Statistical Designs and Methods for Double-Sampling for HIV/AIDS

Associate Name: Constantine Frangakis

Funding Source/Period of the Grant: NIAID R01 01/15/13-12/31/17


Accurate evaluation of programs that treat and monitor HIV/AIDS patients around the world is central for fighting the epidemic. A major obstacle for program evaluation is patient dropout. An important such program is the President's Emergency Plan for AIDS Relief (PEPFAR). The US has been sponsoring PEPFAR ($63 billion for 2004-2013), and methods to accurately estimate patient survival are central to guide US management of the program. However, PEPFAR experiences high dropout rates (e.g., 39% in two years). Standard survival methods use only the observed non-dropout data, with no objective information for the dropouts. Such methods can be severely biased when dropout patients differ from nondropouts after accounting for observed information. To provide valid evaluation, in earlier work we have used a richer design known as "double-sampling". This design re-allocates increased resources to target, intensively pursue and find a subset of the dropouts. These double-sampled dropouts are intended to represent the non-double-sampled dropouts, and to provide objective information for the entire cohort. Although standard methods have been known for double-sampling in surveys, we have shown earlier that standard survival methods fail when double-sampling is used in continuous enrollment programs such as PEPFAR. Also, we have shown earlier that standard evaluation without double-sampling can dramatically underestimate mortality in PEPFAR by a factor of 5. The proposed methods will build on our earlier work with the framework of "principal stratification". The success of that framework increases the potential impact of this proposal. The proposed new methods are developed for three specific aims, motivated by PEPFAR in East Africa. (Aim 1). Develop methods to estimate the performance of follow-up programs by using data from a given double- sampling design. In this aim we will develop methods to estimate survival from double-sampling designs that select patients based on their history characteristics before dropout. This is also important for the next two aims. (Aim 2). Develop methods to create double-sampling designs that produce most accurate estimation of a pro- gram's performance given fixed resources. Evidence shows that information specific to a patient is important for what double-sampling designs provide best information about a program. Here, we will create patient-dependent double-sampling designs that maximize the accuracy given resources to estimate survival in such programs. (Aim 3). Develop double-sampling designs to best target clinical goals. Aim 1 can use the dropout patients' clinical history to predict those with highest mortality risk. These predictions can constrain the design to ensure to double-sample all such patients to better serve them medically. In Aim 3, we will create designs that maximize the accuracy of estimation and best benefit patients (From NIH Report).