Impact of maternal mental health interventions on child-related outcomes in low- and middle-income countries: a systematic review and meta-analysis

Aims Observational studies have shown a relationship between maternal mental health (MMH) and child development, but few studies have evaluated whether MMH interventions improve child-related outcomes, particularly in low- and middle-income countries. The objective of this review is to synthesise findings on the effectiveness of MMH interventions to improve child-related outcomes in low- and middle-income countries (LMICs). Methods We searched for randomised controlled trials conducted in LMICs evaluating interventions with a MMH component and reporting children's outcomes. Meta-analysis was performed on outcomes included in at least two trials. Results We identified 21 trials with 28 284 mother–child dyads. Most trials were conducted in middle-income countries, evaluating home visiting interventions delivered by general health workers, starting in the third trimester of pregnancy. Only ten trials described acceptable methods for blinding outcome assessors. Four trials showed high risk of bias in at least two of the seven domains assessed in this review. Narrative synthesis showed promising but inconclusive findings for child-related outcomes. Meta-analysis identified a sizeable impact of interventions on exclusive breastfeeding (risk ratio = 1.39, 95% confidence interval (CI): 1.13–1.71, ten trials, N = 4749 mother–child dyads, I2 = 61%) and a small effect on child height-for-age at 6-months (std. mean difference = 0.13, 95% CI: 0.02–0.24, three trials, N = 1388, I2 = 0%). Meta-analyses did not identify intervention benefits for child cognitive and other growth outcomes; however, few trials measured these outcomes. Conclusions These findings support the importance of MMH to improve child-related outcomes in LMICs, particularly exclusive breastfeeding. Given, the small number of trials and methodological limitations, more rigorous trials should be conducted.

Despite demonstrated links in the epidemiological literature, few systematic investigations have been conducted to examine whether maternal mental health (MMH) interventions can reduce potential negative impacts on children's outcomes. The aim of this study was to conduct a systematic review and meta-analysis on this topic. Specifically, our research question was: do interventions with a dedicated psychiatric or psychosocial component delivered to pregnant women and mothers during the perinatal period improve children's health and development in LMICs relative to standard antenatal care or interventions lacking a dedicated psychiatric or psychosocial component?
Randomised controlled trials (RCTs) were eligible for our systematic review if the study: (1) described interventions delivered during the perinatal period, defined as pregnancy through 1-year post-partum; (2) incorporated an MMH intervention component; (3) included a MMH outcome; (4) was conducted in an LMIC (http://data.worldbank.org/about/country-and-lendinggroups) and (5) included a child health, nutrition or development outcome. We retained the child outcomes for inclusion broad since this is (to our knowledge) the first systematic review and meta-analysis on this topic. All non-randomised, non-controlled studies were excluded. We did not limit our results to studies that restricted their samples to women with mental health problems.
Two independent reviewers assessed titles and abstracts from all searches. English and Spanish full texts were retrieved for potentially relevant articles and assessed by two reviewers independently to evaluate eligibility. Inter-rater reliability in the full text review was 74.4%. Articles and abstracts in other languages (two in Farsi) were assessed by a single reviewer that was fluent in the language. This reviewer worked with another member of the research team to review eligibility criteria. Discrepancies were resolved through discussion or consultation with a third reviewer.

Data collection, risk of bias assessment and GRADE certainty of evidence
Two reviewers independently extracted data on study design, sample, study conditions, child-related outcomes, results and risk of bias for each included trial (MCG, MEL, see 'Acknowledgements'). Quantitative results were extracted using the unadjusted means and standard deviations for continuous outcomes and the number of events and denominator for dichotomous outcomes. The risk of bias assessment followed the Cochrane Risk of Bias tool where reviewers rated several potential sources of bias as 'high', 'low' or 'unclear' risk in relation to random sequence generation, allocation concealment, masking of participants/personnel, masking of outcome assessors, attrition, reporting and any other sources of bias of each trial (Higgins and Greene, 2011). We considered overall risk of bias to be high if trials displayed high risk of bias in two or more of these seven domains. Discrepancies were resolved through discussion.
We employed the GRADE approach to assess the overall certainty of evidence and to interpret findings (Barbui et al., 2010). We adhered to the standard methods for the preparation and presentation of results outlined in the Cochrane Handbook for Systematic Reviews of Interventions and PRISMA guidelines (Higgins and Greene, 2011). We included the following outcomes in the GRADE evidence profiles: exclusive breastfeeding, cognitive development, psychomotor development, low birth weight, weight (continuous), height (continuous), underweight (i.e. weight-forage z-score <−2), stunting (i.e. height-for-age z-score <−2) and weight-for-height.

Data analysis
Narrative synthesis: included trials were compared with respect to population, intervention, measurement and methodological features that may contribute to clinically relevant heterogeneity in the synthesis of the results. Reporting of these results followed PRISMA recommendations.
Quantitative synthesis: data from included trials were pooled using a random effects model for outcomes reported in at least two trials and expressed as relative risk (RR) for categorical data, and standardised mean difference (SMD) for continuous data. For categorical outcomes with evidence supporting an intervention effect across more than one study, we calculated the number needed to treat (NNT) to provide benefit (Furukawa et al., 2002). Review Manager was used for all analyses (The Nordic Cochrane Center, 2014). Data from cluster RCTs were adjusted with an intracluster correlation coefficient (ICC). If the ICC was not available, we assumed it to be 0.05 (Higgins and Greene, 2011). Below, we report intention-to-treat analyses including all randomised patients.
We conducted a sub-group analysis by intervention type: (1) focused MMH interventions (i.e. interventions mainly aimed at improving MMH) and (2) integrated interventions (i.e. interventions that included a mental health focused component, but also focused on other outcomes). We evaluated publication bias for outcomes that included more than ten studies.

Results
Searches yielded 13 918 results, with an additional 48 records identified through cross-referencing and expert recommendation (Fig. 1). After removal of duplicates (n = 1921), 12 045 articles were screened. Reviewers identified 273 articles that were potentially relevant and thus included in full text screening. Thirty-six articles representing 21 randomised trials met criteria for inclusion in this systematic review and seven articles were classified as awaiting assessment because eligibility could not be adequately evaluated given available information (Aracena et al., 2011;Aracena et al., 2012;Akbarzadeh et al., 2016;Shirazi et al., 2016;Frith et al., 2017;Kahalili et al., 2019;Tran et al., 2019). The most common reasons for exclusion were studies that described an intervention that did not aim to improve MMH and studies that did not include a child outcome (Fig. 1). The 36 included articles represent data from 21 RCTs and 28 284 mother-child dyads.

Risk of bias and GRADE certainty of evidence
Few studies showed high risk of bias on two or more of the seven domains assessed in this review. While all included trials were RCTs, three trials did not describe how the randomisation sequence was generated leading to unclear risk of bias. Similarly, the method of allocation concealment was not well described in eight trials. Only ten trials described acceptable methods for blinding outcome assessors. Attrition and selective outcome reporting were common sources of bias that could compromise the validity of trials (Fig. 2). Certainty of evidence ranged from very low to high using the GRADE methodology. Downgrading was due to the high level of heterogeneity across studies (i.e. I 2 above 55%), lack of information on masking of outcome assessors and attrition (online Supplementary File 1).

Narrative synthesis and meta-analyses
A summary of the results from meta-analyses is provided in Table 3. Growth indicators: the earliest growth indicator, low birth weight, was reported in four publications representing three trials. Findings were inconclusive as one trial reported a lower prevalence of low birth weight in infants of mothers in the intervention v. control , while others found marginal (Le Rotheram-Borus et al., 2014b) or no difference in the prevalence of low birth weight between groups  (online Supplementary File 2). Standardised measures of weight-for-age and height-for-age were evaluated in five trials. Three trials reported weight-or height-for-age on a continuous scale Fuhr et al., 2019;Sikander et al., 2019). The observed effect of the intervention on greater height-for-age in the trial by Rahman and colleagues (2008) was nullified after adjusting for baseline covariates at 6-and 12-months. However, the pooled effect of three trials of the Thinking Healthy Program found a small effect of the intervention on greater height-for-age at 6 months (SMD = 0.13, 95% confidence interval (CI): 0.02-0.24; online Supplementary File 3). Two additional trials measured weight on a continuous scale (Mohd Rajeswari and SanjeevaReddy, 2020), and when combined with the three Thinking Health Program trials, we did not find an effect of these interventions on child weight (online Supplementary File 4).
Several publications transformed height-for-age and weightfor-age into a binary variable indicating whether a child was stunted or underweight Aracena et al., 2009;Le Roux et al., 2013;Rotheram-Borus et al., 2014a;Tomlinson et al., 2015). Le  found that infants in the home visit intervention group were less likely to be stunted at 6-months, but found no between-group differences for underweight. Tomlinson and colleagues found that infants of depressed mothers in the intervention group were comparable to infants of non-depressed mothers under intervention and control conditions in terms of height-for-age; whereas, infants of depressed mothers under control conditions had lower height-for-age at 6-months. Weight-for-age did not differ by condition or maternal depression . At 18-months, there was no difference in the odds of stunting between intervention conditions among children of mothers with elevated symptoms of antenatal depression, yet the odds of being underweight were greater under control conditions . In the same trial, weight-for-height findings were complex: children of depressed mothers under intervention conditions were at WHO recommended weight-for-height scores (i.e. weight-for-height z-score = 0), but children of non-depressed mothers (intervention and control conditions) and children of depressed mothers under control conditions, were above WHO recommended weight-for-height scores (i.e. weight-for-height z-score > 0). The authors suggest that these findings can be explained by the intervention children being taller and less likely to be stunted, whereas children of depressed mothers under control conditions were shorter and similar in weight to children of depressed mothers under intervention conditions and children of non-depressed mothers under intervention and control conditions . A separate trial that identified a main effect of the intervention on the odds of not being underweight (odds ratio (OR) = 1.08, 95% CI: 1.01-1.16), but no intervention effects on stunting from birth to 12-months (OR = 0.99, 95% CI: 0.90-1.08) . Meta-analyses of categorical growth indicators did not find evidence of pooled intervention effects for being underweight, being stunted, and severe acute malnutritionweight-for-height (online Supplementary Files 5-7). Child health status: newborn health status was reported in seven trials and was operationalised as a function of growth and development indicators , infant/foetal complications (Rajeswari and SanjeevaReddy, 2020), incidence of illness Gureje et al., 2019), Apgar score Rajeswari and SanjeevaReddy, 2020), duration of hospitalisation  or neonatal mortality . Heterogeneity in outcome definitions precluded meta-analysis of child health status, but independent studies reported positive intervention effects on Apgar scores and neonatal mortality . In contrast, one study found that psychological intervention was associated with more hospital and NICU days, mixed findings related to postpartum complications (Rajeswari and SanjeevaReddy, 2020), and no effect of interventions on the incidence of child illness Carvalho et al., 2009;Gureje et al., 2019).   Breastfeeding: ten trials included breastfeeding as an outcome. Results of the meta-analysis (n = 4749) including data across ten comparisons indicated a sizeable overall impact in favour of intervention with moderate certainty according to the GRADE assessment: RR of 1.39, 95% CI: 1.13-1.71, NNT = 22.00, 95% CI: 15.00-40.90) Rahman et al., 2008;Le Roux et al., 2013;Rotheram-Borus et al., 2014a;Zhao et al., 2017;Fuhr et al., 2019;Gureje et al., 2019;Nabulsi et al., 2019;Sikander et al., 2019;Zhao et al., 2020) (Fig. 3). Heterogeneity was significant (I 2 = 61%) indicating substantial variation between interventions in their impacts on the outcome. Sub-group analyses revealed slightly larger effect sizes for integrated MMH interventions compared to focused MMH interventions, however uncertainty was high in these subgroups.
Maternal-child relationship outcomes: one trial focusing on the mother-child relationship found more secure attachment of infants of mothers under the intervention relative to the control conditions (74 v. 63%), which was driven by a higher probability of avoidant attachment in control infants (19 v. 11%) . In contrast, results from another trial found the proportion of infants with 'normal bonding' similar under intervention (98%) and control (98.9%) conditions .
Developmental outcomes: seven publications representing four trials evaluated one or more of the following domains of child development: cognitive development, language development, socio-emotional development, motor development, physical development, aggressive and prosocial behaviour and executive functioning. When evaluating development as one broad outcome, there were no differences between infants of mothers under the intervention relative to the control conditions in the short-and long-term Rotheram-Borus et al., 2014a;Rotheram-Borus et al., 2014b;Maselko et al., 2015). Results focusing on specific domains of child development were mixed Maselko et al., 2015;Murray et al., 2015). Cognitive and psychomotor developments were the only indicators measured in more than one study. We did not observe an impact of MMH interventions on cognitive development (3 trials, 1256 participants, SMD = 0.07, 95% CI: −0.04 to 0.18, I 2 = 0%; online Supplementary File 8) Murray et al., 2015;Tomlinson et al., 2018). Similarly, there was no effect of MMH interventions on psychomotor development (2 trials, 496 participants, SMD = 0.05, 95% CI: −0.13 to 0.23, I 2 = 0%; online Supplementary File 9) Le Roux et al., 2013).

Discussion
The aim of this systematic review and meta-analysis was to summarise existing experimental knowledge regarding the impact of MMH interventions on child-related outcomes. We identified 21 RCTs reporting on more than 28 000 participants. All trials focused on common mental disorders and most were conducted in middle-income countries.
The most commonly included outcome across these trials was exclusive breastfeeding. A recent meta-analysis found breastfeeding to be protective against child infections and malocclusion, associated with higher intelligence, and probable reductions in overweight and diabetes (Victora et al., 2016). Nevertheless, only 37% of children under 6-months are exclusively breastfed in LMICs (Victora et al., 2016). In our study, meta-analysis of ten comparisons with a combined number of 4749 women showed that with intervention 39% more children are exclusively breastfed than under control conditions. Given the varied nature of the interventions, it is challenging to single out the unique influence of the MMH components on improved rates of exclusive breastfeeding. However, one broad observation supports the contribution that MMH components can make in improving rates of exclusive breastfeeding. Future studies can be improved in two ways to clarify the impact of mental health components on exclusive breastfeeding. First, trials could be designed specifically so that mediation analyses can be conducted to assess whether improvements in MMH are in turn associated with exclusive breastfeeding. Second, head-tohead comparisons of interventions with and without a mental health component would be helpful to estimate the additional contribution of MMH components in integrated interventions.
Meta-analyses on other outcomes did not identify sizeable benefits of intervention and there was high heterogeneity between studies. These meta-analyses were limited by fewer available publications relative to the exclusive breastfeeding meta-analysis and should be interpreted with caution. There was a significant pooled effect of intervention on child height, but the effect size was small and only incorporated findings from three trials. There were trends favouring intervention for cognitive and psychomotor development, low birth weight, weight-for-age and height-for-age, but these did not reach statistical significance. It is possible that MMH interventions may have impacts on particular development domains, but not on broad indicators of child development. Similarly, MMH interventions may have impacts on particular growth indicators at specific developmental stages.
Before discussing implications of this systematic review and meta-analysis, we note the strengths and limitations of the existing literature. Overall, few trials included in this systematic review showed high risk of bias. Attrition and lack of masking were the greatest sources of potential bias. We presented conservative intention-to-treat analyses, but attrition introduced significant uncertainty in estimates. A substantive limitation to the generalisability of this review is that all interventions were focused on common mental disorder. It would be helpful for future studies to also evaluate whether interventions for other mental health outcomes (e.g. psychosis) are associated with improvements in child-related outcomes. Additionally, only one trial was conducted in a low-income country. Scaling of interventions may be particularly challenging in such settings, so further studies assessing impacts in low-income countries would be useful. Finally, there was substantial variation in how outcomes were defined and assessed, which limited the possibility to conduct meta-analyses for some outcomes.
Results from this review should be considered in light of several limitations in the review process. First, we included trials with diverse populations, who may respond differently to MMH interventions. Second, we included child outcomes that were reported across different studies, but did not prespecify primary v. secondary outcome measures, increasing the risk for selective reporting. However, we attempted to report on all available outcomes, without focusing only on those that were included in statistical re-analysis. The data used for our meta-analysis were primarily extracted from unadjusted results (means, standard deviations for continuous outcomes; n, percentage for binary outcomes), which in few instances resulted in marginally different measures of associations compared to adjusted models reported in the original trial publications. However, restriction of our searches to RCTs should reduce concerns of confounding and selection bias and thus these differences in outcome-specific inferences are not expected to result in substantial bias in our meta-analyses. Third, we did not specify sub-group analyses a priori as we were not sure which different intervention types had been studied and our review protocol was not pre-registered. While we aimed to report on the complete set of studies, outcomes and interventions that met our eligibility criteria, it is possible that not having published the study protocol prior to conducting the review may have introduced meta-bias. It is also possible that due to publication bias, our review does not reflect a fully representative synthesis of the evidence on the effect of MMH interventions on child development outcomes (Bender et al., 2018). To mitigate this potential for publication bias, we searched eight non-academic databases to include unpublished literature meeting our eligibility criteria.
Notwithstanding these limitations, results of this systematic review and meta-analysis are promising and have implications for policy and practice. We identified a sizeable number of RCTs that evaluated the impact of MMH interventions on childrelated outcomes in LMICs. Whereas impacts of these interventions on most child outcomes were uncertain, we identified a promising sizeable impact of MMH interventions on rates of exclusive breastfeeding, an outcome of vital public health importance globally. Evidence from this review further supports the importance of improving MMH, which has similarly been recommended by the WHO, as a strategy to further the critical effort to improve child health in LMICs.
Supplementary material. The supplementary material for this article can be found at https://doi.org/10.1017/S2045796020000864 Data. Data extracted from included studies for the narrative review and meta-analysis are available online: https://osf.io/qwdet/.