Effect of a mass radio campaign on family behaviours and child survival in Burkina Faso: a repeated cross-sectional, cluster-randomised trial

Background: Media campaigns can potentially reach a large audience at relatively low cost but, to our knowledge, no randomised controlled trials have assessed their effect on a health outcome in a low-income country. We aimed to assess the effect of a radio campaign addressing family behaviours on all-cause post-neonatal under-5 child mortality in rural Burkina Faso. Methods: In this repeated cross-sectional, cluster randomised trial, clusters (distinct geographical areas in rural Burkina Faso with at least 40 000 inhabitants) were selected by Development Media International based on their high radio listenership (>60% of women listening to the radio in the past week) and minimum distances between radio stations to exclude population-level contamination. Clusters were randomly allocated to receive the intervention (a comprehensive radio campaign) or control group (no radio media campaign). Household surveys were performed at baseline (from December, 2011, to February, 2012), midline (in November, 2013, and after 20 months of campaigning), and endline (from November, 2014, to March, 2015, after 32 months of campaigning). Primary analyses were done on an intention-to-treat basis, based on cluster-level summaries and adjusted for imbalances between groups at baseline. The primary outcome was all-cause post-neonatal under-5 child mortality. The trial was designed to detect a 20% reduction in the primary outcome with a power of 80%. Routine data from health facilities were also analysed for evidence of changes in use and these data had high statistical power. The indicators measured were new antenatal care attendances, facility deliveries, and under-5 consultations. This trial is registered with ClinicalTrial.gov, number NCT01517230. Findings: The intervention ran from March, 2012, to January, 2015. 14 clusters were selected and randomly assigned to the intervention group (n=7) or the control group (n=7). The average number of villages included per cluster was 34 in the control group and 29 in the intervention group. 2269 (82%) of 2784 women in the intervention group reported recognising the campaign’s radio spots at endline. Post-neonatal under-5 child mortality decreased from 93·3 to 58·5 per 1000 livebirths in the control group and from 125·1 to 85·1 per 1000 livebirths in the intervention group. There was no evidence of an intervention effect (risk ratio 1·00, 95% CI 0·82–1·22; p>0·999). In the first year of the intervention, under-5 consultations increased from 68 681 to 83 022 in the control group and from 79 852 to 111 758 in the intervention group. The intervention effect using interrupted time-series analysis was 35% (95% CI 20–51; p<0·0001). New antenatal care attendances decreased from 13 129 to 12 997 in the control group and increased from 19 658 to 20 202 in the intervention group in the first year (intervention effect 6%, 95% CI 2–10; p=0·004). Deliveries in health facilities decreased from 10 598 to 10 533 in the control group and increased from 12 155 to 12 902 in the intervention group in the first year (intervention effect 7%, 95% CI 2–11; p=0·004). Interpretation: A comprehensive radio campaign had no detectable effect on child mortality. Substantial decreases in child mortality were observed in both groups over the intervention period, reducing our ability to detect an effect. This, nevertheless, represents the first randomised controlled trial to show that mass media alone can change health-seeking behaviours. Funding: Wellcome Trust and Planet Wheeler Foundation. This study was done using a repeated cross-sectional cluster randomised design by an independent team from the London School of Hygiene & Tropical Medicine and Centre Muraz in Burkina Faso. The intervention and evaluation design have been described previously.8, 9, 10, 11, 12 Briefly, women of reproductive age and caregivers of children younger than 5 years were the main targets of the campaign, which covered 17 behaviours along the continuum of care. Women were told the surveys were about their children's health, without any mention of the radio campaign, and they recorded their consent to participate in the survey on a Personal Digital Assistant. The study was approved by the ethics committees of the Ministry of Health of Burkina Faso and the London School of Hygiene & Tropical Medicine. Of 19 distinct geographical areas, 14, each centred around a community FM radio station, were selected by DMI based on their high listenership (above 60% of women listening to the radio in the past week) and minimum distances between radio stations to exclude population-level contamination. For evaluation purposes, clusters around each radio station were identified using the last national census to provide an evaluation population of about 40 000 inhabitants per cluster. We included villages located around the selected community radio station, with a good radio signal but limited access to television (and thus more likely to listen to the radio). We did this by excluding communities likely to be served by the electricity grid—ie, the towns from which the selected radio stations were broadcast, villages within 5 km of these towns, other villages with electricity or with a population larger than 5000 inhabitants (and likely to be a priority for the national electrification programme). Seven clusters were then randomly allocated to receive the intervention or control using pair-matched randomisation based on geography and radio listenership (figure 1). Specifically, we defined three radio listenership strata (61–70%, 71–80%, and >80%), and within each stratum we paired the areas geographically closest with each other, one of which was randomly assigned to receive the intervention. Randomisation was done by SS and SC, independently of DMI. The randomisation sequence was generated using computer-generated random numbers (Stata version 13). Because of time constraints, randomisation was done before the baseline survey. The nature of the intervention precluded formal masking of respondents and interviewers. The average number of villages included per cluster was 34 in the control group and 29 in the intervention group. In all clusters the government was the main health service provider and, with the exception of Kantchari (intervention cluster), a regional or district hospital was located in the town with the community radio station. The trial population also had access to primary health facilities in villages across each cluster. Pair-matched randomisation based on geography and radio penetration rate DMI’s radio campaign launched in March, 2012, and ended in January, 2015. A description of the theory of change and the Saturation+ methodology used to design and implement the campaign is provided elsewhere.12 Short spots, of 1 min duration, were broadcast approximately ten times per day, and 2 h, interactive long-format programmes were broadcast 5 days per week. All materials were produced in the predominant local languages spoken in each intervention cluster. The dramas were based on message briefs that DMI drew up for each target behaviour. The long-format programmes were followed by phone-ins to allow listeners to comment on the issues raised. Behaviours covered by spots changed weekly, while the long-format programme covered two behaviours a day and changed daily. Table 1 shows the campaign resources allocated to each of the target behaviours. Target behaviours and broadcasting intensity up to the month preceding the endline survey (October, 2014) During the trial period, no other radio campaigns related to child survival and of comparable intensity were broadcast in any of the clusters included in the trial. Various health programmes operated in similar numbers of clusters per group (appendix p 1). From 2010 to 2013, community case management for malaria, pneumonia, and diarrhoea was supported by the Catalytic Initiative to Save a Million Lives in one of the intervention clusters and one of the control clusters, although the independent evaluation of this rapid scale-up programme concluded that it did not result in changes in coverage or mortality.13 Cross-sectional household surveys were performed in all clusters at three time points: at baseline, from December, 2011, to February, 2012; at midline, in November, 2013, after 20 months of campaigning; and at endline, from November, 2014 to March, 2015, after 32 months of campaigning. At baseline and endline surveys, a census of villages selected for the survey was performed with GPS coordinates recorded. For all households with at least one woman aged 15–49 years, the household head was interviewed to collect socioeconomic data and all women aged 15–49 years were interviewed on their pregnancy history. At baseline, due to time and cost constraints, the census took place in a simple random sample of villages covering half of the population in each cluster (about 20 000 inhabitants) and pregnancy history data collection was truncated to cover the period from January, 2005, to the date of the interview. At endline, the census took place in all villages included in the trial and a full pregnancy history was recorded. At each survey, about 5000 mothers with at least one child younger than 5 years living with them were interviewed regarding their demographic characteristics, radio listenership, and family behaviours of relevance to child survival.9 To test recognition of the campaign at midline and endline, the two spots broadcast in the last 2 weeks of the previous month were played at the end of the interview and women were asked whether they had heard them on the radio. With respect to long format programmes, recognition was tested by referring to its title. At baseline and endline, mothers for the behavioural interviews were selected using systematic random sampling of all women interviewed about their pregnancy history. At midline, a two-stage sampling procedure was used.9 Before each survey, fieldworkers received 2 weeks’ training. At baseline and endline, 84 fieldworkers were deployed across clusters in teams of six fieldworkers. At midline, 56 fieldworkers were deployed in teams of four fieldworkers. Each team was managed by a supervisor. Interviews were performed using Trimble Juno SB Personal Digital Assistants using Pendragon forms software. Data were backed up twice a week by a team of seven data managers and checked for consistency and completeness. Re-interviews were requested in cases of missing or inconsistent responses (for 7% of pregnancy history interviews and 3% of behavioural interviews at endline). The trial was designed to detect a 20% reduction in the primary outcome (all-cause post-neonatal under-5 child mortality) with a power of 80%.8 We assumed a baseline mortality rate of 25 per 1000 per year, a coefficient of variation between clusters of 0·18, that mortality would decline in all clusters by 5% over the course of the study, and that the analysis would be based on cluster-level summaries with adjustment for pre-intervention mortality. Simulations indicated that, given a total of 14 clusters, a sample size of 7000 under-5 children per cluster would be required. The sample size of 5000 mothers was calculated assuming a design effect of 2 with a view to providing an absolute precision of within 10% or better for all behaviours. Routine health facility data from January, 2011, to February, 2016, were obtained from the Direction Générale des Etudes et des Statistiques Sanitaires of the Ministry of Health. For 78 primary health facilities located in trial clusters (41 in control clusters), monthly numbers were provided for: pregnant women attending for a first antenatal consultation, facility deliveries, and all-cause under-5 child consultations. The primary outcome was all-cause post-neonatal under-5 child mortality, the secondary outcome was all-cause under-5 child mortality, and intermediate outcomes included the coverage of the campaign (as measured by the proportion of mothers who reported listening to the campaign) and family behaviours targeted by the campaign as listed in table 1 (as measured by the proportion of mothers who reported a given behaviour during interviews and the number of attendances at primary health facilities). Primary analyses were performed on an intention-to-treat basis, and followed an analysis plan agreed in advance with the trial’s Independent Scientific Advisory Committee. All analyses were performed on cluster-level summary measures14 and adjusted for pre-intervention levels to control for imbalances between groups and improve precision. The matching procedure was ignored, as recommended for trials with fewer than ten clusters per group.15 All clusters were given equal weight in all analyses. All analyses were done with Stata (versions 13 and 14). The pre-intervention period was defined as the 2 years before the campaign, from March, 2010, to February, 2012. The post-intervention (March, 2012, to October, 2014) period was split into three periods: from March, 2012, to December, 2012 (the first 10 months of campaigning), January, 2013, to October, 2013 (next 10 months), and November, 2013, to October, 2014 (the 12 months preceding the start of the endline survey; figure 2). Full pregnancy history data collected at the endline survey were used to calculate both pre-intervention and post-intervention cluster-level mortality estimates. Cluster-level estimates of post-neonatal under-5 child mortality and under-5 child mortality were computed using the Demographic and Health Survey (DHS) synthetic cohort life-table approach. Missing months of birth (for 2% of livebirths across all pre-intervention and post-intervention periods) were randomly imputed according to the DHS method.16 This method relies on the construction of logical ranges for each date, which are refined in three steps, resulting in successively narrower ranges. In the final step, months of birth are randomly imputed within the final constrained logical range. Pre-intervention and post-intervention periods for mortality analysis An analysis of covariance was performed on a log-risk scale to estimate the risk ratio for the effect of the intervention adjusted for pre-intervention mortality. The Wild bootstrap test, recommended when there are few clusters,17 was used to test for evidence of an intervention effect and for evidence of effect modification by post-intervention period. We did a cluster-level difference-in-difference analysis to assess the change from baseline to follow-up survey (either at midline or endline) in self-reported behaviours. For each of the 17 target behaviours, cluster-level differences in prevalence from baseline to follow-up survey were calculated and regressed on the intervention status of clusters and the cluster-level baseline prevalence to account for regression to the mean. Wild bootstrap tests were done to test the null hypothesis of no intervention effect. No formal adjustment was made to account for multiple testing. Analyses of maternal and newborn related behaviours at midline and endline were restricted to pregnancies ending after June, 2012 (with at least 4 months of exposure to the campaign). At baseline, the mean post-neonatal under-5 mortality risk during the 2 years preceding the intervention was estimated at 112·3 per 1000 children in the intervention group versus 82·9 per 1000 children in the control group. Three covariates, expected to predict mortality, were particularly imbalanced between groups at baseline: the distance to the capital, Ouagadougou, as a proxy for general level of development (158 km in the control group vs 232 km in the intervention group); the median distance to the closest health facility (2·5 km vs 6·3 km, respectively); and the facility delivery prevalence (82% vs 56%, respectively). These covariates were combined using principal component analysis to produce a single cluster-level summary confounder score. After controlling for the confounder score, the pre-intervention mortality risk difference between groups estimated at baseline was reduced from 30·9 to 6·8 per 1000 children. To control for imbalance between groups, analyses of behaviour and mortality were adjusted for the confounder score. Three categories of radio ownership were defined: no radio in the compound (or household), radio in the compound but not in the household, and radio in the household. The Wild bootstrap test was used to test for evidence of effect modification by radio ownership. With respect to care-seeking behaviours, three categories of distance to the closest health facility were also defined (<2 km, 2–5 km, and >5 km) to look for evidence of effect modification by distance on service-dependent behaviours, using the same analysis as described above. In a first analysis, the absolute number of attendances was calculated by yearly period (March, 2011, to February, 2016) and by cluster. For each post-intervention period, the ratio of the absolute number of attendances over the absolute number in the year before the intervention was then calculated in each cluster, a mean ratio to baseline was then computed by group, and a Wild bootstrap test, adjusted for confounder score, was used to compare the mean ratio to baseline between groups. In addition, an interrupted time-series analysis was also done using mixed-effects Poisson regression of monthly counts of attendances per cluster, from January, 2011, to February, 2016. The model included fixed effects allowing for a long-term secular trend, for month of the year to account for seasonal variation, for intervention status of the cluster to account for systematic differences between groups at baseline, for confounder score, and for intervention effect by period, with cluster treated as a random effect.18 To obtain 95% CI and p values, we used bootstrap resampling (using the BCa method and 1000 bootstrap replications).19 This trial is registered with ClinicalTrial.gov, number {“type”:”clinical-trial”,”attrs”:{“text”:”NCT01517230″,”term_id”:”NCT01517230″}}NCT01517230. The funders of the study had no role in study design, in the collection, analysis, and interpretation of data, in the writing of the report, and in the decision to submit the paper for publication. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication.