Supporting adolescent girls to stay in school, reduce child marriage and reduce entry into sex work as HIV risk prevention in north Karnataka, India: protocol for a cluster randomised controlled trial

The overall goal of Samata is to improve the quality of life of adolescent girls from vulnerable and marginalized communities in Bijapur and Bagalkot districts of north Karnataka, by supporting entry into and retention of adolescent girls in secondary education, and by delaying age at marriage and entry into sex work. Specifically, the intervention aims to bring levels of these outcomes among low caste girls in-line with those of all girls in Karnataka. The project comprises a comprehensive, multi-level intervention that works with key stakeholders (low caste adolescent girls and their families; adolescent boys; village communities; schools; school development and management committees (SDMC); and policy makers and policy implementers), to change social norms regarding gender, child marriage and girls’ education, as well as to link low caste families to government schemes that provide scholarships, bicycles and other incentives to support girls’ retention in school. A summary of the main intervention activities are detailed in Table 1. The intervention will be implemented in 119 intervention villages in Bijapur and Bagalkot districts in northern Karnataka, and will run for a period of three years (2015–2018).

The study will employ a cluster-randomized controlled trial design. A cluster-randomized design was chosen because the intervention is at a cluster level, i.e. many components of the intervention will not be delivered to specific individuals, but to the schools and the communities where the individuals study and reside. All high-schools in Bijapur and Bagalkot districts have been enumerated (n = 1075). For practical and logistical reasons, schools with the following criteria were excluded from the study: schools in urban areas (n = 319); schools that are private and unaided by the government (n = 87); schools that are only for boys (n = 13); and schools in villages where the total number of low caste (scheduled caste/scheduled tribe) girls enrolled in 8th standard was less than 10 (n = 431). This left 225 high-schools eligible for the study, located in 121 village clusters; a ‘village cluster’ comprises one ‘main’ village with one or more eligible high-schools plus the surrounding ‘feeder’ villages, which do not contain a high-school but have low caste children living there who attend the high-school in the ‘main’ village. The village cluster was chosen as the unit of randomization instead of the school so as to minimize contamination between schools in the same village.

Sample size calculations were based on estimates of: (i) the harmonic mean cluster size; (ii) the refusal/loss to follow-up rate; (iii) the between cluster variation (κ); and (iv) the estimated levels in the control villages of the following four co-primary outcomes:

The expected effect sizes were guided by the aim of the intervention, i.e. to bring the primary outcomes in-line with all girls in Karnataka, and are detailed in Table 2. We used the prevalence of these outcomes from the most recent district and state wide surveys to estimate plausible effect sizes. We have sized the trial to detect a risk ratio of 0.70 for transition from standard 7 to 8 among low caste girls in the intervention villages compared with control villages. This is based on a state-level dropout rate among low caste girls of 9% during higher primary grades, compared to 5.6% among all girls, corresponding to a risk ratio of RR = 0.62 [36]. We have sized the trial to detect the weaker RR of 0.70, as district level data were not available, and it is plausible that the intervention will not be able to fully bring low caste drop-out during transition to the average among all girls in Karnataka. It is assumed that retention in school until the end of 10^th standard will be harder to influence than transition from standard 7–8, as the barriers to staying in school increase over time. For this outcome, we have assumed a minimum detectable effect size of RR = 0.75-0.70, based on a state-wide school drop-out rate from standard 8 to standard 10 of 22.5% among low caste girls, compared to 16.5% among all girls, corresponding to a RR of 0.73 [36]. For age at marriage and age at sexual debut, we have assumed a minimum detectable effect size of RR = 0.75, based on a district-level ‘marriage before age 15′ rate of 25.9% among low caste girls compared to 19.3% among other girls (Unpublished data, District Level Household and Facility Survey (DLHS-3), 2007–8, Government of India). We aim to decrease the proportion married before age 15 by 25-30% (bringing the proportion in-line with all girls), corresponding to a risk ratio of 0.75-0.70. We have assumed the proportion who have sexual intercourse before age 15 is equal to the proportion who marry before age 15, and therefore also aim to decrease the proportion who have sexual intercourse before age 15 by 25-30%.

We are uncertain about many of the parameters underlying the sample size calculation, especially the value of k – the between village variation in the outcomes. Therefore, we explored the relationship between the number of clusters and the value of K for a range of realistic estimates of the above parameters, shown in Table 3. In a large proportion of the scenarios shown, we will have sufficient numbers of clusters to detect an effect. However, there are plausible conditions where as many as 54 clusters per arm would be required to detect an effect.

For our main power calculations we have assumed a modest refusal rate and loss to follow-up rate of around 7%. We have included conservative estimates of the control outcome proportions, and reported calculations for a range of K from 0.1 to 0.25 (Table 3). The selection of 80 village clusters with a harmonic average of 30 low caste girls in 7^th standard per cluster, enrolled in two sequential cohorts of 15 girls per year, should be sufficient to detect the effect estimates for the four co-primary outcomes. With minimal loss to follow-up, and a minimum of 28 girls per cluster followed-up at endline, we will have 80% power to detect a 33% reduction in the risk of not making the transition from 7-8^th standard (i.e. a risk ratio of 0.67 between the arms), and a 25% reduction in the risk of not remaining in school until the end of 10^th standard (i.e. a risk ratio of 0.75 between the arms).

80 village clusters were selected using a systematic random sampling method from the sample frame of 121 village clusters, using STATA 11. The selected village clusters were allocated randomly to either receive the intervention immediately (experimental condition) (n = 34), or at a later date (waitlist control group) (n = 34) (using STATA 11). We notified the selected village clusters of their selection into the study and have begun outreach and advocacy work with community leaders in the intervention villages. After allocation, we realised that there were 12 village clusters where there was a possibility of contamination, because these villages were contiguous with a village cluster in a different arm of the study. To compensate for potential loss of statistical power due to contamination, we selected an additional 12 village clusters (6 intervention and 6 control arm) for inclusion in the study as follows: of the 53 village clusters remaining in the original sampling frame of 121 village clusters, we excluded 24 village clusters due to their close proximity to villages that had already been selected for the study. From the remaining 29 village clusters in the sampling frame, 12 village clusters were randomly selected to be included in the study and these were randomly allocated to intervention and control arms using STATA 11. Thus the trial will be implemented in 80 village clusters (40 in each study arm), encompassing 296 villages (119 intervention and 177 control arm) and 129 high schools (69 intervention schools and 60 control schools) (Figure 1).

The study population will be all low caste girls (defined as belonging to scheduled caste or scheduled tribes) who have completed 7th standard. Most of them will be 13 years of age. We estimate that approximately 2,400 low caste girls (aged 13 years) will be enrolled in the trial in two cohort waves one year apart: 1,200 low caste girls in year one (600 control and 600 intervention arms) and 1,200 low caste girls in year two (600 control and 600 intervention)) (Table 4). Each low caste adolescent girl enrolled in the trial will be profiled and the main decision maker in her family will be identified and invited to participate in the study at the family level. All high-schools in the 80 study village clusters will be profiled and line-lists provided of all staff and school development and monitoring committee (SDMC) members. Each principal, 2 SDMC members and 2 teachers (1 male and 1 female) per school will be randomly selected for participation in the study. Adolescent boys (from the low caste community neighbourhood) (n = 20) will also be recruited purposively for qualitative interviews.

Girls in the control arm (40 village clusters) will be provided with the standard level of support as currently provided by the government and other NGOs. Project staff will visit each cohort member every 6 months so as to minimise loss to follow up at the endline survey. Where girls have migrated as a result of family migration or marriage, information regarding the location of migration will be recorded and kept so that the interviewers can go to their new location to administer the questionnaire at the endline survey.

Girls in the intervention arm (40 village clusters), along with their families, adolescent boys, village communities and their leaders, and staff and SDMC committee members at government-run high-schools, will receive the intervention described in the section above.

The trial has four main components:

We plan to conduct our primary analysis with simple between-arm comparisons in the co-primary outcomes of entering 8th standard, completing tenth standard, being married before age 15 and having sexual debut before age 15. In the course of the analysis, we will present simple cluster summaries for the primary outcomes, and report the empirical value of k with updated power calculations. Given the large number of clusters, and to increase statistical efficiency, individual-level logistic regression models will be used to estimate differences between intervention and control arms [37]. We will use random effects to account for the clustering by village, and include dummy variables for the strata. We will also estimate effects adjusting for cluster-level or individual-level variables that appear to be imbalanced at baseline. Stata 11 will be used for all analyses.

We will conduct a survival analysis for marriage, sexual debut, pregnancy and entry into sex work. We will conduct sub-group analyses as follows: (i) the two cohorts of girls; (ii) per district; (iii) for non-devadasis and (iv) for socio-economic strata. For each sub-group analysis, we will test for interaction with the complete model, and report the interaction test p-values. We will explore differences between the arms of a range of parameters measured, using the questionnaires of the girls and their families.

The qualitative data will be transcribed and translated into English. Qualitative analysis packages (eg., NVivo9, ATLAS TI) will be used to identify key issues and themes emerging over time.

In the intervention arm, schools will use the monitoring system to track all participants on specific input, output and outcome indicators, as part of the intervention. These will include recording all services provided to each participant and their family, school attendance, school drop-out, marriage, academic scores, and mobility (for example due to migration, marriage or sex work). In the control arms, so as not to influence the outcomes through the effect of direct observation and data collection, the monitoring system will be limited to tracking only a few outcome indicators (school drop-out and marriage).

The monitoring data will be analysed and used to understand school attendance, drop-out rates, the proportion of girls marrying and the quality of education (academic scores). A unique identifier will be used for each participant to enable triangulation of the outcomes of interest between monitoring and survey data.

An economic evaluation of the intervention will be conducted, from a provider perspective. This will follow established evaluation approaches, including both the start-up and operational costs of the intervention delivery, and valuing both the monetary and non-monetary inputs into the intervention delivery.

Interviews will be conducted in Kannada (the local language) in private settings in a sensitive and non-judgmental manner. Since some of the participants will be minors, we will seek both informed consent from the girl’s parent or legal guardian (if she is unmarried) and informed assent of the girls. Many of the parents in this region are illiterate and reluctant to sign documents. Therefore we will accept either written consent or verbal, witnessed consent. We will seek independent informed consent from girls who are married and living with their husband. Girls who are illiterate will be asked to assent in front of a literate witness of their choice (but not a parent or guardian). The girl will acknowledge her assent with a thumbprint and the witness will sign the document. The Kannada/English version of the consent/assent form will be given to the participants to read and will also be read out and explained before the beginning of the interviews. As part of the consenting/assenting procedure, participants will be assured that their participation is voluntary, and their decision to participate will not affect any benefits they receive from the school or the intervention.

The main potential area of distress for girls relates to answering sensitive questions around sexual activity, pregnancy, harassment or sexual coercion. To limit embarrassment and encourage disclosure, sensitive questions will be asked through an anonymous pen and paper questionnaire, administered at the end of the face to face interview. If girls cannot read, the interviewer will read the questions aloud and allow the girls to tick the appropriate box privately on the paper. After completing the sensitive questions, girls will fold the paper ballot and place it in a see-through bag with the answers from many other respondents. The questionnaire asks about eve teasing, transactional sex, sex work and sexual coercion, but it does not collect information on who is responsible for harassing or sexually abusing the girl. Interviewers will be trained how to respond if the girl becomes upset while answering these questions or at any other time during the interview, including offering referral and transport to local support and counselling services.

Anonymity will be maintained by using proxy names on qualitative interview transcripts and in any publication that quotes a participant. Quantitative survey data will be anonymised and a unique participant identification number will be used to link survey rounds. The identity of the participants and the information shared by them will not be revealed to anyone who does not work in the research study. Unique identifying numbers will be used to identify the questionnaires; no identifying names will be entered with the computer data. Paper copies of the data will be kept for 5 years after completion of the end of the study before being shredded.

In order to ensure equity to all schools enrolled in the trial, we are applying for funding to enable the school-based interventions to be provided in these schools at the end of the intervention period if the trial indicates that the intervention was effective.

This study has been approved by the Institutional Ethical Review Board of St. John’s Medical College, Bangalore, India, and the Observational/Interventions Research Ethics Committee of the London School of Hygiene and Tropical Medicine and the University of Manitoba.