to estimating a population average, there is also interest in estimating the effects of demographic factors or insurance status on this indicator.
To estimate such an indicator requires information on date of measurement and level of CD4 count as well as date of ART prescriptions given or filled. Some data sources, such as health maintenance organizations (HMOs; e.g., Kaiser Permanente), the Department of Veterans Affairs (VA), and federal prisons provide all of the relevant information needed, permitting a relatively straightforward estimation for subsets of the population. However, analytic issues arise from the fact that patients may leave these systems at any point—possibly after a CD4+ cell count <500 cells/mm3 is measured but before the prescription is provided or 6 months have elapsed. Furthermore, delays in reporting (e.g., of HIV/AIDS cases, CD4 counts) must be taken into account, particularly if the goal is to investigate trends over time. In addition, patients may die within 6 months of receiving a CD4 count—a situation that makes it impossible to obtain the indicator. For patients who leave a system before their contribution to the indicator can be assessed, it is important to make use of the available partial follow-up information in an attempt to avoid, or at least reduce, bias. This is fairly straightforward using methods for failure-time data if the loss to follow-up is not informative (i.e., unassociated with greater or lower risk of starting treatment). If it is informative, appropriate methods must be used to minimize bias; however, unbiased estimation is possible only if all potentially confounding variables are available (a very unlikely situation). To investigate the effect of demographic and other factors on the risk of not receiving appropriate ART, regression methods can be used. Limitations arise from losses to follow-up, as described above, as well as from the fact that with the exception of the NHSS, which captures data on the vast majority of people identified with HIV/AIDS in the United States, none of the data sources is representative of either the American population as a whole or any particular demographic group.
The limitation of representativeness can be addressed by making use of other sources of data that have broader coverage. To do so, however, one must make use of data systems that provide only part of the necessary information by combining them in some way. For example, the NHSS provides dates of measurements and CD4 counts but not (reliably) the time of receiving ART. By contrast, Medicare and Medicaid databases provide information about dates of ART prescriptions filled but not CD4 counts. In the absence of unique identifiers, no direct linkage between databases can be made. However, combining across sources is still feasible through linkage by demographic factors. For example, suppose one knew that for one demographic group in a given state, 400 people had CD4+ cell counts <500 cells/mm3 at some point in 2011 among 600 people who had CD4+ cell counts drawn. Suppose one also knew for this group that 300 people