*3.2. Sampling*

The sample included a total of 45 adult women (18 years and above) who go<sup>t</sup> married before the age of 18 years, 22 and 23 of whom were from the northern and central regions of Uganda, respectively. To capture the diverse lived experiences of survivors of child marriage, the women respondents were selected purposively on the basis of various criteria, including age at first marriage, residence (rural or urban), current marital status and whether they were still married to the same men or left the relationship. Eligible participants were identified with the help of community leaders and community-based organisations working with GBV survivors.

### *3.3. Data Collection*

Data were collected between August and September 2018, with the support of 10 research assistants proficient in the local languages spoken in the central and northern regions of Uganda, notably Luganda and Acholi, respectively. In-depth interviews were held with each of the selected women. These were held in the local language using a guide with a list of open-ended questions to enable the participants to express their views. Each interview lasted between 40 min and 1 h and explored a range of issues pertaining to the participants' experience of child marriage. There were questions on participants' socio-demographic characteristics such as age at first marriage and highest level of education attained, the events that preceded the marriage, experiences during the union including IPV and GBV in general, their responses to abuse and available support services, among others. Probes and prompts were used to motivate participants to provide more detailed information on topics of interest where necessary. The flexibility o ffered by in-depth interviews enabled us to gain deep insights into the participants' lived experiences of child marriage, while keeping the interview focused on a specific range of topics (Russell 2002).

### *3.4. Data Management and Analysis*

All the interviews were audio recorded. They were then transcribed verbatim and translated into English by a team of experts. Each transcript was compared with the original audio interview by a member of the core research team proficient in the local language to ensure consistency in the translation. The transcripts were then word processed and imported into NVivo.12 qualitative data analysis software for further management. Coding was conducted by two members of the core research team. To enhance the accuracy of the process, the two members regularly reviewed each other's work to minimise contradictions in the interpretation and assignment of codes to specific data. In addition, where there was doubt on the appropriate code to assign to specific statements, the two members always discussed the coding to build consensus.

The analysis was conducted thematically. The process involved reading and re-reading the transcripts several times and coding relevant sections, words, paragraphs and sentences according to the identified themes and categories. The themes for the article were generated deductively and inductively (Vaismoradi et al. 2013). We started o ff with two broad themes derived from GBV literature: experience of intimate partner violence and leaving violent relationships. Intimate partner violence was defined as physical, sexual, economic and emotional abuse and controlling behaviour subjected to the survivor by their partner. The data were scrutinised for evidence of IPV and the identified forms assigned the relevant categories, that is, physical, sexual and emotional abuse and controlling behaviour. Data were further examined to identify patterns of leaving violent child marriages with the intent of understanding the underlying facilitative factors. The analysis generated the four main sub-themes on which the article is based: having a secure base to return to, reaching a tipping point in the relationship, financial independence and intervention of a significant other.

### *3.5. Ethical Considerations*

The study was granted ethical approval by the Institutional Research Ethics Committee of Makerere University's School of Social Sciences (MAKSS REC 09.18.217) and the Uganda National Council for Science and Technology (SS 4939). Written informed consent was obtained from all the selected participants before involving them in the study. Participants who could not write due to illiteracy and other factors signed using a thumb print. The informed consent forms were printed in Luganda, Luo and English. Literate participants were given the opportunity to read and interpret the consent form before signing, except where they opted to be read to. The research assistants read and interpreted the informed consent forms for illiterate and semi-illiterate participants. The consent process involved informing the participants of the study purpose, how they had been selected, the benefits of participation and their right to voluntary participation and to withdraw from the study at any point, as well as assuring them of confidentiality.

We ensured confidentiality by making sure that participants were interviewed in private spaces, where the conversation could not be heard by others. The participants were not asked for their real names during the interviews. Unauthorised access to interview transcripts was restricted by locking computer files with passwords. In addition, the digital audio recordings were deleted after transcription. Only pseudonyms are used in this article. Participants who exhibited signs of trauma during the interviews were referred to our partner agencies for psychosocial support.
