5.6; SmartGene, Zug, Switzerland) which contains all genotypic HIV resistance tests performed by the four authorized laboratories in Switzerland [19]. A GSS was defined for each NRTI, NNRTI and PI using the Stanford algorithm (version 6.0.3), such that 0 denotes full resistance to a given drug, 0.5 denotes intermediate resistance C646 supplier and 1 denotes full susceptibility. Raltegravir and enfuvirtide were deemed fully susceptible if no mutation

in the International AIDS Society (IAS)-USA mutation list was detected in integrase and glycoprotein 41 (gp41) tests, respectively [20]; or in the absence of these tests, full susceptibility was assumed for these drugs (and for maraviroc) unless these drugs had already been used in a failed regimen. To derive an overall GSS for therapy, we summed the scores of each drug in the regimen. We also considered a number of alternatives to selleckchem this overall GSS, to see if these alternatives suggest some simple rules for clinical practice. First, we replaced the overall GSS with two components – a GSS for darunavir and a GSS for background therapy. Secondly, we considered whether each of these component GSS values can be approximated by simple clinical

measures. As rough measures of existing resistance to darunavir, we assessed whether the patient failed on both amprenavir and saquinavir and counted the number of failed PI regimens. As rough measures TCL of the potency of background therapy, we assessed whether the patient had at least one other second generation antiretroviral in the regimen in addition to darunavir and counted the number of de novo drugs in the regimen in addition to darunavir. With limited data for analysis, we took a Bayesian approach to fitting Cox proportional hazards models for time to virological failure. Given

that we assessed failure in each of three periods, we used a discrete time version of the Cox model with an offset that adjusts for variation in the time between assessments [21]. For each predictor in our model, we asserted a ‘vaguely informative’ prior where ‘the percentiles of the prior distribution would be viewed as at least reasonable if not liberally inclusive by all those working in the research topic’ [22]. Each prior was represented by a lognormal distribution for a hazard ratio, data that reproduced this distribution were added to the observed data, and standard software was then used to estimate an approximate posterior hazard ratio by a weighted averaging over observed and prior data with each set of prior data assigned to a separate stratum [23]. A priori, we classified each predictor into one of five categories. First we rescaled continuous predictors age, viral load and CD4 cell count into clinically meaningful units (per 10 years, log10 copies and 100 cells/μL, respectively) and centred each about its median.