Background:
Existing estimates of the effect of ART on survival have come from interval and clinical cohorts that differ on patient demographic and clinical characteristics from recently HIV-diagnosed persons in the United States (US) (the target population). If the effect of ART varies with respect to these same patient characteristics, then the magnitude of existing estimates of the effect of ART on survival may not be directly generalizable to the recently diagnosed US population.
Methods:
Patients (n=12,547) initiating HIV care in eight academic medical centers in the Centers for AIDS Research Network of Integrated Clinical Systems (CNICS) after January 1, 1998 were followed until death from any cause, administrative censoring at 5 years after therapy initiation or the end of follow up on December 31, 2011. The target population was persons diagnosed with HIV in the US between 2009 and 2011, which was provided by the Centers for Disease Control and Prevention from national HIV surveillance data. We estimated 5-year mortality from the complement of standardized Kaplan-Meier survival functions and described the relative reduction in the hazard of mortality due to ART using a marginal structural Cox proportional hazards model. Bias due to confounding and drop out were controlled using inverse probability weights for treatment and drop out. The effect of ART was estimated within subgroups defined by a priori selected patient demographic and clinical characteristics using stratified analyses. We standardized our final estimate of the effect of ART on survival to the recently HIV-diagnosed population using inverse probability weights for generalizability.
Results:
The hazard of all-cause mortality among ART initiators was 0.33 (95% CI: 0.25, 0.43) times the hazard among non-initiators. The protective effect of ART was stronger among patients with no history of injection drug use, lower CD4 cell count at baseline, no prior AIDS diagnosis, and non-Hispanic white race/ethnicity. Although conditions were present that might preclude generalizability of the study estimate, the HR among persons recently HIV-diagnosed in the US was similar (HR=0.32, 95% CI: 0.23, 0.45).
Conclusions:
We demonstrate the use of formal methods for generalizing an estimate of effect from one cohort to a different target population. The similarity of the results in both the CNICS cohort and the recently HIV-diagnosed provides reassuring evidence for the external validity of research conducted in the CNICS.