A non-probabilistic incidental sampling method was used. The inclusion criteria were: being female, at least 18 years old, and in the postpartum period (between 2 and 12 months after delivery). The exclusion criteria were: having a visual impairment that prevented completion of the questionnaire, being underage, being less than two months or more than 12 months postpartum, and being pregnant. Participants who met any of the exclusion criteria did not complete the questionnaire or the additional items. Consequently, missing data were handled by excluding those cases. Mothers within the first six weeks postpartum were excluded to obtain a sample reflecting pure postpartum depression, uncontaminated by the postpartum blues syndrome (; ).

The invitation to participate in the research was sent via social networks (Instagram, Facebook and WhatsApp) to specific groups of mothers with different interests, such as breastfeeding, sales, complementary feeding, parenting, food allergies, and baby dance classes. The invitation, disseminated by the second author, included details about the study and a link to access the questionnaire. The most challenging sample to obtain was that of women engaged in unpaid work (i.e. home duties), so the second author visited a health center in the municipality of General Escobedo to collect data in person.

The questionnaire was completed by 165 women; however, 21 were excluded for meeting one or more exclusion criteria, and two additional cases were removed based on the elimination criterion. The final sample therefore consisted of 142 participants.

We considered this sample size to be adequate, as it allows for approximately 14 participants per item (142:10 ≈ 14:1) and 7 participants per parameter to be estimated in a single-factor model (142:20 ≈ 7:1) with high measurement weights ( ranging from .641 to .934, with an average of .758).

The electronic form included the informed consent, questions on sociodemographic, occupational, and pregnancy history, as well as the Edinburgh Postnatal Depression Scale (EPDS; ) and the Perceived Social Support Scale from Partner and Significant Others (PSSS; Cienfuegos-Martínez, 2010 ). The form was designed using the Google Forms platform.

The EPDS, developed by , is a self-report scale created in health centers in Livingston and Edinburgh to assist primary care professionals in detecting postpartum depression in mothers. It is recommended that the scale be administered after the sixth postpartum week, as symptoms during the initial stage (the first six weeks after delivery) may reflect postpartum blues associated with hormonal changes. The scale is available in both English and Spanish; the Spanish version was used in this study. See Appendix.

Each EPDS item offers four response options. Mothers are asked to select the option that best reflects how they felt during the previous seven days. Items 1 and 2 are scored from 0 (first option) to 3 (last option), with higher scores indicating greater symptom severity. Items 3 through 10 are reverse-keyed and scored from 3 (first option) to 0 (last option). The total score is obtained by summing the scores for all items, resulting in a possible range from 0 to 30. Scores from 0 to 11 indicate no depression, 12 to 19 suggest mild postpartum depression, 20 to 29 indicate moderate depression, and a score of 30 indicates severe depression ( Alvarado-Esquivel et al., 2006 ; Macías-Cortés et al., 2020 ).

The PSSS, developed by Cienfuegos-Martínez (2010) , consists of 44 Likert-type items with five response options (“never,” “almost never,” “sometimes,” “almost always,” and “always”), scored from 1 to 5. The total score ranges from 44 to 220, with higher scores indicating greater perceived social support. The scale shows excellent internal consistency (44 items; Cronbach’s α = .90). It comprises two subscales: support from the partner (24 items; Cronbach’s α = .87) and support from significant others (20 items; Cronbach’s α = .88).

The research design was non-experimental and cross-sectional. The study was approved by the Research Sub-directorate of the School of Psychology at the Autonomous University of Nuevo León.

To comply with ethical standards, the Code of Ethics of the Mexican Society of Psychology ( Sociedad Mexicana de Psicología, 2010 ), and the guidelines of the American Psychological Association ( APA, 2017 ), were used as a basis for informed consent procedures and privacy notices.

This study classifies it as risk-free research, as no intervention or intentional modification was made to the physiological, psychological, or social variables of the participants, in accordance with Article 17 of the General Health Law on Health Research ( Cámara de Diputados del H. Congreso de la Unión, 2014 ).

The analysis of the items began with a description of their distributions, followed by an assessment of their discriminative capacity, and concluded with an evaluation of their reliability. The distributions of the items, which are ordinal variables, were visualized using bar plots. Additionally, they were described using measures of central tendency (arithmetic mean, median, and mode), variability (mean absolute deviation, semi-interquartile range, and quartile coefficient of variation), and shape—specifically, Bowley’s ( Bowley, 1901 ) quartile coefficient of skewness and Kelley’s ( Kelley, 1923 ) percentile coefficients of skewness and kurtosis.

To assess the discriminating power of each item, high and low scoring groups were created based on the EPDS total score (the sum of the 10 items). The low-score group consisted of scores less than or equal to the 27th percentile of the total score, while the high-score group included scores greater than or equal to the 73rd percentile ( Kelley, 1939 ). Differences in central tendency for each item between these two groups were tested using the Mann-Whitney U test (). In addition, mean differences were computed along with 95% bootstrap confidence intervals using the bias-corrected and accelerated (BCa) percentile method, based on 1,000 resamples with replacement. A significant mean difference of at least one quarter of the item range ([3 – 0]/4 = .75) was considered evidence of discriminative power ().

The reliability of each item was assessed based on three criteria: 1) its correlation with the sum of the remaining items, 2) the reduction in overall internal consistency when the item was removed, and 3) its communality, or the variance it shared with the other items ( Furr, 2021 ; Moral, 2006 ).

The correlation between each item and the sum of the remaining items was estimated using the polyserial correlation coefficient when the assumption of bivariate normality was met ( Zheng & Cao, 2022 ), or the Spearman rank-order correlation coefficient () otherwise. The normality assumption was tested using the chi-square goodness-of-fit test ().

Due to the violation of the tau-equivalence assumption, internal consistency was estimated using McDonald’s omega coefficient () and Green-Yang’s categorical omega coefficient ( Green & Yang, 2009 ). The assumption of tau-equivalence—or equal factor loadings on a general factor—was tested via confirmatory factor analysis (CFA). A one-factor model with 10 indicators was specified with the constraint that all factor loadings were equal. The model was estimated using the Weighted Least Squares Mean and Variance adjusted (WLSMV) method, which is appropriate for ordinal data and relies on polychoric correlations. These correlations assume an underlying bivariate normal distribution for the ordinal variables ( Han, 2022 ). This assumption was tested using the chi-square goodness-of-fit test ( Li, 2021 ).

Communality was estimated using the squared multiple correlation obtained by predicting each item from the remaining items. The predictive model was specified through path analysis, with parameters estimated using the WLSMV method. Communalities below .16 are generally considered low. Ideally, communalities should be at least .25, and preferably .50 or higher ( Furr, 2021 ; Moral, 2006 ).

Regarding the second objective, a one-factor model with freely estimated factor loadings (with one loading fixed for identification) was tested using the WLSMV method. The significance of the factor loadings was assessed using the z-test. In addition, 95% bootstrap confidence intervals for the loadings were computed using the percentile method based on 1000 bootstrap samples ( Efron & Narasimhan, 2020 ).

Model fit is considered good when the null hypothesis of exact fit is not rejected by the likelihood ratio chi-square test at the .05 significance level. Good fit is also indicated by a standardized likelihood ratio chi-square statistic (χ²/df) value less than or equal to 2, Comparative Fit Index (CFI) and the Tucker-Lewis index (TLI) —or Normed Fit Index (NFI) —values greater than .95, a Root Mean Square Error of Approximation (RMSEA) value less than or equal to .05, and a Standardized Root Mean squared Residual (SRMR) value less than or equal to .08. An acceptable fit corresponds to values such as p ≥ .01 for χ², χ²/df ≤ 3, CFI and TLI ≥ .90, RMSEA ≤ .08, and SRMR ≤ .1. In contrast, a poor fit is defined by values such as p < .01 for χ², χ²/df > 3, CFI and TLI > .90, RMSEA > .08, and SRMR > .1 ( McNeish & Wolf, 2023 ; Rosseel, 2012 ; Rosseel et al., 2024 ). An Average Variance Extracted (AVE) greater than .5 and a McDonald’s omega coefficient value equal to or higher than .7 indicate convergent validity in a measurement model ().

Regarding the third objective, the internal consistency of the scale was estimated using both McDonald’s omega coefficient ( McDonald, 1999 ) and the categorical omega coefficient proposed by Green and Yang (2009) . The latter was computed using the MBESS package in R ( Kelley, 2023 ). Analogous to Cronbach’s alpha, omega values in the range [.5, .6) were considered indicative of unacceptable internal consistency; values in the range [.6, .7), questionable; [.7, .8), acceptable; [.8, .9), good; and [.9, 1], excellent ( Izah et al., 2023 ). However, omega values ≥ .95 and χ²/df < 1 may indicate the presence of redundant items and an overfitted model ( Byrne, 2016 ; Kline, 2016 ).

With respect to the fourth objective, the distribution of the EPDS total score was described using measures of central tendency (arithmetic mean, median, and Grenander’s ( Grenander, 1965 ) mode), variability (variance, standard deviation, and mean absolute deviation) and shape (coefficients of skewness and kurtosis based on Fisher’s ( Fisher, 1930 ) cumulants). Outliers were tested using Grubbs’ ( Grubbs, 1969 ) test, symmetry was examined using D’Agostino’s ( D’Agostino, 1970 ) test, and kurtosis was assessed using test. The distribution was visualized through a table of class intervals, a box plot, and a histogram with the estimated density and normal curves overlaid. Normality was assessed using the tests by D’Agostino, Belanger, and D’Agostino ( D’Agostino et al., 1990 ); Shapiro and Wilk (1965) , with Royston’s () standardization; Anderson and Darling (1952) ; and Kolmogorov (1933) and Smirnov (1948) , with p-values computed via Lilliefors’ Monte Carlo simulation ().

Regarding the fifth objective, convergent construct validity was assessed by examining whether there is a negative correlation between the EPDS and the EAP using Pearson’s product-moment correlation. To test its significance, the bootstrap BCa confidence interval and a one-sided bootstrap p-value (H₀: ρ ≥ 0; H₁: ρ < 0) were computed. The significant difference in central tendency of the EPDS total score between groups of women with and without depression during pregnancy was tested using a one-sided Mann–Whitney U test ( Mann & Whitney, 1947 ) at a 5% significance level. Tie correction was applied, but not Yates’ continuity correction ( Yates, 1934 ). Both asymptotic and bootstrap p-values were calculated ( Conover, 1999 ).

The effect size was estimated using Rosenthal’s () r coefficient, Spearman’s ( Spearman, 1904 ) rank correlation coefficient (rₛ), and Kerby’s () difference, which corresponds to Curenton’s () biserial rank correlation coefficient. This last measure represents the probability that a randomly selected value from the group with depression during pregnancy exceeds a randomly selected value from the group without depression. It takes values between –1 and 1, where negative values indicate superiority of the second group (without depression) over the first group (with depression), and positive values indicate superiority of the first group over the second. It can be interpreted analogously to McGraw and Wong’s ( McGraw and Wong, 1992 ) Common Language Effect Size (CLES) measure for comparing means between two independent groups. Considering the correspondence between CLES and Cohen’s d statistic ( Cohen, 1988 ), given by CLES = Φ(d / √2), and the conventional cut-off points of .2, .5, and .8 for d stipulated by , absolute values of Kerby’s d in the interval [.5, .579) indicate a trivial effect size, [.579, .691) small, [.691, .788) medium, and [.788, 1] large.

The difference between the marital status groups was tested using the Kruskal & Wallis’ ( Kruskal & Wallis, 1952 ) test. The effect size was estimated by the epsilon squared (ε²) and eta squared (η²) coefficients. According to Cohen (1988) , values of ε² and η² less than .01 are considered trivial, values between .01 and .06 are considered small, values between .06 and .14 are considered medium, and values of .14 or greater are considered large. Pairwise comparisons were conducted with Dunn’s ( Dunn, 1964 ) test applying Bonferroni’s ( Bonferroni, 1936 ) correction.

The mean age of the mothers was 31.56, 95% BCa CI [30.73, 32.34] ranging from 20 to 44 with a sample standard deviation of 5.049. Among the 142 women, 110 (77.5%) were married, 22 (15.5%) were cohabiting, 6 (4.2%) were single without a partner, and 4 (2.8%) were single with a partner. In terms of educational level, one out of 142 women (.7%) reported having a primary education; 11 (7.7%) reported having a secondary education; seven (4.9%) reported having a high school education; nine (6.3%) reported having technical studies; 80 (56.3%) reported having an undergraduate or engineering degree; and 34 (23.9%) reported having postgraduate studies. Regarding work activity, 38% of the women were housewives, 43% had paid work in a company, and 19% were entrepreneurs, either through catalog sales or owning their own business.

Among the 142 participants, 71 (50%) were primiparous, while 71 (50%) were multiparous. Among the multiparous women, 42 (59.2%) had two children, 23 (32.4%) had three children, 4 (5.6%) had four children, and 2 (2.8%) had five children. When asked whether they had experienced at least one previous abortion, 38 (26.8%) responded yes, while 104 (73.2%) responded no. Additionally, 10 (7%) of the 142 women reported having been diagnosed with depression during pregnancy, whereas 132 (93%) had not.

The infants’ ages ranged from 2 to 12 months. At the time the mothers completed the questionnaire, 21 (14.8%) were 2 months old, 15 (10.6%) were 3 months, 16 (11.3%) were 4 months, 14 (9.9%) were 5 months, 13 (9.2%) were 6 months, 13 (9.2%) were 7 months, 12 (8.5%) were 8 months, 9 (6.3%) were 9 months, 9 (6.3%) were 10 months, 12 (8.5%) were 11 months, and 8 (5.6%) were 12 months.

Table 1 shows the measures of central tendency, variability, and shape of the 10 items that make up the EPDS, whose distributions are represented by a bar chart in Figures 1-10 .

The central tendency of item 1 (“I have been able to laugh and see the funny side of things”), as assessed by mode and median, was “as much as always” (value 0). The arithmetic mean was .394, which rounded also is 0. Absolute variability (MAD = .394 and RSI = .5), and relative variability (QCV = 1/3) indicate a moderate-to-low variability. The distribution exhibited a long right tail (PCS = 1), and platycurtosis, characterized by widened shoulders (PCK = .237). Thus, the interquartile range coincided with the percentile range. See Figure 1 and Table 1 .

Items 2 and 7 were similar to item 1 (see Figure 2 and Figure 3 ). Additionally, the median and mode coincided at 0 for item 10. The arithmetic mean of item 10 was very small (m = .239). Its variability was null, as measured by the semi-interquartile range and the coefficient of quartile variation, and very low, as measured by the mean absolute deviation (MAD = .239). It concentrated more on 0 than the previous three. Its profile also showed platykurtosis, but it was symmetrical (see Figure 4 ).

If a floor effect is defined as more than half of the responses being concentrated in the lowest category, then these four items exhibited a floor effect. If the threshold is raised to three-quarters, only item 10 would show a floor effect. In the remaining six items, the first category accounted for less than half of the responses. On the other hand, none of the items showed a ceiling effect.

Note. Measures of central tendency: mdn = median, mo = mode, and m = arithmetic mean. Measures of variation: MAD = median absolute deviation (from the median), SQR = semi-quartile range, and QCV = quartile coefficient of variation. Measures of shape: QCS = quartile coefficient of skewness, PCS = percentile coefficient of skewness, and PCK = percentile coefficient of kurtosis centered at 0. Criteria of high value (case): P80 = percentile 80 and P90 = percentile 90.

The median and mode values of Item 3 were both 2, as was its rounded arithmetic mean (m = 2.014). The variability was medium-low (MAD = .507, SQR = .5, and QCV = .143). The distribution was symmetric (PCS = 0) with mean tails (PCK = −.013) ( Figure 5 ). Item 3 was similar to items 4 and 6 in terms of central tendency and variability. However, the latter two showed negative asymmetry and a long tail towards the left side (PCS = −.333), as well as slight pointing (PCK = −.096). See Figure 6 and Figure 7 .

The central tendency of item 5 was concentrated between 1 and 2 (mdn = 1, mo = 2 and m = 1.246). It exhibited the highest variability among the ten items (MAD = .810, SQR = 1, and QCV = .5). Its profile was leptokurtic, that is, it had widened shoulders (PCK = .237), where the interquartile range coincided with the percentile range. Item 5 was the only symmetrical item (QCS = 0 and PCS = 0). See Figure 8 .

The central tendency of items 8 and 9 was 1, with medium-low variability. Item 8 was symmetrical (PCS = 0) and mesokurtic (PCK = −.013), whereas item 9 exhibited a long tail towards the left side (PCS = −1) and widened shoulders (PCK = .237) due to a shift in responses from category 2 to category 1. However, both items had a longer left tail than right tail, as indicated by the quartile coefficient of skewness (QCS = −1). Thus, their profiles were highly similar. See Figures 9-10 .

Assessed using the Mann-Whitney’s U test, the difference in central tendency was significant for all ten items (p < .001). The mean difference was greater than .75 (one-quarter of the item’s range) for eight items and greater than .5 (one-sixth of the item’s range) for items 2 and 10 ( Table 2 ).

Note. n_LSG = size sample of the low score group: EPDS total score ≤ the 27th percentile of EPDS total score = 6 and n_HSG = size sample of the high score: EPDS total score >= the 73rd percentile of EPDS total score = 12, md = mean difference, U_stat = U statistic, exact_p_value = exact probability value, Z_stat = standardized U statistic, asympt_p_value = asymptotic probability value.

The correlation between each item and the rest of the scale was significantly greater than or equal to .5 in all cases (bootstrap p-values > .05), ranging from .592 to .833, with an average of .690, according to the polyserial correlation coefficient ( Table 3 ). The null hypothesis of bivariate normality was not supported for items 1, 2, 3, and 6, based on the chi-square goodness-of-fit test (p < .05), making Spearman’s rank correlation coefficient a more appropriate choice for these items. For these four correlations, the null hypothesis of a correlation greater than or equal to .5 was also supported (bootstrap p-values > .05), with values ranging from .481 to .672 and an average of .572 ( Table 3 ).

Note. Chi-square goodness-of-fit test: χ² = test statistic, df = degrees of freedom, and p-value = probability of obtaining a value equal to or greater than the observed statistic under the chi-square distribution with df degrees of freedom, for testing the null hypothesis of goodness of fit to the bivariate normality model. Correlations: rₚₛ = polyserial correlation; rₛ = Spearman rank correlation coefficient; 95% BCa CI = 95% confidence interval based on the Bias-Corrected and accelerated (BCa) percentile method with 1000 bootstrap replications (except for item 1, for which 250 replications were used, and item 2, for which 900 replications were used in the rₚₛ estimation); boot p-value = bootstrap probability for the null hypothesis that ρ ≥ .5, based on 1000 replications.

Consistency can be assessed using the ordinal coefficient alpha ( Zumbo et al., 2007 ) if the tau-equivalence assumption holds. If this assumption is not met, consistency can be evaluated using omega coefficients, either McDonald’s ( McDonald, 1999 ) or Green and Yang’s ( Green and Yang, 2009 ).

It is said that tau-equivalence exists when the weights of measurement of a general factor that determines the items are equivalent. Thus, a general factor model with ten indicators was defined with the restriction that the measurement weights are homogeneous ( Figure 11 ). It was estimated by the WLSMV method that appears by default for ordinal variables in the R Lavaan package.

The iterative estimation process converged after the third iteration. The estimated homogeneous factor loading was .777, with 39.7% of the variance in each item explained by the factor. The factor variance was significant (s² = .603, se = .033, z = 18.526, p < .001). Although the Comparative Fit Index (CFI = .971) and the Tucker–Lewis Index (TLI = .970) exceeded the .95 threshold, the null hypothesis of tau-equivalence was rejected based on several fit criteria: the likelihood ratio chi-square test (χ²[44] = 133.6, p < .001), the standardized chi-square statistic (χ²/df = 3.036 > 3), the Root Mean Square Error of Approximation (RMSEA = .120 > .08, 95% CI [.097, .144], p < .001 for H₀: RMSEA ≤ .05), and the Standardized Root Mean squared Residual (SRMR = .110 > .08). Consequently, omega coefficients were used.

Internal consistency, as estimated by McDonald’s omega coefficient, was very high (ω = .932, 95% BCa CI [.908, .947]). The item-deleted omega coefficients were also high, ranging from .916 to .930, with an average of .925. In all ten cases, removing an item led to a decrease in overall internal consistency. The largest decrease (.017) occurred with item 8, and the smallest (.002) with item 3.

However, these decreases were not statistically significant in two-tailed tests at the 5% significance level, as the 95% confidence intervals with and without the item overlapped ( Table 4 ).

Note. total - [i] = the sum of the 10 items excluding the item in square brackets (i = 1, 2, …, 10). ω^M = McDonald’s omega and ω^GY = Green-Yang categorical omega, with theirs point estimates and 95% bootstrap confidence intervals; the latter were calculated using the bias-corrected and accelerated percentile method based on 1,000 resamples with replacement (95% BCa CI). Diff. = difference between the internal consistency estimate with all 10 items and the estimate obtained after removing the item in square brackets. Int.: ↓ if ωₜ > ωₜ₋ᵢ, ↑ if ωₜ < ωₜ₋ᵢ; ns = non-significant change (confidence intervals for ωₜ and ωₜ₋ᵢ overlap); s = significant change (confidence intervals do not overlap).

Overall internal consistency was also very high, as estimated by Green and Yang’s omega coefficient (ω = .912, 95% BCa CI [.882, .926]). In all ten cases, item removal led to a decrease in omega values, ranging from .003 to .028, with an average decrease of .013. However, these decreases were not statistically significant in two-tailed tests at the 5% significance level, as the 95% confidence intervals with and without the item overlapped ( Table 4 ).

The proportion of variance in each item, explained by the predictive model (which includes the other nine items), ranged from .423 (Item 3) to .786 (Item 8), with a mean of .566. Seven out of ten items (70%) had explained variances greater than .5. The predictive model was specified through path analysis and estimated using the WLSMV method ( Table 5 ).

Note. R² = squared multiple correlation or proportion of variance explained by the model (item commonality). The model was calculated through path analysis with its parameters estimated using the Weighted Least Squares with Means and Variances adjusted (WLSMV) method.

The 45 correlations among the 10 EPDS items were estimated using polyserial correlation via the two-step method. In all 45 cases, the null hypothesis of bivariate normality was retained according to the chi-square test; therefore, the 95% confidence intervals were computed using the asymptotic approximation. Correlations ranged from .287 to .878, with a mean of .552. One correlation was below .3, 13 were in the range (.3, .5), 28 in the range (.5, .7], and 3 were above .7 ( Table 6 ).

Note. χ²[df, n = 142] = test statistic value, df = degree of freedom, p-value = probability value calculated towards right tail of a chi-square distribution, r_pc = point estimate of the polychoric correlation, 95% LL = lower limit and UL = upper limit of the 95% asymptotic confidence interval, and ase = asymptotic standard error.

The one-factor model with 10 indicators was estimated using the WLSMV method, based on the previously reported polychoric correlations ( Table 6 ). The iterative process converged after 38 iterations. All measurement weights were statistically significant, ranging from .641 to .934, with a mean of .758. The factor variance was significant (s² = .615, se = .071, z = 8.609, p < .001). See Figure 12 . Bootstrap 95% confidence intervals for the standardized measurement weights (factor loadings) were calculated using the percentile method based on 1000 bootstrap samples ( Table 7 ). The BCa method could not be applied to these data.

The fit of the one-factor model with 10 indicators was adequate according to the likelihood ratio chi-square test (χ²[35] = 54.252, p = .020), the Root Mean Square Error of Approximation (RMSEA = .062 < .08, 95% CI [.025, .094], p = 249 for Ho: RMSEA <= 05), and the Standardized Root Mean squared Residual (SRMR = 076 < 08). Model fit was good according to the remaining indices: standardized likelihood ratio chi-square statistic (χ²/df = 54.252/35 = 1.550 < 2), Comparative Fit Index (CFI = 994 > 95), and Tucker-Lewis Index (TLI = 992 > 95).

The average variance extracted (AVE) exceeded the recommended threshold of .5 (AVE = .582), and McDonald’s omega coefficient was above .7 (ω = .932, 95% BCa CI [.908, .947]). These results indicate that the measurement model demonstrated convergent validity ( Cheung et al., 2024 ).

Note. Parameters: λ(F→EPS) = measurement weight of the indicator EPSi (i = 1, 2, …, 10) and σ² = variance. F = depression factor with ten indicators. Estimates: = unstandardized point estimate, se = standard error, z-value = test statistic, p-value = 2 × (1 - P(Z ≤ |z|)) = two-tailed probability for the null hypothesis: Λ = 0, [95% CI] = standardized measurement weight, including its point estimate and its bootstrap confidence interval obtained using the percentile method.).

No item warranted elimination based on distributional characteristics that deviated markedly from those of the other items, pronounced ceiling or floor effects, poor discriminating power, lack of internal consistency, or low measurement weight. The ten items were discriminative, internally consistent, and contributed significantly to the one-factor model, which demonstrated an acceptable-to-good fit and evidence of convergent validity (AVE > .5 and ω > .7). Given the good fit of the one-factor model with ten indicators, the hierarchical three-factor model proposed by Alfayumi-Zeadna et al. (2022) was not examined.

In this sample of 142 women, the mean of the scale was 9,782 (95% BCa CI [8,917, 10,642]), the median was 10 (95% BCa CI [9, 10]), and the mode, estimated using Grenader’s ( Grenader, 1960 ) method, was 8,995 (95% BCa CI [6,841, 10,590]). The mode based on the highest frequency sample value was 10; this was unique, with a frequency of .12. Thus, the measures of central tendency were very similar.

The sample minimum value was 0, the maximum was 28, and the sample standard deviation was 5.31. The coefficient of variation was 54.3% (95% BCa CI [48.52, 61.87]), which exceeds the expected value for a normal distribution. For this type of data, values of around one quarter or less are usually observed ().

Grubbs’ (1969) outlier test identified the sample maximum of 28 as an outlier (G = 3.431, U = .916, p = .033, < α = .05). Thus, the distribution presented positive asymmetry with a long tail towards the right side (√b₁ = .431, 95% BCa CI [.127, .994]) as tested by D’Agostino (1970) test (z[√b₁] = 2.115, p = .034), indicating non-normality. However, the null hypothesis of mesokurtosis (H₀: β₂ = 3) was maintained (b₂ = 3.026, 95% BCa CI [2.279, 4.822]) according to Anscombe & Glynn (1983) test (z[b₂] = .345, p = .730), which is one of the characteristics of the normal distribution ( Figure 13 ).

The null hypothesis of normality was rejected by the Shapiro & Wilk’s ( Shapiro & Wilk, 1965 ) test with the Royston’s ( Royston, 1992 ) standardization: W = .976, p = .014 < α = .05. This result was confirmed by the Anderson and Darling’s ( Anderson and Darling, 1952 ) test: A² = .825, p = .032 < α = .05, and of Kolmogorov-Smirnov’s test with the probability calculated via Lilliefors’ ( Lilliefors, 1967 ) Monte Carlo simulation: D = .096, p = .003 < α = .05. The histogram, with the normal density curve overlaid, showed that the peak of the sample distribution is higher, the left tail was shorter, and the right tail was heavier than the peak and the tails corresponding to the normal density curve ( Figure 14 ). Since the distribution of the EPDS total score deviated from normal, the scale should be scaled by percentile scores rather than by standard deviation units from the mean. Table 8 shows the sample deciles estimated using R’s type-8 rule.

Note. The sample deciles were estimated using the type-8 rule in R.

In the present sample, a correlated two-factor model was specified for the 44 items of PSSS, and its parameters and fit were estimated using the WLSMV method. The first factor, support from partner (items 1 to 24), showed an average variance extracted (AVE) above .50 (AVE = .712) and high composite reliability (McDoanld’s ω = .983). The second factor, support from significant others (items 25 to 44), exhibited an AVE close to .50 (AVE = .586) and similarly high composite reliability (McDoanld’s ω = .966). The correlation between the two factors was low (r(F1, F2) = .281, 95% BCa CI [.065, .409]), yet statistically significant (z = 4.297, p < .001), indicating a shared variance of 7.9%. The overall composite reliability of the scale was very high (McDoanld’s ω = .988).

Model fit was good according to several indices: χ²/df = 1535.969/901 = 1.705 < 2; GFI = .980, AGFI = .978, CFI = .954, TLI (or NNFI) = .952, Bollen’s IFI = .954, and the Relative Noncentrality Index (RNI) = .954, all exceeding the .95 threshold. Fit was considered adequate based on Bentler-Bonett’s NFI = .896 and Bollen’s RFI = .890, both above .85, and RMSEA = .071, with a 90% confidence interval of [.065, .077], below the .08 cutoff. However, model fit was not supported by the likelihood ratio chi-square test (χ²[901] = 1535.969, p < .001), and the Standardized Root Mean Square Residual (SRMR = .103) exceeded the recommended threshold of .08, indicating poor fit on this index.

The distribution of EPDS total scores deviated from normality according to the Shapiro-Wilk test (W = .979, z = 1.878, p = .030 < .05), as did the distribution of the couple support factor scores (W = .851, z = 6.328, p < .001). However, the null hypothesis of normality was not rejected for the significant other support factor scores (W = .982, z = 1.507, p = .066 > .05).

Correlations between EPDS total scores and PSSS total scores, as well as their two factor scores, were calculated using Pearson’s product-moment coefficient with 95% BCa confidence intervals. Statistical significance was assessed using two-sided bootstrap probabilities. EPDS total scores were negatively correlated with PSSS total scores and the partner support factor scores, both showing strong associations, and with the significant other support factor scores, showing a moderate association ( Table 9 ).

Note. PSSS = Perceived Social Support Scale, r = Pearson’s product-moment correlation, 95% BCa CI = 95% bootstrap confidence interval using the bias-corrected and accelerated percentile method, bias = difference between bootstrap estimate (mean of the 1000 estimates from the 1000 bootstrap samples) and original sample estimate, se = bootstrap standard error, p_boot = two-sided bootstrap probability for the null hypothesis of zero correlation.

As expected, the mean of the group with depression during pregnancy (m₁ = 15.5) was higher than the mean of the group without depression during pregnancy (m₂ = 9.348) with a mean difference of 6.152, 95% BCa CI [3.962, 8.825].

When comparing the central tendency using the one-tailed Mann-Whitney’s U test, the difference was statistically significant: U₁ = 1119.5, U₂ = 200.5, U = 200.5, z(U) = −3.672, p (left-tailed asymptotic probability) = .00012 < α = .05, p (left-tailed bootstrap probability) = .0001 < α = .05. The median EPDS total score of the 10 women with depression during pregnancy was 15 and that of the 132 women without depression during pregnancy was 9.

The effect size was medium, according to both Rosenthal’s ( Rosenthal, 1991 ) correlation (r = |z|/√n = .308 > .3) and Spearman’s ( Spearman, 1904 ) rank correlation (r_s = −.309 < −.3). Likewise, Kerby’s ( Kerby, 2014 ) d, equivalent to Cureton’s ( Cureton, 1968 ) rank-biserial correlation, also indicated a medium effect size: d = r_rb = .696; CI ∈ [.691, .788]. A d value of .696 suggests that in 69.6% of cases, the values in group 1 are higher than those in group 2. In other words, there is a .696 probability that a randomly selected participant from the group with depression during pregnancy (group 1) will score higher on the EPDS than a randomly selected participant from the group without depression during pregnancy (group 2). This value is comparable to the Area Under the ROC Curve (AUC).

When comparing the central tendency in the EPDS total scores among the four marital status groups using the Kruskal-Wallis’s test, significant differences were found at a 5% significance level: H = 8.698, p = .034 ( Table 10 ). Pairwise comparisons using the Dunn’s test with Bonferroni’s correction showed that single women without partners had a significantly higher mean rank than married and cohabiting women ( Table 10 and Table 11 ). As expected, women without partners experience the highest levels of depression. The size of the effect of marital status on depression level was medium according to the epsilon squared (.06 < ε² = H / (n - 1) = 8.698 / (142 - 1) = .062 < .14), but small according to the eta squared (.01 < η² = (H - k + 1) / (n - k) = (8.698 - 3 + 1) / (142 - 3) = .048 < .06).

Due to the small number of single women without partner (n = 4) and single women with partners (n = 6), and the absence of divorced or separated women, this analysis should be interpreted with caution and considered exploratory.

Note. n = subsample size, m [95% BCa CI] = arithmetic mean with a 95% BCa confidence interval, MR = mean rank, H = value of the test statistic, df = degrees of freedom, p = P(χ²[3] ≥ H) = probability at the right tail of a chi-square distribution with three degrees of freedom.

Note. RM = mean rank of the group: 1 = married women, 2 = single with a partner, 3 = single without a partner and 4 = cohabiting, d = difference between mean ranks, se = standard error of the difference, z = d/se = standardized value of the difference, p = 2 × (1 – P(Z ≤ |z|) = two-tailed probability value in a standard normal distribution, p_aj. = min(p × 6, 1), where 6 represents the number of tested differences or probability value with the Bonferroni adjustment for type I error (false positives) in multiple comparisons.