A principal component analysis (PCA) of the 12 acts of violence and their frequency (never, sometimes, often), corresponding to the 12 items in the ‘perpetrated IPV’ questionnaire section was conducted. The Promax rotation technique, which takes into account correlation between factors, was implemented to improve the fit of the data [37]. The PCA resulted in three scores of IPV, which were standardized and computed using values ranging from 0 to 1. These three scores corresponded to three forms of IPV which we defined as follows (S1 Table): i) psychological and physical IPV (PPV), which included 5 items (humiliation, insults or belittlement, threats, shoving or pushing or object throwing, slapping; eigenvalue = 3.2, Cronbach’s α = 0.83); ii) severe physical IPV (SPV), which included 5 items (shoving or pushing or object throwing, slapping, arm twisting or hair pulling, punching or hitting, kicking or dragging or beating up; eigenvalue = 2.3, Cronbach’s α = 0.69); iii) sexual IPV (SV) which included 2 items (forced sexual intercourse and any forced sexual act; eigenvalue = 1.8, Cronbach’s α = 0.62). No respondent reported choking or burning a partner, or using or threatening to use a gun, knife or other weapon against a partner. The standardized scores were considered very reliable and reliable, respectively, when the Cronbach α value was ≥0.7 and [0.5; 0.7] [38, 39]. The three IPV scores explained 73% of the cumulative variance.

For each standardized IPV score a three-class variable (corresponding to our three study outcomes) reflecting the level of violence perpetrated was built using the following individual score cut-offs: score = 0 (no PPV, SPV or SV, as relevant), score <median among non-zero values (moderate level, as relevant) and score ≥median among non-zero values (high level, as relevant).

The proportions of MLHV with unstable aviremia and HIV transmission risk were described overall and according to each IPV outcome (PPV, SPV, PV). With regard to the latter, we used univariate ordinal logistic regressions to test for significant differences between the proportions for the three outcomes.

Univariate ordinal logistic regressions were performed to investigate the associations between each outcome and the explanatory variables listed above. Covariates with a p-value <0.2 in the univariate analysis were considered eligible for the multivariate ordinal logistic regression models. A backward selection method was used to select the covariates for the final multivariate model with a p-value <0.05. All statistical analyses were performed using SAS 9.4 and RSTUDIO 1.1.453.