In these studies, BED testing was performed on specimens from the same cohorts where incidence was conventionally estimated. The use of the same specimens is an advantage over study designs where BED-estimated incidence is determined on specimens that are related to, but separate from, those of the referent cohort. For instance, the true incidence in the pre-enrollment screening for a cohort study or in a separate cross-section of the same population may not be equivalent to the incidence measured in the related cohort because of selection bias, recruitment bias, or both. The design also allows evaluation of the window and allows subanalysis of features that affect assay performance. We were alerted to two such features in this study: the effect of unstable incidence and the influence of specimens from long-term infected participants.
Incidence estimates from cohorts and from cross-sectional analysis differ fundamentally. Cohort data are collected during prescribed periods, whereas the cross-sectional method produces an estimate at a given point in time. Cohort data measure the number of seroconversions that occur during a given period of follow-up and are frequently used as the criterion standard. The cross-sectional analysis is dependent on the number of seroconverters who are within the recency period (in this instance, 152 days) at the time of specimen collection. The incidence rate that is actually measured is the number of recent infections per 152 days. Extrapolation of this value to a longer period (e.g., 365 days for an annualized estimate) is based on the assumption that the rate remains the same. If the rate is not constant, the BED estimate, extrapolated to a period longer than the recency period, will be in error. An example of this occurred during the first 8 months of follow-up in the BMA study, when more seroconversions during the early part of the period biased the BED estimate, resulting in an underestimate of the conventionally estimated incidence (). This observation highlights the importance of understanding the relationship of recency period to the sampling period of a cohort study and how fluctuations in incidence during the sampling frame may bias BED results.
The recency period used for calculating BED incidence in this study was 152 days. This period was based on analysis of seroconversion panels from 190 seroconverters representing subtypes B′ and CRF01_AE 
. The recency period that would have given precise agreement between conventionally and BED-estimated incidences can be calculated by entering the conventionally estimated incidence into the BED incidence formula and solving for the recency period. This calculation would result in a recency period of 140 days (range, 130–153 days). Thus, the recency period used and the recency period that would have given perfect concordance in the conventionally estimated and BED-estimated incidences are similar.
In the Vax003 cohort, the effect of long-term infections was observed. The BED estimate was based on analysis of specimens collected at the end of each year from those who seroconverted in that year (). As the study progressed, specimens collected in the designated year from participants who had seroconverted in previous years (longer-term infected persons) became available. These specimens should, in theory, be classified as long-term infections in the BED assay, but approximately 5% of these specimens register false-recent BED results 
. As more and more specimens from participants infected for longer periods of time accumulated, the prevalence of seropositive persons in the cross-sectional samples rose. Consequently, the small portion of false-recent BED results inflated the BED-estimated incidence (). This analysis is most relevant to the context in which the BED assay is currently widely used: cross-sectional populations with a predominance of longer-term infected people.
The inflation of the incidence estimate related to false-recent results for long-term infected subjects can be substantial, rises with increased prevalence in the test population, and has been noted before 
. The need for further studies in this regard has been pointed out by many 
. There are several potential ways of mitigating these effects. To some extent, persons who are known to have long-standing infection can be classified as having long-term infection as part of case-based surveillance 
. Participants who self-report or otherwise are known to be long-term HIV-1–seropositive 
, patients with AIDS 
, or patients receiving antiretroviral therapy 
are unlikely to be recently infected and likely to register recent by the assay. This history may be available or can be included in the design of the cross-sectional study and can complement the testing classification. A more stringent testing algorithm could be used, one that requires confirmation of BED-recent specimens with a second test for recent infection and the addition of testing for the presence of antiretroviral drugs in specimens that are BED-recent.
The use of posttest mathematical adjustments that correct for misclassification have been proposed by several investigators 
. These adjustments rely on an accurate estimate of the anticipated false-recent rate in long-term infected participants 
. If the false-recent rate is accurate and relevant to the population being screened, the correction works quite well, as shown in the analysis of the VAX003 data (). However, relevant data may not be available, or special screening over time in the test population may be required to generate the data 
Our results suggest that the BED estimate of incidence, when determined on specimens from prospective cohort studies of initially HIV-1–seronegative persons, is comparable to the independently estimated conventional incidence from the same cohorts in Thailand, both for CRF01_AE and subtype B′.
The cohort design allows one to identify, model, and quantify factors that perturb the estimate, two of which are noted here (unstable incidence and the significant impact of long-term prevalent specimens that may register false-recent in the assay). The availability of tests for determining incidence has multiple potential advantages, not the least of which is that an incidence testing program can easily be superimposed on surveillance programs for HIV-1 prevalence. Several countries, Thailand included, are supplementing national surveillance for HIV infection by using BED-based incidence estimation while incorporating elements of case-based surveillance. A recent survey of Thai military conscripts during 2005 and 2006 found a BED-estimated incidence of 0.14 to 0.20% per year 
, a significant decline compared with the estimate of 1.19 per 100 PY for 1991–1993 ().
Many of the survey design, data collection, and sampling issues related to prevalence estimates also apply to incidence estimates. However, there will be issues, expected or otherwise, that are unique to incidence testing and particularly to population-based versus cohort-based settings. In implementing the BED method for population-based surveillance, it will be important to be aware of the biases, assumptions, and limitations of making incidence estimates and to mitigate their impact by careful survey design, testing, analytic adjustment, and extrapolation.