Violence prevention accelerators for children and adolescents in South Africa: A path analysis using two pooled cohorts

Background The INSPIRE framework was developed by 10 global agencies as the first global package for preventing and responding to violence against children. The framework includes seven complementary strategies. Delivering all seven strategies is a challenge in resource-limited contexts. Consequently, governments are requesting additional evidence to inform which ‘accelerator’ provisions can simultaneously reduce multiple types of violence against children. Methods and findings We pooled data from two prospective South African adolescent cohorts including Young Carers (2010–2012) and Mzantsi Wakho (2014–2017). The combined sample size was 5,034 adolescents. Each cohort measured six self-reported violence outcomes (sexual abuse, transactional sexual exploitation, physical abuse, emotional abuse, community violence victimisation, and youth lawbreaking) and seven self-reported INSPIRE-aligned protective factors (positive parenting, parental monitoring and supervision, food security at home, basic economic security at home, free schooling, free school meals, and abuse response services). Associations between hypothesised protective factors and violence outcomes were estimated jointly in a sex-stratified multivariate path model, controlling for baseline outcomes and socio-demographics and correcting for multiple-hypothesis testing using the Benjamini-Hochberg procedure. We calculated adjusted probability estimates conditional on the presence of no, one, or all protective factors significantly associated with reduced odds of at least three forms of violence in the path model. Adjusted risk differences (ARDs) and adjusted risk ratios (ARRs) with 95% confidence intervals (CIs) were also calculated. The sample mean age was 13.54 years, and 56.62% were female. There was 4% loss to follow-up. Positive parenting, parental monitoring and supervision, and food security at home were each associated with lower odds of three or more violence outcomes (p < 0.05). For girls, the adjusted probability of violence outcomes was estimated to be lower if all three of these factors were present, as compared to none of them: sexual abuse, 5.38% and 1.64% (ARD: −3.74% points, 95% CI −5.31 to −2.16, p < 0.001); transactional sexual exploitation, 10.07% and 4.84% (ARD: −5.23% points, 95% CI −7.26 to −3.20, p < 0.001); physical abuse, 38.58% and 23.85% (ARD: −14.72% points, 95% CI −19.11 to −10.33, p < 0.001); emotional abuse, 25.39% and 12.98% (ARD: −12.41% points, 95% CI −16.00 to −8.83, p < 0.001); community violence victimisation, 36.25% and 28.37% (ARD: −7.87% points, 95% CI −11.98 to −3.76, p < 0.001); and youth lawbreaking, 18.90% and 11.61% (ARD: −7.30% points, 95% CI −10.50 to −4.09, p < 0.001). For boys, the adjusted probability of violence outcomes was also estimated to be lower if all three factors were present, as compared to none of them: sexual abuse, 2.39% to 1.80% (ARD: −0.59% points, 95% CI −2.24 to 1.05, p = 0.482); transactional sexual exploitation, 6.97% to 4.55% (ARD: −2.42% points, 95% CI −4.77 to −0.08, p = 0.043); physical abuse from 37.19% to 25.44% (ARD: −11.74% points, 95% CI −16.91 to −6.58, p < 0.001); emotional abuse from 23.72% to 10.72% (ARD: −- 13.00% points, 95% CI −17.04 to −8.95, p < 0.001); community violence victimisation from 41.28% to 35.41% (ARD: −5.87% points, 95% CI −10.98 to −0.75, p = 0.025); and youth lawbreaking from 22.44% to 14.98% (ARD −7.46% points, 95% CI −11.57 to −3.35, p < 0.001). Key limitations were risk of residual confounding and not having information on protective factors related to all seven INSPIRE strategies. Conclusion In this cohort study, we found that positive and supervisory caregiving and food security at home are associated with reduced risk of multiple forms of violence against children. The presence of all three of these factors may be linked to greater risk reduction as compared to the presence of one or none of these factors. Policies promoting action on positive and supervisory caregiving and food security at home are likely to support further efficiencies in the delivery of INSPIRE. We use secondary data with self-reported measures to examine how variation in protective factors between individuals is associated with multiple forms of violence. All of the protective factors examined are possible targets of INSPIRE interventions. Identifying which protective factors predict lower rates of violence is intended to inform efforts to identify and prioritise the most efficient subset of interventions. The study setting was South Africa—a country with high levels of government commitment to preventing violence against children but also facing implementation challenges shared by many low- and middle-income contexts: limited service delivery capacity, infrastructure, and financial resources. The study was reported according to the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) checklist for cohort studies (S1 Checklist) [9]. Study data were pooled from two large prospective cohorts (Young Carers and Mzantsi Wakho) spread across three South African Provinces: Eastern Cape, Mpumalanga, and Western Cape. The Young Carers and Mzantsi Wakho cohorts were designed to allow data merging, with shared investigators, measures, data collection procedures, and sampling in urban and rural sites. The Young Carers recruitment took place between 1 January 2010 and 7 December 2011, with follow-up between 5 January 2011 and 15 December 2012. Participants included N = 3,515 children and adolescents in Mpumalanga and Western Cape Provinces. In each province, census enumeration areas were randomly selected within one urban and one rural health district, and all households with a resident 10- to 17-year-old were recruited. The cohort had 97% uptake during recruitment and 97% retention at 12- to 18-month follow-up [10]. The Mzantsi Wakho (‘Our South Africa’) recruitment took place between 4 March 2014 and 25 September 2015, with follow-up between 10 November 2015 and 5 April 2017. Participants included N = 1,519 children and adolescents in the Eastern Cape Province, living with and without HIV. Across two urban and rural health districts, all children and adolescents (aged 10–19 years) who had ever initiated HIV care in 53 health facilities were recruited, alongside their closest child or adolescent neighbours. The cohort had 90% uptake and 94% retention at 12- to 18-month follow-up (with 2.4% mortality) [11]. For both cohorts, interviews took approximately 45–70 minutes and were conducted in a location of the adolescent’s choice. For Young Carers, interviews were administered using self-interviewing on paper forms, and for Mzantsi Wakho, interviews were administered using audio mobile-assisted self-interviewing on electronic tablets. All interviews were supported by trained local interviewers, and the level of assistance was adjusted according to the age and literacy level of participants. Questionnaires were available in the languages of the adolescent’s choice, including Xhosa, Swati, Tsonga, Pedi, and English. There was no financial payment for participation, but all participants received certificates and snacks, regardless of whether they completed interviews. All participants were asked whether they would be happy to be contacted or visited again for a follow-up at baseline. For those consenting, telephone numbers, addresses, or GPS location was collected for tracing. Ethical approvals were gained for Young Carers and Mzantsi Wakho from the University of Oxford (SSD/CUREC2/11-40; SSD/CUREC2/12-21) and University of Cape Town (REC REF: CSSR 389/2009; REC REF CSSR 2013/04) and provincial South African Departments of Health, Basic Education, and Social Development. In both cases, ethical approval included examination of predictors of violence with collected data. Both studies obtained written consent from all adolescents and their primary caregivers at both baseline and follow-up. Owing to varying levels of literacy, informed consent information was read aloud to potential participants. Confidentiality was maintained except when participants disclosed serious risk of harm to themselves or others. In such circumstances, safeguarding processes were followed. For those reporting current abuse, recent rape, or suicidality, immediate responses included support to access postexposure prophylaxis, pregnancy prevention, and child protection measures in conjunction with government services. Findings of the studies are reported back to communities and health facilities in research areas as part of ongoing local knowledge-sharing. Children and adolescents were involved in the conceptualisation, design, conduct, and dissemination of these two studies and of this paper. This took place through 11 years of Adolescent Advisory Group weekend workshops in urban and rural areas of South Africa, in which children and adolescents (10–20 years old) engaged in planning research studies, designing content and appearance of questionnaires, planning dissemination, and engaging with the South African National Ministry of Health who attended these workshops. In addition, Young Carers and Mzantsi Wakho questionnaires were piloted with 20 and 25 adolescents, respectively. Adolescents were from local areas involved in the respective studies. All comments were taken into consideration and errors amended. Input to questionnaire design in both studies was given by the South African National Departments of Health, Basic Education, and Social Development; the South African National AIDS Council; UNICEF; PEPFAR; USAID; and local NGOs. All questionnaires are available at www.youngcarers.org.za: Young Carers baseline [12] and follow-up [13] and Mzantsi Wakho baseline [14] and follow-up [15]. Six violence outcomes were identified in the data, including five measures of violence victimisation and one of violence perpetration. For the analysis, all six outcomes were coded by collapsing responses to multiple self-reported items into binary indicators of any experience. The six outcomes were (1) sexual abuse, measured as reporting either sexual assault and/or completed rape in the last year using items from the Sexual Victimization module of the Juvenile Victimization Questionnaire [16]; (2) transactional sexual exploitation, measured as reporting any lifetime receipt of money, drinks, clothes, cell phone airtime, a place to stay, lifts in a car/taxi, better marks at school, school fees, food, or anything else for having sex with someone (this question was drawn from the National Survey of Risk Behaviour Amongst Young South Africans [17]); (3) physical abuse, measured as reporting a caregiver or other adult to have either used a stick, belt, or other hard item to hit you or slapped, punched, hit, pinched, or pulled your ear/hair so that it hurt or left marks in the last year, as in UNICEF’s Psychosocial Vulnerability and Resilience Measures For National-Level Monitoring of Vulnerable Children [18]; (4) emotional abuse, defined as reporting a caregiver or other adult to have said they would call ghosts or evil spirits or harmful people, said you would be sent away or kicked out of the house, and/or called you dumb, lazy, or other names in the last year, as in the same UNICEF Psychosocial Vulnerability and Resilience Measures [18]; (5) community violence victimisation, defined as reporting having had things stolen in the last year, having ever been hit or attacked outside, or having ever seen someone stabbed or shot, as in the Child Exposure to Community Violence checklist [19] (these three items reflect the most common community traumas in South Africa, as identified by police statistics); (6) youth lawbreaking as reporting hanging around with kids that get in trouble, stealing at home, stealing things from places other than home, fighting a lot, and carrying a gun and/or knife for protection in the past 6 months, as in the Child Behaviour Checklist Delinquency subscale [20]. Items about gang affiliation and carrying of a knife and/or gun were added during questionnaire piloting. As data collection was initiated before the publication of the INSPIRE framework in 2016, we retrospectively identified seven protective factors that were aligned with its strategies. (1) We measured positive parenting/caregiving using a continuous sum of items 2, 13, 16, and 27 from the child form of the Alabama Parenting Questionnaire [21], which consider warmth and praise from a primary caregiver (range: 0–16). Given high rates of informal fostering in Southern Africa, we note that ‘parents’ were defined as any primary caregiver and there was no requirement for a biological relationship. (2) We measured parental/caregiver monitoring and supervision using a continuous sum of items 10, 17, and 19, from the child form of the Alabama Parenting Questionnaire, which include setting rules about coming home in evenings, and knowing who an adolescent is friends with (range: 0–12). (3) We measured food security at home as a binary indicator of 6 or 7 days in the past week with enough food in the home. (4) We measured basic economic security at home as a binary indicator of access to all six of the top socially perceived necessities for youth in the South African Social Attitudes Survey: a doctor when needed, school uniform, basic clothing, toiletries to wash, school equipment, and a pair of shoes [22]. (5) We measured free schooling as a binary indicator of enrolment in a no-fees school, or receipt of a fee exemption. (6) We measured free school meals as a binary indicator of daily free lunches or breakfasts at school [23]. Exposure to these protective factors was measured as consistent reporting at both baseline and follow-up, based on evidence that such factors need to be predictable and sustained in order to protect children and adolescents in high-risk settings [24, 25]. For this, continuous variables were combined additively. (8) We also describe the prevalence of adolescent engagement with response services for children who are victims of violence, measured as any social or justice service response at follow-up related to emotional, physical, or sexual abuse (Young Carers data) and sexual abuse (Mzantsi Wakho data). Because experiencing study outcomes of abuse was a condition for being asked about access to response services, we were unable to evaluate it as a protective factor. Control variables included 10 key sociodemographic variables and violence-associated factors: participant sex, age, HIV status, rural/urban household location, household size, informal/shack housing, maternal orphanhood, paternal orphanhood, and province of residence. Participant HIV status was assessed using self-report (Young Carers data) and clinical records (Mzantsi Wakho data). Baseline measures of self-reported sexual abuse, emotional abuse, physical abuse (all last year), and youth lawbreaking in the last 6 months were also included as covariates in regression models investigating associations between hypothesised protective factors and these forms of violence at follow-up. Analysis took place in seven steps in Stata 15, all stratified by sex. First, we compared baseline characteristics of participants retained and lost to follow-up. Second, we described baseline and follow-up characteristics of retained participants using count (percent) for binary and categorical variables and mean (standard deviation [SD]) for continuous variables. Third, we evaluated univariable associations between the six hypothesised protective factors and six binary violence outcomes. Fourth, we investigated multivariable associations between hypothesised protective factors and violence outcomes. For this, we used path analysis consisting of six single-outcome multivariable logistic regression models, each regressing one of the six violence outcomes at follow-up on the six hypothesised protective factors, controlling for outcome-specific baseline exposure to violence and for common sociodemographic factors. Missing data were handled using listwise deletion. Fifth, to account for risk of type I error from multiple-hypothesis testing, we adjusted estimated p-values using the Benjamini-Hochberg procedure specified with a false discovery rate of 5% [26]. The analysis had six regression models, each containing six predictors of scientific interest. Hence, a family of tests consisted of six values, one for each protective factor, across regression models. Sixth, hypothesised protective factors significantly associated (for either sex or both) with three or more violence outcomes were defined as protective factors that had potential to indicate development ‘accelerators’. Seventh, we estimated adjusted probabilities for experiencing each violence outcome under three scenarios: (1) experiencing no ‘accelerator’ protective factors, (2) experiencing each ‘accelerator’ protective factor, and (3) experiencing all ‘accelerator’ protective factors together. Positive parenting and parental monitoring and supervision were the only hypothesised protective factors modelled as continuous scales. For these two variables, adjusted probabilities were estimated at the sample mean value and at their maximum scores (32 for positive parenting and 24 for parental monitoring and supervision). Finally, adjusted risk ratios (ARRs) and adjusted risk differences (ARDs) were used to compare scenarios 2 and 3 relative to scenario 1. All confidence intervals (CIs) are given at 95% confidence level. Although the specific analysis reported in this study was not prespecified, it followed a prespecified methodological approach that has been developed to investigate factors associated with multiple SDG outcomes in observational data as part of the UKRI GCRF ‘Accelerate Hub’ [27]. This approach is laid out on Open Science Framework [28]. Multiple-outcome models that include correlation between the error terms of outcomes can be more accurate when modelling highly correlated outcomes with different sets of predictors per outcome [29]. As a sensitivity analysis, we calculated adjusted probabilities of experiencing each violence outcome from a multiple-outcome probit model that correlated the error terms of our six different regression models. This was done using the mvprobit command in Stata 15 set at 100 random draws [30]. As in our main analysis, each regression regressed one of the six violence outcomes at follow-up on the six hypothesised protective factors, controlling for outcome-specific baseline exposure to violence and for common sociodemographic factors.