Elevated depressive symptoms among childbearing women are linked with increased morbidity in both the mother (Wisner et al. Reference Wisner, Chambers and Sit2006; Yonkers et al. Reference Yonkers, Wisner, Stewart, Oberlander, Dell and Stotland2009; Meltzer-Brody & Stuebe, Reference Meltzer-Brody and Stuebe2014) and her offspring (O'Connor et al. Reference O'Connor, Monk and Burke2016). The peripartum period is an especially high-risk period for depression, where as many as 19.2% of new mothers may be diagnosed with either minor or major depression in the first 3 months (Gavin et al. Reference Gavin, Gaynes, Lohr, Meltzer-Brody, Gartlehner and Swinson2005). Baseline prevalence rates for postpartum depression range from 6.5% to 12% (Gavin et al. Reference Gavin, Gaynes, Lohr, Meltzer-Brody, Gartlehner and Swinson2005) and for antenatal depression from 7% to 13% (Bennett et al. Reference Bennett, Einarson, Taddio, Koren and Einarson2004). However, when examining patterns of depressive symptoms over time, the picture may be more complex. First, most postpartum (in the period after childbirth and beyond) women exhibit a stable course of low levels of depressive symptoms during pregnancy and after childbirth (Affonso et al. Reference Affonso, De, Horowitz and Mayberry2000; Luoma et al. Reference Luoma, Korhonen, Salmelin, Helminen and Tamminen2015). Second, some women report elevated depressive symptoms between 18 and 32 weeks of gestation but later report remitted symptoms in the 8 months after delivery, and therefore may not require intervention (Heron et al. Reference Heron, O'Connor, Evans, Golding, Glover and Team2004). For other women, elevated depressive symptoms in the antepartum (before parturition) period may be the beginning of a chronic state (McCall-Hosenfeld et al. Reference McCall-Hosenfeld, Phiri, Schaefer, Zhu and Kjerulff2016) (see also Putnam et al. (Reference Putnam, Wilcox, Robertson-Blackmore, Sharkey, Bergink and Munk-Olsen2017)). Documenting heterogeneous trajectories among childbearing women increases understanding of the varying course of depressive symptoms and may identify who to target for intervention during a period of heightened vulnerability. Identifying modifiable risk factors also has important public health implications for reducing or mitigating the effects of elevated depressive symptoms on both the mother and her offspring.
With the development of statistical modeling approaches capable of calculating depressive trajectories, there have been numerous studies on longitudinal patterns of depressive symptoms associated with childbirth, but only 13 have followed women prospectively from pregnancy to after childbirth (Santos et al. Reference Santos, Tan and Salomon2017). Two general approaches to statistical trajectory modeling have been used in these studies. Most prospective studies have employed latent class growth analysis (LCGA), a special case of growth mixture modeling, which restricts the variance of the intercept and growth model parameter to zero (Jung & Wickrama, Reference Jung and Wickrama2008). While this approach has the advantage of reducing computational burden, it assumes that individual growth trajectories have homogeneous intercept and slope parameters. Of these studies, some have found evidence for anywhere from two (Parade et al. Reference Parade, Blankson, Leerkes, Crockenberg and Faldowski2014) to five distinct classes (Hammerton et al. Reference Hammerton, Mahedy, Mars, Harold, Thapar and Zammit2015) of depressive trajectories among childbearing women. Other approaches using LCGA have shown different subtypes of postpartum depression based on onset and chronicity of depressive symptoms [Postpartum Depression: Action Towards Causes and Treatment (PACT) Consortium, 2015].
In contrast to LCGA, latent growth mixture modeling (LGMM) (Muthén, Reference Muthén and Kaplan2004) allows growth parameters to be estimated separately for each class, thereby avoiding the assumption that intercept and slope terms are invariant. Because the assumption of a single homogeneous distribution is relaxed, models can be specified with unusual precision because growth parameters are tailored to fit the data. Studies using this approach have found evidence for more heterogeneity. For example, one study identified an additional trajectory of depressive symptoms that emerged 2 years after childbirth (Mora et al. Reference Mora, Bennett, Elo, Mathew, Coyne and Culhane2009), whereas studies among women of low socioeconomic class suggest that either two (Glasheen et al. Reference Glasheen, Richardson, Kim, Larkby, Swartz and Day2013) or three classes best characterize population heterogeneity (Christensen et al. Reference Christensen, Stuart, Perry and Le2011). The present report employed an LGMM approach using specifications that had not been used in prior work, including freely estimating both the intercept and slope parameters and the influence of covariates on these terms, as well as flexibly modeling time-varying changes on depressive symptoms across several measurement intervals.
Predictors of depressive trajectories
There are two categories of variables known to predict depressive symptoms among childbearing women that were of interest. First, sociodemographic factors such as level of education of the mother (Buttner et al. Reference Buttner, Brock and O'Hara2015), socioeconomic status (Beck, Reference Beck2001), maternal age (Rubertsson et al. Reference Rubertsson, Waldenström and Wickberg2003), and offspring gender (Murray et al. Reference Murray, Kempton, Woolgar and Hooper1993) have been shown to influence risk for depression.
Second, exposure to adversity during pregnancy increases risk for elevated depressive symptoms both during pregnancy (Lancaster et al. Reference Lancaster, Gold, Flynn, Yoo, Marcus and Davis2010) and after childbirth (Robertson et al. Reference Robertson, Grace, Wallington and Stewart2004). Exposure to interpersonal stressors is most closely associated with depressive symptoms among antenatal and postpartum women (Xie et al. Reference Xie, He, Koszycki, Walker and Wen2009; Wright et al. Reference Wright, Hill, Pickles and Sharp2015), although a diverse range of major life events also increases risk (Stone et al. Reference Stone, Diop, Declercq, Cabral, Fox and Wise2015; Fredriksen et al. Reference Fredriksen, von Soest, Smith and Moe2017; Mukherjee et al. Reference Mukherjee, Coxe, Fennie, Madhivanan and Trepka2017). Prior studies have examined stressor exposure in a yes/no self-report format, leaving open the question of outcomes associated with the perceived impact of exposure. This is an important distinction as prior research has suggested that perceptions about exposure have a substantial effect on outcomes (Bonanno et al. Reference Bonanno, Westphal and Mancini2011).
Current study
The primary aim of the present study was to distinguish patterns of chronically elevated levels of depressive symptoms v. trajectories that are elevated in the antepartum phase but then remit after childbirth, v. patterns that increase after childbirth. Our secondary aim was to identify sociodemographic and adversity exposure factors that, when measured prospectively, distinguish trajectories of depression antepartum, postpartum, and beyond. The present report applies a comparative individual differences framework to a large, prospective population-based cohort using a statistical modeling technique that offers three distinct advantages over prior studies. (1) Prior work has not used an approach that allows for the flexible modeling of variation in growth parameters across different classes. (2) Second, we modeled time parameters freely as the increments of change across measurement intervals are not equivalent. (3) Third, we evaluated whether sociodemographic factors, child gender, and exposure to adversity while pregnant predicted membership in distinct classes.
Methods
Procedure
Data are drawn from the Avon Longitudinal Study of Parents and Children (ALSPAC), a longitudinal prospective observational study investigating influences on health and development across the lifespan (Boyd et al. Reference Boyd, Golding, Macleod, Lawlor, Fraser and Henderson2013). This birth cohort population study focused on women residing in a circumscribed geographic region in the South West of England with an expected date of delivery between April 1, 1991 and December 31, 1992 [see Fraser et al. (Reference Fraser, Macdonald-Wallis, Tilling, Boyd, Golding and Davey Smith2013) for complete details on study procedures and Golding et al. (Reference Golding, Pembrey, Jones and Team2001) on study recruitment]. Ethical approval for the study was obtained from the ALSPAC Ethics and Law Committee and the Local Research Ethics Committees. Subjects provided written informed consent after receiving a complete description of the study. The study website contains all details through a fully searchable data dictionary (http://www.bris.ac.uk/alspac/researchers/data-access/data-dictionary/). Participants included in the present report were assessed across seven waves from 1990 to 1995 at 18 and 32 weeks of gestation, 8 weeks after the birth of the baby, and thereafter at 8, 21, 33, and 61 months.
Participants
The complete study cohort consists of 14 541 pregnancies and 13 998 children still alive at 12 months of age. Our sample in the present report consisted of 12 121 mothers who completed questionnaires assessing depressive symptoms. Participation rates for each wave of data collection were: 18 weeks gestation (n = 12 121, 100%); 32 weeks gestation, n = 12 096 (99.8%); 8 weeks postpartum, n = 11 710 (96.6%); 8 months postpartum, n = 11 195 (92.3%); 2 years postpartum, n = 10 259 (84.6%); 3 years postpartum, n = 9595 (79.2%); and 5 years postpartum, n = 8950 (73.8%). As reported in previous studies, mothers lost from the cohort were significantly more likely to endorse higher levels of depression and anxiety (Heron et al. Reference Heron, O'Connor, Evans, Golding, Glover and Team2004), to have lower education levels, and to report greater social adversity (Boyd et al. Reference Boyd, Golding, Macleod, Lawlor, Fraser and Henderson2013). Complete details regarding sample distribution of ethnicity, social class, and maternal and paternal education can be found in online Supplementary material 1.
Measures
Demographics
Maternal age at delivery was calculated by subtracting the offspring delivery date from the maternal date of birth. Sex of the child was obtained from both the birth notification and postmortem examinations in the absence of birth notification. Maternal social class was determined by calculating a composite variable based on the mother's occupation, primary tasks, and job role status (e.g. foreman, supervisor) using the UK Office of Population Census & Surveys (OPCS) classification system (Office of Population Census & Surveys, 1991), resulting in six categories: I (professional), II (managerial and technical), IIIa (skilled manual), IIIb (skilled non-manual), IV (semi-skilled), and V (unskilled). To aid in model interpretation and convergence, we collapsed maternal class into three categories by combining professional and managerial categories (I and II), skilled (IIIa and IIIb), and semi-skilled and unskilled (IV and V). Maternal and paternal education corresponded to the highest attainment at late pregnancy; 1, certificate of secondary education/none; 2, vocational; 3, ordinary level; 4, advanced level; 5, university degree.
Adversity variables
Mothers completed a life-events inventory of 42 items at 18 weeks of gestation (Brown & Harris, Reference Brown and Harris1978; Barnett et al. Reference Barnett, Hanna and Parker1983). Mothers were asked, ‘Listed below are a number of events that may have brought changes in your life. Have any of these [e.g. friend or relative died] occurred since you became pregnant? If so, please assess how much effect it had on you.’ Responses were recorded on a 1–5 scale, with 1, affected me a lot; 2, moderately affected me; 3, mildly affected me; 4, no effect; 5, did not happen. ALSPAC pilot studies suggest that both the impact score and a count tally of the number of events are highly correlated, with the former being more sensitive to emotional well-being. For the purposes of the present report, variables were first recorded (1,2,3,4, = 4,3,2,1) (5 = 0). Then, we combined variables into conceptually related domains according to theory and a principle components analyses. This generated five variables of perceived impact across several domains: bereavement (loss of a friend, relative, child, or partner); medical (illness in a partner, child, friend, or family member; mother's hospitalization; fear of miscarriage; fetal abnormality); interpersonal (arguing with partner, friends, or family; partner cruelty to mother or child; partner physically hurting mother or child; divorce; separation; partner absence; partner not wanting child); financial and employment (partner or mother loosing job, partner or mother having problems at work, mother experiencing reduction in income or financial difficulty); and housing (becoming homeless, moving, or being burglarized). Before summing items within each domain, scores were first standardized to equalize response variability.
Maternal depression
The Edinburgh Post-Natal Depression Scale (EPDS) (Cox et al. Reference Cox, Holden and Sagovsky1987) is a 10-item questionnaire that assesses depressive symptomatology in women during pregnancy and after delivery. Each item is scored from 0 to 3 with 0 representing no endorsement and 3 representing full endorsement (note: some items are reversed scored), yielding a sum score ranging from 0 to 30. The scale has been validated with standardized psychiatric interviews (Cox et al. Reference Cox, Chapman, Murray and Jones1996), and has shown adequate reliability and validity in large, diverse international samples (Affonso et al. Reference Affonso, De, Horowitz and Mayberry2000).
Statistical analysis
We used Mplus, Version 7.4 (Muthén & Muthén, Reference Muthén and Muthén1988–2015) to identify latent depression trajectories. We specified a full-information maximum-likelihood (FMIL) estimation procedure to handle missing data (Enders, Reference Enders2001). FMIL assumes that missing data are unrelated to the outcome variable (missing at random) and its use is widely accepted (Graham, Reference Graham2009). Most participants had complete data at a minimum of three time points, 86.3%. For all time points, the percentages of missing data were: one time point, 16.5%; two time points, 9%; three time points, 7%; four time points, 4.7%; five time points, 4.7%; and six time points, 4.4%. To assess the effects of missing data patterns, Mplus estimates covariance coverage for all pairs of variables; values should be above a minimum threshold of 0.10 for model convergence. In the present study, covariance coverage for each pair of variables did not fall below 0.77.
Analyses proceeded in three steps (Jung & Wickrama, Reference Jung and Wickrama2008). First, we determined the number of classes that optimally characterized individual differences in depressive symptoms by identifying a univariate single-class growth model without covariates. Second, we compared fit of progressive one- to six-class models, using conventional indices to assess fit. Third, we included covariates in a multinomial logistic regression on class membership (Muthén, Reference Muthén and Kaplan2004). We inspected modification indices to determine retention of default specifications. In an iterative process, we tested models with different covariates, comparing log-likelihood ratios and using χ2 tests to adjudicate relative improvement in fit. To ensure that our final model did not represent a local solution, we varied the number of random starts and replicated the two best log-likelihoods using the OPTSEED procedure, confirming that model parameter estimates were replicated (Jung & Wickrama, Reference Jung and Wickrama2008). We regressed covariates on class membership and on growth parameters, and used log-likelihood ratio χ2 tests to optimize model fit (Muthén, Reference Muthén and Kaplan2004). Finally, we examined the effect of exposure to adversity at baseline by including these variables as covariates within the best-fitting LGMM.
Results
Unconditional models
We began by estimating a simple growth model without covariates. To identify factor loadings for the time points, we reasoned that the increment of change in depressive symptoms was likely to vary between pregnancy and the 5 years after childbirth, and thus would assume a non-linear form. We sought to confirm this using several model fitting procedures, examining models with linear-only, linear and quadratic, and a freely estimated slope parameter (first wave set to 0 and last wave set to 1, and the remaining freely estimated). Based on log-likelihood χ2 testing, we retained the model with freely estimated factor loadings. We allowed the intercept variance to be estimated across classes, but we fixed the slope variance at 0 across classes to address problems with non-convergence.
We then compared successive solutions from one to six classes in unconditional models. Table 1 summarizes the fit statistics for the one- to six-class solutions. To adjudicate model fit, we used the Aikaike information criterion (AIC), Bayesian information criteria (BIC), the sample size-adjusted Bayesian information criteria (SSBIC), the Vuong-Lo-Mendell-Rubin likelihood ratio test (VLMRT) (Lo et al. Reference Lo, Mendell and Rubin2001), and the bootstrapped likelihood test (BLRT) (Nylund et al. Reference Nylund, Asparouhov and Muthén2007). We sought a model with a significant p value for the k-1 v. k class models (VMLRT and BLRT), a higher value for entropy, and lower values for the BIC, AIC, and SSBIC, as well as theoretical coherence and interpretability (Muthén, Reference Muthén and Kaplan2004; Nylund et al. Reference Nylund, Asparouhov and Muthén2007). Each model from one to six classes provided successive improvements in fit according to the AIC, BIC, SSBIC. While the five-class solution showed slightly improved fit over the four-class solution, it fragmented the stable low trajectory into two close but parallel trajectories, which carries little theoretical relevance. Additionally, the VLMRT p value for the k-1 class solution at five and six classes was not significant. Based on fit indices as well as theoretical interpretability, we adopted the four-class solution for subsequent examination in conditional models.
Table 1. Fit indices for one- to six-class growth mixture models (unconditional)
LL, log likelihood; AIC, Akaike information criterion; BIC, Bayesian information criterion; SSBIC, sample size-adjusted Bayesian information criterion; VLMRT, Vuong-Lo-Mendell-Rubin likelihood ratio test; BLRT, parametric bootstrapped likelihood ratio test.
Conditional model
Next, we sought to extend the model to include relevant covariates. We regressed class membership on offspring gender, maternal age and education, paternal education, and maternal social class, improving model fit. Entropy for the conditional model increased from 0.75 in the unconditional four-class model to 0.77 in the conditional model, indicating improved classification accuracy. Finally, we regressed the intercept and slope growth parameters (class-invariant) on maternal age, further improving model fit. Inclusion of covariates resulted in trajectory patterns and class membership probabilities that were nearly identical to the unconditional model.
Figure 1 shows that most participants in the conditional model (77.6%) were assigned to a class with stable low depressive symptoms, which we labeled resilient. This class had the lowest intercept (b = 6.25, s.e. = 0.37, p < 0.001) and a significant negative slope (b = −1.79, s.e. = 0.27, p < 0.001). The next largest class (10.7%) had individuals with high levels of depressive symptoms both during pregnancy and after childbirth, declining slightly over the 5 years after childbirth, a group we labeled chronic depression. This group had a high intercept (b = 13.55, s.e. = 0.46, p < 0.001) and a significant negative slope (b = −1.83, s.e. = 0.35, p < 0.001). A third class of individuals (7.7%) had elevated levels of depressive symptoms during pregnancy that declined after the birth, a group we labeled improving. This class had a higher intercept (b = 13.79, s.e. = 0.48, p < 0.001) and a significant negative slope (b = −6.52, s.e. = 0.66, p < 0.001). Finally, we identified a group of participants who had moderate levels of depressive symptoms during pregnancy followed by a sharp elevation after childbirth (4.0%). This emergent group had a moderate intercept (b = 7.98, s.e. = 0.52, p < 0.001) and a significant positive slope (b = 3.53, s.e. = 0.43, p < 0.001).
Fig. 1. Four-class conditional trajectory model of EPDS depression scores (n = 9111). B, weeks before childbirth; A, weeks after childbirth; standard error bars represent standard error of the mean (s.e.m.).
Next, we examined covariate predictors of trajectory class assignment, as shown in Table 2. In the first analyses, the resilient class served as the referent. Compared with the resilient class, mothers in the chronic class had fewer years of education (Est = −0.14, s.e. = 0.05, p = 0.008), as did their partners (Est = −0.08, s.e. = 0.04, p = 0.043). Mothers in the improving group were younger at delivery (Est = −0.06, s.e. = 0.02, p = 0.001) and were more likely to be in a lower social class (Est = 0.67, s.e. = 0.17, p = 0.000). Finally, compared with the chronic group, the emergent group was more likely to have a female child (Est = −0.35, s.e. = 0.16, p = 0.033).
Table 2. Multinomial regression estimates for covariate predictors of class membership
Est., parameter estimate; s.e., standard error.
a High maternal class served as the referent.
b Adversity was measured at 18 weeks of gestation.
†p < 0.10; *p < 0.05; **p < 0.01; ***p < 0.001.
In addition to demographic factors, we included appraised adversity exposure as additional covariate predictors of trajectory class assignment (see Table 3). Compared with mothers in the resilient group, those in the chronic, emergent, and improving groups reported that they were more severely affected by every category of stressor considered in this analysis, except for bereavement. Specifically, when compared with the resilient group, mothers in the chronic, emergent, and improving groups all reported greater exposure to medical adversity, interpersonal adversity, and financial difficulty. Again, compared with the resilient group, only mothers in the chronic and improving groups experienced greater difficulty with housing. Compared with mothers in the chronic group, mothers in the improving group were less likely to report medical adversity, interpersonal difficulty, and financial adversity. There were no significant differences between mothers in the chronic and improving groups in terms of appraised exposure to adversity. Finally, compared with the improving group, mothers in the emergent group were less likely to experience medical, interpersonal, and financial adversity.
Table 3. Descriptive statistics by depression trajectory class
Gest, gestation; Dep. sx's, depressive symptoms. Standard deviations are in parentheses.
a Percentages represent % within total gender. 48.7% of the total offspring sample was female.
b Percentages calculated using trajectory by social class/trajectory total.
c Adversity values for all variables except bereavement are reported in summed z-scores with negative values representing fewer exposures relative to the mean, and positive values indicating more exposure relative to the mean.
*p < 0.05; **p < 0.01; *** p < 0.001.
Discussion
This study had two aims: (1) to distinguish trajectories of chronically elevated levels of depressive symptoms from trajectories that are either elevated during pregnancy but then remit after childbirth, or increase after childbirth; (2) to identify sociodemographic and appraised adversity exposure factors that, when measured prospectively, distinguish trajectory membership. Using LGMM, a four-class solution best characterized patterns of depressive symptoms during pregnancy, after childbirth, and beyond. A clear majority of women (77.6%) demonstrated a resilient pattern characterized by stable, low levels of depressive symptoms. However, 10.7% displayed a chronic pattern characterized by consistently elevated levels of depressive symptoms both before and after childbirth, while 4.0% evidenced an emergent pattern, consisting of low depressive symptoms prior to childbirth, which became elevated after childbirth. Finally, 7.7% evidenced a rarely documented improving pattern whereby elevated levels of depressive symptoms prior to the birth steadily declined in the postpartum period.
Our study findings highlight two particularly high-risk patterns – emergent and chronic, where treatment can potentially have an impact on the 14–15% of pregnant women who may require more intensive screening and monitoring. The critical issue is to distinguish women during pregnancy who have low symptoms and will continue to do so (resilient) from emergent – and then to distinguish those who will improve from those who are chronic. In terms of prediction, the chronic group had lower maternal and partner education compared with the resilient group, as well as greater perceived exposure to adversity across medical, interpersonal, financial, and housing domains. The effects of adversity on depression among perinatal women (Seng et al. Reference Seng, Sperlich, Low, Ronis, Muzik and Liberzon2013; Yonkers et al. Reference Yonkers, Smith, Forray, Epperson, Costello and Lin2014) are well established, and data presented here demonstrated that interpersonal stressors exert a threefold increased risk for depression among the chronic and emergent groups compared with other types of adversity, consistent with prior work (Wright et al. Reference Wright, Hill, Pickles and Sharp2015).
The second high-risk group is the emergent group for whom depressive symptoms do not manifest until after childbirth; they are clearly distinct from the chronic group. This group may not report elevations in depressive symptoms during pregnancy, but could be confused with women with chronically elevated profiles if assessed following childbirth. What factors distinguish these women? First, having a female offspring increased odds of experiencing emergent compared with chronic trajectories and, marginally, differentiated resilient from emergent groups. Why would female gender increase risk for membership in the emergent group v. the chronic group? Prior work has demonstrated mood-destabilizing effects associated with sensitivity to hormonal events of pregnancy and childbirth in a subgroup of women (Bloch et al. Reference Bloch, Schmidt, Danaceau, Murphy, Nieman and Rubinow2000; Pedersen et al. Reference Pedersen, Leserman, Garcia, Stansbury, Meltzer-Brody and Johnson2016), and there is sufficient evidence to suggest that fluctuations in hormones during pregnancy and the postpartum period may trigger postpartum depression (PPD) in susceptible women [see Schiller et al. (Reference Schiller, Meltzer-Brody and Rubinow2015) for a review]. Though largely unexplored in the extant literature and a speculative interpretation, fetal gender may moderate neuroendocrine pathophysiology for PPD given the higher loading of gonadal hormones associated with female fetuses (Yaron et al. Reference Yaron, Lehavi, Orr-Urtreger, Gull, Lessing and Amit2002). Alternatively, non-biological variables may also account for the association between offspring gender and risk for depressive symptoms.
Second, the emergent group was less likely than the chronic group to report stressful experiences. This suggests that membership in the emergent class is likely a consequence of childbirth itself. Third, our model suggests that for women who experience an increase in depressive symptoms after childbirth, this elevated pattern may persist for up to 5 years. These findings are consistent with prior population-based cohort studies that have found evidence for an increased prevalence of depressive symptoms after the first postpartum year (Aasheim et al. Reference Aasheim, Waldenstrom, Hjelmstedt, Rasmussen, Pettersson and Schytt2012; Woolhouse et al. Reference Woolhouse, Gartland, Mensah and Brown2015).
Results also offer some clues as to the factors that distinguish the improving group from the chronic, a critical distinction as these two groups could be confused during pregnancy, even though some will ultimately recover on their own. Though of marginal significance, mothers in the improving group were younger compared with those in the chronic group. Second, compared with the resilient class, the improving class was younger at age of delivery, of lower social class, and reported greater exposure to medical, interpersonal, and financial strain. A key point here is that the improving group is distinct from the resilient group, and likely to engage in different coping patterns during pregnancy and after childbirth (Bonanno, Reference Bonanno2004). What could explain the remission of depression post-childbirth among these mothers? Though the research on postpartum recovery is sparse, this pattern of post-adversity improvement has been well documented in other exposure groups (Mancini et al. Reference Mancini, Littleton and Grills2016). While it is possible that the improving mothers in the present study could have accessed mental health services after delivery, a study in Ghana and the Côte d'Ivoire found evidence of the improving pattern among new mothers not exposed to treatment, noting associations with reduced marital and economic stress after childbirth, as well as parity (Barthel et al. Reference Barthel, Kriston, Barkmann, Appiah-Poku, Te Bonle and Esther Doris2016). Further research on this improving group promises to identify yet undiscovered pathways to postpartum adjustment among mothers.
Finally, most women in this study reported a pattern of stable, low depressive symptoms during pregnancy, after childbirth, and beyond, suggesting that low depressive symptoms may be normative. A substantial body of work has demonstrated that resilience is often the predominant response trajectory after exposure to stressors (Bonanno, Reference Bonanno2004), and the present report is among the first to demonstrate its presence in the context of childbirth [see also Barthel et al. (Reference Barthel, Kriston, Barkmann, Appiah-Poku, Te Bonle and Esther Doris2016) and Galatzer-Levy et al. (Reference Galatzer-Levy, Mazursky, Mancini and Bonanno2011)]. Resilient mothers were more likely to be older than the mothers in the improving class, consistent with prior work noting that older age is a protective factor after childbirth (Rubertsson et al. Reference Rubertsson, Waldenström and Wickberg2003; Buttner et al. Reference Buttner, Brock and O'Hara2015). Second, mothers in this class reported reduced exposure to adversity compared with those in the emergent, chronic, and improving categories. The key implication of this finding is that rather than being unusually healthy, mothers is this class have effective strategies to maintain equilibrium, even if these strategies are not necessarily universally advantageous. For example, studies have shown that self-enhancing biases are linked to coping with exposure to extreme adversity (Bonanno et al. Reference Bonanno, Field, Kovacevic and Kaltman2002), even though self-enhancement can evoke negative impressions from others. The critical challenge for future research on peripartum resilience is to identify the complex multiple pathways to resilience that new mothers are likely to demonstrate.
Several limitations should be considered when interpreting the primary findings. Most importantly, although we controlled for baseline depressive symptoms, we did not have data on prior family or personal history of depression. Thus, we were unable to evaluate whether trajectories of depression symptoms differed by prior depression or a family history of depression. Nor could we adjust for these factors in examining the association of adversity and other risk factors with depression trajectories. This limits the interpretation we can make from these data; for example, familial loading for depression could be associated with increased exposure to adversity and risk of being in the chronic trajectory, thus inflating that association. Second, depressive symptoms were assessed via self-report and were not adequate markers of clinical depression. This has two major implications. First, we are limited in the extent to which we can generalize our findings to defined clinical criteria for perinatal depression and major depression as established in the Diagnostic and Statistical Manual of Mental Disorders (DSM) 5 (American Psychiatric Association, 2013) and International Classification of Diseases (ICD)-9 given that our postpartum measurement fell outside the 4-week postpartum window per the DSM 5, and outside the 6-week window per the ICD-9. Second, prior work has shown that prevalence estimates tend to be higher when symptoms, rather than clinical criteria are used (Bennett et al. Reference Bennett, Einarson, Taddio, Koren and Einarson2004) and we may therefore have inflated estimates of depressive experiences among this group. Third, adverse event chronicity is unknown. This is important given that the type of adversity and its timing can impact the course of subsequent mood symptoms (Class et al. Reference Class, Abel, Khashan, Rickert, Dalman and Larsson2014; Keyes et al. Reference Keyes, Pratt, Galea, McLaughlin, Koenen and Shear2014). Fourth, this study did not consider other known risk factors for depression in the modeling, such as high parity, low social support, and low self-esteem. Fifth, this sample was notable for its lack of representation from non-white ethnicities; the generalizability of our findings to other ethnicities is limited. Finally, data were collected several decades prior to this report and caution should be exercised when interpreting findings given the potentially limited generalizability to contemporary settings. These limitations notwithstanding the current report builds on a body of literature suggesting the presence of heterogeneous depressive symptoms both during pregnancy and in the 5 years after childbirth, offering some insight into variables that distinguish resilient, chronic, improving, and emergent depressive symptom trajectories.
Supplementary material
The supplementary material for this article can be found at https://doi.org/10.1017/S0033291717002549.
Acknowledgements
The authors are extremely grateful to all the families who took part in this study, the midwives for their help in recruiting them, and the whole ALSPAC team, which includes interviewers, computer and laboratory technicians, clerical workers, research scientists, volunteers, managers, receptionists, and nurses. The UK Medical Research Council and Wellcome (Grant ref: 102215/2/13/2) and the University of Bristol provide core support for ALSPAC. The authors would also like to thank Lori Chibnik for her assistance with generating the graphics. This publication is the work of the authors and Christy A. Denckla, Anthony D. Mancini, Nathan S. Consedine, Snezana Milanovic, Archana Basu, Soraya Seedat, Georgina Spies, David C. Henderson, George A. Bonanno, and Karestan C. Koenen will serve as guarantors for the contents of this paper.
Conflict of Interest
None.
