Effect of perinatal depression on risk of adverse infant health outcomes in mother-infant dyads in Gondar town: a causal analysis

This analysis forms part of a large mother-child health cohort study designed to examine the incidence and prevalence of perinatal depression and its effects on birth and infant health outcomes. The required sample size was determined using Epi Info version 7 [32], with the following assumptions: a type-1 error rate of alpha = 0.05, 90% power, an exposed to unexposed (perinatal depression) ratio of 1:2, and an odds ratio of low birth weight of 1.5 for infants born to women with depression compared to those without depression. A sample size of 809 was estimated and 20% was added for expected losses during follow up, giving a final recruitment target of 970.

Ethics approval was obtained from the Institutional Review Board of the University of Gondar and the Social and Behavioral Research Ethics Committee (SBREC) of Flinders University in South Australia [33]. A support letter was provided by the mayor’s office for Gondar town. Participants were informed of the study’s aims and objectives and their right to withdraw from the research during follow-up. Each volunteer was asked to provide written consent and confidentiality was maintained throughout the study. Women with an overall Edinburgh Postnatal Depression Scale (EPDS) score of 13 to 16, and a score 1, 2, or 3 on item ten (thought of suicide) were referred to University of Gondar Specialized Hospital [34] for further diagnosis and treatment, whilst those with an overall EPDS ≥17 were excluded from the study.

Trained nurse data collectors conducted face-to-face interviews with women in their home using a structured electronic-based questionnaire (supplementary material 1) to collect data on the exposure, outcomes, and potential confounders. The Open Data Collection Kit (ODK) was used to collect the data online using a Lenovo 7 tablet after being checked for validity using Enketo [35] and uploaded to the Google cloud platform.

The Edinburgh Postnatal Depression Scale (EPDS) developed by Cox [34] and adapted for use in the Ethiopian context [36] was used to assess the mother’s depression. The tool measures the extent of stress that pregnant women experienced during the previous week [37‐39] and has been validated in the urban population with a sensitivity, specificity, and misclassification rate of 78.9, 75.3, and 24.0% respectively. Women were considered to be depressed if they had an EPDS score of ≥12 during pregnancy (antenatal depression) and ≥ 6 during the postnatal period (postnatal depression) [40]. The Cronbach’s alpha for internal consistency was 0.74 in this study.

The primary infant outcomes assessed were malnutrition, diarrhea, and ARI. Malnutrition was assessed using the measurement of Middle-Upper Arm-Circumference (MUAC) and defined as infant with MUAC of ≤110 mm [41, 42]. The Integrated Management of Newborn and Childhood Illnesses (IMNCI) guideline was used to identify infants for diarrhea and ARI [43]. Diarrhea was defined as three or more episodes of loose stools in 24 h [44] and ARI was defined as a cough/cold accompanying fever or rapid breathing [45].

Potential confounding variables were identified using the modified disjunctive cause criterion. According to this criteria, covariates were considered to be potential confounders if: (1) they had significant associations with the exposure, the outcome or both; (2) they were not an instrumental variable; and (3) they were not likely to be on the causal pathway between the primary exposure and the outcome [46, 47]. Accordingly, the following confounders for both exposure and outcome were identified: family food access, maternal age, infant age, maternal education and occupational status, pregnancy intention, maternal service uptake, parity, maternal nutritional status, fear of giving birth, a history of chronic mental disorder, level of partner support, quality of partner relationship, social support, and stress coping ability. The “targeting step” in TMLE involves the use of these confounders to estimate the predicted probability of exposure (maternal depression) for each participant given these confounders. This probability is then used to update the estimated risk of the outcome, which is modelled using the observed exposure (depression/no depression) and the same set of confounders. The updated estimates of the risk of the outcome are then used to generate updated pairs of potential outcomes. The “average treatment effect (ATE)”, which here is the risk difference, is finally calculated as the average difference between these pairs across individual [28].

Social support during pregnancy was measured using the Oslo Social Support Scale (OSSS-3) [48]. The three items from OSSS-3 Likert scales were summed to a possible 14 points and women were categorized as having either ‘poor’ (total score < 9) or moderate to strong (overall rating 9–14) support. The Cronbach’s alpha for internal consistency was 0.76 in this study. The support that participants received from their partner was assessed using a 5-point Likert scale via the question ‘my partner helps me a lot’, which had possible responses; ‘always,’ ‘most of the time,’ ‘some of the time,’ ‘rarely,’ and ‘never’. Quality of partner relationship was assessed using a 3-point Likert scale via the question “How do you rate your relationship with your partner in day to day life?” with response categories, ‘very good’, ‘good’, and ‘poor’.

The maternal Middle-Upper Arm Circumference (MUAC) measured maternal nutritional status. The MUAC is validated for measuring nutritional status in the postnatal period and a cutoff score of 18-22 mm rated as ‘underweight’ and 22.5 to 31 mm as ‘normal” [49]. Participants were asked about their pregnancy intention via the question ‘at the time you became pregnant with this pregnancy, did you want to become pregnant, did you want to wait until later, or did you not want to have any more children?’. Their responses were categorized as ‘wanted now’, ‘wanted later’, and ‘not wanted at all’. The wanted now or later options were combined and labelled as ‘planned’ and ‘not wanted at all’ was labelled as ‘unplanned’. LBW was classified as a birth weight less than 2500 g [50].

The Perinatal Coping Inventory (PCI-4) was developed to assess maternal stress coping ability during pregnancy [51]. Coping styles within this tool included: (1) preparation for motherhood, ‘planned how you would handle the birth’; (2) avoidance ‘avoided being with people in general’; (3) positive appraisal ‘felt that being pregnant has enriched your life’; and (4) prayer ‘prayed that the birth would go well’. Participants were asked to report how often they used each of the above coping styles and responses were recorded using a 4-point Likert scale as 0 (never), (1) rarely, (2) sometimes, (3) most of the time [52]. The Cronbach’s alpha for internal consistency was 0.50 in this study.

Completed survey data were downloaded from the Google cloud platform in an Excel spreadsheet, checked for completeness, and imported to Stata version 14 [53] for analysis. Descriptive statistics including mean (SD), median (IQR), frequency (percentage) were used as appropriate. Targeted maximum likelihood estimation (TMLE) was used to investigate the causal effects of perinatal depression on the risk of infant diarrhea, ARI and malnutrition using the estimated average treatment effect (ATE) [54, 55] which was reported as a risk difference (RD). The ATE estimates the average difference in the outcome between participants had they all been exposed and had they all been unexposed, adjusting for potential confounders [56]. TMLE applies G-computation and propensity score methods that involve both exposure and outcome mechanisms [28, 57], and is a doubly robust estimator, providing unbiased estimates when either the exposure model or the outcome model are miss-specified [58]. TMLE is also unbiased in the presence of outliers, unmeasured confounding, sparsity, and other modeling challenges [59]. Assumptions of the model includes no loss to follow up, similar to that for a randomized trial design [60]. We also assessed the potential mediating effects of LBW, postnatal depression, and early initiation of breast feeding using Generalized Structural Equation Models (GSEM) [61]. We included these same potential mediator variables in the causal model as a sensitivity analysis. Interaction terms for antenatal and postnatal depression with social support, partner support, and stress coping ability were also assessed for inclusion in the causal model.

A generalized estimating equation (GEE) model with a Poisson link function and exchangeable correlation structure was used as a comparison model to the TMLE model. Robust standard errors were used for the Poisson GEE given the clustering for the incidence of diarrhea, ARI and malnutrition within districts [62, 63]. Multicollinearity was assessed using correlation coefficients and the Variance Inflation Factor (VIF) with cut-off values of ≥0.8 and ≥ 10, respectively [64].

The socio-demographic characteristics of the study participants are described in Table 1. The mean (SD) age of the mothers and the infants were 26.5 (4.5) years and 4.8 (1.3) months, respectively. There was a significant age difference for infants with and without diarrhea (p = 0.024) and malnutrition (p = < 0.001). Nearly half of the participants (n = 422, 48.7%) reported low-income, 322 (37.2%) completed high school, and 700 (80.8%) were Orthodox Christians. Education, income, and religion were associated with infant malnutrition and ARI (p < 0.001 for each). Most participants were engaged in home duties (n = 617, 71.3%), partnered (n = 832, 96.1%), and did not have difficulties in accessing food in the previous 3 months (n = 832, 96.1%).

Maternal and infant characteristics of the study participants are described in Table 2. Most of the pregnancies were planned, 738 (85.2%), and 333 (38.4%) were first pregnancies. Almost all (95.7%) of the study participants had engaged with antenatal care (ANC), and most had attended postnatal care service (76.7%). Twenty-six (3.0%) and 128 (14.9%) were low birth weights and preterm births, respectively. Seventy-one (8.2%) infants were underweight, 529 (61.4%) mothers initiated early breastfeeding, and 577 (66.6%) strongly agreed that their infants were satisfied with breastfeeding. The incidence (95% CI) of diarrhea, ARI, and malnutrition were 17.0% (14.5–19.6), 21.6% (18.89–24.49), and 14.4% (12.2–16.9) respectively. During the follow up period, the episodes of ARI and diarrhea ranged from 1 to 5 and 1 to 4 respectively.

Univariate associations were observed between diarrhea and LBW (p = 0.016) and infant care (p = 0.004). Univariate associations were also observed between ARI and ANC (p = 0.014), postnatal care (p = 0.016), LBW (p < 0.001), and early initiation of breastfeeding (p < 001). Early initiation of breast feeding (p < 0.001), postnatal care (p < 0.001), breastfeeding satisfaction (p < 0.001), and infant care (p = 0.024) were univariately associated with the risk of malnutrition.

The psycho-social characteristics of the study participants are described in Table 3. Fifty-six (6.5%) and 74 (8.5%) of the mothers had antenatal and postnatal depression respectively. In univariate analysis, neither antenatal or postnatal depression were associated with diarrhea, ARI or malnutrition (p > 0.1 for each). More than half of the study participants, 564 (66.6%), had a good relationship with their partners and 691 (79.8%) had good social support. About 403 (46.5%) women had frequent support from their partner and 547 (63.2%) had poor stress coping abilities. Diarrhea (p = 0.042) and ARI (p < 0.001) were associated with stress coping ability.

Table 4 shows the results from the TMLE analysis. In the fully adjusted models, there was no association between antenatal depression and the risk of diarrhea, ARI or malnutrition. Specifically, compared to those without antenatal depression, the risk difference for antenatal depression and diarrhea was 0.8% (95%CI: − 9.2, 10.9), − 1.3% (95%CI: − 21.0, 18.5) for ARI, and − 7.3% (95%CI: − 22.0, 21.8) for malnutrition. Similarly, postnatal depression was not associated with diarrhea (risk difference = − 2.4%; 95%CI: − 9.6, 4.9), ARI (risk difference = − 3.2%; 95%CI: − 12.4, 5.9) or malnutrition (risk difference = 0.9%; 95%CI: − 7.6, 9.5). In comparison to the TMLE models, the point estimates for the IRR’s for each of the GEE models were generally much further from the null association i.e. an incidence rate ratio equal to one, although none reached statistical significance (Table 4).

In the mediation analysis using GSEM, results showed that: (i) There was no indirect effect of antenatal depression via postnatal depression on the risk of either diarrhea (p = 0.213), ARI (p = 0.660) or malnutrition (p = 0.182); (ii) There was no indirect effect of antenatal depression on the risk of malnutrition via LBW (p = 0.551) or early initiation of breastfeeding (p = 0.705). There were also no significant interactions between antenatal and postnatal depression with social support, partner support, and stress coping ability on the risk of diarrhea, ARI, and malnutrition (p > 0.2 for each). In the sensitivity analysis of the TMLE models, including the potential mediators did not change any of the substantive conclusions.

This is the first prospective cohort study, conducted in an urban setting of Ethiopia, investigating the causal effect of antenatal and postnatal depression (measured using the EPDS) on the risk of diarrhea, ARI and malnutrition among infants aged under 6 months. We used Targeted Maximum Likelihood Estimation (TMLE) to assess the causal effects. Whilst there were several potential unobserved confounders such as personal hygiene, water access, latrine availability and cleanliness, various forms of violence, and immunization status, the method we used is doubly robust, such that model misspecification for either the exposure or outcome due to unobserved confounding reduces the potential for bias [58, 59].

The results of this study need to be considered under the assumptions for valid causal inference. This includes the positivity assumption, which states that all individuals have a positive probability of exposure. Since women with a high likelihood of having severe depression were excluded, this might have introduced a potential selection bias. In addition, the use of MUAC to assess malnutrition has not been validated in Ethiopia in infants of age under 4 months, and this might have either under or overestimated the true infant nutritional status. Similarly, we employed proxy indicators or symptoms (as established in the IMNCI guideline) to identify infants with diarrhea and ARI, which may have led to either under or over-diagnosis of cases. Nonetheless, the IMNCI guideline is recommended by the World Health Organization for use in primary health facilities and demonstrates high sensitivity and specificity.