Methodological Challenges in Biomedical HIV Prevention Trials (2008)

Below is the uncorrected machine-read text of this chapter, intended to provide our own search engines and external engines with highly rich, chapter-representative searchable text of each book. Because it is UNCORRECTED material, please consider the following text as a useful but insufficient proxy for the authoritative book pages.

8 Estimating HIV Incidence O ne of the most important aspects of site preparation is accurately estimating the HIV incidence rate in the trial population. As Chap- ter 2 notes, studies are powered based on the number of HIV infections investigators expect to occur during the trial. Small errors in estimates of HIV incidence can have a significant impact on study power and sample size. Overestimating HIV incidence can have particularly disastrous effects. For example, several recent late-stage microbicide and PrEP studies were stopped prematurely because HIV incidence was much lower than originally estimated. In the Savvy microbicide trials in Ghana and Nigeria, estimated HIV incidence at both sites was 5 percent, while observed incidence was 1.9 percent and 1 percent, respectively. Other multisite trials (HPTN 035, CS-CONRAD) have closed some individual sites because HIV incidence was lower than expected. Canceling trials or closing individual sites because of lower-than-expected incidence ultimately delays the ability to identify effective agents, wastes scarce resources, and disrupts the local community (van de Wijgert and Jones, 2006). These experiences underscore the need to accurately estimate HIV incidence before a trial starts. Although HIV prevalence—the proportion of people who are infected at a single point in time—can be measured through cross-sectional studies, estimating HIV incidence, or the number of people who become infected with HIV over a given amount of time, is much more difficult, because HIV infection is a silent event. There are three major approaches for estimating HIV incidence for the purpose of designing a trial: direct longitudinal fol- low-up of populations; biomarkers that indicate recent infection on cross- 175 

176 METHODOLOGICAL CHALLENGES IN HIV PREVENTION TRIALS sectional samples; and mathematical models that rely on serial prevalence data. Other indirect methods used in the past to estimate HIV incidence are not appropriate for estimating incidence for trial design. Back-calculation— a method that infers HIV incidence from AIDS case-reporting data and the distribution of “incubation periods” (the time from initial infection to the onset of AIDS) (Brookmeyer and Gail, 1986; Brookmeyer, 1991; Bacchetti et al., 1993)—is less relevant for several reasons. First, back-calculation requires an accurate and complete AIDS case-reporting system, which is not available in most resource-poor settings. Second, with the advent of effec- tive therapies, the incubation period has been difficult to predict, making this approach unusable in areas where antiretroviral therapy is available (McDougal et al., 2005). Third, while back-calculation methods can be used to estimate past HIV incidence, they cannot provide statistically reli- able estimates of current HIV incidence rates, because recent HIV infections are not reflected in AIDS case data because of the long incubation period. Surrogate markers and behavioral risk factors have sometimes been used to indirectly infer information about HIV incidence rates during the pretrial period. For instance, high incidence of reported bacterial sexually transmitted infections (STIs), such as syphilis or gonorrhea, can suggest geographical areas and groups in which sexual transmission of HIV may be occurring. Similarly, high incidence of hepatitis B and hepatitis C (or, more likely, high prevalence in the case of hepatitis C) in injection drug users can suggest where parenteral transmission of HIV may be occurring. Behavioral surveys may also help guide investigators to geographical areas, demographic groups, or risk groups with an elevated likelihood of HIV transmission. However, although surrogate markers and behavioral risk factors can be used to target further investigations of HIV incidence, they cannot provide quantitative estimates of HIV incidence, and they should not substitute for direct measurement of HIV incidence in trial design. DIRECT LONGITUDINAL FOLLOW-UP: COHORT STUDIES A method often used to estimate HIV incidence is a prospective cohort study, in which a well-defined cohort of at-risk, uninfected individuals is followed over time and serially tested for HIV infection to identify serocon- versions. Cohort studies may follow individuals in the same geographic area as a planned trial, or they may employ a run-in design, in which the popula- tion or a subset of individuals selected for the trial is followed in the period leading up to the trial. Cohort studies are advantageous in that they can provide a direct and unbiased measure of HIV incidence at or near the trial site. As discussed in Chapter 7, a run-in design also has other advantages, 

ESTIMATING HIV INCIDENCE 177 in that it simulates the conditions of the trial and can provide valuable information about factors such as adherence, retention, and pregnancy. The drawbacks of cohort studies are that they are costly, time con- suming, and can be logistically difficult to implement. Also, while cohorts provide unbiased results, cohort studies must have large sample sizes to obtain a tight confidence interval (CI) on the true HIV incidence rate. For example, if the true incidence rate were 5 percent, a 1-year cohort study would require about 1,900 subjects to provide a 95% CI of width 2 per- cent (for example, 4–6 percent). If the cohort study were based on 500 subjects, the CI width would be 4 percent (for example, 3–7 percent), with an underlying 5 percent incidence rate, which would likely be too wide for confidently planning a randomized trial. Another drawback is that estimates of HIV incidence from cohort studies may not necessarily reflect the HIV incidence that would occur during a prevention trial. The differences could result from differences in study populations, participation rates, follow-up rates, or secular changes in incidence (Brookmeyer and Quinn, 1995; McDougal et al., 2005). Fur- thermore, as trial participants are exposed to repeated prevention counsel- ing and education over time, HIV incidence could decline within the study population itself. If secular changes in incidence are occurring in the com- munity apart from the trial itself, HIV estimates could misrepresent actual incidence. Researchers need to assess such factors, though it recognizes that such an assessment may in part be qualitative. BIOMARKERS OF RECENT INFECTION A second approach that has been investigated since the 1990s entails using laboratory-based assays, which can distinguish recent from long- standing HIV infections based on virologic or immunologic markers of HIV disease progression, to estimate population-level HIV incidence (Brookmeyer and Quinn, 1995). The development of an accurate bio- marker of recent infection that could be used to estimate incidence would be a major advance. The key advantage of this approach is that investiga- tors can estimate incidence by testing blood samples at a single point in time from a cross-sectional sample, thus avoiding the problems of recruitment, follow-up bias, or secular changes in incidence, and reducing the cost and time required in cohort studies. In practice, this method involves a two-stage process. First, antibody status of blood samples collected through surveillance studies is deter- mined by using standard HIV-1 serological tests. A second assay is then   Thissection does not address the use of laboratory assays for testing and diagnosing acute HIV infection in individuals. 

178 METHODOLOGICAL CHALLENGES IN HIV PREVENTION TRIALS applied to the HIV-1-positive samples (or seronegative samples depending on the specific tests used and the testing algorithm) to determine if they represent “recent” or “established” infections based on defined parameters (McDougal et al., 2005). Incidence is then calculated using standard epi- demiological relationships between prevalence and mean duration. In this case, mean duration refers to the mean duration of the “window” period— the time it takes newly infected individuals to pass from “recent” infection to “established” infection according to the biomarkers. The window peri- ods are not the same for all samples; rather, there is inherent variability. The window periods are random and have a probability distribution. The mean window period is the average of these window periods, which represents the “typical” duration of time it takes to move from recent to established infection, as shown in Equation 1 below (Quinn et al., 2000; McDougal et al., 2005). Incidence = prevalence/mean duration of the window period (1) The accuracy of the incidence estimate depends on the accuracy of both factors: prevalence and mean duration of the window period (McDougal et al., 2005). Prevalence is calculated as the proportion of those identified as recent infections divided by the number of individuals in the susceptible (uninfected) population. As a result, the sensitivity and specificity of the tests designed to identify recent infection are important factors in the accu- racy of prevalence. The mean window period is needed in order to convert the data col- lected from a cross-sectional sample into an incidence rate. The mean window period is assessed in advance based on serological panels of known seroconverters (obtained from cohort or other studies) who are serially tested over time. All else equal, tests with longer window periods will provide more statistically stable and reliable HIV incidence estimates (McDougal et al., 2005). The biomarker-based assays used in algorithms to detect recent infec- tion can be classified into two groups: viral tests and antibody tests. Viral Tests Viral tests can detect HIV infection in individuals who are acutely infected but who have not yet seroconverted. Although individuals with acute infection will not test positive for antibodies, markers of HIV repli- cation, such as RNA and DNA will generally begin to appear during the second week following exposure. Estimating incidence using viral tests involves a two-stage algorithm that first tests all samples for antibodies and then tests all antibody negative samples with a viral preseroconversion test, 

ESTIMATING HIV INCIDENCE 179 such as an HIV-1 RNA assay (nucleic acid amplification test or RNA PCR test) or the HIV-1 antigen (p24) assay. Those who are negative on the anti- body test but positive on the viral test are classified as recent infections. Although viral preseroconversion tests are used in the United States and abroad for blood bank screening and clinical testing algorithms (see, for example, Pilcher et al., 2002), these tests are not practical for use in estimating HIV incidence in resource-constrained settings for two reasons. First, the mean duration of the window period—the period of time from the appearance of viral products to the development of antibodies—is very short (on the order of one to two weeks). Thus very large sample sizes are required to estimate incidence with any statistical precision. Second, the viral tests are expensive and must be performed on all antibody-negative samples. Identifying and testing large cohorts of seronegative individuals can be costly and logistically difficult, particularly in resource-constrained settings with high HIV prevalence. Antibody Tests Antibody tests can be used to estimate incidence by distinguishing anti- body responses in recently infected individuals from antibody responses in those with established infection. The two most commonly used antibody based methods are: the Serological Testing Algorithm for Recent HIV Seroconversion (STARHS) (Janssen et al., 1998) and BED capture enzyme immunoassay (BED-CEIA) (Parekh et al., 2002). The STAHRS algorithm involves testing blood samples drawn from a cross-sectional population for HIV antibodies using standard antibody tests. In a second step, researchers test those samples that were positive for antibodies in the first test with a less sensitive or “detuned” assay (Janssen et al., 1998). Those samples that test positive on the first antibody test, but negative on the second test, are classified as recent infections (Parekh et al., 2002). The BED-CEIA can be used to estimate HIV incidence by measuring increasing levels of HIV-1-specific immunoglobulin (IgG) as a proportion of total IgG following seroconversion (Parekh et al., 2002). Following seroconversion, the proportion of HIV-specific IgG to total IgG rises over the first year of seroconversion. Both the United States and South Africa have used BED to develop national estimates of HIV incidence for surveil- lance purposes (Rehle et al., 2005; Lee and McKenna, 2007) but not for trial design. Because these antibody assays have a longer window period than the viral assays (p24 antigen or RNA tests), they are theoretically more suit- able for use in estimating HIV incidence (McDougal et al., 2005). However, these assays have three major limitations. First, the STAHRS was developed 

180 METHODOLOGICAL CHALLENGES IN HIV PREVENTION TRIALS and optimized for estimating the incidence of subtype B HIV-1, using sero- samples from the Americas, Australia, Japan, the Caribbean, and Europe. These assays do not perform as well when applied to other HIV subtypes, such as those found in sub-Saharan Africa (A, C, D, E), India (C), and South East Asia (C, E) (McDougal et al., 2006). As a result, they need to be recalibrated using seropanels of other subtypes. A concern with both the BED-CEIA, along with LS-EIA, is that it misclassifies as recent infections some individuals with longer-term infec- tions (more than 1 year), individuals with AIDS, and individuals receiving antiretrovirals (McDougal et al., 2006; Karita et al., 2007; Sakarovitch et al., 2007). Furthermore, because BED-CEIA is based on the proportion of anti-HIV immunoglobulins (IgG) to total IgG, misclassification may be greater in areas with high HIV prevalence and high chronic coinfection that can result in elevated background rates of IgG. The concern is that these false positives lead researchers to overestimate HIV incidence. To correct for this overestimation, at least two different statistical adjustments have been proposed. McDougal and colleagues (2006) incor- porate various misclassification rates that provide sensitivity and specificity corrections. Hargrove and colleagues (2006) proposed a second, simpler statistical adjustment. Using data from a study in Zimbabwe, Hargrove et al. estimate that the BED assay would classify 5.2 percent of persons who have been infected for 12 or more months as recent infections. He uses this number to correct the incidence estimates by essentially subtracting 5.2 percent of all HIV positives from the number identified as recent infections by the BED assay. Some empirical studies indicated that these adjustments performed well in specific settings. Nevertheless, the committee identified some theoreti- cal questions regarding these adjustments. Equation 1 relies on the mean or average window duration. For example, the reported 160-day mean window for the BED assay (Parekh et al., 2002) included people who have windows much longer than 160 days as well as people with windows much shorter than 160 days. As such, the calculation is already accounting for persons with very long windows periods. Thus, theoretically, no other statistical adjustments should be needed, so the theoretical foundation for the adjustments is unclear. An alternative approach to enhancing the performance of the BED assay is to improve the accuracy of the mean window period in Equation In Sakarovitch et al., 2007, most of the specimens that were misclassified as incident cases were from individuals infected longer than 6 months but less than 1 year.   The BED-CEIA detects levels of anti-HIV IgG relative to total IgG and is based on the ob- servation that the ratio of anti-HIV IgG to total IgG increases with time after HIV infection. 

ESTIMATING HIV INCIDENCE 181 1. That improvement could eliminate the need for any additional ad hoc statistical adjustments. The BED test could be useful in estimating HIV incidence for prevention trials, provided that scientists adequately adjust for uncertainty in the window period. MacDougal et al. (2006) have also proposed another potential method for reducing false positives in sequential testing algorithms. This approach involves taking specimens that BED classifies as recent infections and retest- ing them with a different assay for recent infection. The specimen must be classified as recent infection by both assays to be considered positive for recent infection. McDougal and colleagues (2006) applied such a sequential testing algorithm—using the BED assay followed by the avidity assay—to speci- mens obtained from a longitudinal cohort study, the AIDSVAX B/B vaccine trial, which had a direct measure of HIV incidence. Although the sequen- tial testing algorithm reduced false positives by 41 percent, it also slightly decreased (by 11 percent) the number of true positives that registered as recent. Sequential testing algorithms using a different combination of tests may further reduce misclassification (McDougal et al., 2006). Overall, researchers’ efforts to identify better biomarker-based methods of incidence have been limited by inadequate understanding of antibody formation and insufficient numbers of specimens to test new approaches. To understand the generation of immune responses, a massive cross-sectional population screening program is required, which would recruit subjects pre- seroconversion and follow them for at least 2 years. The CHAVI 001 proto- col (CHAVI, 2005) is one example of work in this field that is designed to address some of the current shortcomings in the development of antibody formation. CHAVI 001 is a multicenter, prospective, observational, cohort study that will collect biological specimens for up to 1300 enrollees fol- lowed for 2 years to study the HIV-1 virus, the host response, the genetic factors that determine HIV transmission, and viral set point. Although this effort is primarily directed at vaccine discovery and characterization, this type of research could potentially support discovery of markers of acute infection. Overall, these methods can provide quick and inexpensive estimates of HIV incidence rates. However, the results are not only imprecise but may also be biased. Further validation studies are under way to address concerns that the BED test may produce biased estimates of HIV incidence. MATHEMATICAL MODELS A third approach to estimating HIV incidence in study populations relies on mathematical models of serial cross-sectional data on HIV preva- lence. Although this method for estimating incidence can be relatively 

182 METHODOLOGICAL CHALLENGES IN HIV PREVENTION TRIALS inexpensive and quick, it is the least direct of the three approaches, and the most uncertain, because it requires input parameters for which there is sometimes relatively little information. Mathematical models have been most often used to estimate population- level trends in the HIV epidemic. For example, UNAIDS and the World Health Organization use modeling to produce annual country-specific esti- mates of adult HIV prevalence, incidence, and mortality (Walker et al., 2003). These groups use a two-step process to estimate incidence. First, the UNAIDS/WHO Estimation and Project Package (EPP) develops prevalence curves by modeling HIV data collected over time from national surveillance systems and ad hoc research studies (Walker et al., 2003). Second, Spectrum software (Stover, 2002) estimates incidence by applying assumptions (such as survival time after HIV infection and sex ratio of HIV prevalence) to the EPP prevalence curves (Walker et al., 2003). The quality of these and other model estimates is affected by the avail- ability, quality, and representativeness of the underlying data. For example, one of the major weaknesses of the EPP model is its reliance on prevalence data from national surveillance systems, which vary in quality. Coun- tries with generalized HIV epidemics base prevalence estimates among all adults on sentinel surveillance of pregnant women tested at prenatal clinics. Assumptions are required to translate prevalence among pregnant women to prevalence in the adult population. In addition, much of the informa- tion is collected in urban locations and little in rural settings (Walker et al., 2003). More complex models incorporate many input parameters in an effort to more closely mirror reality. However, the data for those parameters may be so unreliable that a more complex model is actually less reflective of reality than a simple model. As a result, it is important to understand the source and validity of the data, the methods used to develop estimates, and how well those estimates match reality. Overall, while incidence estimates derived from mathematical models can be particularly useful in tracking changes in the HIV epidemic at the population level (partly because the relative change rather than the abso- lute estimate is important), such models do not have enough precision to estimate HIV incidence for a prevention trial. Investigators should use them only as a corroborating source of evidence. In sum, each of the three methods for estimating HIV incidence has strengths and weaknesses. The methods can be broadly classified into those that involve the direct longitudinal follow-up of individuals, and those that indirectly infer incidence rates using other methods. The latter include   ountries C with generalized epidemics are those where HIV prevalence in pregnant women is greater than 1 percent. 

ESTIMATING HIV INCIDENCE 183 those based on biomarkers of recent infection derived from cross-sectional samples and mathematical models of serial prevalence data. Recommendation 8-1: Investigators should base their estimate of HIV incidence on at least one source of data from the direct longitudinal fol- low-up of individuals in the trial setting. Given the importance of accu- rate estimates and the inherent uncertainties of any single approach, the direct estimate of HIV incidence should be corroborated by at least one other source. This corroborating source could be based on any of a number of approaches, including direct longitudinal follow-up in the proposed set- ting, follow-up in related populations in another setting, or any of the indirect approaches. Researchers should also critically assess factors that could change estimates of HIV incidence, such as the impact of sustained counseling and education on a cohort or trial participants, secular changes in incidence as the epidemic evolves, and attrition rates that vary by risk level of HIV infection. Recommendation 8-2: Donors and appropriate U.S. and international agencies should make development of a reliable, accurate biomarker- based test for recent HIV infection that can be run with blood from a single draw a high priority. They should provide the necessary fund- ing and laboratory resources to conduct a substantial cross-sectional screening program. This will require recruiting subjects from countries with low-level, concentrated, and generalized epidemics during the pre- seroconversion period and following them for several years. Recommendation 8-3: Although further validation studies are being conducted to examine concerns that the STAHRS and BED tests may produce biased estimates of HIV incidence, investigators should not rely solely at this time on these or other biomarker assays of recent infection to estimate HIV incidence for the specific purpose of design- ing a prevention trial. REFERENCES Bacchetti, P., M. Segal, and N. Jewell. 1993. Backcalculation of HIV infection rates. Statistics in Medicine 8(2):82-101. Brookmeyer, R. 1991. Reconstruction and future trends of the AIDS epidemic in the United States. Science 253(5015):37-42. Brookmeyer, R., and M. H. Gail. 1986. Minimum size of the acquired immunodeficiency syndrome (AIDS) epidemic in the United States. Lancet 2(8519):1320-1322. 

184 METHODOLOGICAL CHALLENGES IN HIV PREVENTION TRIALS Brookmeyer, R., and T. C. Quinn. 1995. Estimation of current human immunodeficiency virus incidence rates from a cross-sectional survey using early diagnostic tests. American Journal of Epidemiology 141(2):166-172. CHAVI (Center for HIV/AIDS Vaccine Immunology). 2005. CHAVI 001: Acute HIV-1 infec- tion prospective cohort study. Bethesda, MD: NIAID. Hargrove, J. K., J. Humphrey, and K. Mutasa. 2006. Back to BED: Resurrecting a method for estimating HIV incidence from a single cross-sectional survey. Paper presented at STARHS workshop in association with the 16th International Conference on AIDS, Toronto, Canada. Janssen, R. S., G. A. Satten, S. L. Stramer, B. D. Rawal, T. R. O’Brien, B. J. Weiblen, F. M. Hecht, N. Jack, F. R. Cleghorn, J. O. Kahn, M. A. Chesney, and M. P. Busch. 1998. New testing strategy to detect early HIV-1 infection for use in incidence estimates and for clinical and prevention purposes. Journal of the American Medical Association 280(1):42-48. Karita, E., M. Price, E. Hunter, E. Chomba, S. Allen, L. Fei, A. Kamali, E. J. Sanders, O. Anzala, M. Katende, and N. Ketter. 2007. Investigating the utility of the HIV-1 BED capture enzyme immunoassay using cross-sectional and longitudinal seroconverter speci- mens from Africa. AIDS 21(4):403-408. Lee, L. M., and M. T. McKenna. 2007. Monitoring the incidence of HIV infection in the United States. Public Health Reports 122(Suppl 1):72-79. McDougal, J. S., C. D. Pilcher, B. S. Parekh, G. Gershy-Damet, B. M. Branson, K. Marsh, and S. Z. Wiktor. 2005. Surveillance for HIV-1 incidence using tests for recent infection in resource-constrained countries. AIDS 19(Suppl 2):S25-S30. McDougal, J. S., B. S. Parekh, M. L. Peterson, B. M. Branson, T. Dobbs, M. Ackers, and M. Gurwith. 2006. Comparison of HIV type 1 incidence observed during longitudinal follow-up with incidence estimated by cross-sectional analysis using the BED capture enzyme immunoassay. AIDS Research and Human Retroviruses 22(10):945-952. Parekh, B. S., M. S. Kennedy, T. Dobbs, C. P. Pau, R. Byers, T. Green, D. J. Hu, S. Vanichseni, N. L. Young, K. Choopanya, T. D. Mastro, and J. S. McDougal. 2002. Quantitative detection of increasing HIV type 1 antibodies after seroconversion: A simple assay for detecting recent HIV infection and estimating incidence. AIDS Research and Human Retroviruses 18(4):295-307. Pilcher, C. D., J. T. McPherson, P. A. Leone, M. Smurzynski, J. Owen-O’Dowd, A. L. Peace- Brewer, J. Harris, C. B. Hicks, J. J. Eron, Jr., and S. A. Fiscus. 2002. Real-time, universal screening for acute HIV infection in a routine HIV counseling and testing population. Journal of the American Medical Association 288(2):216-221. Quinn, T. C., R. Brookmeyer, R. Kline, M. Shepherd, R. Paranjape, S. Mehendale, D. A. Gadkari, and R. Bollinger. 2000. Feasibility of pooling sera for HIV-1 viral RNA to diagnose acute primary HIV-1 infection and estimate HIV incidence. AIDS 14(17):2751-2757. Rehle, T., O. Shisana, V. Pillay, and K. Zuma. 2005. National HIV incidence measures—new insights into the South African epidemic. South African Medical Journal 95(3):6. Sakarovitch, C., F. Rouet, G. Murphy, A. K. Minga, A. Alioum, F. Dabis, D. Costagliola, R. Salamon, J. V. Parry, and F. Barin. 2007. Do tests devised to detect recent HIV-1 infection provide reliable estimates of incidence in Africa? Journal of Acquired Immune Deficiency Syndromes 45(1):115-122. Stover, J. 2002. AIM: A computer program for making HIV/AIDS projections and examining the social and economic impacts of AIDS. Produced for USAID by the POLICY Project. Washington, DC. 

ESTIMATING HIV INCIDENCE 185 van de Wijgert, J., and H. Jones. 2006. Challenges in microbicide trial design and implementa- tion. Studies in Family Planning 37(2):123-129. Walker, N., K. A. Stanecki, T. Brown, J. Stover, S. Lazzari, J. M. Garcia-Calleja, B. Schwartlander, and P. D. Ghys. 2003. Methods and procedures for estimating HIV/AIDS and its impact: The UNAIDS/WHO estimates for the end of 2001. AIDS 17(15):2215-2225.