Skip to main content
NCBI home page
NIHPA Author Manuscripts logoLink to NIHPA Author Manuscripts
. Author manuscript; available in PMC: 2026 May 31.
Published in final edited form as: Multivariate Behav Res. 2026 May 1;61(5):849–864. doi: 10.1080/00273171.2026.2661244

Beyond Linear Risk: A Machine Learning Approach to Understanding Perinatal Depression in Context

Phillip Sherlock a, Maxwell Mansolf b, Julie Hofheimer c, Christine W Hockett d,e, Thomas G O’Connor f, Danielle Roubinov g, J Carolyn Graff h, Jin-Shei Lai b, Nicole R Bush i, Rosalind J Wright j, Yueh-Hsiu Mathilda Chiu j, on behalf of program collaborators for Environmental Influences on Child Health Outcomes
PMCID: PMC13221726  NIHMSID: NIHMS2166893  PMID: 42063395

Abstract

The goal of this study was to investigate the contextual nature of prenatal depression (PND) and postpartum depression (PPD). We report an investigation of maternal PND and PPD using nonrandomly clustered data from 8,936 mothers in 16 cohorts in the Environmental influences on Child Health Outcomes (ECHO) Cohort. We used a recursive partitioning algorithm (ctree) to account for differential effects arising from predictor and effect heterogeneity across cohort to identify contexts, and the characteristics comprising these contexts, associated with PND and PPD. Consistent with Bronfenbrenner’s bioecological systems framework, risk factors for PND and PPD are heterogeneously and synergistically present in context: the same variables did not consistently present as risk factors across groups. Findings from this study support the idea that the presence or absence of medical (and sociodemographic) risks do not have simple associations with perinatal depression. To that end, no univariate predictors of PND and PPD can be said to constitute risk absent the broader context of mothers’ lives. Thus, a more useful research question than “which variables increase risk for PPD?” is “in what contexts, or in which subgroups, does a particular risk or set of risks contribute to perinatal depression?”

Keywords: Postpartum depression, regression, conditional inference, recursive partitioning

Introduction

Postpartum depression (PPD) is a form of major depressive disorder occurring in the perinatal period, affecting about 27.6% of women globally on average across reviews (Al-Abri et al., 2023). PPD is among the most common childbirth complications (Flynn et al., 2004). Furthermore, prenatal depression (PND) and PPD are associated with maternal and child adverse outcomes, including maternal mortality and morbidity (Gavin et al., 2005), increased risk for infanticide (Lindahl et al., 2005), poor maternal-infant attachment, and impaired parenting behaviors (Flynn et al., 2004). Prior work has highlighted the need to identify women at risk for PPD to potentially target early risk factors and emerging indicators (Guintivano et al., 2018). Recent syntheses reinforce the substantial prevalence and heterogeneity of perinatal depressive symptoms and the importance of early detection and prevention (Al-Abri et al., 2023; Mitchell et al., 2023; Yin et al., 2021). However, much of the work on PND and PPD has relied on variable-centric statistical approaches and consequently focused on individual risk factors, such as prenatal perceived stress and single parenthood (Norhayati et al., 2015; Yang et al., 2022). Although these approaches shed light on individual risk factors, they fall short with respect to identifying the contexts (i.e., environments described by biopsychosocial characteristics) in which women are susceptible or resilient to PND and PPD as functions of these risks (Gastaldon et al., 2022; Waqas et al., 2023). Mothers live as members of families, peer groups, and communities, wherein each subgroup brings its own set of protective factors and risks (Bronfenbrenner, 2005; Elder, 1998; Patterson et al., 1989). Within public health and other fields, Bronfenbrenner’s bioecological systems framework characterizes combinations of exposures as differentially associated with outcomes of interest (Bronfenbrenner & Morris, 2006). Evidence in support of Bronfenbrenner’s model indicates that risks do not manifest homogeneously over populations, but instead are heterogeneously and synergistically present in context (Barra & Coo, 2023; Wilkinson, 2023; Tudge et al., 2016). Moreover, risks and protective factors associated with PND and PPD act conditional on their placement in broader constellations of factors, specific to subgroups of individuals, rather than applying generally across a population. Therefore, situations may exist in which an experience or condition may prove deleterious to one mother but not to another due to differences in context.

In this study, rather than investigating the influence of individual variables or two-/three-way interactions, we identified contexts, and the characteristics comprising these contexts, associated with PND and PPD using conditional inference trees (ctrees; Hothorn et al., 2006). The ctree approach eclipses generalized linear models (GLMs) for complex biopsychosocial inquiry in its ability to detect complex differential effects. In ctree, the analysis sample is recursively partitioned (i.e., each partition is further partitioned, et cetera) based on observed predictor values using a statistical significance criterion, yielding subgroups defined by combinations of predictor values (Figure 1). Therefore, ctree is not limited by the number of multi-way interactions and nonlinearities (as in GLMs) that can be uncovered. Instead, ctree enables data-driven identification of subgroups with differential relationships between predictors and outcomes, maintaining specificity of heterogeneous effects. By specificity, we refer to preserving separation between conceptually distinct variables rather than collapsing them into composite indices or summary risk scores. For example, instead of aggregating education, marital status, and BMI into a single socioeconomic or sociodemographic risk index, we retain these as distinct features within the multivariate model. This allows the tree to identify subgroups defined by particular combinations of characteristics without obscuring distinctions between individual variables.

Figure 1.

Figure 1.

Open in a new tab

Illustration of ctree. The predictor matrix (1) is considered column-by-column. Each column is examined at all possible splits of its values, here columns a, b, and c (2a, 2b, 2c), and values in the higher category for each split are shaded in grey. The split that yields the lowest Bonferroni-corrected p-value, here the second candidate split on c (3), is selected to partition the sample (4) into distinct subgroups, Subsample A (c < 1) and B (c ≥ 1). Steps 1–3 are then repeated within each subsample. If a significant split exists within a subsample (5), that split is applied and that subsample is further subdivided (6), repeating until no significant splits are found. After ctree, all subsamples are significantly different from neighboring terminal nodes (e.g., 7a and 7b).

In practice, GLMs may be augmented, often post-hoc, with interaction terms and/or piecewise subdivisions of the larger model to account for the complexities of data. In theory, proper a priori application of these tools can better account for these features of the data; however, the problem is when and how to apply these modifications in complex datasets. Making these decisions a priori requires considerable up-front application of effort and risks model misspecification if an important feature or interaction is errantly assumed to exist or, more problematically, is omitted. Thus, it is no surprise that such explorations often occur post-hoc. In these explorations, generalizability is threatened by data-dredging approaches, and attempts to make sense of the data can be easily led astray. As a result, attempts to model a given data set are prone to severe and often irrecoverable misspecification error. It is not clear how a GLM-based approach could be “redeemed” given these shortcomings.

To illustrate, we conducted an investigation using data from the Environmental influences on Child Health Outcomes (ECHO) Cohort, which includes 69 existing longitudinal observational cohorts. ECHO was created to investigate early-life environmental exposures (e.g., biological, chemical, physical/built environment, social) on five primary child health outcomes: pre-, peri-, and postnatal outcomes; obesity; airways (e.g., asthma); neurodevelopment; and positive health (e.g., life satisfaction). The ECHO cohorts represent approximately 50,000 children and their primary caregivers from 44 U.S. states and Puerto Rico, with participants of diverse race, ethnicity, and socioeconomic status (Gillman & Blaisdell, 2018). One of the strengths of ECHO is the availability of data from 69 participating cohorts, but advanced algorithms, such as ctree, are required to reveal the full complexities of such data.

Ctree offers many benefits over GLMs in this type of observational research, particularly with respect to partitioning predictor and effect heterogeneity across cohorts. Here, the analytic sample was not representative of mothers in the U.S. but was instead an aggregation of data from several extant cohort studies (see Supplementary Table 1 for cohort descriptions). Cohorts differ greatly with respect to predictors of interest, increasing the likelihood that meaningful data partitions would be masked by cohort effects in multiple regression a la Simpson’s paradox, wherein between-cohort associations can differ from within-cohort associations (Blyth, 1972). Traditional pooled regression can therefore produce aggregated associations that misrepresent within-subgroup associations. Furthermore, by treating cohort membership as a partitioning variable within a conditional inference tree rather than solely as a fixed covariate or random effect, the algorithm empirically uncovers subgroup-specific relationships, mitigating masked or null effects that arise due to Simpson’s paradox. Including cluster membership in a conventional regression model can yield different interaction patterns across cohorts only if those interactions are specified a priori. Instead, ctree allows such interactions to emerge from the data, which is particularly appropriate when ecological-system relationships may vary by cohort. Because the ECHO cohorts were nonrandomly sampled, the combined dataset contains segmented portions of the broader population. Including cohort as a predictor in the tree helps to account for information gaps in the multivariate predictor space and to reveal where predictor-outcome relationships differ. For these reasons, this sampling design does not preclude meaningful inference—complex, nonrandom data simply require statistical methods that can detect context-dependent, high-order associations, such as ctree. Here, we demonstrate this type of inquiry.

Lastly, prenatal and postpartum depression are related but non-identical phenomena whose association can vary across social and clinical contexts. To capture this complexity, we modeled PND and PPD jointly as a multivariate outcome rather than forcing a univariate or strictly sequential framing. This joint approach is agnostic to the relationship between these conditions, permitting discovery of contexts where PND and PPD co-occur and contexts in which they diverge. This person-centered strategy therefore better reflects heterogeneous lived experiences and supports more targeted, context-sensitive interpretation of risk.

Conditional inference trees

Conditional inference trees (ctree) are a form of recursive partitioning that iteratively divide the sample into increasingly homogeneous subgroups based on observed predictor values (Hothorn et al., 2006; 2025; Hothorn & Zeileis, 2015). At each node, the algorithm first examines every eligible predictor to identify candidate splits. For continuous and ordered categorical predictors, candidate splits are considered at each potential cut point; for unordered categorical predictors, all possible two-way partitions of category sets are evaluated. For each candidate split, the algorithm computes a test statistic (for example, a test of the reduction in mean squared error for continuous outcomes) and an associated p value. The candidate split producing the smallest p value—i.e., the most statistically significant improvement in node homogeneity—is selected, if it remains significant after applying a Bonferroni correction that adjusts for the multiplicity of tests performed at that node. Partitioning proceeds recursively in the same manner within each resulting subsample until no candidate split meets the Bonferroni-corrected significance threshold, at which point splitting stops and the node becomes a terminal node. The algorithm continues in this manner until no more potential splits satisfy this statistical significance threshold. Ctree reports terminal-node sample sizes so that readers may assess balance and the precision of node-specific estimates. Nodes that are too small to support further splitting automatically terminate, which limits spurious partitioning and reduces the risk of overinterpretation. Importantly, in the ctree framework these p values function as a statistically principled stopping rule for recursive partitioning rather than as classical hypothesis tests used for inferential claims. The permutation-based testing procedure evaluates whether any association between a predictor and the outcome is sufficiently strong to justify further partitioning of the data. This design distinguishes ctree from traditional impurity-based tree algorithms (e.g., CART), which select splits solely based on measures such as variance reduction and can exhibit biased variable selection (e.g., favoring predictors with more possible cut points). By conditioning splits on statistically supported associations, ctree reduces selection bias and helps limit overfitting while preserving the exploratory, descriptive nature of the model.

Because ctree evaluates splits across all predictors at every step, it naturally uncovers high-order and nonlinear interactions without requiring a priori specification of interaction terms. Whereas multiple regression produces predictions from a linear combination of predictors, ctree assigns each observation to a terminal node based on its observed predictor values and predicts outcomes using the node-specific estimated outcome means. In this way, ctree is a person-centered approach that identifies profiles of individual characteristics as subgroups defined by a sequence of data splits. Figure 1 illustrates the ctree algorithm and its recursive splitting logic (technical and implementation details: Hothorn et al., 2006; Hothorn & Zeileis, 2015; Hothorn et al., 2025).

Although ctree can only be supervised (i.e., the training features/predictors are provided to the model directly), it is related to latent class regression (Vermunt & Magidson, 2003), in that both can be used to identify subgroups with differential associations. The most notable difference relates to how the groups are described. In LCR, predictor and outcome distributions are used to identify latent variables that define class membership, as it is assumed that class membership cannot be directly determined from the data. In order to recover these latent distributions and probabilities, larger samples tend to be required depending on effect size, complexity, and class separation (Van Horn et al., 2012; Dziak et al., 2014; Lamont et al., 2016; Sherlock et al., 2022). Moreover, issues remain related to nonrandomly sampled heterogeneous clusters, for example, ECHO cohorts. At least two possibilities arise in practice: (1) the recovered classes will merely represent cohort membership, and (2) the LCR is misspecified and/or its assumptions are violated. For the latter, direct tests of model assumptions are impossible because the unconditional distribution within each class is unobservable (Van Horn et al., 2012).

In contrast to LCR where true class membership is unknown, subgroups identified by ctree (denoted by a terminal node) are described by values of observed predictors for all the splits that lead to each terminal node. Furthermore, ctree is more appropriate for secondary data analysis because we are more interested in identifying intervenable contextual features of individuals’ lives (i.e., levels of variables used to define splits in ctree) rather than describing the variability of those features across cohorts (mixed-effects regression) or deriving subgroups based on unobserved variables (LCR). Most importantly, ctree is not limited by the number of multi-way interactions (as in GLMs) that can be uncovered or the identifiability of latent subgroups (as in LCR).

Given the conceptual and practical differences among person-centered and variable-centered approaches, we provide a focused comparison to clarify why ctree was selected for this application. Table 1 contrasts conditional inference trees (ctree) with generalized linear/hierarchical linear models (GLM/HLM) and latent class regression/regression mixture models (LCR/RMM) across a set of criteria of importance for multicohort, high-dimensional data: centering (person vs. variable), parametric assumptions, how subgroups are defined, covariate specification, interaction detection, sample-size requirements, overfitting control, and interpretability. We present this comparison not to argue that one approach is universally superior, but to make transparent the practical tradeoffs and researcher degrees of freedom that influence model choice in this context.

Table 1.

Conceptual analytic comparisons.

Feature Conditional Inference Tree (ctree) Generalized Linear Model / Hierarchical Linear Model (GLM/HLM) Latent Class Regression/Regression Mixture Model (LCR/RMM)
Analytic focus Person-centered: discovers subgroups by recursively splitting on observed predictor values Variable-centered: estimates overall average effects of predictors (GLM) with optional random-effect deviations (HLM) Hybrid: uncovers unobserved classes among people (person-centered), then estimates predictor–outcome relationships within each class (variable-centered)
Parametric assumptions No distributional form required for predictors or outcome; relies on valid test/statistics at each split. Recursive piecewise function allows interactions and nonlinearities. Requires correct link function, error distribution (GLM), and random-effects distribution (HLM) Parametric within-class distributions (often normal) and correct number of classes
Subgroup definition Observed: terminal nodes defined by explicit combinations of predictor values None by default: subgroups only arise if the analyst explicitly specifies interactions or random clusters, or with dummy predictors Unobserved: class membership inferred probabilistically from joint distributions of latent class predictors and outcome
Covariate vs. predictor No need to pre-label covariates vs. predictors—all variables are candidates for splitting and thus can include interactions and nonlinearities Must pre-specify which variables are main predictors, which are covariates for adjustment, and which enter the random structure Requires careful a priori designation of covariates in both class-membership and within-class models
Interaction detection Automatic: uncovers any number of high-order, nonlinear interactions via data-driven splits, limited only by available power in each node Manual: only interactions the analyst explicitly includes are tested Implicit: interactions are embedded in class structure but not directly tested; classes may reflect sampling clusters rather than substantive subpopulations
Sample size requirements Needs enough observations per node; tree growth halts when nodes become too small Many predictors/interactions (GLM) or many clusters/levels (HLM) require substantial sample sizes Stable class separation and reliable parameter estimation typically demand large n, especially as class count increases
Overfitting control Built-in: Bonferroni-corrected p-value stopping rule prunes spurious splits External: requires penalization (e.g., AIC/BIC, ridge/lasso) or cross-validation External: uses information criteria (e.g., BIC) to choose class count, but overfitting remains a risk if classes are poorly separated
Interpretability High: transparent decision rules (“if X > a & Y = b then …”), intuitive visualization Moderate: coefficients interpretable, but complex interactions and variance components (ICCs) require careful explanation Moderate–low: latent classes are abstract; require post-hoc labeling, detailed narrative to convey class meanings, and speculation on the source of hidden heterogeneity (i.e., unmeasured variance)
Code availability and specification complexity Mature, open-source (R party/partykit; Python statsmodels); little specification needed Universally available (R: glm(), lme4; Python: statsmodels, pymer4; SAS, Stata), well-documented, detailed specification needed Available in specialized software (Mplus; R poLCA, lcmm), but model specification and convergence can be more challenging
Open in a new tab

Methods

Sample

The study population included mothers whose data were collected in an ECHO cohort and who subsequently provided consent for their data to be shared with ECHO. Of this set, the only participant inclusion and exclusion criteria was that mothers completed a harmonizable (to the PROMIS® metric, see next section) depression measure during both the prenatal and postpartum periods and that children were not adopted. Other inclusion and exclusion criteria were cohort-specific (see Supplementary Table 1 for descriptions of cohorts).

Prenatal and postpartum depression

PPD and PND were measured on the PROMIS Depression metric (Cook et al., 2016; Pilkonis et al., 2011) with scores linked from several extant measures; see (Choi et al., 2014; Blackwell et al., 2021; Kaat et al., 2017). PND was measured during pregnancy, and PPD was measured from 46 weeks after the mother’s last menstrual period to 11 months and 30 days of child age. When multiple measurements were available, scores were averaged within each time period separately for each mother.

PROMIS Depression T scores were norm referenced to the general population, with a mean of 50 and a standard deviation of 10. For this analysis, T scores were dichotomized at 59.7, corresponding to the clinical cutoff for the Edinburgh Postnatal Depression Scale, which is the most commonly used postpartum depression measure in ECHO (Blackwell et al., 2021; Cox & Holden, 2003), and is approximately one standard deviation above the norm-referenced population mean, which optimizes sensitivity and specificity for case detection (Blackwell et al., 2021; Cox & Holden, 2003). This threshold-based outcome aligns with clinical decision-making and makes our subgroup prevalence estimates directly actionable for practitioners, without compromising ctree’s ability to discover context-dependent associations (Hothorn et al., 2006). Estimated subgroup means of these dichotomous depression indicators were thereby used in ctree to quantify the proportions of women with PND and/or PPD. By linking multiple depression measures to a common PROMIS metric, we ensure comparability across cohorts and instruments, as recommended by Cook et al. (2016). While continuous PROMIS scores could in principle be modeled, our focus on a clinically meaningful binary outcome ensures that identified subgroups reflect risk categories of practical relevance.

Predictors

Unlike GLMs, ctree is agnostic to the analyst’s conventional conception of core “predictors” versus adjusted-for “covariates,”, as all predictors are treated equivalently in the algorithm. Additionally, all variables in ctree are potential moderators. Here, we included the following predictors: a cohort membership indicator; child sex, race, and ethnicity (Hispanic vs. Non-Hispanic); gestational age at birth; singleton versus multiple delivery; maternal education, employment status, marital status, insurance status, and pre-pregnancy body mass index (BMI); maternal birth outside of the U.S.; a set of indicators for pregnancy-related health conditions (Supplementary Table 2); and six items from the Childhood Trauma Questionnaire (CTQ; Bernstein et al., 1998), which queries maternal adverse childhood experiences. These variables were identified from prior meta-analyses and literature reviews as important predictors of postpartum depression (Beck, 2001; Guintivano et al., 2018; Yim et al., 2015) and were available broadly within the ECHO data. We also considered recent meta-analytic work when selecting candidate predictors to ensure coverage of factors shown to be robust across studies (Gastaldon et al., 2022; Mitchell et al., 2023). Predictors span sociodemographic, reproductive, psychosocial, and health domains and were harmonized across cohorts to permit person-centered partitioning by the ctree algorithm. Measurement levels for each indicator are reported in Supplementary Table 2, excluding the marital statuses widowed, divorced, or separated, which were combined with single, never married due to the small number of mothers with these marital statuses.

Conditional inference trees (ctree)

Ctree recursively partitions a sample based on observed predictor values such that each partition yields two subgroups with the most statistically significant differences possible (lowest p-value). A Bonferroni correction is used at each step to account for the multiplicity of tests required at each step. Partitioning stops when no further split passes the Bonferroni criterion; therefore, the subgroups identified by ctree (terminal nodes) are each significantly different, after a Bonferroni correction, from their neighboring terminal nodes.

Conditional inference trees and missing data

To handle item-missingness we used multivariate imputation by chained equations (MICE; Van Buuren & Groothuis-Oudshoorn, 2011) and the stacked-imputation workflow recommended for analyses of recursive partitioning (Rodgers et al., 2021). In this workflow, multiply imputed datasets are concatenated vertically or “stacked”, such that, for a sample of size n with m imputations, the stacked data set has m × n rows, as each observation is repeated m times across the m imputations. Rodgers et al. (2021) showed that this approach yields unbiased recursive partitioning results. However, as with any application of multiple imputation, care must be taken in conducting and analyzing the results of statistical tests, such as those used by ctree. Specifically, stacking allows the recursive partitioning algorithm to encounter the variability induced by the imputation process, but it also increases the nominal sample size artificially and can inflate test statistics. To counteract this, we rescaled each node-level test statistic by dividing by the number of imputations before recalculating p values and pruning the final tree, thereby attempting to preserve the intended type-I error control for node selection. To evaluate the performance of our pruning adjustment, we conducted Monte Carlo simulations under missing completely at random (MCAR) with multiple imputation (M = 30). At the node level, the stacked-imputation χ2 statistic divided by M (“Stack ÷ M”) was compared against two permutation-based approaches. With n = 9000 to mirror the applied sample, Stack ÷ M yielded sub-nominal type-I error rates (≈0.03, 0.02, 0.01 for 10%, 30%, 50% MCAR, respectively) and reduced power relative to the permutation methods (e.g., 0.72 vs. 0.91 at 50% MCAR and effect size δ = 0.05). This confirms that our chosen approach is conservative, although it sacrifices sensitivity to weak effects. However, this also reduces the chance of spurious splits, making it well-suited for an exploratory analysis such as the current analysis, while still providing a single, interpretable tree structure across imputations. Full simulation details and additional results are provided in the Supplement.

Importantly, this stacked approach is preferable to fitting trees separately within each imputation and attempting to pool the results. Because tree-based models are hierarchical and path-dependent, even small differences across imputations can lead to different variables, split points, and orders of branching. Such differences preclude straightforward pooling: once the root split diverges, all downstream nodes describe different subgroups, and the targets of inference are no longer aligned across imputations. Stacking circumvents this structural mismatch by producing a single, coherent tree that incorporates imputation variability into one partitioning solution, avoiding the instability and interpretive ambiguity that arise when trying to combine structurally different trees. Because our substantive goal was an interpretable, unified tree structure to characterize subpopulations, this conservative approach represents the only tenable way to reconcile imputation variability within the recursive partitioning framework.

MICE and various approaches combining multiple imputation with decision-tree methods have been well studied under missing-at-random assumptions and shown via simulation to preserve bias, coverage, and interaction effects in many settings (Doove et al., 2014; Rodgers et al., 2021). While fewer studies have evaluated recursive partitioning with node-level test rescaling (such as dividing χ2 tests by the number of imputations), our supplementary simulation shows that doing so yields a conservative α under the null in our pruning context. This means that although the method trades off power, the resulting subgroups are less likely to be artifacts of chance variation, which is an acceptable and even desirable property when the interpretability of a single tree is prioritized. For these reasons, we retained the stacked-imputation approach for the primary analyses rather than relying on a complete-case analysis, which is known to produce biased estimates and reduced efficiency unless data are missing completely at random (MCAR) (Rubin, 1976).

All analyses were conducted in R (R Core Team, 2022).

Presentation of results

When power is high, ctrees can split the data many times, becoming cumbersome to interpret. We included color-coded dendrograms for all trees in Figures 2–5. In a dendrogram, one follows a path from the root node to a terminal node to determine how that node was defined (i.e., which variables were used to split and at what values). The left side of each dendrogram displays the tree structure: the root node, containing the entire sample, is partitioned according to the most significant predictor’s split, and each of those nodes is then partitioned similarly, and so on. The right side of each dendrogram displays sample size, the proportion of mothers with prenatal and postpartum depression, and the strength of association between prenatal and postpartum depression (φ for each 2 × 2 contingency table) in each node, with darker shading indicating higher proportions or stronger associations.

Figure 2.

Figure 2.

Open in a new tab

Cohort-level overview of ctree. Note: n = sample size in each node. φ = mean square contingency coefficient (Pearson correlation) for prenatal and postpartum depression indicators in each node. Horizontal and vertical lines and indentation in the first column depict the hierarchical structure of the fitted ctree. Cells containing sample statistics (n, prevalence, φ) are indented according to their position within the nesting structure of the ctree. Prevalence and u values are color-coded based on their values for visual emphasis, with darker colors corresponding to higher values. Depression indicators were constructed by dichotomizing PROMIS Depression T scores at 59.7, corresponding to the clinical cutoff for the Edinburgh Postnatal Depression Scale. PROTECT, URECA, MADREs, ACCESS, PRISM, MINNIE, BYS, PASS, CANOE, PGS, CANDLE, Rochester, Magee, Project Viva, MARCH, and NYU CHES indicate membership in the corresponding ECHO cohorts. Nodes 4, 12, and 37, are pruned in this Figure, indicated by italics; see Figures 3–5 for expansions of these nodes.

Figure 5.

Figure 5.

Open in a new tab

Branch 4 of ctree. Note. n = sample size in each node. φ = mean square contingency coefficient (Pearson correlation) for prenatal and postpartum depression indicators in each node. Horizontal and vertical lines and indentation in the first column depict the hierarchical structure of the fitted ctree. Cells containing sample statistics (n, prevalence, φ) are indented according to their position within the nesting structure of the ctree. Prevalence and φ values are color-coded based on their values for visual emphasis, with darker colors corresponding to higher values. Depression indicators were constructed by dichotomizing PROMIS Depression T scores at 59.7, corresponding to the clinical cutoff for the Edinburgh Postnatal Depression Scale. BYS, Rochester, Magee, MARCH, PGS, and Project Viva indicate membership in the corresponding ECHO cohorts. MarrOrCohab: married or cohabitating with partner; NotMarrOrCohab: not married or cohabitating with partner.

Comparison with alternative methods

To illustrate the advantage of ctree in this context, we also applied linear modeling approaches, specifically GLMs, HLM, and LCRs (i.e., regression mixture models) to the same data. Given the latitude researchers have when defining these models, we elected to evaluate the most basic representations of these comparators. For GLMs, this is a model with all predictors entering as direct effects. For HLMs, the most basic representation is a multilevel linear regression model with all level-1 predictors entering as direct effects with only a random intercept for cohort. This model is simpler than what a researcher may typically specify in practice, but that’s our point: there are additional researcher DOF involved in defining a model, and these decisions are often based on the researcher’s guiding interest, rather than representing the data well, hence, in our view, leading to misspecification a priori. Regarding the latent class regression, we chose a model with no latent class predictors or covariates, such that it reduces to a regression mixture model. As with the GLM, choices for what is a latent class predictor or covariate are often made based not on a detailed understanding of the complexity of the problem, but on the researcher’s interests, again leading to potential misspecification when data are more complex than the model can accommodate.

Results

Results for alternative methods

Estimates for GLMs with and without cohort fixed effects that regressed prenatal and postpartum depression on all the aforementioned predictors in the Methods section are presented in Supplemental Tables 3–6. The GLMs without cohort fixed effects were only able to identify education, marital status, and gestational age at birth as having significant associations with the outcomes for both prenatal and postpartum depression. The GLMs that included cohort were able to identify education and marital status as significant predictors of PND, and education, marital status, gestational age at birth, infection, and body mass index (BMI) as significant predictors of postpartum depression. In each of the single-level GLMs, lower education, lower gestational age at birth, single/non-cohabitating motherhood, presence of infection, and higher BMI were all associated with a higher risk for clinically significant depression.

Perhaps not unexpectedly, the multilevel GLMs for PND and PPD failed to converge in 46.7% and 80% of the 30 imputations, respectively; hence, estimates for these models were not reported, as nonconvergence in some imputations means some parts of the distribution of the missing data given the observed data were not captured by these models (Enders, 2022). Similar to the HLMs, the RMMs for PND and PPD, each estimated with and without cohort as a fixed-effects predictor, failed to converge in 43.3% to 60% of the 30 imputations, respectively, when two mixture components were specified, and in 90% to 100% of imputations when three mixture components were specified. Therefore, estimates were also not reported for these models.

Although we could find any (large but finite) number of theoretical justifications to augment the HLM and RMM comparators, such as additional random effects in HLMs or latent class predictors in RMMs, this presents an intractable problem for the purpose of comparison because the models become more dissimilar as more of these features are added. One could also augment the linear portion of these models by adding interactions or removing predictors, but as noted in the literature this type of specification search in high-dimensional data is burdened by additional researcher degrees of freedom and vulnerable to pitfalls including Simpson’s paradox (Kievit et al., 2013). To that end, we present the simplest models only. And, as we can see with HLMs and RMMs, there are immediate convergence issues even with the “simplest” versions of these models, which does not bode well for potential modifications.

Conditional inference trees

We focus on terminal nodes with the highest and lowest proportions of PND and PPD to highlight contexts that are associated with an elevated and reduced risk of depression. In this way, we reach similar conclusions as would be obtained with GLMs (i.e., which predictors have the strongest associations with the outcomes), within the contextualized framework of trees, including higher-order interactions, rather than assumed-homogeneous associations as in GLMs.

The ctree model with cohort as a predictor first detected differences among cohorts comprising four separate branches. Figure 2 contains a cohort-level overview of the conditional inference tree. These cohort groupings were not specified a priori but were empirically found by ctree. The first branch partitioned the PROTECT, URECA, and MADREs cohorts from others, with no further splits. The proportions of PND and PPD in these cohorts were 0.23 and 0.18, respectively.

Branch two

Figure 3 displays the second branch, including the ACCESS and PRISM cohorts. Mothers who were not married, cohabitating, or employed had the second-highest proportion of PND (0.34) and PPD (0.17). Similarly, among mothers who were not married or cohabitating and who were working, the proportions of PND and PPD were 0.14 and 0.10, respectively. Among mothers who were married or cohabitating, two terminal subgroups were identified—those with a bachelor’s degree or less (PND = 0.20; PPD = 0.08) versus those with more than a bachelor’s degree had the highest rate of PND (0.36) and almost non-existent PPD (0.01).

Figure 3.

Figure 3.

Open in a new tab

Branch 2 of ctree. Note: n = sample size in each node. φ = mean square contingency coefficient (Pearson correlation) for prenatal and postpartum depression indicators in each node. Horizontal and vertical lines and indentation in the first column depict the hierarchical structure of the fitted ctree. Cells containing sample statistics (n, prevalence, φ) are indented according to their position within the nesting structure of the ctree. Prevalence and u values are color-coded based on their values for visual emphasis, with darker colors corresponding to higher values. Depression indicators were constructed by dichotomizing PROMIS Depression T scores at 59.7, corresponding to the clinical cutoff for the Edinburgh Postnatal Depression Scale. MarrOrCohab: married or cohabitating with partner; NotMarrOrCohab: not married or cohabitating with partner; BS: Bachelor’s degree.

Branch three

Figure 4 displays the third branch, including the MINNIE, PASS, CANOE, CANDLE, and NYU CHES cohorts. In these cohorts, the first split occurred at a BMI of 26.2; further splits are described in the next section.

Figure 4.

Figure 4.

Open in a new tab

Branch 3 of ctree. Note: n = sample size in each node. φ = mean square contingency coefficient (Pearson correlation) for prenatal and postpartum depression indicators in each node. Horizontal and vertical lines and indentation in the first column depict the hierarchical structure of the fitted ctree. Cells containing sample statistics (n, prevalence, φ) are indented according to their position within the nesting structure of the ctree. Prevalence and φ values are color-coded based on their values for visual emphasis, with darker colors corresponding to higher values. Depression indicators were constructed by dichotomizing PROMIS Depression T scores at 59.7, corresponding to the clinical cutoff for the Edinburgh Postnatal Depression Scale. MINNIE, PASS, NYU CHES, CANOE, and CANDLE indicate membership in the corresponding ECHO cohort. MarrOrCohab: married or cohabitating with partner; NotMarrOrCohab: not married or cohabitating with partner; SC: Some college; no 4-year degree.

BMI less than or equal to 26.2

Among mothers with a BMI less than or equal to 26.2 and at least a bachelor’s degree, the proportions of PND and PPD were 0.02 and 0.03, respectively. Among mothers with a BMI less than or equal to 26.2, those with less than a bachelor’s degree in the MINNIE, PASS, and NYU CHES cohorts were subdivided based on insurance status during pregnancy. Among mothers with insurance, the proportion of PND was 0.03. This group had the lowest PPD (0.01). In contrast, the proportions of PND and PPD in mothers without insurance were 0.09 and 0.07, respectively. Among mothers with a BMI less than or equal to 26.2 and less than a bachelor’s degree in the CANDLE cohort, the proportions of PND and PPD were 0.02 and 0.04, respectively. Mothers with BMI less than or equal to 26.2 who had at least a bachelor’s degree had PND and PPD equal to 0.02 and 0.03, respectively.

BMI greater than 26.2

Among mothers with a BMI greater than 26.2, ctree identified eight subgroups. Mothers in the MINNIE, PASS, and NYU CHES cohorts who were married or cohabitating, had no infection during pregnancy, and had a BMI greater than 26.2 but less than 39.3 were partitioned based on insurance status, where mothers without insurance had PND and PPD proportions of 0.02 and 0.01, respectively. Mothers with insurance during pregnancy were partitioned again by cohort. Mothers from the MINNIE and NYU CHES cohorts had PND and PPD proportions of 0.06 and 0.02, respectively. Mothers from PASS had PND and PPD proportions of 0.07 and 0.06, respectively.

Among mothers in the MINNIE, PASS, and NYU CHES cohorts who were married or cohabitating, had no infections during pregnancy, and had a BMI greater than 39.3, the proportions of PND and PPD were 0.09 and 0.10, respectively. Among mothers in the MINNIE, PASS, and NYU CHES cohorts who were married or cohabitating and who experienced an infection during pregnancy, the proportions of PND and PPD were 0.08 and 0.10, respectively. Mothers in the MINNIE, PASS, and NYU CHES cohorts who were not married or cohabitating had PND and PPD proportions of 0.12 and 0.12, respectively. Mothers in the CANOE and CANDLE cohorts had PND and PPD proportions of 0.03 and 0.06, respectively.

Branch four

Figure 5 displays the fourth ctree branch, including BYS, PGS, Rochester, Magee, Project Viva, and MARCH. In these cohorts, the most important predictor of PND and PPD was marital status.

Married or cohabitating

Among mothers who were married or cohabitating and with gestational ages at delivery less than or equal to 36 wk, the proportions of PND and PPD were 0.13 and 0.12, respectively. Married or cohabitating mothers with gestational ages at delivery greater than 36 wk were further split based on cohort. Among mothers in the PGS and Project Viva cohorts, the proportions of PND and PPD were 0.07 and 0.07, respectively. Mothers from the BYS, Rochester, Magee, and MARCH cohorts were divided based on maternal education into those with some college or less (PND = 0.11; PPD = 0.07) versus those with at least a bachelor’s degree (PND = 0.04; PPD = 0.02).

Not married or cohabitating

Mothers not married or cohabitating were split again based on cohort. Among mothers from the BYS, PGS, Rochester, and Magee cohorts, the proportions of PND and PPD were 0.12 and 0.09, respectively. Mothers from the Project Viva and MARCH cohorts were divided based on hypertension during pregnancy (PND = 0.45; PPD = 0.07) versus none (PND = 0.22; PPD = 0.12).

Discussion

Complications arise when using GLMs to model complex social phenomena. This is particularly important for investigating complex conditions, such as PND and PPD, with manifestations that are context-dependent. We estimated and presented results from a recursive partitioning algorithm (ctree) to account for differential effects arising from predictor and effect heterogeneity across cohorts. Several findings from recent high-quality syntheses align with and help contextualize our results. Globally, pooled prevalence estimates for perinatal depression fall in the high-teens to high-20s. For example, one meta-review reported pooled estimates of 28.5% for prenatal depression (PND) and 27.6% for postnatal depression (PPD) (Al-Abri et al., 2023), while another found a pooled prenatal prevalence of 20.7% (95% CI 19.4%–21.9%) (Yin et al., 2021). These ranges illustrate both the substantial overall burden and the variation introduced by measurement and setting. Furthermore, the meta-review estimates are much higher than the overall-sample rates in this study (10% and 7% for PND and PPD, respectively), which likely reflects sampling from the constituent ECHO cohorts (see Supplement). To that end, recent work emphasizing heterogeneity in perinatal depression phenotypes and trajectories reinforces the person-centered perspective that PND and PPD can manifest in multiple, context-dependent patterns rather than as a single homogeneous condition (Waqas et al., 2023). Meta-analytic reviews of risk factors further highlight appreciable associations with psychosocial and reproductive adversities, for instance, OR = 2.20 for premenstrual syndrome, OR = 2.07 for violent experiences, and OR = 1.53 for unintended pregnancy. These results are broadly consistent with our observation that nodes characterized by social risk and reproductive/health stressors tend to show higher rates of clinically significant PND and PPD (Gastaldon et al., 2022), although this research assesses these risk factors across the population, whereas our approach identifies contexts of (in)activation of these risk factors.

Consistent with Bronfenbrenner, risk factors for PND and PPD are heterogeneously and synergistically present in context: the same variables did not consistently present as risks across groups. Thus, a more useful research question than “which variables increase risk for PPD?” is “in what contexts, or in which sub-groups, does a particular risk or set of risks contribute to PPD?” Answering this question requires statistical tools to search for interactions between variables often relegated to covariates in GLMs, such as socioeconomic and demographic factors, alongside more typical “predictors.” Furthermore, ctree avoids the common pitfall of creating risk indices (e.g., (Evans et al., 2013)), which aggregate information across multiple risk factors, by allowing all variables to act as candidates for partitioning the data rather than muddling context-specific risk by aggregation.

The results from this study indicate that “risk” of PND and PPD depends on the context of mothers’ lives. Were risks primarily determined linearly by one or multiple predictors, as in GLMs, ctree would repeatedly split on those predictors, whereas splits in this study were diverse, spanning study design (ECHO cohort), social, psychological, structural, and developmental predictors. Furthermore, several predictors demonstrated differential effects across groups. As can be seen in Figures 3–5, associations with education, BMI, insurance status, marital status, infection, and gestational age constitute complex non-linear interactions. None of these univariate predictors can be said to constitute risk absent this context. For example, lower education has been associated with higher PPD (Di Florio et al., 2017; Matsumura et al., 2019). However, mothers from the MINNIE, PASS, and NYU CHES cohorts with less than a bachelor’s degree, a BMI ≤ 26.2, and no health insurance had the lowest proportion of PPD in our study.

Interpretation of regression trees

In contrast to regression coefficients, which apply across an entire sample, node-level estimates describe participants who meet all splitting conditions on the path to that leaf (for example: cohort = CANDLE; education < Bachelor’s; BMI > 30; not married). A node-level PND prevalence is the proportion of that specific subgroup who meet the PND cutoff. Because nodes are subsets defined by multiple characteristics, node-level prevalences or effect patterns can be much higher or lower than the pooled average. For example, the pooled PND prevalence might be 25%, but a node defined by social disadvantage and pregnancy complications could have a PND prevalence of 45%, whereas another node defined by social support and infection might have only 7%.

This difference, between subsample-specific and aggregate summaries of data, matters for two reasons. First, aggregate summaries can mask heterogeneity, especially in complex nonrandom-cluster structure such as in the ECHO project: averaging across heterogeneous subgroups may suggest weak or null associations even when strong associations exist within particular contexts (a classical example of Simpson’s paradox). Second, node-level descriptions can be more actionable: identifying a subgroup with 45% PND prevalence points to a high-risk population that may benefit from targeted prevention, whereas the pooled 25% figure gives no information on where that burden is concentrated, instead only speaking to the raw magnitude of the problem in the aggregate.

As further evidence for the importance of context, recent syntheses point to several measurement and coverage limitations that illuminate why some node-level findings (from ctree) may differ from pooled results. Meta-analytic reviews and meta-analyses draw on studies that vary substantially in exposure measurement (for example, different operationalizations of trauma or social adversity), outcome ascertainment (self-report scales versus diagnostic interviews), and timing of assessments across pregnancy and the post-partum period. As a result, certain high-risk pathways identified in the broader literature cannot be consistently harmonized across multicohort data and may therefore be under-represented in pooled analyses (Al-Abri et al., 2023; Gastaldon et al., 2022). A pooled estimate summarizes information across the whole dataset: e.g., a pooled PND prevalence of 25% means that, averaged over all participants and contexts, one quarter met the threshold for clinically significant prenatal depression. Pooled regression coefficients similarly express the average association between a predictor and outcome across everyone in the sample. As noted, these aggregate relationships and statistics lack the granularity needed to identify specific, actionable interventions, or the most vulnerable subgroups that warrant those interventions, and instead, implicitly, suggest widespread problems (which may not exist in many contexts) and broad-strokes solutions (which may be unnecessary or counterproductive in many contexts).

Furthermore, examining risk and resilience for complex conditions like PND and PPD in context using methods like ctree can improve our implementation of interventions to address these problems. Reviews show that preventive interventions show small-to-moderate effects: pooled overall standardized mean difference (SMD) = 0.29 (95% CI: 0.20–0.38); psychological interventions = 0.28 (95% CI: 0.16–0.39); exercise/physical activity interventions = 0.43 (95% CI: 0.09–0.77). Reviews of mixed ‘any type’ interventions report SMD = 0.36. However, the underlying evidence is generally low-to-very-low quality (Motrico et al., 2023). Together, these limitations both support our focus on within-sample heterogeneity and caution against extrapolating any single node’s pattern beyond similar measurement and cohort contexts (Al-Abri et al., 2023; Gastaldon et al., 2022; Motrico et al., 2023). As we work to improve our interventions and our implementation of them, we must ask the question, and use methodologies such as ctree that can help answer the question: which interventions are appropriate for whom?

While examples of this risk (in)activation are numerous in this study, they necessitate further inquiry, speaking to the utility of frameworks, such as the Perinatal Maternal Health Promotion Model (PMHPM), developed around the core principle that maternal well-being goes beyond merely the absence of medical complications (Fahey & Shenassa, 2013). Findings from this study support the idea that the presence or absence of medical (and sociodemographic) risks do not have simple associations with perinatal depression. Moreover, just as mental health cannot be viewed as completely independent from physical health, the well-being of mothers cannot simply be understood as a function of observable physical and sociodemographic circumstances, independent of mothers’ abilities to navigate those circumstances and meet their needs.

Limitations and future directions

Although conditional inference trees (ctree) are well suited to detecting context-dependent, high-order associations in multicohort data, several important limitations should be acknowledged. First, ctree is a greedy algorithm: at each step it selects the predictor that yields the greatest immediate improvement in node homogeneity rather than searching for a globally optimal tree. As a result, the final tree is one of many plausible ways of partitioning the data and may not represent the absolute optimal prediction structure (Hothorn et al., 2006). Methods that search for globally optimal trees (e.g., evolutionary algorithms) exist, but these approaches have not yet been fully integrated with the conditional-inference testing framework used here (Grubinger et al., 2014).

Second, although ctree implements internal controls for overfitting (for example, node-level hypothesis testing with multiplicity adjustment and pruning rules), the method remains sensitive to sample size within nodes. Small terminal nodes yield imprecise prevalence estimates and unstable split selection; in addition, nodes that are too small simply cannot split further, which limits overinterpretation but also means that some fine-grained heterogeneity may go undetected. Although it was beyond the scope of the current study, researchers have used random forests within terminal nodes to explore remaining heterogeneity after using ctree to investigate differential associations within terminal nodes (Knopf et al., 2025).

Third, missing-data handling and imputation introduce additional sources of complexity. We used multivariate imputation by chained equations (MICE) with a stacked-imputation workflow following recent methodological recommendations (Rodgers et al., 2021), and we rescaled node-level test statistics by the number of imputations to correct for power inflation. This conservative adjustment, as confirmed by our supplementary simulation, reduces type-I error below nominal levels in the pruning step. Even so, imputation models may fail to fully capture complex, high-order interactions present in the substantive data; if interactions are poorly represented in the imputation model, estimated splits can be biased. For this reason, researchers should carefully specify imputation models, consider substantive-model-compatible imputation strategies, and report sensitivity checks (for example, comparing stacked-imputation results with alternate analysis strategies) when feasible (van Buuren, 2018; Beesley & Taylor, 2021). Here, we note that our use of MICE, which does not account for interactions, may downwardly bias our estimates of group differences in context (i.e., further from the root node of the tree). Combined with the reduced type-I error rate, the tests conducted by ctree are further conservative, which is preferable to liberal tests in such an exploratory analysis. Our approach is also relatively straightforward to implement, as pruning using the Stack ÷ M method can be done post hoc on an estimated ctree grown from the stacked data using freely available software. More sophisticated approaches to this testing problem will likely require rewriting the ctree procedure from scratch. That said, future work to render regression trees compatible with multiple imputation and obtain proper type-I error rates, with unbiased associations, will be a valuable contribution to this methodology.

Finally, interpretation of complex trees requires caution. Terminal nodes describe empirical subgroups but do not on their own establish causal effects; node-level prevalences and associations are descriptive and can be influenced by unmeasured confounding and cohort-specific sampling. For example, regression and path analysis frameworks can permit interpretation of effects with respect to the causal role of predictors (e.g., direct, mediated, and total effects in mediation analysis) within a delineated causal model; our application of ctree, which does not distinguish variables by these causal roles, likewise does not allow interpretation with respect to these roles. As in regression, results from ctree should only be interpreted causally if the research design permits this interpretation (e.g., an experiment); our application to perinatal depression risk in ECHO is purely observational and does not permit these interpretations. In this sense, ctree is used in this study as an inherently exploratory model, and we encourage researchers to follow up these sorts of inquiry with targeted research designs.

Here, note that our analytic objective was not to establish temporally ordered predictors or estimate causal effects. Rather, we conceptualize the ctree model as describing patterns within a perinatal ecological system in a manner that is agnostic to causal mechanisms and temporal directionality. Specifically, our goal was exploratory: to identify subgroups of individuals who are systematically different with respect to PND and PPD risk. In this framework, variables are treated as observed features of the broader ecological context rather than as temporally ordered predictors intended to estimate directional effects. Accordingly, the inclusion of gestational age at delivery as a predictor of PND (despite being observed after PND) does not imply that this feature temporally precedes or causally influences PND. Rather, its appearance in the tree reflects its utility, operationalized in the ctree algorithm, in partitioning the population into subgroups with distinct conditional outcome prevalence, not evidence of etiologic influence. Differences on this or similar variables can be interpreted by understanding them to be proxies for unmeasured characteristics that are associated with both gestational age at birth and PND. However, it deserves mentioning that the same caveat applies to regression models involving observational data; wherein, unobserved variables may underly observed associations.

Importantly, modeling the multivariate system without enforcing temporal ordering also has strengths. By allowing all observed characteristics to be considered simultaneously, the approach captures complex interrelations across demographic, clinical, and contextual features that jointly characterize risk profiles. This enables a holistic understanding of the ecological system surrounding perinatal mental health rather than isolating single predictors in a linear framework. The resulting tree structure therefore reflects conditional heterogeneity across combinations of characteristics, which may generate hypotheses about subgroup-specific vulnerability without asserting causal pathways.

Lastly, to realize the wider potential of ctree for perinatal mental-health research, future work should pursue external validation and extension. Applying ctree to other large, multicohort datasets (and to different outcome constructs and settings) will clarify which node-level patterns replicate and which are sample-specific. Methodological extensions include, but are not limited to stability assessment via bootstrap, integration with globally optimal tree searches, and ensemble techniques that combine tree-based partitioning with other flexible learners (for example ctree plus random forest; Knopf et al., 2025). Taken together, these steps will help ensure that person-centered discoveries about PND and PPD are robust, interpretable, and useful for advancing targeted prevention and intervention. Researchers from all areas of biopsychosocial inquiry are encouraged to explore the use of recursive partitioning algorithms to explore context amidst high dimensionality in situations where they might otherwise default to GLMs, HLMs, LCRs, or composite indices.

Supplementary Material

Supplementary Information
Simulation Study

Supplemental data for this article can be accessed online at https://doi.org/10.1080/00273171.2026.2661244.

Acknowledgments:

The ideas and opinions expressed herein are those of the authors alone, and endorsement by the authors’ institutions or the National Institutes of Health is not intended and should not be inferred.

Funding:

This work was supported by the Environmental influences on Child Health Outcomes (ECHO) Program, Office of the Director, National Institutes of Health, under Award Numbers U2COD023375 (Coordinating Center), U24OD023382 (Data Analysis Center), U24OD023319 with co-funding from the Office of Behavioral and Social Science Research (PRO Core), UH3OD023251 (Alshawabkeh), UH3OD023313 (Koinis Mitchell), UH3OD023328 (Duarte), UH3OD023279 (Elliott), UH3OD023282 (Gern), UH3OD023287 (Breton), UH3OD023244 (Hipwell), UH3OD023271 (Karr), UH3OD023349 (O’Connor), UH3OD023286 (Oken), UH3OD023285 (Kerver), UH3OD023305 (Trasande), UH3OD023337 (Wright).

Role of the Funders/Sponsors:

None of the funders or sponsors of this research had any role in the design and conduct of the study; collection, management, analysis, and interpretation of data; preparation, review, or approval of the manuscript; or decision to submit the manuscript for publication.

Footnotes

Ethical principles: The authors affirm having followed professional ethical guidelines in preparing this work. These guidelines include obtaining informed consent from human participants, maintaining ethical treatment and respect for the rights of human or animal participants, and ensuring the privacy of participants and their data, such as ensuring that individual participants cannot be identified in reported results or from publicly available original or archival data.

Conflict of interest disclosure: Each author signed a form for disclosure of potential conflicts of interest. No authors reported any financial or other conflicts of interest in relation to the work described.

Data availability statement:

Select de-identified data from the ECHO Program are available through NICHD’s Data and Specimen Hub (DASH). Information on study data not available on DASH, such as some Indigenous datasets, can be found on the ECHO study DASH webpage.

References

  1. Al-Abri K, Edge D, & Armitage CJ (2023). Prevalence and correlates of perinatal depression. Social Psychiatry and Psychiatric Epidemiology, 58(11), 1581–1590. 10.1007/s00127-022-02386-9 [DOI] [PMC free article] [PubMed] [Google Scholar]
  2. Barra L, & Coo S (2023). Preterm-born children’s development: A bioecological perspective. Infant and Child Development, 32(1), e2384. 10.1002/icd.2384 [DOI] [Google Scholar]
  3. Beck CT (2001). Predictors of postpartum depression: An update. Nursing Research, 50(5), 275–285. 10.1097/00006199-200109000-00004 [DOI] [PubMed] [Google Scholar]
  4. Beesley LJ, & Taylor JMG (2021). A stacked approach for chained equations multiple imputation that incorporates the substantive model. Biometrics, 77(4), 1342–1354. 10.1111/biom.13372 [DOI] [PMC free article] [PubMed] [Google Scholar]
  5. Bernstein DP, Fink L, Handelsman L, & Foote J (1998). Childhood trauma questionnaire. In Olson DH(Ed.), Assessment of family violence: A handbook for researchers and practitioners. Sage. [Google Scholar]
  6. Blackwell CK, Tang X, Elliott AJ, Thomes T, Louwagie H, Gershon R, Schalet BD, & Cella D (2021). Developing a common metric for depression across adulthood: Linking PROMIS depression with the Edinburgh Postnatal Depression Scale. Psychological Assessment, 33(7), 610–618. 10.1037/pas0001009 [DOI] [PMC free article] [PubMed] [Google Scholar]
  7. Blyth CR (1972). On Simpson’s paradox and the sure-thing principle. Journal of the American Statistical Association, 67(338), 364–366. 10.1080/01621459.1972.10482387 [DOI] [PubMed] [Google Scholar]
  8. Bronfenbrenner U (2005). Making human beings human: Bioecological perspectives on human development. Sage. [Google Scholar]
  9. Bronfenbrenner U, & Morris PA (2006). The bioecological model of human development. In Lerner RM (Ed.), Handbook of child psychology: Vol. 1. Theoretical models of human development. (6th ed., pp. 793–828). John Wiley & Sons. [Google Scholar]
  10. Choi SW, Schalet B, Cook KF, & Cella D (2014). Establishing a common metric for depressive symptoms: Linking the BDI-II, CES-D, and PHQ-9 to PROMIS depression. Psychological Assessment, 26(2), 513–527. 10.1037/a0035768 [DOI] [PMC free article] [PubMed] [Google Scholar]
  11. Cook KF, Jensen SE, Schalet BD, Beaumont JL, Amtmann D, Czajkowski S, Dewalt DA, Fries JF, Pilkonis PA, Reeve BB, Stone AA, Weinfurt KP, & Cella D (2016). PROMIS measures of pain, fatigue, negative affect, physical function, and social function demonstrated clinical validity across a range of chronic conditions. Journal of Clinical Epidemiology, 73, 89–102. 10.1016/j.jclinepi.2015.08.038 [DOI] [PMC free article] [PubMed] [Google Scholar]
  12. R Core Team (2022). R: A language and environment for statistical computing. R Foundation for Statistical Computing. https://www.R-project.org/ [Google Scholar]
  13. Cox JL, & Holden JM (2003). Perinatal mental health: A guide to the Edinburgh Postnatal Depression Scale (EPDS). Royal College of Psychiatrists. [Google Scholar]
  14. Di Florio A, Putnam K, Altemus M, Apter G, Bergink V, Bilszta J, Brock R, Buist A, Deligiannidis KM, Devouche E, Epperson CN, Guille C, Kim D, Lichtenstein P, Magnusson PKE, Martinez P, Munk-Olsen T, Newport J, Payne J, … Meltzer-Brody S (2017). The impact of education, country, race and ethnicity on the self-report of postpartum depression using the Edinburgh Postnatal Depression Scale. Psychological Medicine, 47(5), 787–799. 10.1017/S0033291716002087 [DOI] [PMC free article] [PubMed] [Google Scholar]
  15. Doove LL, Van Buuren S, & Dusseldorp E (2014). Recursive partitioning for missing data imputation in the presence of interaction effects. Computational Statistics & Data Analysis, 72, 92–104. 10.1016/j.csda.2013.10.025 [DOI] [Google Scholar]
  16. Dziak JJ, Lanza ST, & Tan X (2014). Effect size, statistical power, and sample size requirements for the boot-strap likelihood ratio test in latent class analysis. Structural Equation Modeling, 21(4), 534–552. 10.1080/10705511.2014.919819 [DOI] [PMC free article] [PubMed] [Google Scholar]
  17. Elder GH (1998). The life course as developmental theory. Child Development, 69(1), 1–12. 10.1111/j.1467-8624.1998.tb06128.x [DOI] [PubMed] [Google Scholar]
  18. Enders CK (2022). Applied missing data analysis (et al. ed.). Guilford Press. [Google Scholar]
  19. Evans GW, Li D, & Whipple SS (2013). Cumulative risk and child development. Psychological Bulletin, 139(6), 1342–1396. 10.1037/a0031808 [DOI] [PubMed] [Google Scholar]
  20. Fahey JO, & Shenassa E (2013). Understanding and meeting the needs of women in the postpartum period: The perinatal maternal health promotion model. Journal of Midwifery & Women’s Health, 58(6), 613–621. 10.1111/jmwh.12139 [DOI] [PubMed] [Google Scholar]
  21. Flynn HA, Davis M, Marcus SM, Cunningham R, & Blow FC (2004). Rates of maternal depression in pediatric emergency department and relationship to child service utilization. General Hospital Psychiatry, 26(4), 316–322. 10.1016/j.genhosppsych.2004.03.009 [DOI] [PubMed] [Google Scholar]
  22. Gastaldon C, Papola D, Carta MG, & Haro JM (2022). Risk factors for postpartum depression: An umbrella review of systematic reviews and meta-analyses. British Journal of Psychiatry, 221(4), 591–602. 10.1192/bjp.2021.222 [DOI] [PubMed] [Google Scholar]
  23. Gavin NI, Gaynes BN, Lohr KN, Meltzer-Brody S, Gartlehner G, & Swinson T (2005). Perinatal depression: A systematic review of prevalence and incidence. Obstetrics and Gynecology, 106(5 Pt 1), 1071–1083. 10.1097/01.AOG.0000183597.31630.db [DOI] [PubMed] [Google Scholar]
  24. Gillman MW, & Blaisdell CJ (2018). Environmental influences on child health outcomes, a research program of the NIH. Current Opinion in Pediatrics, 30(2), 260– 262. 10.1097/MOP.0000000000000600 [DOI] [PMC free article] [PubMed] [Google Scholar]
  25. Grubinger T, Zeileis A, & Pfeiffer K-P (2014). evtree: Evolutionary learning of globally optimal classification and regression trees in R. Journal of Statistical Software, 61(1), 29. 10.18637/jss.v061.i01 [DOI] [Google Scholar]
  26. Guintivano J, Manuck T, & Meltzer-Brody S (2018). Predictors of postpartum depression: A comprehensive review of the last decade of evidence. Clinical Obstetrics and Gynecology, 61(3), 591–603. 10.1097/GRF.0000000000000368 [DOI] [PMC free article] [PubMed] [Google Scholar]
  27. Hothorn T, & Zeileis A (2015). partykit: A modular toolkit for recursive partitioning in R. Journal of Machine Learning Research, 16, 3905–3909. [Google Scholar]
  28. Hothorn T, Hornik K, & Zeileis A (2006). Unbiased recursive partitioning: A conditional inference framework. Journal of Computational and Graphical Statistics, 15(3), 651–674. 10.1198/106186006X133933 [DOI] [Google Scholar]
  29. Hothorn T, Seibold H, Zeileis A (2025). partykit: A toolkit for recursive partitioning (Version 1.2–24) [Computer software]. Comprehensive R Archive Network. https://cran.r-project.org/web/packages/partykit/ [Google Scholar]
  30. Kaat AJ, Newcomb ME, Ryan DT, & Mustanski B (2017). Expanding a common metric for depression reporting: Linking two scales to PROMIS® depression. Quality of Life Research, 26(5), 1119–1128. 10.1007/s11136-016-1450-z [DOI] [PMC free article] [PubMed] [Google Scholar]
  31. Kievit RA, Frankenhuis WE, Waldorp LJ, & Borsboom D (2013). Simpson’s paradox in psychological science: A practical guide. Frontiers in Psychology, 4, 513. 10.3389/fpsyg.2013.00513 [DOI] [PMC free article] [PubMed] [Google Scholar]
  32. Knopf HT, Sherlock P, Mansolf M, Chapman R, Li J, Daley D (2025). An ecological study of literacy growth among VPK attendees. Early Childhood Policy Research Group, Anita Zucker Center for Excellence in Early Childhood Studies, University of Florida. https://thesunshineportal.org/research/child-outcome/ [Google Scholar]
  33. Lamont AE, Vermunt JK, & Van Horn ML (2016). Regression mixture models: Does modeling the covariance between independent variables and latent classes improve the results? Multivariate Behavioral Research, 51(1), 35–52. 10.1080/00273171.2015.1095063 [DOI] [PMC free article] [PubMed] [Google Scholar]
  34. Lindahl V, Pearson JL, & Colpe L (2005). Prevalence of suicidality during pregnancy and the postpartum. Archives of Women’s Mental Health, 8(2), 77–87. 10.1007/s00737-005-0080-1 [DOI] [PubMed] [Google Scholar]
  35. Matsumura K, Hamazaki K, Tsuchida A, Kasamatsu H, & Inadera H (2019). Education level and risk of postpartum depression: Results from the Japan Environment and Children’s Study (JECS). BMC Psychiatry, 19, 419.[ [Google Scholar]
  36. Mitchell AR, Gordon H, Lindquist A, Walker SP, Homer CSE, Middleton A, Cluver CA, Tong S, & Hastie R (2023). Prevalence of perinatal depression in low- and middle-income countries: A systematic review and meta-analysis. JAMA Psychiatry, 80(8), 857. 10.1001/jamapsychiatry.2023.0069 [DOI] [PMC free article] [PubMed] [Google Scholar]
  37. Motrico E, Mayora M, Gil M, Salgado A, & Déniz C (2023). Interventions to prevent perinatal depression: A systematic review and meta-analysis. General Hospital Psychiatry, 82, 47–61. 10.1016/j.genhosppsych.2023.03.007 [DOI] [PMC free article] [PubMed] [Google Scholar]
  38. Norhayati M, Hazlina NN, Asrenee A, & Emilin WW (2015). Magnitude and risk factors for postpartum symptoms: A literature review. Journal of Affective Disorders, 175, 34–52. 10.1016/j.jad.2014.12.041 [DOI] [PubMed] [Google Scholar]
  39. Patterson GR, DeBaryshe BD, & Ramsey E (1989). A developmental perspective on antisocial behavior. In Cicchetti D(Ed.), The emergence of a discipline: Rochester symposium on developmental psychopathology (Vol. 1, pp. 83–105). Erlbaum. [Google Scholar]
  40. Pilkonis PA, Choi SW, Reise SP, Stover AM, Riley WT, & Cella D (2011). Item banks for measuring emotional distress from the Patient-Reported Outcomes Measurement Information System (PROMIS®): Depression, anxiety, and anger. Assessment, 18(3), 263–283. 10.1177/1073191111411667 [DOI] [PMC free article] [PubMed] [Google Scholar]
  41. Rodgers DM, Jacobucci R, & Grimm KJ (2021). A multiple imputation approach for handling missing data in classification and regression trees. Journal of Behavioral Data Science, 1(1), 127–153. 10.35566/jbds/v1n1/p6 [DOI] [Google Scholar]
  42. Rubin DB (1976). Inference and missing data. Biometrika, 63, 581–592. 10.1093/biomet/63.3.581 [DOI] [Google Scholar]
  43. Sherlock P, DiStefano C, & Habing B (2022). Effects of mixing weights and predictor distributions on regression mixture models. Structural Equation Modeling, 29(1), 70– 85. 10.1080/10705511.2021.1932508 [DOI] [PMC free article] [PubMed] [Google Scholar]
  44. Tudge JRH, Payir A, Merçon-Vargas E, Cao H, Liang Y, Li J, & O’Brien L (2016). Still misused after all these years? A reevaluation of the uses of Bronfenbrenner’s bioecological theory of human development. Journal of Family Theory & Review, 8(4), 427–445. 10.1111/jftr.12165 [DOI] [Google Scholar]
  45. Van Buuren S, & Groothuis-Oudshoorn K (2011). mice: Multivariate imputation by chained equations in R. Journal of Statistical Software, 45(3), 1–67. 10.18637/jss.v045.i03 [DOI] [Google Scholar]
  46. Van Horn ML, Smith J, Fagan AA, Jaki T, Feaster DJ, Masyn K, Hawkins JD, & Howe G (2012). Not quite normal: Consequences of violating the assumption of normality in regression mixture models. Structural Equation Modeling, 19(2), 227–249. 10.1080/10705511.2012.659622 [DOI] [PMC free article] [PubMed] [Google Scholar]
  47. Vermunt JK, & Magidson J (2003). Latent class models for classification. Computational Statistics & Data Analysis, 41(3–4), 531–537. 10.1016/S0167-9473(02)00179-2 [DOI] [Google Scholar]
  48. Waqas A, Raza A, Sharif A, Ahmed MU, & Ullah I (2023). Exploring heterogeneity in perinatal depression: A systematic review of latent class, trajectory, and mixture modeling approaches. BMC Psychiatry, 23, 643. 10.1186/s12888-023-05121-z [DOI] [PMC free article] [PubMed] [Google Scholar]
  49. Wilkinson K (2023). Maternal (perinatal) mental health and employment: An agenda for research and practice. Human Resource Management Journal, 33(2), 346–361. 10.1111/1748-8583.12434 [DOI] [Google Scholar]
  50. Yang K, Wu J, & Chen X (2022). Risk factors of perinatal depression in women: A systematic review and meta-analysis. BMC Psychiatry, 22(1), 63. 10.1186/s12888-021-03684-3 [DOI] [PMC free article] [PubMed] [Google Scholar]
  51. Yim IS, Tanner Stapleton LR, Guardino CM, Hahn-Holbrook J, & Dunkel Schetter C (2015). Biological and psychosocial predictors of postpartum depression: Systematic review and call for integration. Annual Review of Clinical Psychology, 11(1), 99–137. 10.1146/annurev-clinpsy-101414-020426 [DOI] [PMC free article] [PubMed] [Google Scholar]
  52. Yin X, Sun N, Jiang N, Xu X, Gan Y, Zhang J, Qiu L, Yang C, Shi X, Chang J, & Gong Y (2021). Prevalence and associated factors of antenatal (prenatal) depression: A systematic review and meta-analysis. Clinical Psychology Review, 83, 101932. 10.1016/j.cpr.2020.101932 [DOI] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Supplementary Information
NIHMS2166893-supplement-Supplementary_Information.docx (53.9KB, docx)
Simulation Study
NIHMS2166893-supplement-Simulation_Study.docx (429.3KB, docx)

Data Availability Statement

Select de-identified data from the ECHO Program are available through NICHD’s Data and Specimen Hub (DASH). Information on study data not available on DASH, such as some Indigenous datasets, can be found on the ECHO study DASH webpage.

RESOURCES