Participants included pregnant women and mothers of children aged 0–2 years residing in Spain. Recruitment combined paid online advertisements with outreach through midwifery and breastfeeding associations. The first wave (T1) was conducted in April 2021, with 1242 eligible women completing the baseline survey. The second wave (T2) occurred three months later and included 182 participants (14.6% retention), while the final wave (T3) occurred approximately three years after T2 and included 23 participants (1.9% retention). The extended follow-up interval was intended to capture long-term changes in parenting experiences, child development, and family dynamics.

Inclusion criteria were identical across waves: participants had to be aged 18 years or older, reside in Spain, be fluent in Spanish, and have internet access to complete the online questionnaires. Both pregnant women and mothers of children born within the two years preceding participation were eligible. Information on ethnicity and socioeconomic status (SES) was not collected, which consequently limits generalizability; however, given that Spain is predominantly a White European population, substantial ethnic variability was unlikely. Recruitment through online and community channels may also have resulted in a self-selected sample of women who were more motivated or health-conscious than average, potentially narrowing variability in ambivalence scores.

Each participant received a unique alphanumeric code to link responses across waves while preserving anonymity. Electronic informed consent was obtained via the survey platform before participation, and consent records were automatically archived. Procedures adhered to the Declaration of Helsinki and the General Data Protection Regulation (GDPR; EU 2016/679). The study protocol was approved by the Ethics Committee of Universitat Jaume I (CD/22/2021; 15 April 2021). Participants could withdraw at any time by closing the survey, and all data were stored securely on encrypted university servers.

Maternal ambivalence was assessed using the 14-item Maternal Ambivalence Scale [8], which captures three dimensions: doubts about oneself or motherhood, rejection of motherhood, and suppression of ambivalent feelings. Items are rated on a 4-point Likert scale (1 = completely disagree–4 = completely agree), with higher scores indicating greater ambivalence. The instrument has demonstrated satisfactory internal consistency ($\alpha$ = 0.70–0.83) and convergent validity with measures of anxiety, depression, and life satisfaction in Spanish samples [8].

All analyses were conducted in R (v4.5.1., R Core Team, R Foundation for Statistical Computing, Vienna, Austria, https://www.r-project.org) and IBM SPSS Statistics 26 (IBM, Armonk, New York, USA, https://www.ibm.com/products/spss-statistics). Descriptive statistics summarized sociodemographic characteristics and ambivalence scores at each time point (T1, T2, T3). All tests were two-tailed with a significance level of $\alpha$ = 0.05.

The primary analyses employed linear mixed-effects models (LMMs) to examine changes in maternal ambivalence across time while accounting for unequal intervals between waves (~3 months between T1–T2; ~36 months between T2–T3) and missing data due to attrition. Time was coded continuously in months since baseline (0, 3, 36). Each model included random intercepts for participants and fixed effects for time, maternal age, parity (primiparous vs. multiparous), perinatal status (pregnant vs. mother; time-varying), and education. Time $\times$ perinatal status interactions were estimated to explore differential trajectories between pregnant and postpartum participants. Parameters were obtained using restricted maximum likelihood (REML) with Satterthwaite’s approximation for denominator degrees of freedom, assuming an unstructured residual covariance matrix. Model assumptions were evaluated by inspecting residuals and random-effect distributions. Sensitivity analyses using inverse-probability weighting (IPW) and the T1–T2 completer subsample were performed to assess the robustness of the results.

To triangulate findings, repeated-measures general linear models (RM-GLMs) were conducted as a complementary analytic strategy. RM-GLMs were fitted with time (T1, T2, T3) as a within-subjects factor treated categorically, allowing direct pairwise comparisons despite unequal time intervals. When Mauchly’s test indicated violations of sphericity, Greenhouse–Geisser corrections were applied. Partial $\eta$² with 95% CIs was reported as an effect-size measure, and Bonferroni-adjusted pairwise contrasts identified specific wave-to-wave differences. Perinatal status was entered as a time-varying covariate, and maternal age and parity were included as between-subjects covariates.

Rank-order (relative) stability of ambivalence dimensions was assessed using Pearson correlations across waves, with 95% CIs computed via Fisher’s z transformation.

Attrition bias was examined by comparing baseline characteristics of completers and dropouts (T1–T2) using independent-samples t tests for continuous variables and $\chi$² tests for categorical variables. Standardized mean differences (Hedges’ g) and risk differences (RDs) with 95% CIs were calculated to quantify group differences. Little’s Missing Completely at Random (MCAR) test indicated that missingness was consistent with a missing-at-random (MAR) mechanism. Additionally, logistic regression models predicting retention from baseline age, parity, perinatal status, education, and ambivalence scores were used to generate IPW values, which were applied in all sensitivity analyses.

At baseline (T1), 1242 women completed the survey. At the 3-month follow-up (T2), 182 participants (14.6%) provided data, and 23 participants (1.9%) remained at the 36-month follow-up (T3). Attrition was entirely passive, with no withdrawals of consent. At T1, 33.2% of participants were pregnant, and 66.8% were mothers of children under two years [Mean (M)_age = 34.70, SD = 4.87, range 18–50]. At T2, 23% were pregnant (M_age = 36.12, SD = 4.51), and at T3, all participants were mothers (M_age = 35.83, SD = 3.93), with 81.6% being multiparous.

Baseline comparisons between T2 completers and drop-outs are presented in Supplementary Table 1. T2 completers were slightly older than drop-outs [Hedges’ g = –0.35, 95% CI (–0.50, –0.19)], but no significant differences were observed for parity, perinatal status, or T1 ambivalence subscale scores (all $|$g$\mathrm{|}$ $\le$ 0.03, p values $\ge$ 0.63). Education differed modestly between groups [Cramer’s V = 0.15, 95% CI (0.10, 0.21)]. Logistic regression indicated that age predicted T2 retention [odds ratio (OR) = 1.07, 95% CI (1.03, 1.11); p 0.001], whereas parity, perinatal status, subscale scores, and education were not significant predictors. Little’s MCAR test was non-significant (p = 0.21), suggesting that missing data were largely at random.

To address potential attrition bias, IPW was applied to T1–T2 analyses. Weighted LMMs yielded results highly consistent with complete-case analyses (Supplementary Tables 2,3), confirming that differential attrition did not meaningfully bias short-term inferences. T1–T3 findings were interpreted cautiously due to the extremely small T3 sample (n = 23), and the post hoc power analyses indicated that only medium-to-large effects (Cohen’s d $\approx$ 0.50–0.55) could be reliably detected, meaning smaller effects may have remained undetected.

Internal consistency across waves was acceptable to good: doubts ($\alpha$ = 0.86/0.87/0.85), rejection ($\alpha$ = 0.83/0.84/0.84), and emotional suppression ($\alpha$ = 0.79/0.80/0.81). Rank-order stability from T1 to T2 was high for doubts [r = 0.78, 95% CI (0.71, 0.84)] and rejection [r = 0.83, 95% CI (0.76, 0.88)], and moderate for suppression [r = 0.67, 95% CI (0.58, 0.75)]. Stability decreased at T3, with suppression showing negative T1–T3 and T2–T3 correlations [r = –0.60, 95% CI (–0.82, –0.24); r = –0.56, 95% CI (–0.78, –0.21)], suggesting potential inversions in relative suppression scores over time. This pattern may reflect true individual differences, selective retention of participants, or measurement variability due to the very small T3 sample, and should be interpreted as exploratory.

Descriptive statistics, rank-order correlations, and LMM estimates for all waves are summarized in Table 1 and Supplementary Tables. Effect sizes were consistently reported: Hedges’ g for pairwise contrasts, partial $\eta$² for analyses of variance (ANOVAs), and $\beta$ (95% CI) for LMMs. Covariates contributed minimally to observed changes over time, and residual confounding from unmeasured variables (e.g., broader SES or partner support) is unlikely to account for the null T1–T2 effects.

LMMs with random intercepts for participants (1$|$ID) were fit, adjusting for age, parity, education, and perinatal status at T1. Time was coded in months since T1 (~3 months), and time $\times$ perinatal status interactions were included to test for differential trajectories between pregnant and postpartum women. Residual diagnostics indicated acceptable normality and homoscedasticity.

Across subscales, mean-level change from T1 to T2 was negligible (Table 1):

$\bullet$ Doubts: $\beta$ = 0.07, 95% CI (–0.71, 0.85); p = 0.859

$\bullet$ Rejection: $\beta$ = 0.02, 95% CI (–0.43, 0.47); p = 0.929

$\bullet$ Suppression: $\beta$ = –0.03, 95% CI (–0.69, 0.64); p = 0.940

Time $\times$ perinatal status interactions were also non-significant:

$\bullet$ Doubts: $\beta$ = –0.05, 95% CI (–0.23, 0.13); p = 0.591

$\bullet$ Rejection: $\beta$ = 0.04, 95% CI (–0.07, 0.15); p = 0.454

$\bullet$ Suppression: $\beta$ = 0.06, 95% CI (–0.10, 0.21); p = 0.482

Effect sizes for T1–T2 paired comparisons were small (Hedges’ g = –0.03 to 0.09), and marginal R² indicated that fixed effects explained 1–2% of variance. In contrast, conditional R², including random intercepts, ranged from 0.67 to 0.84, reflecting substantial between-subject variability. RM-ANOVA triangulation confirmed negligible mean-level changes across subscales (doubts: $\mathrm{\Delta }$M = 0.11, t(180) = 0.62; p = 0.536; rejection: $\mathrm{\Delta }$M = –0.18, t(180) = –1.77; p = 0.078; suppression: $\mathrm{\Delta }$M = –0.22, t(180) = –1.44; p = 0.152; partial $\eta$² = 0.003–0.015). Any potential masking of short-term changes by selective attrition was minimal, given the very small effect sizes.

Weighted LMMs mirrored complete-case models (Supplementary Table 3).

$\bullet$ Doubts: No significant change was observed over time ($\beta$_time 0.01; p = 0.932). Primiparous participants reported higher baseline doubts compared to non-primiparous participants ($\beta$ = 1.02; p = 0.003). Higher education (University or Master/PhD) was associated with lower doubts ($\beta$_university = –0.71; p = 0.042; $\beta$_Master/PhD = –0.73; p = 0.044). No significant associations were observed for the other covariates.

$\bullet$ Rejection: Time did not significantly predict rejection ($\beta$_time = 0.01; p = 0.330). Older age was associated with higher rejection scores ($\beta$ = 0.07; p 0.001). No significant associations were observed for the other predictors, including parity and education.

$\bullet$ Suppression: Suppression increased significantly over time ($\beta$_time = 0.16; p 0.001). Primiparous participants showed lower baseline suppression ($\beta$ = –0.49; p = 0.046). No significant associations were observed for the other covariates.

Random effects suggested moderate between-person variability for doubts (ICC $\approx$ 0.35) and lower variability for rejection and suppression, indicating that attrition was unlikely to bias the observed short-term trajectories.

Weighted LMMs including all available data from T1 to T3, adjusted for age, parity, education, and perinatal status at T1, indicated the following:

$\bullet$ Doubts: There was no significant change over time [$\beta$_time 0.01, 95% CI (–0.02, 0.03); p = 0.932], and the trajectory did not differ between pregnant and postpartum women [time $\times$ perinatal status = 0.15, 95% CI (–0.42, 0.72); p = 0.613].

$\bullet$ Rejection: Scores remained stable [$\beta$_time = 0.01, 95% CI (–0.01, 0.02); p = 0.330], with no significant time $\times$ perinatal status interaction [–0.22, 95% CI (–0.59, 0.16); p = 0.252].

$\bullet$ Suppression: There was a moderate, statistically significant increase over time [$\beta$_time = 0.16, 95% CI (0.14, 0.19); p 0.001], which did not differ by perinatal status [time $\times$ perinatal status = 0.07, 95% CI (–0.34, 0.48); p = 0.728].

Primiparity predicted higher baseline doubts and lower baseline suppression; older age predicted slightly higher rejection. Triangulation with RM-ANOVA showed similar patterns: negligible changes in doubts and rejection, and a moderate increase in suppression (partial $\eta$² = 0.622). However, the effect size was likely inflated due to the very small T3 sample.

Given the extremely limited T3 sample (n = 23), negative rank-order correlations for suppression (T1–T3 r = –0.60, T2–T3 r = –0.56) should be interpreted with caution. These may reflect true reversals in relative suppression scores, selective retention of participants, or measurement variability. Therefore, the T3 results are exploratory, and small changes may have gone undetected.

Despite the notable strengths of this research, this study also has several limitations. Attrition across waves was substantial, particularly at T3 (1.9% retention), severely limiting statistical power and rendering long-term findings exploratory. Recruitment via community and online channels may have favored women already engaged in psychological well-being, potentially underrepresenting those experiencing greater distress or limited access to resources. Negative rank-order correlations for suppression suggest variability or potential measurement issues in the very small T3 sample, which may reflect either true reversals in relative scores or artifacts of coding and selective retention.

The absence of detailed SES and partner support data constrains interpretation, as both factors can influence perinatal mental health and maternal ambivalence [20]. Higher SES and stronger partner support are generally protective, so observed trajectories may underestimate risk among women with lower SES or less support. Education was included as a proxy in LMMs and IPW sensitivity analyses, and results were largely unchanged, suggesting that primary inferences are robust. Overall, missing data were largely random (Little’s MCAR; p = 0.21), and complete-case and weighted analyses converged. Thus, future research should replicate these findings in larger, more diverse samples and incorporate multimethod assessments, including behavioral and partner-reported measures, to improve ecological validity [20]. Finally, T3 participants were all postpartum, so time $\times$ perinatal status interactions are effectively limited to T1–T2, and interpretations regarding perinatal status at T3 should be made cautiously.