). A CFI greater than 0.90 was used to indicate an acceptable model fit, with values equal to or exceeding 0.95 representing a good fit (Škodová et al., 2021). TLI values greater than 0.90 were regarded as indicative of a well-fitting model (Škodová et al., 2021). All statistical tests were two-tailed with a significance level of α = 0.05. The semPlot v1.1.6 package was used to draw the path diagram.

Finally, the reliability of EPDS was assessed using Cronbach’s alpha (α), McDonald’s omega total (ωₜₒₜₐₗ), and McDonald’s omega hierarchical (ωₕ) coefficients. Values of α, ωₜₒₜₐₗ, and ωₕ ≥ 0.7 are considered indicative of acceptable reliability, while values ≥0.8 indicate good reliability (Revelle and Zinbarg, 2009; Revelle and Condon, 2019).

All analyses were conducted using R version 4.3.3.

The study enrolled a total of 905 participants. Among them, 56 individuals were excluded due to missing sociodemographic information. Therefore, the final sample comprised 849 individuals, all of whom were surveyed in the early stages of pregnancy. The sociodemographic and reproductive characteristics of the study population are detailed in Table 1.

The number of factors identified by the parallel analyses with principal component analysis (PCA), exploratory factor analysis (EFA) and squared multiple correlation (SMC) for the EPDS across different time points were presented in Table 2. Based on the results of parallel analysis and the scree plot, combined with existing research recommendations on factor structures, we adopted a 2- and 3-factor modeling approach.

Exploratory factor analyses (EFAs) were conducted for the EPDS to compare the two-factor and three-factor models suggested by parallel analysis at six time points (Table 3). Table 3 displayed the factor loadings obtained from various factor extraction and rotation techniques for each item of the EPDS scale. Bolded values in the table indicate the highest loadings for each item on its corresponding factor, signifying the strongest associations. For example, in the two-factor model at T1, items 1–2, items 7–9 were loaded on F1, and items 3–6 were loaded on F2. Item 10 was not retained in the antepartum factor analyses as it did not meet the criterion of >0.30 for significant factor loading. However, in the postpartum analyses, item 10 was loaded, indicating its relevance in the postpartum period.

Table 4 presents CFA fit indices for the 2- and 3-factor models of the EPDS across the six specific groups. The results indicated a consistent superior fit for the three-factor model compared to the two-factor model at each time point, except that the two-factor was slightly better at T5. For instance, at T1, the three-factor model demonstrated a CFI of 0.98 and a RMSEA of 0.05, outperforming the two-factor model which had a CFI of 0.95 and an RMSEA of 0.07. Longitudinal comparison of the fit of the three-factor model at different time points showed that the three-factor model was always well fitted. CFI was always above 0.95 and RMSEA was always less than 0.08.

The EFAs identified three-factor models for the EPDS, however, factor structure varied slightly at the six time points. Therefore, we grouped T1-T3 as the antepartum period and T4-T6 as the postpartum period to test measurement invariance. We obtained a three-factor model: (1) anhedonia (items 1 and 2), (2) anxiety (items 3,4,5,6, and 7), (3) low mood (items 8 and 9). For the antepartum group, the three-factor model demonstrated good fit, with statistical indices as follows: χ² (df = 24) = 179.23, p < 0.001; CFI = 0.97, RMSEA = 0.06, TLI = 0.95. For the postpartum group, the model fit was enhanced: χ² (df = 24) = 169.326, p = 0.001; CFI = 0.99, RMSEA = 0.05, TLI = 0.98. The path diagram illustrating the three-factor model was presented in Figure 1.

Table 5 presents the fit indices for the multigroup confirmatory factor analysis models of the data as the antepartum and postpartum groups, each imposing stricter equality constraints. The table includes chi-square difference tests to compare the fit of each subsequent model with the preceding one. The baseline model of the EPDS showed good model fit [χ²(48) = 231.92, CFI = 0.97, RMSEA = 0.06, SRMR = 0.03]. Equating the factor loadings across groups did not significantly affect the model fit (p = 0.25). However, when item thresholds were also constrained to be equal, the model fit deteriorated significantly according to the chi-square test (p < 0.001), although the fit indices remained within an acceptable range (CFI = 0.97, RMSEA = 0.06, SRMR = 0.04). It provided support for measurement invariance across groups based on the three-factor model of the EPDS.

As shown in Table 6, the EPDS presented a Cronbach’s alpha of 0.84 and an Omega Total of 0.86, supporting adequate internal consistency and strong overall reliability by accounting for both common and specific factors influencing item responses. The Omega Hierarchical was 0.73, indicating a substantial degree of variance attributable to the general factor in the hierarchical three-factor model. The average inter-item correlation for the EPDS was 0.13, suggesting a relatively low association between individual items, which reflected the diversity in the items’ contribution to the overall scale score.