Bravery of giving birth in the homeland: pregnant women’s and new mothers’ decisions on staying in Ukraine during the war Authors: Liudmyla Krupelnytska, Nazar Yatsenko, Magdalena Chrzan-Dętkoś, Natalia Murawska, Alona Vavilova, Olha Morozova-Larina, Helena S García López, Maria F Rodríguez-Muñoz Published: 2026 ABSTRACT AbstractBackgroundThe full-scale Russian invasion of Ukraine has profoundly disrupted maternal care systems, exposing pregnant women and new mothers to extreme contextual adversity. While many relocated abroad, others remained to give birth in a war zone. The psychosocial and health-related factors shaping these decisions remain poorly understood. The purpose of the study is to identify the psychological, contextual, and dispositional factors associated with women’s wartime decisions to remain in Ukraine or relocate abroad.MethodsCross-sectional data were collected online between December 2022 and October 2023 from 730 Ukrainian women who were pregnant or postpartum during the full-scale invasion. Decision tree modelling, supported by random forest analysis, was used to explore predictors of relocation. Regression models and equivalence testing examined group differences in personality traits. Missing data were addressed using multiple imputation.ResultsPerceived disruption of daily life and geographic proximity to hostilities were the strongest predictors of relocation decisions. These contextual factors consistently outperformed dispositional traits, including Big Five personality dimensions. Personality traits did not distinguish women who relocated abroad from those who remained in Ukraine overall. However, higher Extraversion appeared as a secondary factor in the decision-tree models, suggesting that it may be associated with remaining in Ukraine under specific contextual conditions. Women who gave birth abroad exhibited higher rates of postnatal depression than those who remained.ConclusionsRelocation decisions among pregnant women and new mothers during war are driven primarily by acute contextual adversity rather than individual personality traits. These findings align with evidence from non-war migration contexts and highlight the need for trauma-informed, context-sensitive approaches in perinatal care and policy during armed conflict.Trial registrationClinicalTrials.gov, NCT05654987, 2022-12-16.Keywords: Perinatal mental health, Women, Maternity, Displacement, Non-displaced persons, Relocation decision, War, Ukraine Background The full-scale Russian invasion of Ukraine, launched on 24 February 2022, has resulted in widespread exposure to traumatic events, including forced or violent displacement, sexual violence, restrictions on freedom, torture, and loss [1]. These conditions have contributed to elevated rates of anxiety, depression, and post-traumatic stress disorder (PTSD) across the Ukrainian population [2, 3]. While the war has affected all segments of society, women have been disproportionately impacted by its psychological and social consequences. Extensive evidence from previous armed conflicts indicates that women are significantly more likely than men to develop mental disorders in response to war-related adversity [4, 5]. For example, data from the 1998–1999 Kosovo War indicate that 19.7% of women met the criteria for PTSD, compared to 12% of men, even after controlling for trauma load [6]. This gender disparity remained evident two years later, with women continuing to report higher PTSD symptom severity [7]. Similar patterns have been documented following the 2006 Lebanon War [8], the 2012–2013 M23 rebellion in the Democratic Republic of Congo [9], and the 2020–2022 Tigray War in Ethiopia [10]. Notably, these disparities often persisted despite women reporting fewer direct traumatic exposures, with relational losses and social disruption emerging as salient predictors [11, 12]. Recent studies in Ukraine, the Gaza Strip, and Israel have confirmed similar patterns of women’s heightened vulnerability to war-related mental health consequences [13–15]. Within this broader gendered impact, pregnant women and new mothers represent an especially at-risk subgroup. In both clinical and humanitarian contexts, pregnancy and the postpartum period are recognised as times of heightened biological, psychological, and social vulnerability [16]. Armed conflict exacerbates these vulnerabilities by disrupting access to antenatal and perinatal care, increasing risks of obstetric complications, and intensifying exposure to traumatic stressors [17]. The collapse or degradation of healthcare infrastructure further undermines maternal safety, heightens the risk of perinatal mortality, and limits access to emergency obstetric interventions [18]. A similar trend has been observed since 2014 [19]. It persists under the current full-scale war, affecting the mental and physical health of pregnant and new mothers [20]. Since the onset of the full-scale invasion in 2022, the Russian Federation has systematically targeted Ukraine’s healthcare infrastructure, with especially severe implications for maternal care. By 2025, approximately 10% of medical facilities had sustained direct damage, including numerous maternity hospitals [21]. Notably, a maternity hospital in Mariupol was bombed on 29 March 2022, within the first weeks of the full-scale invasion [22], and more than three years later, on 11 June 2025, a maternity hospital in Kharkiv was destroyed [23]. Such targeted attacks on maternal healthcare facilities have persisted throughout the years of war, affecting multiple cities across the country [24]. In response to these threats, a substantial number of pregnant women and new mothers were compelled to relocate either within Ukraine or across international borders [25]. For instance, in 2022, a baby boom was recorded in Poland, largely attributed to births among Ukrainian refugees [26]. The overall scale of displacement is staggering: more Ukrainians were uprooted than the entire population of Austria or Switzerland, and approximately 63% of refugees are women [27]. Nevertheless, despite the scale of trauma and disruption, some women remained in Ukraine and gave birth while under siege. Emerging data indicate that remaining in place was not necessarily associated with worse psychological outcomes. According to Lushchak et al. [28], women who did not flee exhibited a slightly lower prevalence of probable PTSD (50.8%) compared to those who relocated (62.2%). Similarly, a review by Stevenson et al. [29] concluded that forced relocation was associated with an elevated risk of perinatal depression, affecting roughly one in three displaced women. These findings challenge the intuitive assumption that relocation offers psychological relief and raise important questions about the factors guiding such critical decisions. Understanding what motivates women to stay or leave in the face of danger thus necessitates a closer examination of the underlying decision. Migration has often been conceptualised as a product of rational deliberation grounded in long-term cost–benefit analyses [30]. However, empirical evidence increasingly challenges this assumption, suggesting that the rational-choice framework provides, at best, an oversimplified account [31, 32]. Decisions in crisis contexts are rarely purely rational. Instead, they are shaped by emotional distress, time constraints, individual circumstances, and uncertainty or inconsistency in available information [33, 34]. Among pregnant women and new mothers, this decision-making is further influenced by specific psychosocial and health-related considerations [35]. Studies on crisis-driven relocation commonly draw on the framework of “push” and “pull” factors, drivers that either compel individuals to leave or attract them to alternative location [36, 37]. Exposure to war-related violence and threats for personal security typically constitute strong push factors [38]. However, as Biehler [39] notes, studies often focused on examining the reasons for leaving, with less attention given to understanding why some choose to stay. Among the key deterrents to relocation is the complex factor of “preference for home,” rooted in longstanding social networks, place-based identity, and accumulated local economic opportunities [37]. For women, the choice to migrate may indeed offer safety and access to healthcare. Yet it often entails a separation from partners and family members – a particularly acute burden during the postpartum period [40, 41]. These patterns were notably confirmed in recent studies examining migration decision-making and mobility trajectories during the Russian war in Ukraine [42]. This study seeks to expand understanding of the women who decide to give birth in war-affected Ukraine. Specifically, it aims to identify the psychological, contextual, and dispositional factors that are related with their decisions. In line with this objective, the following research questions (RQ) were formulated: RQ1: What factors are associated with pregnant women’s and new mothers’ decisions to stay in Ukraine during the war? RQ1: What factors are associated with pregnant women’s and new mothers’ decisions to stay in Ukraine during the war? RQ2: Are personality traits associated with the war-driven relocation decisions among pregnant women and new mothers? RQ2: Are personality traits associated with the war-driven relocation decisions among pregnant women and new mothers? Methods Design and procedure This study forms part of a broader observational cohort investigation, with the study protocol published in advance [43]. The project was preregistered as a clinical trial (NCT05654987) and received ethical approval from the Research Ethics Committee of the Universidad Nacional de Educación a Distancia (Ref: 21-PSI-2022), the Research Ethics Committee of the Faculty of Psychology, Taras Shevchenko National University of Kyiv (Date: August 2022; Decision No. 04; Protocol No. 01), and the Ethics Board for Research Projects at the Faculty of Social Sciences, University of Gdańsk (Approval No. 13-2022). Data collection Data were collected between 1 December 2022 and 1 October 2023. During this period, all participants were either pregnant or in the postpartum period. The survey was administered online through an open survey link. Recruitment was carried out through multiple channels, including Facebook groups, Instagram, personal networks, and Ukrainian support centres across European countries. Because the invitation was disseminated through a general link, the number of women who received, viewed, or opened the invitation could not be determined. Therefore, a formal response rate could not be calculated. This recruitment strategy should be understood as non-probability, convenience-based online sampling, which may have limited the representativeness of the sample. However, this approach was considered necessary given the war-related constraints [44, 45] on direct recruitment, participant safety, mobility, and access to pregnant women and new mothers across different locations [46]. Before analysis, all raw survey submissions were reviewed individually to determine eligibility, completeness, and possible duplication. Eligibility was assessed based on the study inclusion criteria, completion of the sociodemographic questionnaire, and sufficient completion of the outcome measures. Potential duplicate or repeated submissions were identified using IP address and then evaluated against the consistency of response content. Repeated entries reflecting the same participant were merged where possible. Submissions that did not meet the eligibility or completeness criteria were excluded during data preparation. After this screening process, six additional cases were excluded from the present analysis because relocation status was missing, which precluded classification into the comparison groups. Sample The final sample comprised 730 women, all of whom were residing in Ukraine at the outbreak of the full-scale war on 24 February 2022. Of these, 617 remained in Ukraine, while 113 relocated abroad as refugees. Inclusion criteria were as follows: (a) being pregnant or the biological mother of an infant aged up to 12 months; (b) aged 18 years or older; (c) residing in Ukraine during or after the onset of war, or having refugee status following displacement to the European Union from 24 February 2022 onwards; and (d) provision of informed consent to participate. Measures Assessment, Care, and Trust – in Pregnant and New Mothers (ACT-PNM) The questionnaire was developed specifically for the present study to assess various aspects of women’s backgrounds [43]. It was designed as a structured data-collection instrument based on the World Health Organization’s framework for social determinants of health [47] and captured sociodemographic, birth-related, and war-related variables relevant to women’s experiences during pregnancy and the postpartum period in wartime Ukraine. In the present study, questionnaire variables were analysed as separate observed indicators potentially influencing relocation decisions; no total score or latent construct was derived. Accordingly, psychometric validation procedures were not applicable. Relocation status was treated as a multinomial outcome, with participants categorised as either displaced or non-displaced, the latter group comprising both those who stayed in their original settlements and those who became internally displaced persons. Ten-Item Personality Inventory (TIPI) The Ten-Item Personality Inventory is a brief self-report measure of the Big Five personality traits [48]. It comprises 10 items, with two items assessing each of the five traits: Extraversion, Agreeableness, Conscientiousness, Emotional stability, and Openness. Responses are rated on a 7-point Likert scale ranging from 1 to 7, with higher scores indicating stronger expression of the respective trait. In the present study, the Ukrainian version of the measure developed by Klimanska and Haletska was employed [49]. As reported by the authors, internal consistency of the scales ranged from poor to acceptable (α = 0.44–0.74), test–retest reliability and convergent validity were satisfactory. The Edinburgh Postnatal Depression Scale (EPDS) The EPDS is a widely used self-report measure of the depressive symptoms in peripartum women [50]. It comprises 10 items, covering subscales for anhedonia, anxiety, and depressive symptoms. Responses rated on a 4-point Likert scale ranging from 0 to 3, with higher scores reflecting greater severity of postnatal depressive symptoms. In the present study, the Ukrainian version of the EPDS was developed. This version demonstrated excellent internal consistency for the overall scale (α = 0.87) and good for its subscales (α = 0.75–0.76), as reported by Chrzan-Dętkoś et al. [51]. A cut-off score of 13 was applied to identify women with clinically significant postnatal depressive symptoms [52]. Data analysis Relocation decisions modelling To investigate factors associated with the decision to relocate, we constructed a random forest models using candidate predictors selected as theoretically relevant [53]. This step was used to identify which variables most strongly distinguished women who relocated abroad from those who remained in Ukraine. Predictors were considered eligible if they reflected experiences or characteristics present before or at the onset of the full-scale invasion, or relatively stable dispositional traits unlikely to be affected by subsequent events. This variable selection strategy was necessitated by the retrospective study design, in which data were collected only after participants had either relocated abroad or remained in Ukraine. The inclusion of post-relocation variables would introduce substantial bias and conflate predictors with outcomes. The candidate set comprised war-related contextual factors, namely perceived wartime impact on daily life and proximity of the participant’s place of residence to hostilities; sociodemographic and resource-related factors, including pre-war financial condition, educational level, and perceived social support; and dispositional characteristics, represented by the Big Five personality traits. Maternal status was addressed as a stratifying variable rather than treated only as one predictor within a single model, with separate models conducted for pregnant women and new mothers. This decision reflected the central role of pregnancy and postpartum status in the study design, as well as the possibility that relocation decisions may be shaped by different constraints during pregnancy and early motherhood. Variable importance was evaluated using two random forest metrics: mean decrease in accuracy and mean decrease in the Gini index, to assess the consistency of ranking across methods. Given that the relocation outcome was imbalanced, with a substantially larger proportion of the sample remaining in Ukraine, the random over-sampling examples (ROSE) algorithm was applied to the dataset to create a balanced synthetic training samples for the classification models [54]. ROSE generates synthetic observations through a smoothed bootstrap procedure, thereby increasing the representation of the minority class. This approach was used to reduce the tendency of the model to favour the majority class and to improve the identification of profiles associated with relocation abroad. Subsequently, based on the random forest results, a classification and regression tree (CART) model was constructed to provide an interpretable representation of how the selected predictors jointly classified participants into relocation groups. The CART algorithm recursively partitioned the dataset into internally homogeneous subgroups according to the selected predictors [55]. The minimum terminal node size was set at 30 observations for the full sample and 10 for subgroup analyses to reduce overfitting and ensure interpretability. Variable importance within the CART model was quantified by the total reduction in the Gini index. Evaluation of the complexity parameter indicated minimal influence on the resulting tree structure. Model performance was evaluated using relative and cross-validated error, sensitivity, specificity, area under the curve (AUC), and proportional reduction in classification error relative to the root node. Because ROSE changes the class distribution used for model training, the terminal-node probabilities in the CART models should be interpreted as model-derived classification probabilities used descriptively to compare relocation profiles across combinations of predictors. Comparative analysis Group comparisons were conducted to describe differences between displaced and non-displaced women in sociodemographic, war-related, and mental health characteristics. Comparisons of categorical outcomes, such as demographic characteristics and whether participants met the EPDS clinical cut-off, were conducted using chi-squared tests. Continuous outcomes related to postnatal depressive symptoms were compared across groups using Welch’s t-tests, which allow for unequal variances. Group differences in personality traits were tested using ordinary least squares regression models, with relocation status entered as the independent variable. Each Big Five trait served as a separate dependent variable. All between-group comparisons were conducted using multiply imputed datasets. All p-values were two-tailed and adjusted using Holm-Bonferroni correction for multiple imputed datasets. To further assess magnitude of between-group differences, equivalence testing was conducted using the two one-sided tests (TOST) procedure. Equivalence bounds set at 0.25 standard deviations and were converted into trait-specific raw-score bounds for interpretation. Equivalence was evaluated using 90% confidence intervals and TOST p-values. Standardised effect sizes were reported as Cohen’s d with corresponding confidence intervals. For both conventional group-difference tests and TOST analyses, unadjusted p-values were reported alongside Holm-Bonferroni-adjusted values. Missing data handling Prior to analysis, six cases were excluded due to missing data on relocation status, which precluded their classification into comparison groups. For the construction of the random forest model, missing values were imputed using a non-iterative, single-pass procedure based on median substitution. This approach was considered appropriate given that the model was employed solely for assessing variable importance, not for inferential purposes. Accordingly, this imputation step was used only to support predictor ranking in the random forest model. In contrast, personality trait variables exhibited item-level missingness ranging from 26% to 29%, warranting a more rigorous approach. Different missing-data approaches were used because the analyses served different purposes: the random forest model was used only for exploratory variable ranking, whereas the group comparisons required pooled statistical estimates. Multiple imputation by chained equations was applied using predictive mean matching [56]. Twenty imputed datasets were generated. All subsequent analyses were performed within each dataset and pooled using Rubin’s rules [57]. The fraction of missing information ranged from 0.20 to 0.29, with relative efficiency exceeding 95% for all variables. Software All statistical analyses were conducted using RStudio (version 2025.05.1 + 513) for macOS Sequoia 15.5. A two-tailed significance threshold of α = 0.05 was applied. The complete analysis script, along with the dataset used in the present study, is available via the Open Science Framework [58]. Results Participants were drawn from virtually all major regions of Ukraine; however, several regions are absent from the sample (see Fig. 1). No respondents were recorded from the Russian-occupied territories, specifically, from the Luhansk and Donetsk Oblasts (Oblast refers to an administrative region in Ukraine, similar to a province), and the Autonomous Republic of Crimea. Also no respondent recruited from Kherson or Chernihiv Oblasts. Under-representation of Kherson region is plausibly linked to its partial occupation at the time of data collection, whereas the absence of respondents from Chernihiv may be attributable to heightened security risks and ongoing aerial bombardment in the region [59]. Geographic distribution of respondents’ settlements of residence within Ukraine. Note: Administrative boundaries correspond to the 1991 oblast structure of sovereign Ukraine, prior to territorial occupation. Alt text: Respondent residences are distributed across Ukraine, with denser clusters in central and eastern regions and no data from occupied territories An overall comparison between women who stayed in Ukraine and those who relocated abroad is presented in Table 1. No significant differences were observed in proximity to hostilities of their original residence (p = .078), pre-war financial condition (p = .5), or highest level of education (p = .9). However, perceptions of wartime impact on daily life differed markedly (p < .001), with displaced women more frequently (32–40%) reporting their daily life as very highly disrupted by outbreak of war in contrast to 10–20% among non-displaced. A similar pattern emerged for postnatal depression, with significantly lower prevalence among women staying in Ukraine (p = .004). This group also reported lower levels of depression, anxiety, and anhedonia symptoms (p < .05). Table 1Sociodemographic characteristics of the study sampleCharacteristicsResided in Ukraine(n = 617)Displaced Abroad(n = 113)p-value1Pregnant womenn = 205New mothersn = 412Pregnant womenn = 29New mothersn = 84Proximity to hostilities, n (%)0.078 Distal50 (25%)99 (24%)5 (17%)7 (8.4%) Intermediate94 (46%)185 (45%)14 (48%)44 (53%) Proximal60 (29%)125 (31%)10 (34%)32 (39%) n/d21 (1.2%)3 (0.7%)1 (0.5%)Pre-war financial condition, n (%)0.5 High32 (17%)59 (16%)7 (25%)20 (26%) Low14 (7.6%)25 (6.8%)2 (7.1%)4 (5.3%) Moderate139 (75%)281 (77%)19 (68%)52 (68%) n/d20 (9.8%)47 (11%)1 (3.4%)8 (9.5%)Wartime impact on daily life, n (%)< 0.001 High34 (18%)68 (19%)8 (29%)20 (26%) Low19 (10%)37 (10%)1 (3.6%)3 (3.9%) Moderate73 (39%)106 (30%)9 (32%)11 (14%) Rather high40 (22%)71 (20%)1 (3.6%)12 (16%) Very high19 (10%)72 (20%)9 (32%)31 (40%) n/d20 (9.8%)58 (14%)1 (3.4%)7 (8.3%)Highest level of education, n (%)0.9 Postgraduate111 (54%)228 (55%)14 (48%)47 (56%) Undergraduate94 (46%)184 (45%)15 (52%)37 (44%)Postnatal depression3, n (%)0.004 No15 (54%)129 (72%)31 (43%)195 (57%) Yes13 (46%)49 (28%)41 (57%)149 (43%)EPDS score, M (SD)13 (6)10 (6)13 (6)11 (6)< 0.001 Depressive symptoms, M (SD)5.4 (3.3)3.9 (3.2)5.4 (3.5)4.5 (3.3)0.003 Anxiety symptoms, M (SD)4.86 (2.07)4.06 (2.33)5.19 (2.01)4.59 (2.26)0.002 Anhedonia, M (SD)2.39 (2.06)1.71 (1.78)2.60 (1.97)2.09 (1.85)0.0061 p-values reflect chi-squared tests for categorical variables and Welch’s t-tests for continuous variables 2 n/d – no data; this category denotes item-level nonresponse for the corresponding variable and is shown only where applicable3 postnatal depression defined as meeting the EPDS clinical cut-off of 13 Sociodemographic characteristics of the study sample 1 p-values reflect chi-squared tests for categorical variables and Welch’s t-tests for continuous variables 2 n/d – no data; this category denotes item-level nonresponse for the corresponding variable and is shown only where applicable 3 postnatal depression defined as meeting the EPDS clinical cut-off of 13 To examine factors associated with relocation decisions among pregnant women and new mothers in Ukraine during wartime, a classification tree model was constructed using five predictors identified through random forest variable selection: perceived wartime impact on everyday life, proximity of residence to hostilities, perceived pre-war financial condition, and Emotional stability and Extraversion trait scores from the Big Five personality model (see Fig. 2). Decision tree modelling relocation probability in the full sample. Note: The tree should be read from the root node downward. Each split separates participants with different probabilities of relocation abroad. Decimal value always represents probabilities of relocation abroad; for nodes classified as non-displaced, it represents the proportion of cases expected to belong to the displaced group and therefore also indicates the expected loss. The percentage reflects node weights – the proportion of participants assigned to node. Alt text: Lower perceived wartime disruption, higher extraversion, and moderate financial status are associated with remaining in Ukraine, as shown by primary splits in the decision tree The classification model was trained on a balanced sample of 515 respondents and pruned using a complexity parameter (cp.) of 0.002, with a minimum terminal node size of 30. The final tree consisted of five splits. The root node misclassification rate was 46.8%, with a relative error of 0.63 and a cross-validated error of 0.76 (xstd = 0.045), corresponding to a 37.3% apparent reduction and a 23.7% cross-validated reduction in classification error relative to the root node. Additional performance indices indicated limited-to-moderate discriminatory performance, with sensitivity of 62.2%, specificity of 78.1%, and an AUC of 0.74. Despite limited performance gain compared to simpler models, this solution was retained because it provided an interpretable representation of the main predictor combinations. The most informative split occurred on the root node based on perceived wartime impact on daily life, with an improvement value of 21.11 and variable importance (vi) of 39. It was followed by proximity to hostilities (vi = 22), Extraversion (vi = 19), pre-war financial condition (vi = 16), and Emotional stability (vi = 4). Overall, the tree distinguished two broad patterns. First, women reporting low to rather high levels of daily disruption caused by war (52.6%, n = 271) were predominantly classified as non-displaced (66.8%). Within this lower-disruption branch, Extraversion was the next splitting variable, with higher scores strengthening the probability of staying in Ukraine. Among respondents with Extraversion scores above 4.5 (28.9%, n = 149), 77.9% were classified as non-displaced. Among those with lower Extraversion score (23.7%, n = 122), classification depended further on pre-war financial condition, with the overall predicted probability of remaining was 53.3%. Women reporting moderate pre-war income (17.1%, n = 88) were still more likely to be classified as non-displaced (63.6%). Conversely, those reporting either low or high pre-war income (6.6%, n = 34) were predicted to have relocated abroad in 73.5% of cases. Among respondents who reported high or very high levels of wartime disruption (47.4%, n = 244), 61.9% were classified as displaced abroad. Within this high-disruption branch, proximity to hostilities further differentiated relocation patterns. Participants who did not reside in immediate proximity to hostilities (27.0%, n = 139) were almost evenly divided, with 50.4% classified as displaced abroad. However, this subgroup was further separated, participants reporting high disruption of daily life (13.2%, n = 68) were more likely to stay in Ukraine (61.8%), whereas those with very high disruption caused by war (13.8%, n = 71) were more likely to be classified as displaced (62.0%). In contrast, participants residing in proximal regions (20.0%, n = 103) were more likely to be classified as displaced (77.1%), without further differentiation by perceived disruption intensity. Given the centrality of maternal status to the study design, we first assessed whether this variable would emerge as a primary splitter in the overall CART. It did not, with a mean decrease in accuracy of -3.05 and a mean decrease in Gini of 5.66, ranking sixth in variable importance compared to 36 for perceived wartime impact on everyday life. Acknowledging the distinct psychological and contextual factors associated with pregnancy versus the postpartum period, we therefore estimated separate trees for pregnant women and new mothers. The classification tree of 354 new mothers included five predictors as well: perceived wartime impact on everyday life, highest level of education accomplished by women, proximity of region of residence to hostilities, perceived income level, and Emotional stability trait score (see Fig. 3). Decision tree modelling relocation probability among new mothers. Note: The tree should be read from the root node downward. Each split separates participants with different probabilities of relocation abroad. Decimal value always represents probabilities of relocation abroad; for nodes classified as non-displaced, it represents the proportion of cases expected to belong to the displaced group and therefore also indicates the expected loss. The percentage reflects node weights-the proportion of participants assigned to node. Alt text: Lower perceived wartime disruption and emotional stability, undergraduate education, and residence distant from hostilities are associated with the decision to stay in Ukraine The model demonstrated a relative error of 0.67 and a cross-validated error of 0.88 (xstd = 0.056), corresponding to a 33.1% apparent reduction and a 12.9% cross-validated reduction in classification error relative to the root node. Additional performance indices indicated limited-to-moderate discriminatory performance, with sensitivity of 51.5%, specificity of 84.3%, and an AUC of 0.71. The initial split was again made on perceived wartime disruption to daily life with an improvement value of 18.16 (vi = 52), indicating that this variable was the primary discriminator among new mothers. Among participants reporting low to rather high disruption (51.1%, n = 181), 69.6% were predicted to stay in Ukraine. For those experiencing high or very high levels of disruption (48.9%, n = 173), relocation patterns became more differentiated, with subsequent splits occurred based on educational attainment (vi = 14) and Emotional stability scores (vi = 19), both of which contributed more strongly in this model than in the full-sample tree. Within this high-disruption branch, participants with postgraduate education (30.5%, n = 108), the majority (70.4%) were classified as displaced abroad. Emotional stability further separated this group: those with lower Emotional stability (7.1%, n = 25) were more likely to be classified as non-displaced (56.0%), whereas those with higher Emotional stability (23.4%, n = 83) showed the highest probability of displacement in this tree (78.3%). In contrast, participants with undergraduate education (18.4%, n = 65) were almost evenly divided, with 50.8% predicted to stay and 49.2% predicted to relocate. For this group, proximity to hostilities emerged as the key differentiator (vi = 10). 61.3% of those residing in proximal areas (8.8%, n = 31) were classified as displaced, while 61.8% of those living in distal or intermediate regions (9.6%, n = 34) were predicted to have stayed to give birth in Ukraine. The subsequent model focused on the subsample of pregnant women (n = 163), representing the smallest group analysed (see Fig. 4). Decision tree modelling relocation probability among pregnant women. Note: The tree should be read from the root node downward. Each split separates participants with different probabilities of relocation abroad. Decimal value always represents probabilities of relocation abroad; for nodes classified as non-displaced, it represents the proportion of cases expected to belong to the displaced group and therefore also indicates the expected loss. The percentage reflects node weights – the proportion of participants assigned to node. Alt text: Lower perceived wartime disruption and higher extraversion are associated with the decision to stay in Ukraine The model yielded a relative error of 0.48 and a cross-validated error of 0.57 (xstd = 0.072), corresponding to a 51.9% apparent reduction and a 43.0% cross-validated reduction in classification error relative to the root node. Additional performance indices indicated acceptable discriminatory performance despite the limited sample size, with sensitivity of 86.1% and specificity of 67.9% and an AUC of 0.75. Consistent with the previous trees, the first and most informative split occurred on perceived wartime disruption to daily life, with an improvement value of 6.30 (vi = 35). Among participants reporting very high levels of disruption (23.1%, n = 38), 73.7% were classified as displaced abroad. The remaining participants (76.9%, n = 125), those with low to high, but not very high, disruption, had a predicted probability of remaining in Ukraine of 59.2%. Although the initial split did distinguish between high and very high levels of disruption, subsequent patterns were consistent with previous models. Notably, even within the presumed non-displaced subgroup (those with undergraduate education and high Extraversion), those reporting high disruption (9.2%, n = 15) had a predicted displacement rate of 85.7%, underscoring previous results on the salience of perceived wartime impact on daily life. Subsequent splits among participants not reporting very high disruption were based on Extraversion scores and education level, each with a variable importance of 28. Among pregnant women with lower Extraversion scores (26.4%, n = 43), 65.1% were classified as displaced abroad, whereas among those with higher scores (50.3%, n = 82), only 28.0% were predicted to be displaced, replicating the pattern observed in the overall and new mothers’ models. Within the high Extraversion branch, educational attainment further differentiated relocation patterns. Participants with postgraduate education (29.4%, n = 48) were classified as remaining in Ukraine in 91.7% of cases, forming the most stable profile across all models. This outcome contrasts with the new mothers’ model, where postgraduate education was linked with higher displacement rates in 70.4% of cases. To clarify this apparent divergence, we inspected the original, non-ROSE sample stratified by maternal status. In the observed data, displacement rates were similar across educational levels within both subgroups: among pregnant women, 11.2% of those with postgraduate education and 13.8% of those with undergraduate education were displaced abroad; among new mothers, the corresponding proportions were 17.1% and 16.7%, respectively. Overall, these descriptive proportions indicate that education alone did not differentiate relocation status within either maternal-status group; the contrast between the two tree models emerged only after participants had been separated by preceding variables, including perceived disruption and Extraversion. Among new mothers, postgraduate education appeared within the branch characterised by high or very high disruption, whereas among pregnant women it appeared only after preceding splits on less-than-very-high disruption and higher Extraversion. By contrast, participants with similarly high Extraversion but undergraduate education (20.9%, n = 34) were more evenly distributed, with the overall predicted probability of displacement was 55.9%. Within this subgroup, perceived disruption again differentiated outcomes. Those reporting high disruption (9.2%, n = 15) had a predicted displacement rate of 85.7%, whereas those with lower disruption (12.3%, n = 20) were predicted to be displaced in only 34.9% of cases. Personality traits did not emerge as primary splitting variables in any of the three classification models. However, they consistently served as surrogate splits, reflecting partial explanatory overlap with the main contextual factors. In the classification tree based on the full sample of pregnant women and new mothers (n = 515), Extraversion served as the first-ranked surrogate at the root node, replicating the split on perceived wartime disruption with 56.3% agreement, while Emotional stability achieved 55% as the second-ranked surrogate. Across the full tree, agreement for Extraversion ranged from 57% to 78%, and for Emotional stability from 57% to 64%, reflecting moderate consistency with the main model structure. In the model based on new mothers (n = 354), Emotional stability was again the top-ranked surrogate at the root node, with 55.1% agreement, and maintained agreement rates of 63% to 64% at subsequent nodes. In the model for pregnant women (n = 163), Openness, though not retained in the final tree, was the most informative first-ranked surrogate, replicating the root split with 78% agreement and maintaining 61% to 82% agreement at other nodes. While the analysis of surrogate splits indicated that certain personality traits tracked the structure of relocation decisions to a moderate extent, their role appeared to be indirect and context-dependent. To further assess the relevance of personality in shaping relocation outcomes, we conducted a direct comparison of Big Five trait scores between women who stayed in Ukraine and those who were displaced abroad (see Fig. 5). Visualisation of Big Five personality traits scores among relocation groups. Note: Violin and box plots represent weighted imputed data, dots represent observed values. Alt text: Smoothed distributions of trait scores across relocation groups, with no visible divergence Descriptive statistics indicated comparable central tendencies across groups, with mean scores ranging from 4.1 to 5.6 across all traits. No statistically significant differences were observed for any of the five traits following multiple comparison adjustment. Point estimates were uniformly small in magnitude (|β| ≤ 0.21), and all 95% confidence intervals encompassed zero, indicating an absence of statistically and practically meaningful effects. The largest unadjusted difference was observed for Openness score (β = -0.21, 95% CI [-0.47, 0.05]), with displaced women reporting marginally higher scores; however, the adjusted p-value of 0.54 provided no support for rejecting the null hypothesis. Equivalence testing was conducted using predefined ± 0.25 standard deviation bounds, corresponding to raw-score bounds ranging from ± 0.24 to ± 0.34 across personality traits. In the imputed analyses, the 90% confidence intervals for Cohen’s d were fully contained within the standardised equivalence bounds for Conscientiousness (d = 0.02, 90% CI [-0.16, 0.21]), Agreeableness (d = -0.05, 90% CI [-0.25, 0.15]), and Emotional stability (d = 0.03, 90% CI [-0.15, 0.22]), with unadjusted TOST p-values ranging from 0.023 to 0.049. Evidence for equivalence was less conclusive for Openness (d = -0.19, 90% CI [-0.37, 0.00], p = .285) and Extraversion (d = 0.16, 90% CI [-0.03, 0.35], p = .209). After Holm-Bonferroni adjustment, however, formal equivalence was not retained for any trait, with adjusted TOST p-values ranging from 0.117 to 0.418. Complete-case analyses yielded comparable results, supporting the robustness of findings obtained from the imputed dataset. Discussion The decisions faced by Ukrainian women during the war are marked by profound difficulty. Staying in the homeland entails living under constant threat, frequent air-raid alarms, and disrupted healthcare infrastructure – albeit within a familiar environment and established social support networks [60]. Conversely, relocation, especially during pregnancy or early motherhood, entails separation from family, navigation of linguistic and cultural barriers, and persistent concern for those left behind [61]. Although distinct in nature, both choices reflect an act of considerable bravery. This study aimed to identify the factors associated with wartime relocation decisions among pregnant women and new mothers in Ukraine. The findings indicate that perceived wartime disruption to daily life and proximity to active hostilities were primary factors, associated with relocation decisions. Across the full sample and subsamples, the classification trees consistently identified perceived disruption of daily life as the principal discriminator. Women reporting high or very high levels of disruption were markedly more likely to relocate, whereas those experiencing lower levels of disruption tended to remain. The shared motivation to protect oneself and one’s (unborn) child rendered participants especially sensitive to contextual adversity as a salient “push” driver [35]. These findings are consistent with previous studies conducted in other sociocultural contexts [62, 63]. Geographic proximity to active hostilities further shaped these decisions. Among those who perceived severe disruption of daily life in proximal for hostilities regions produced the highest displacement probabilities (77% overall, 35%–78% among subgroups), in distal or intermediate regions was less decisive (62% overall, 30%–63% among subgroups), even taking in account perceived disruption everyday life. These patterns support prior evidence that proximity to active hostilities amplifies perceived risk and increases displacement likelihood [64]. Across diverse conflicts, measures of distance to violent events consistently emerge among the strongest predictors of displacement, though effect sizes vary with conflict type and duration [65]. While wartime impact and proximity to hostilities area emerged as dominant factors for pregnant women and new mothers, other characteristics also demonstrated predictive relevance. Notably, pre-war financial conditions exhibited a U-shaped association with relocation decisions, suggesting that both resource scarcity, as a “push” factor, and resource sufficiency, as a “pull” factor, can facilitate displacement. In contrast, women with moderate financial resources likely face fewer economic constraints yet also experience weaker incentives to leave. This pattern diverges from findings in non-war contexts, where disaster-related displacement in Nigeria [66] and Ethiopia [67] followed different economic dynamics. Still, our findings on women’s migration decisions highlights the ambivalence within middle-income groups: such women often remain due to family obligations and sufficient local opportunities, while experiencing less financial “pushes” to flee. It should be noted that no significant group differences in personality traits were observed between displaced and non-displaced women in our sample. However, these traits emerged as relevant discriminators and surrogate split variables in the decision tree models, suggesting a potential, albeit secondary, role in shaping relocation decision-making. Specifically, higher levels of Extraversion were associated with a greater likelihood of remaining in Ukraine within specific circumstances, especially when perceived wartime disruption was not at its highest level. This finding aligns with the study by Olowojolu and Ettang [68], conducted in north-eastern Nigeria during the Boko Haram conflict, where individuals with higher risk aversion were more likely to flee, while those exhibiting greater risk tolerance tended to stay. This association may reflect the established link between Extraversion and risk-taking tendencies [69]. This pattern stands in contrast to economically motivated voluntary migration, where individuals with higher risk tolerance are typically more inclined to migrate, as described in the study by Otonneli and Toresu [70]. Subgroup analysis revealed additional stage-specific patterns in relocation decisions. Among pregnant women who did not report very high disruption of daily life, higher Extraversion and postgraduate education were associated with a greater likelihood of remaining in Ukraine. This pattern may reflect the specific decision context of pregnancy. Relocation before childbirth can disrupt continuity of maternal care, while legal, financial, language-related, or administrative barriers may limit access to essential prenatal services in a host country [71, 72]. Therefore, for pregnant women who did not perceive wartime disruption as extreme, higher education may have supported a more informed decision to remain within a familiar antenatal-care context. A different pattern emerged among new mothers. Among those, who experienced high or very high disruption of daily life, women with higher levels of education and greater emotional stability were more likely to relocate abroad. This may suggest that, after childbirth, higher education functioned less as a factor supporting the decision to stay and more as a “push” resource facilitating relocation under severe wartime disruption. After birth, the decision context may shift from preserving antenatal-care continuity to securing safer and more stable long-term conditions for the infant. This interpretation is consistent with evidence that migrant women’s access to reproductive and maternal health services is often constrained by lack of information, language barriers, and administrative procedures [73]. These are precisely the domains in which higher educational resources may help women navigate new systems and adapt to relocation. Similarly, Salelew et al. [74], in a study of war-survivor women in Northern Ethiopia, reported an interaction between psychological distress and educational level, further supporting the importance of both personal and contextual factors in shaping displacement-related decisions. However, Openness to experience appeared to shape decision-making in a manner similar to perceived wartime disruption – more open individuals tended to relocate. Flynn and Smith [75] also found that higher Openness is associated with a preference for active decision-making, a finding consistent with broader literature on this trait [76]. While relocation may remove individuals from immediate physical danger, it should not be regarded as a comprehensive solution. Across diverse conflict contexts and host countries, post-migration stressors have been shown to maintain or even intensify war-related mental health difficulties. Factors such as poverty, discrimination, social isolation, and uncertain legal status contribute to prolonged psychological distress, sustaining conditions such as PTSD, depression, and anxiety [77, 78]. These patterns have also been observed among Ukrainian refugees, as documented by Bilewicz et al. [79]. Our findings reflect a similar trend, with particular concern for maternal and infant well-being. Women who gave birth abroad exhibited a higher prevalence of postnatal depression compared to those who remained in Ukraine. This observation is consistent with previous studies conducted among Syrian [80] and Palestinian [81] refugee populations. Additionally, our results correspond with findings from Lushchak et al. [28], who reported that remaining in one’s place of residence was associated with comparatively lower levels of psychopathological symptoms. However, it is important to note that most existing studies were not conducted specifically on pregnant or postpartum women, which may partly explain observed differences between groups. At the same time, it is critical to avoid pathologising pregnancy or motherhood as causes of psychological distress. Mental health challenges among Ukrainians were evident well before the full-scale invasion of 2022. Following the annexation of Crimea and the occupation of parts of the Donbas region in 2014, Ukraine experienced large waves of internal displacement and external migration. These events had notable psychological consequences. According to the World Bank Group [82], by 2015 the prevalence of depression in Ukraine had reached 6.3%, anxiety 3.2%, and PTSD 8%. Mental disorders became the second most significant cause of disability nationwide, as measured by disability-adjusted life years. Understanding the decisions of pregnant women and new mothers to remain in Ukraine or relocate abroad is therefore essential – not only for researchers, but also for policy-makers, humanitarian organisations, and wider society. As Biehler [39] argues, the choice to stay or move is rarely driven solely by individual characteristics. Rather, it reflects a complex interaction of external constraints and personal circumstances. Some individuals may lack the financial means to leave, possess lower levels of education, or face restrictions on their freedom of movement. This was notably the case in cities such as Mariupol in 2022, where civilians were unable to evacuate due to the threat of violence by Russian soldiers who occupied the city [83]. Similarly, legal restrictions preventing Ukrainian men from leaving the country may have influenced the decisions of women who would otherwise have relocated with their partners. It remains difficult to describe the situation in Ukraine with precision. A substantial portion of the country is under occupation, making data from those territories largely inaccessible. Official statistics fail to reflect the full scale of maternal displacement, as countless women have crossed borders without documentation, been abducted by occupying forces, disappeared from healthcare systems, or given birth without any medical assistance [84, 85]. Only indirect estimates are available regarding the number of pregnant women and new mothers who have been forced from their homes. Although comprehensive data on wartime births are lacking, existing evidence points to a significant decline in pregnancy rates and widespread disruption of maternal healthcare services [86]. The situation in war-affected regions serves as a reminder that, for many Ukrainian women, even the most basic conditions for life – birth, care, and safety – can no longer be guaranteed. Limitations Precise demographic information was not collected due to ethical considerations regarding participant safety, particularly for those with relatives in occupied territories. Consequently, some potentially relevant relocation factors, including partner availability and partner military service or enlistment, were unavailable. Precise demographic information was not collected due to ethical considerations regarding participant safety, particularly for those with relatives in occupied territories. Consequently, some potentially relevant relocation factors, including partner availability and partner military service or enlistment, were unavailable. The geographic distribution of the sample limits the generalisability of the findings. No respondents were recruited from the Russian-occupied territories or from Kherson or Chernihiv Oblasts, which is important given that proximity to hostilities as one of the key predictors in the models. The geographic distribution of the sample limits the generalisability of the findings. No respondents were recruited from the Russian-occupied territories or from Kherson or Chernihiv Oblasts, which is important given that proximity to hostilities as one of the key predictors in the models. Online recruitment through social media and support centres may have introduced selection bias. The findings therefore primarily reflect the experiences of women who were reachable through these recruitment channels and may not represent less accessible or more vulnerable groups. Online recruitment through social media and support centres may have introduced selection bias. The findings therefore primarily reflect the experiences of women who were reachable through these recruitment channels and may not represent less accessible or more vulnerable groups. The retrospective design introduces potential recall bias. As a post hoc analysis, causal inference is limited due to absence of pre-relocation data collected. The findings should therefore be interpreted as identifying factors related to relocation status, rather than causal determinants of relocation decisions. The retrospective design introduces potential recall bias. As a post hoc analysis, causal inference is limited due to absence of pre-relocation data collected. The findings should therefore be interpreted as identifying factors related to relocation status, rather than causal determinants of relocation decisions. Data collection extended from December 2022 to October 2023, covering different phases of the war. Because timing relative to specific war events and relocation trajectories was not modelled, temporal variation may have influenced both decisions and retrospective reports. Data collection extended from December 2022 to October 2023, covering different phases of the war. Because timing relative to specific war events and relocation trajectories was not modelled, temporal variation may have influenced both decisions and retrospective reports. The use of ROSE-balanced training samples limits the interpretation of terminal-node probabilities. These values reflect classification patterns under a balanced modelling scenario and should therefore be interpreted as descriptive indicators of relative relocation profiles rather than population-level estimates. The use of ROSE-balanced training samples limits the interpretation of terminal-node probabilities. These values reflect classification patterns under a balanced modelling scenario and should therefore be interpreted as descriptive indicators of relative relocation profiles rather than population-level estimates. The classification trees demonstrated limited-to-moderate discriminatory performance, suggesting that additional unmeasured factors likely contributed to women’s relocation decisions. These may include immediate family support, family separation, partner availability, partner military service, and caregiving responsibilities for other children or dependents. The classification trees demonstrated limited-to-moderate discriminatory performance, suggesting that additional unmeasured factors likely contributed to women’s relocation decisions. These may include immediate family support, family separation, partner availability, partner military service, and caregiving responsibilities for other children or dependents. Conclusions The findings suggest that relocation decisions among pregnant women and new mothers during the war in Ukraine were shaped primarily by perceived war-related adversity. Perceived disruption of daily life was the most consistent factor associated with relocation decision, while proximity to hostilities further strengthened it. Although pre-war financial status, education, and selected personality traits contributed to additional differentiation within the models, their role was secondary. These factors were not strong enough to outweigh the broader impact of war-related disruption. Thus, war fundamentally alters the framework within which these decisions are made. For some women, staying and giving birth in their homeland appeared to be a meaningful decision made under constrained wartime conditions. Higher Extraversion emerged as a secondary factor associated with remaining in Ukraine in specific decision-tree branches. This finding adds to the literature by highlighting not only why women relocate during war, but also why some remain despite ongoing danger. The higher prevalence of postnatal depression among women who gave birth abroad further underscores that displacement should not be regarded as a fully protective solution. Professionals working with displaced pregnant women and new mothers should recognise that relocation does not necessarily eliminate war-related stress and may be accompanied by new burdens. At the same time, women who remain in Ukraine continue to face ongoing risks and daily-life instability. Trauma-informed, context-sensitive support is therefore needed for both displaced and non-displaced women. Acknowledgements We would like to thank all the women who participated in the study despite the immense challenges brought on by the war. Thank you for your time and bravery. Abbreviations Post-traumatic stress disorder Edinburgh Postnatal Depression Scale (EPDS) Random over-sampling examples Classification and regression tree Area under the curve Two one-sided tests Authors' contributions LK conceptualised the study, led the investigation, provided supervision, and contributed to writing the original draft and to reviewing and editing the manuscript. NY curated the data, developed the methodology, performed the formal analysis and visualisation, contributed to writing the original draft, to reviewing and editing the manuscript. MC-D participated in the investigation, and contributed to writing the original draft and to reviewing and editing the manuscript. NM contributed to validation and to writing the original draft, as well as to reviewing and editing the manuscript. AV participated in the investigation and in reviewing and editing the manuscript. OM-L contributed to conceptualisation, investigation, and to reviewing and editing the manuscript. HSG-L was responsible for data curation and resources. MFR-M managed the project, provided resources and supervision, and contributed to reviewing and editing the manuscript. All authors read and approved the final manuscript. Funding This paper is based upon work from the COST Action Riseup-PPD CA18138. Data availability The datasets supporting the conclusions of this article are available in the Open Science Framework repository, https://doi.org/10.17605/OSF.IO/K2N9M. Declarations Ethics approval and consent to participate The project was preregistered as a clinical trial (NCT05654987) and conducted in accordance with the Declaration of Helsinki. Ethical approval was granted by the Research Ethics Committee of the Universidad Nacional de Educación a Distancia (Ref: 21-PSI-2022), the Research Ethics Committee of the Faculty of Psychology, Taras Shevchenko National University of Kyiv (Date: August 2022; Decision No. 04; Protocol No. 01), and the Ethics Board for Research Projects at the Faculty of Social Sciences, University of Gdańsk (Approval No. 13-2022). Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement The datasets supporting the conclusions of this article are available in the Open Science Framework repository, https://doi.org/10.17605/OSF.IO/K2N9M.