The study was a cluster-randomized controlled trial implemented in the catchment areas of 5 health facilities in Choma and Pemba districts, Southern Province, Zambia.

Households were randomly selected in 3 stages of sampling. In the first stage, 5 rural health centers (RHCs) were purposefully selected. These RHCs were part of a previous research study conducted by the study team [22]. In the second stage, 6 health zones were randomly selected from each of the 5 health centers. In the third stage, villages were randomly selected with probability proportional to size to reach the target sample of 18 eligible caregiver–child dyads in each health zone. The aim was to enroll 524 dyads in the study at baseline. Villages with fewer than 2 eligible children were excluded to ensure a sufficient number of potential participants in group meetings. In selected villages, all households were screened for eligibility. Eligible households were provided with study information as part of the informed consent procedure and decided whether to enroll at that time.

To be eligible for the study, a household had to have a child between 6 and 12 months of age at the time of enrollment. Caregivers reported their child’s birthdate during eligibility screening, and the reported date was confirmed by reviewing the child’s health card when available; 96% of households had the child health card for review. Caregivers younger than 15 years of age were excluded. All caregivers provided informed consent prior to study initiation. The study was approved by the Institutional Review Board at Boston University (protocol number H-32726) and by the ethics board at ERES Converge in Zambia (protocol number 2013-Dec-010) prior to the enrollment of participants.

The initial target sample size of 524 caregiver–child dyads was powered to detect a 0.5–standard deviation (SD) difference in standardized child development scores with 90% power, assuming 30 clusters of equal size, 5% loss to follow-up, and an intracluster correlation coefficient of 0.1.

Health zones (clusters) were randomized prior to baseline enrollment with equal probability to either the intervention or the control group. Prior to randomization, health zones were matched in pairs within RHC catchment areas based on distance to the health center and village population. Within each matched pair, 1 health zone was randomly assigned to the intervention group. Group assignment was masked from all assessors, but masking of participants was not possible.

The original study was funded for 1 year, and all study activities stopped at the end of that year, in November 2015. The study team secured funding for a second-year extension at the end of 2015, and study activities resumed in March 2016, at which time households were also re-consented to participate.

During the original 1-year study, intervention households received 2 services. First, they received a fortnightly visit by a child development agent (CDA), a community-based health worker employed full time by the project. During home visits, CDAs screened and referred children for infections and acute malnutrition, and encouraged caregivers to use routine child health services. Second, caregivers were invited to attend fortnightly parenting group meetings, where they were taught a diverse curriculum that included content on cognitive stimulation and play practices, child nutrition and cooking practices, and self-care for good mental health. During the year 2 study extension, the household visit component of the intervention was dropped, while facilitation of the fortnightly parenting group meetings continued. This change was motivated by the findings from the year 1 assessment, which suggested that the parenting groups were the primary driver of observed positive behavior change [21].

CDAs facilitated the organization of fortnightly parenting group meetings in intervention communities. Each health zone had 2 CDAs who each covered half of its communities. CDAs were selected through consultation with communities, and all had previous experience providing community-based health services, some as part of the formal health system. Prior to the start of the study, CDAs were trained on how to support group meetings. A local child development curriculum, containing age-appropriate activities covering health, nutrition, and early stimulation activities, as well as content related to caregiver mental health, was developed for these meetings. The curriculum was newly developed for the local context by the research team and included elements from existing child development programs, including the Care for Child Development package [23] and the Essential Package from Care International [24]. Content from these existing packages was adapted using an interactive theater approach. The curriculum training manual is provided in S1 Text. During the study period, CDAs were trained on the curriculum 3 meeting rounds at a time every 6 weeks. Groups were also encouraged to meet during intervening weeks without a formal curriculum. Meetings were held at multiple sites within each health zone. Meeting locations were chosen by group members to minimize travel distances. Each group meeting was run by a local “head mother” who was selected by the members of the group. CDAs met with head mothers before each round of meetings and provided them with training and resources according to the planned curriculum for that round. CDAs did not regularly attend the meetings themselves. Each meeting focused on a different topic; the topics included child nutrition, forms of play, cognitive stimulation, and language activities. All female caregivers in study communities with children under 5 years of age were invited to attend meetings. Caregivers were encouraged to bring children to group meetings, as aspects of the curriculum involved interactions with children. Based on consultation with local community leaders, male caregivers were not invited to attend group meetings.

Household survey data were collected at 4 time points: baseline (August/September 2014), year 1 follow-up (October/November 2015), re-consent (March 2016), and year 2 follow-up (November/December 2016). In this paper, we focus on data collected at baseline, re-consent, and the year 2 follow-up; a full analysis of data from the year 1 follow-up has been published previously [21]. Baseline, year 1 follow-up, and re-consent data collection was conducted at children’s homes. To ensure a controlled environment for the administration of the Bayley Scales of Infant and Toddler Development–Third Edition (BSID-III) at the year 2 follow-up, caregivers were invited to bring their child to the nearest health center, where the assessment was conducted one child at a time in a private room.

Primary outcomes were children’s stunting (defined as height-for-age z-score [HAZ] < −2) and neurocognitive development. Children’s height and weight were measured by trained assessors at baseline and the year 2 follow-up using standard anthropometric assessment kits. Length of all children under 2 years of age was collected in a lying position; height of older children was measured in a standing position. Length and height boards were rented from the Zambia National Food and Nutrition Commission (NFNC). Child weight was measured using digital scales also rented from the NFNC. Study interviewers were carefully trained for the anthropometric assessment, and a member of the NFNC attended the training to ensure correct use of the boards and scales. Anthropometric data were converted to z-scores using WHO child growth standards [25]. Child neurocognitive development was assessed using the BSID-III, which has previously been used and validated in Zambia [26]. A team of assessors attended a 2-week training on the BSID-III led by an accredited trainer prior to the start of data collection. Following the training, assessors spent 1 week pilot testing assessment procedures in the field. The BSID-III assesses 5 domains of development: cognition, language, motor, adaptive behavior, and social-emotional. Composite scores were determined for each domain using Bayley-III Scoring Assistant software [27] according to a 3-step procedure: first, raw scores were established by summing the number of items successfully completed for each sub-domain; second, scaled scores were constructed for each sub-domain by age-standardizing raw scores using the BSID-III reference population and rescaling to a range of 1 to 19, a mean of 10, and a standard deviation of 3; and third, composite scores were constructed by summing sub-domain scaled scores within domains and rescaling to a range of 40 to 160, a mean of 100, and standard deviation of 15. In order to show Cohen’s d estimates, composite scores were then converted to z-scores by standardizing within the study population.

Other outcome measures collected included HAZ, caregiver–child interactions, caregiver mental health, and child diet. Data on caregiver–child interactions were collected at baseline, year 1 follow-up, re-consent, and year 2 follow-up, using the 6-item Multiple Indicator Cluster Survey (MICS) module [28]. For child diet, a diet diversity score was constructed, using year 2 data, based on the number of food groups children had consumed the previous day, as per the method described in Steyn et al. [29]. Caregiver mental health was assessed using the 20-item WHO Self Reporting Questionnaire (SRQ) [30]. Caregiver depression was defined as reporting 7 or more symptoms on the SRQ, as per standard procedure [31]. Household demographics and asset information were collected at baseline. Caregivers reported on ongoing attendance at group meetings in the year 2 follow-up survey.

First, we compare baseline characteristics of the intervention and control groups, and present data on attrition by treatment arm. We then estimate the impact of the intervention on the primary outcomes of interest using intention-to-treat analysis. For continuous outcome variables, ordinary least squares regression models were fit to estimate unadjusted and adjusted impacts (β estimates). For dichotomous outcome variables, logistic regression models were fit to estimate unadjusted and adjusted odds ratios (ORs). Unadjusted models included controls for the cluster matching variables (population size and distance from the nearest health center) and, when available, the outcome variable measured at baseline. Adjusted models included a set of controls selected according to the following procedure: first, the outcome measured at the year 2 follow-up was regressed on the outcome measured at baseline, when available, and a set of demographic characteristics measured at baseline following a backward stepwise selection process with a drop threshold set at p = 0.2. The baseline demographic variable set included child HAZ, child weight-for-age z-score, child mid-upper arm circumference, child age, child sex, caregiver age, caregiver’s education, father’s education, and child motor skills according to the CREDI Scale [32]. Interviewer fixed effects were also included in the stepwise selection models. The resulting subset of variables from each model was included in the adjusted model for each outcome. For all models, standard errors were adjusted to account for clustering.

As a robustness check on the primary results, we conducted a per-protocol analysis, stratifying the treatment group according to household attendance at parenting groups. We also present data on caregiver–child interactions at each round of data collection, including the re-consent visit after a 5-month interruption in the delivery of the intervention. Finally, we estimate the impact of the intervention on other outcomes of interest, including caregiver mental health and child diet diversity. Cluster-robust standard errors were used to account for within cluster correlation. All analyses were conducted using Stata statistical software [33]. This trial was registered on ClinicalTrials.gov prior to baseline data collection (NCT02234726). Data were deposited in the Dryad repository: https://doi.org/10.5061/dryad.3340hc4 [34].

Thirty clusters were randomly assigned to intervention (15 clusters, 268 caregiver–child dyads) or control (15 clusters, 258 caregiver–child dyads). Overall, 5% of eligible households approached at enrollment refused to participate in the study. During the first year of the study, 48 dyads in the intervention group and 43 dyads in the control group were lost to follow-up. After a 5-month interruption in the study intervention, 209 intervention dyads and 201 control dyads were re-consented to participate in the second year of the study. At the end of the second year (last visit December 23, 2016), 195 caregiver–child dyads (73%) remained in the intervention group, while 182 dyads (71%) remained in the control group (Fig 1). Baseline characteristics of enrolled caregivers and children were similar in the intervention and control groups (Table 1). Nearly all (97%) caregivers in the study population were the child’s mother.

Children were on average 8 months old at the start of the study, with a relatively uniform distribution within the eligible age range (6 to 12 months). Height was on average well below the international reference median, with a mean HAZ of −1.5 at baseline. More than one-third of children were stunted at baseline. Study caregivers were on average 27 years old at baseline, and slightly fewer than half had completed primary school.

Controlling for a set of baseline characteristics (Table 2), the intervention significantly reduced the odds of stunting (OR 0.45 [95% CI 0.22 to 0.92]; p = 0.028). Probability density functions for HAZ at the year 2 follow-up in the intervention and control groups are presented in Fig 2. Corresponding cumulative density functions are provided in S1 Fig. The intervention had a significant positive impact on child language (β 0.14 [95% CI 0.01 to 0.27]; p = 0.039). There was no impact on cognition (β 0.11 [95% CI −0.06 to 0.29]; p = 0.196), motor skills (β −0.01 [95% CI −0.25 to 0.24]; p = 0.964), adaptive behavior (β 0.21 [95% CI −0.03 to 0.44]; p = 0.088), or social-emotional development (β 0.20 [95% CI −0.04 to 0.44]; p = 0.098). We found no evidence to suggest that physical growth mediates these impacts (S1 Table). Raw and scaled BSID-III subtest scores are provided in S2 Table.

In S2 Fig, we show reported attendance at parenting group meetings. In total, there were 27 parenting groups in intervention communities, and on average 6 caregivers attended each meeting. A majority of caregivers reported attending parenting group meetings either 2 to 3 (22%) or 4 or more times (35%) per month, indicating that many groups took the initiative to meet on their own in the time between fortnightly planned program meetings. Around one-third of caregivers in the intervention group reported that by the end of the study period they did not attend meetings. In Table 3, we present a per-protocol analysis. Treatment effects were significant for stunting, language, and adaptive behavior among compliers, i.e., those who attended 2 or more group meetings per month; no significant impacts were found among non-compliers.

At baseline, no differences were found in reports of caregiver–child interactions (Table 4). The intervention had a positive impact on reported interactions at year 1 follow-up (0.70 SD [95% CI 0.51 to 0.89]; p < 0.001). At the re-consent visit, at the end of a 5-month interruption in the delivery of the intervention, reported interactions remained significantly higher in the intervention group (0.71 SD [95% CI 0.52 to 0.90]; p < 0.001). Significant but somewhat smaller differences were found at year 2 follow-up (0.35 SD [95% CI 0.12 to 0.59]; p = 0.005). Caregiver–child interactions at each survey round are summarized in S3 Table.

Caregivers in the intervention group had slightly lower odds of depression (OR 0.63 [95% CI 0.31 to 1.26]; p = 0.193) at the year 2 follow-up, though these results were nonsignificant (Table 5). The intervention did not significantly impact HAZ (0.15 SD [95% CI −0.12 to 0.41]; p = 0.266) or child diet diversity score (0.20 [95% CI −0.13 to 0.53]; p = 0.218). Child diet data are summarized in S3 Table.