Authors' response: the importance of asking the right question

Abstract

Dear Sirs, We thank Drs Koole and Van den Ende for their comment on our recent study (Keiser et al. 2009). We analysed more than 2000 patients from antiretroviral treatment programmes in Africa and South America that routinely monitor both CD4 cell counts and viral load. We concluded that based on a low positive predictive value of the World Health Organization (WHO) immunological criteria (<30%), these should not be interpreted as indicating virological failure. On the other hand, if immunological criteria were not met, virological failure was highly unlikely: negative predictive values ranged from 88% to 99% depending on the definitions chosen. Surprisingly, Koole and Van den Ende felt that this was ‘a classical example of misuse of predictive values’. We suspect that they misunderstood the objective of our study. Three related questions are of interest: (i) How should CD4 counts be interpreted in a given setting, taking into account the relevant (pre-test) probability of virological failure? (ii) What is the ability of the CD4 count criteria to modify pre-test probabilities? (iii) Should CD4 counts be recommended to detect virological failure? We have worked for several years with antiretroviral treatment programmes that routinely monitor CD4 cell counts but do not have access to viral load testing, and we intended to provide guidance to these programmes (Braitstein et al. 2006; Keiser et al. 2008a). We, therefore, focussed on the first question and calculated predictive values. Based on these probabilities, few people would disagree with our advice that in these settings, the WHO immunological criteria are more appropriate for ruling out than for ruling in virological failure. The ability of the CD4 count criteria to modify pre-test probabilities (question 2) is best captured by likelihood ratios: the further away from one the ratio is, the greater the test result’s ability to modify the pre-test probability (Pewsner et al. 2004). These ratios are presented in Table 1. The ability of a given result to produce a sufficiently high or low post-test probability to rule in or rule out a condition does, however, also depend on the pre-test probability in that setting. For the treatment programmes included in our study, the likelihood ratio of a positive test would need to be very high to rule virological failure in (>200 to arrive at a post-test probability of 80%). In contrast, as we pointed out in our study, because the incidence of virological failure is low, the post-test probability of a negative test was high even with very modest negative likelihood ratios ranging from 0.59 to 0.90 (Table 1). Again, we were interested in the appropriate interpretation of CD4 counts in these settings and focussed on what ultimately matters: the predictive values or post-test probabilities. The third question is important, but it was not the focus of our analysis. We agree with Koole and Van den Ende that CD4 counts should not be measured to detect virological failure; however, CD4 counts also provide essential information on the eligibility for ART, the immunological response to ART and the short-term risk of toxicity, disease progression and death on ART. We concur with Koole and Van den Ende that further research is needed to better define the role of CD4 counts and viral load monitoring, particularly in the context of point of care tests for viral load. We believe such research will also be relevant to high-income settings where a more rational approach to monitoring could probably reduce costs without compromising the effectiveness of ART (Keiser et al. 2008b).