Effects of women’s groups practising participatory learning and action on preventive and care-seeking behaviours to reduce neonatal mortality: A meta-analysis of cluster-randomised trials

Background: The World Health Organization recommends participatory learning and action (PLA) in women’s groups to improve maternal and newborn health, particularly in rural settings with low access to health services. There have been calls to understand the pathways through which this community intervention may affect neonatal mortality. We examined the effect of women’s groups on key antenatal, delivery, and postnatal behaviours in order to understand pathways to mortality reduction. Methods and findings: We conducted a meta-analysis using data from 7 cluster-randomised controlled trials that took place between 2001 and 2012 in rural India (2 trials), urban India (1 trial), rural Bangladesh (2 trials), rural Nepal (1 trial), and rural Malawi (1 trial), with the number of participants ranging between 6,125 and 29,901 live births. Behavioural outcomes included appropriate antenatal care, facility delivery, use of a safe delivery kit, hand washing by the birth attendant prior to delivery, use of a sterilised instrument to cut the umbilical cord, immediate wrapping of the newborn after delivery, delayed bathing of the newborn, early initiation of breastfeeding, and exclusive breastfeeding. We used 2-stage meta-analysis techniques to estimate the effect of the women’s group intervention on behavioural outcomes. In the first stage, we used random effects models with individual patient data to assess the effect of groups on outcomes separately for the different trials. In the second stage of the meta-analysis, random effects models were applied using summary-level estimates calculated in the first stage of the analysis. To determine whether behaviour change was related to group attendance, we used random effects models to assess associations between outcomes and the following categories of group attendance and allocation: women attending a group and allocated to the intervention arm; women not attending a group but allocated to the intervention arm; and women allocated to the control arm. Overall, women’s groups practising PLA improved behaviours during and after home deliveries, including the use of safe delivery kits (odds ratio [OR] 2.92, 95% CI 2.02–4.22; I2= 63.7%, 95% CI 4.4%–86.2%), use of a sterile blade to cut the umbilical cord (1.88, 1.25–2.82; 67.6%, 16.1%–87.5%), birth attendant washing hands prior to delivery (1.87, 1.19–2.95; 79%, 53.8%–90.4%), delayed bathing of the newborn for at least 24 hours (1.47, 1.09–1.99; 68.0%, 29.2%–85.6%), and wrapping the newborn within 10 minutes of delivery (1.27, 1.02–1.60; 0.0%, 0%–79.2%). Effects were partly dependent on the proportion of pregnant women attending groups. We did not find evidence of effects on uptake of antenatal care (OR 1.03, 95% CI 0.77–1.38; I2= 86.3%, 95% CI 73.8%–92.8%), facility delivery (1.02, 0.93–1.12; 21.4%, 0%–65.8%), initiating breastfeeding within 1 hour (1.08, 0.85–1.39; 76.6%, 50.9%–88.8%), or exclusive breastfeeding for 6 weeks after delivery (1.18, 0.93–1.48; 72.9%, 37.8%–88.2%). The main limitation of our analysis is the high degree of heterogeneity for effects on most behaviours, possibly due to the limited number of trials involving women’s groups and context-specific effects. Conclusions: This meta-analysis suggests that women’s groups practising PLA improve key behaviours on the pathway to neonatal mortality, with the strongest evidence for home care behaviours and practices during home deliveries. A lack of consistency in improved behaviours across all trials may reflect differences in local priorities, capabilities, and the responsiveness of health services. Future research could address the mechanisms behind how PLA improves survival, in order to adapt this method to improve maternal and newborn health in different contexts, as well as improve other outcomes across the continuum of care for women, children, and adolescents. Ethical approval for the trials that collected the data for this study came from the UCL Great Ormond Street Institute of Child Health and Great Ormond Street Hospital for Children (UK) and in-country research ethics committees, as previously detailed [7]. We did a meta-analysis of trials of women’s groups practising PLA. Our search strategy and inclusion criteria were similar to those of a previous systematic review and meta-analysis. Briefly, we searched PubMed, Embase, Cochrane Library, CINAHL, African Index Medicus, Web of Science, the WHO Reproductive Health Library, and the Science Citation Index for studies published from the databases’ inception dates until March 1, 2017, with no language restrictions. Search terms included a combination of ‘community mobilisation’, ‘community participation’, ‘participatory learning and action’, ‘women’s groups’, and ‘women’. We also sought unpublished data from researchers known to be active in this area. Studies were included if they were randomised controlled trials, participants were women aged 15–49 years, and the trial tested a PLA cycle with women’s groups and reported information on at least 1 of our chosen outcomes [7]. Six of the 7 studies in the previous review met our inclusion criteria, as did 1 additional study from rural India [13]. In total, our analysis included 7 trials that took place between 2001 and 2012 within socio-economically disadvantaged communities in 4 countries, including rural communities in Bangladesh, Malawi, and Nepal, and rural and urban communities in India [7,10–13,17–19]. We used individual-level data collected during these 7 cluster-randomised controlled trials. Table 1 describes the characteristics of each study, including the number of participants. Two of the trials used a 2-by-2 factorial design. The first Bangladesh trial used a factorial design to assess the effects of the women’s group intervention and of a traditional birth attendant (TBA) training intervention [11]. There was no evidence of interaction between these 2 interventions, so we included data collected from all study participants [11]. The trial in Malawi used a factorial design to assess both the women’s group intervention and an infant feeding intervention. Because there was significant interaction between the 2 interventions and the infant feeding intervention had an independent effect on neonatal mortality, we did not include participants in the infant feeding arm in this analysis [20]. 1Published estimate comparing women’s group intervention to control group adjusting for covariates, unless otherwise specified. 2This number may differ from the number reported in the mortality estimate for the main trial paper as it includes liveborn infants with information collected as part of the survey questionnaire only. 3This number may differ from the number reported in the mortality estimate for the main trial paper as it includes pregnancies with information collected as part of the survey questionnaire only. 4Bangladesh 2005–2007 trial data used in this analysis include both women’s groups and traditional birth attendant training intervention and control areas. 5The Malawi trial was a 2-by-2 factorial cluster-randomised controlled trial of a women’s group intervention and an infant feeding programme. Results are from the women’s group intervention and control arms. OR, odds ratio; RR, risk ratio. We also included 2 studies that took place in the same geographical region of Bangladesh. The initial Bangladesh trial did not find evidence of a reduction in neonatal mortality for the women’s group intervention. This may have been due to very low coverage; only 3% of women reported attending women’s groups. The objective of the second trial was therefore to determine whether scaling up the coverage of women’s groups in the same geographical area would have an effect on neonatal mortality. In all studies except the trials in Nepal and Malawi, the data collection systems involved a female, community-based key informant who reported births and deaths in her area, which covered a population ranging from 250 to 800 households. For the trials in Nepal and Malawi, the key informant identified women in pregnancy. This key informant met with a trained interviewer once a month. The interviewer verified the informant’s reports and paid her an incentive for each correct identification. In the Malawi trial, cluster enumerators, who were similar to key informants, were paid a monthly salary. Four to 6 weeks after delivery, the interviewer visited the home where a birth or death had been identified and collected information on the mother’s and family’s sociodemographic characteristics, as well as events in the antenatal, delivery, and postnatal periods using a structured questionnaire [9–12,17,19,20]. In the event of a maternal death, an interviewer or supervisor conducted a verbal autopsy with a relative or close friend [9,10,19]. We selected outcomes representing a variety of important behavioural indicators in the antenatal, delivery, and postnatal periods, including the following: appropriate antenatal care, facility delivery, use of a safe delivery kit, hand washing by the birth attendant prior to delivery, use of a sterilised instrument to cut the umbilical cord, immediate wrapping of the newborn after delivery, delayed bathing of the newborn, immediate initiation of breastfeeding, and exclusive breastfeeding for the first 6 weeks after delivery. A safe delivery kit was normally available at low cost and typically included the following, at a minimum: soap, a clean string, a razor blade, and a plastic sheet [21]. Information collected in the different surveillance systems did not allow us to understand whether clean delivery practices were used independent of kit use. Although the Malawi trial collected data on clean delivery practices including hand washing by the birth attendant and use of a sterilised blade to cut the cord, the Ministry of Health’s position was to promote facility deliveries, and it was not acceptable for the study’s women’s groups to discuss clean home delivery practices or TBA training. Table 2 lists and defines the outcomes used in the analysis for each trial. We assessed the quality of evidence for each outcome using Grading of Recommendations Assessment, Development and Evaluation (GRADE) criteria, and these results can be found in S1 Table [22]. ‘Yes’ indicates information was collected for this outcome. ‘No’ indicates information was not collected for this outcome The previous meta-analysis assessing the effect of women’s groups on mortality outcomes found that the coverage of groups and the proportion of pregnant women participating in them were key to mortality reduction [7]. As part of an additional analysis to test whether coverage also affected the success of the intervention in improving the behaviours of interest, we created a variable indicating whether a woman attended group meetings. Women who were allocated to the intervention arm and reported attending at least 1 group meeting were considered women’s group attendees. We examined the prevalence of behaviours of interest either at baseline or, when this was not available, in the trial’s control arm. We also tabulated the prevalence of each behaviour by treatment arm and women’s group attendance (S2 Table). We then used 2-stage meta-analysis techniques to estimate the effect of the women’s group intervention on behavioural outcomes. In the first stage, we used individual records to assess the effect of women’s groups on the selected outcomes separately for the different trials. We used logistic regression with random effects (xtmelogit command) in Stata to account for the clustered nature of the data [23]. For trials that used a stratified or paired trial design, we adjusted for the different strata/pairs using a dummy variable that we treated as a fixed effect. These analyses also adjusted for any baseline differences between the intervention and control arms that existed before the inception of any intervention activities (S1 Box). Although the Nepal trial collected information on whether a woman had a facility delivery, due to very few women having a facility delivery and the paired nature of this cluster-randomised trial, these models would not converge. Likewise, for the urban Indian trial, the model assessing the effect of groups on exclusive breastfeeding failed to converge because only 0.9% of women reported a positive response for this outcome. For the second stage of the meta-analysis, we used random effects models via the metan command in Stata [23]. We chose to do a 2-stage meta-analysis rather than use summary estimates from the published trials, as not all trials reported all behaviours of interest for our analysis, and this method also allowed us to adjust for additional confounders that were not accounted for in the original trial. For trials with outcomes or covariates with greater than 10% missing data and significant differences in missingness between the control and intervention arms, we applied multiple imputation by chained equations (MICE) using the MI command in Stata, and assuming data were missing at random (MAR) [24]. Variables included in the MICE models were the outcome of interest, treatment arm, and covariates that were considered to be predictors of missingness [25,26]. We used a weighted sensitivity analysis using the selection model approach with multiple imputed data to test for modest departures from MAR [27–29]. In all instances, there was no evidence that missingness biased our main study findings. For each of the studies, we used logistic regression with random effects (xtmelogit command) in Stata to assess associations between outcomes and the following categories of group attendance and allocation: women attending a group and allocated to the intervention arm, women not attending a group but allocated to the intervention arm, and women allocated to the control arm. Stata’s postestimation command ‘test’ was used to determine if there were significant differences in the ORs between (1) women who attended groups in the intervention arm versus women in the control arm and (2) women who did not attend groups in the intervention arm versus women in the control arm. Models were adjusted using methods similar to those described for the first stage of the meta-analysis in addition to including covariates likely to influence health behaviours and women’s group attendance: parity, maternal age, and maternal educational attainment (S1 Box). We identified these covariates by discussing the intervention with principal investigators and reviewing process evaluations and qualitative research on the women’s group interventions [14–16]. Although the second rural Indian trial (the Jharkhand Odisha Health Action Research [JOHAR] trial), the trial in urban India, and the Malawi trial adjusted for baseline differences, we did not adjust for baseline differences in this analysis as it would not have been possible for women to attend group meetings before their inception [13]. We chose not to do a pooled analysis of the associations between health behaviours and women’s group attendance because we expected both the determinants of women’s group attendance and the types of behaviours discussed at the women’s groups to differ substantially across trials, meaning that a single summary effect would not capture this heterogeneity adequately. All analyses were conducted in Stata 14 [23].