Methods

Data came from the Women’s Interagency HIV Study (WIHS) (Barkan et al. 1998; Adimora et al. 2018) and the Multicenter AIDS Cohort Study (MACS) (Kaslow et al. 1987). Briefly, the WIHS and MACS are ongoing US-based cohort studies of HIV-infected and -uninfected women and men who have sex with men, respectively. MACS began in 1984, with additional

86

urban locations, with additional recruitment waves in 2001, 2011, and 2013, eventually expanding to 10 urban and suburban sites. In both studies, laboratory procedures, clinical examinations, and structured interviews are conducted at entry and every six months thereafter. Information collected through interview includes self-reported medication use along with

demographic, socioeconomic, and behavioral characteristics. The laboratory procedures included measures of, in particular, CD4 cell count, HIV RNA, HCV Ab and RNA, and non-invasive markers of liver fibrosis.

Individuals included in our cohort were HIV-infected at study entry or seroconverted during follow-up. Visits occurring after the opening of WIHS recruitment (October 1, 1994) were included. All participants were ART-naïve and without an AIDS diagnosis prior to their first eligible study visit. Follow-up began at the first eligible study visit after HIV diagnosis and continued until the first of: loss to follow-up, death, 10 years after the first eligible visit, or September 30, 2015. A participant was considered lost to follow-up at the time of their second missed study visit. Time was discretized into 6-month intervals, corresponding to the

approximate interval between planned study visits. Definitions

In both studies, HCV Ab was assessed at baseline by enzyme immunoassay. Specimens with reactive Ab results underwent HCV RNA testing by real-time polymerase chain reaction assays. Those with detectable HCV RNA were considered to have chronic HCV (HCV+).

The definition of ART was guided by the November 2014 US Department of Health and Human Services guidelines (Panel on Antiretroviral Guidelines for Adults and Adolescents 2014). Once a participant reported initiating ART, they were assumed to remain on it for the

87

duration of the study (the intent-to-treat assumption). ART was split into two variables based on time of initiation. ART initiated prior to October 1, 2001 (when tenofovir, a key component of many modern ART regimens (Panel on Antiretroviral Guidelines for Adults and Adolescents 2018), was approved) was considered early ART, while ART initiated after that date was considered modern ART.

Ascertainment of death

Both studies perform death registry searches to obtain information on the mortality of participants. Date and cause of death are obtained either directly from the National Death Index (https://www.cdc.gov/nchs/ndi/index.htm) or through copies of death certificates obtained by study investigators.

Confounders

Confounders were chosen using a causal diagram (Greenland et al. 1999) constructed prior to data analysis. Time-fixed confounders included age, sex, race and ethnicity, injection drug use (IDU), heavy alcohol use, and smoking status. Time-varying confounders included CD4 cell count and HIV RNA. For the effect of DAA treatment (but not HCV infection, see the statistical analysis section), hepatic fibrosis was also included as a time-fixed confounder after

categorization into 3 levels: FIB-4 (Sterling et al. 2006) ≥ 3.25 or AST to Platelet Ratio Index (APRI) (Lin et al. 2011) ≥ 1 was classified as cirrhosis, while FIB-4 < 1.45 and APRI < 0.7 (together) was classified as no significant fibrosis. Other combinations were classified as non- cirrhotic fibrosis. We chose APRI cutoffs based on a meta-analysis (Lin et al. 2011) that suggested the improved performance of these cutoffs for classification of cirrhosis and no

88

significant fibrosis than the commonly used cutoffs of 2 and 0.5, respectively. Further details on variable measurement and operationalization are in Chapter 3.

Statistical analysis

The parametric g-computation algorithm formula (hereafter referred to as the g-formula) (Robins 1986) was used to estimate the causal effects in this study. We estimated the effects on 10-year all-cause mortality of: 1) chronic HCV infection among all PLWH, 2) chronic HCV infection among PLWH+HCV, and 3) DAA treatment among PLWH+HCV. Pooled logistic regression was used to model the discrete-time hazard of mortality conditional on ART, HCV status, and confounders. Pooled logistic and linear regressions were used to model the

conditional distributions of the time-varying confounders. Using these models, the time-varying confounder histories and survival curves were simulated for an enlarged resampled set of the observed population under each HCV scenario (Westreich et al. 2012; Murray et al. 2017). Each effect was estimated after first setting each person to initiate modern ART at study entry.

Confidence intervals were estimated using the nonparametric bootstrap with 1000 samples. Full details of the g-formula estimation procedure are in the Chapter 3.

Multiple imputation was used to handle missing data (Allison 2001) (the amount missing for each variable is presented in Table 4.1, and ranged from none to 30% missing (baseline

fibrosis)). We assumed a multivariate normal imputation model, which is robust to non-

normality in many settings (Allison 2001). Missing time-varying covariates were carried forward from the most recent measurement. The ‘Boot MI’ algorithm was used to incorporate multiple imputation into the bootstrap (Schomaker & Heumann 2018), with 20 imputed datasets per bootstrap sample.

89

To assess the fit of parametric models, the natural course was simulated in one imputed dataset and the distributions of key variables were compared to those observed in the data.

Though most of the period covered in this study predates DAAs, it is still possible to estimate the effect of DAAs using data available from the MACS and WIHS. These effects were

estimated under the assumption that an individual without HCV with a given degree of liver fibrosis at baseline has the same risk of mortality as a similar individual after successful treatment of their HCV infection (conditional on confounders). Intuitively, the survival

experience of HCV-uninfected PLWH served as a proxy for the experience PLWH+HCV would have had if they were treated with DAAs. Technically, PLWH without HCV infection were assumed to be conditionally exchangeable (Hernán & Robins 2017) with treated PLWH+HCV given baseline fibrosis and confounders.

Therefore, baseline fibrosis was included as a confounder to estimate the effect of DAA treatment. In contrast, fibrosis was not included when estimating the effect of HCV infection, because fibrosis is the primary mechanism through which HCV causes mortality. Rather than assuming all PLWH+HCV would achieve SVR with DAAs, a beta-Bernoulli random variable was used to determine DAA treatment success, with average treatment effectiveness of 0.96 and 95% of the distribution falling between 0.93 and 0.98 (Naggie et al. 2015).

Sensitivity analyses

First, the effect of HCV infection was estimated with a marginal structural model fit with inverse probability weighted estimating equations (Robins et al. 2000). The models needed for this method are distinct from those needed for the g-formula, so concordance between the results

90

provides confidence in the model specifications. Further details of the MSM are provided in the Chapter 3 and Appendix A.

Second, subjects may have been enrolled long after HCV acquisition and time of HCV acquisition is unknown in this study. As such, subjects with prolonged infection have more advanced liver fibrosis on average than those with recently acquired infections, and there is also a possibility of selection bias. To provide an estimate less impacted by such biases, we estimated the effect of HCV infection restricted to participants without significant fibrosis, as these

individuals are more likely to have recently acquired HCV.

Third, to address possible confounding by hepatitis B virus (HBV) co-infection, we conducted each of the analyses restricted to those negative for HBsAg at baseline. We used restriction (rather than standardization or adjustment) due to the small number of HBV/HCV co- infected individuals in the study population leading to issues of non-positivity (Westreich & Cole 2010).

4.3 Results