The WIHS is a prospective study of HIV-1 infection in women, conducted in New York City, Washington D.C., Chicago, Southern California and the San Francisco Bay Area. The WIHS methods and baseline cohort characteristics have been previously described [20
]. Briefly, between October 1994 and November 1995, 2056 HIV-1 infected and 569 uninfected women were enrolled. A second enrollment between October 2001 and September 2002, added 737 HIV-infected and 406 HIV-uninfected women. For the second wave of enrollees, medical record abstraction was performed for participants reporting HAART use at enrollment, to determine their pre-HAART CD4 and HIV RNA counts, and to verify date of HAART initiation and regimen [21
]. Study protocols were reviewed and approved by the institutional review boards and informed consent was obtained from the participants.
Every six months, WIHS participants were interviewed and received a physical examination. Multiple gynecologic and blood specimens were collected at each visit. Interviewers assessed self-reported HAART use during the period prior to the study visit. Clinicians performed participants’ skin examinations, including genital and oral exams. The location, description, and diagnosis of verrucae were based on clinical appearance. Beginning in October 2002, the recording of oral warts changed. Due to the change in reporting method and the overall low number of events, we did not examine predictors of incidence for oral warts.
Baseline characteristics examined included HIV status, age, race/ethnicity, marital status, parity, and lifetime number of male partners. Race/ethnicity was categorized as white, African-American, Hispanic or other ethnicity. The definition of HAART was guided by the DHHS/Kaiser Panel guidelines [22
] and categorized as ever initiated HAART use or currently on HAART. The CD4+ cell count, per 100 cells/mm3
, and HIV RNA viral load, rescaled to log10
copies, were obtained from the visit prior if it occurred within 270 days. Exfoliated cells for HPV DNA testing were obtained using cervicovaginal lavage (CVL) fluid [23
]. HPV DNA was detected with L1 consensus primer polymerase chain reaction assays. Details of these laboratory methods have been published previously [25
], and the results were shown to have high reproducibility, sensitivity, and specificity [24
]. Additional variables examined for each study visit included: cigarette use, and incident self-reported diagnosis of clinical AIDS.
Participants who HIV-seroconverted during the study (N=16) were excluded from analysis. An incident event was defined as the first occurrence of either a cutaneous or anogenital wart after the baseline study visit. The time to event was the time from baseline until the visit date when the incident event occurred. Women with oral, skin or anogenital warts at baseline were excluded from incident events. Women who had multiple incident events in the same location were counted at the time of first occurrence only. The development of a verruca in one location did not exclude a participant from contributing to an incident event in another location.
For women in the original cohort, we defined a participant’s baseline visit (time=0) as the first visit that occurred post-January 1, 1996, to correspond to HAART availability. For the 2001-02 recruits, the baseline visit occurs between October 1, 2001 and September 30, 2002. Time to event in years was calculated from time=0 to presentation of an event or last date seen (maximum date March 31, 2004) for participants who did not develop warts. Unadjusted incidence rates for skin and anogenital warts were calculated using the Nelson-Aalen estimate, stratified into one of three groups: HIV-uninfected women; HIV-infected HAART-naïve women; and HIV-infected HAART initiators.
To examine predictors of incidence, univariate and multivariate Cox proportional hazard models with time-dependent covariates were used. Multivariate models were constructed by including all covariates that were considered of clinical interest.
For analyses that assessed risk factors for skin or anogenital warts, we restricted the study sample to those women who were HIV-infected. Because women with the strongest indication for antiretroviral therapy use are more likely to use it [28
], we examined the association with therapy using a version of the marginal structural proportional hazards model [29
]. The weights were constructed at each visit by using pooled logistic regression models to estimate the probability of HAART initiation using both time-fixed and time-varying covariates, including a quadratic polynomial to model the effect of time. HAART initiation can be predicted much more reliably than current use; once a participant initiates HAART, she is treated as always exposed.
The time-varying predictors used to construct the weights for the marginal structural model were: visit, development of clinical AIDS, number of hospitalizations, crack, cocaine and/or heroin (CCH) use, HIV RNA at the prior visit, CD4+ cell count at the prior visit, and, in the prior 4 years, CD4 nadir and highest HIV RNA. The following time-fixed predictors were also used: phase of cohort enrollment, race, and baseline measurements of HPV infection, number of sex partners, smoking status (current/other) at baseline and age. The final, weighted model included only the time fixed covariates and a time-varying variable for initiation of HAART.
The covariates in these models were chosen on the basis of their ability to predict HAART initiation. Participants reporting use of CCH during the previous four years were indicated to have used CCH. Injection drug use (IDU) was coded in a similar way. The highest HIV RNA measurement (rescaled to log10) and the lowest CD4+ cell count (per 100 cells/mm3) from the past four years were also included.
By construction the marginal structural model is focused on the estimation of the effect of HAART initiation, and not useful to examine the effects of other predictors of outcome. For this purpose we used an additional conventional proportional hazards model, including an additional time-varying covariate for injection drug use, an additional risk factor for incidence of skin warts.