L. Labbé (1), S.M. Hammer (2), J.W. Mellors (3), S. Rosenkranz (4), L.B. Sheiner (1), and the ACTG 398 study team
(1) Departments of Laboratory Medicine and Biopharmaceutical Sciences, University of California San Francisco, San Francisco, CA; (2) Division of Infectious Diseases, Department of Medicine, Columbia University College of Physicians and Surgeons, New York, NY; (3) Division of Infectious Diseases, Department of Medicine, University of Pittsburgh School of Medicine, Pittsburgh, PA; (4) Statistical and Data Analysis Center, Harvard School of Public Health, Boston, MA
Objectives: We investigate the effect of adherence to prescribed antiviral drugs on viral response in PI treatment failures randomized to additional medications in clinical trial ACTG 398 using a Markov model, based on that proposed by Vrijens [1], of the bi-monthly viral RNA increment/decrement. The analysis illustrates the use of multiple imputation and bootstrap with a mixed-effects model.
Methods: All bi-monthly viral RNA values within each patient are categorized into one of three classes: low (L: log10RNA ≤ 2.5) / med (M: 2.5 < log10RNA ≤ 4) / high (H: log10RNA > 4). The response (Y) for each inter-observation interval is the change in RNA category over the interval (decreases (D)/remains the same (S)/increases (I)). Covariates (X) are pre-study exposure to NNRTI (N; a baseline variable), duration of 398 therapy (T; early/late), and drug adherence (ADH; see below) during the interval, as measured by questionnaire (AQ) and electronic compliance monitoring caps (MEMS). Different summaries of daily MEMS-based exposure were evaluated: moments of the distribution of inter-dose intervals, as used by Vrijens [1], fraction of inter-dose intervals greater than a specific value, and fraction of days on which medication was taken. All independent variables were dichotomized by finding the cut-point yielding highest explanatory power in the model. The Markov property is conferred by conditioning on starting RNA value (RNASTART). Due to the fact that probabilities must add to unity and certain transitions (e.g, Y=D| RNASTART=L) are impossible, the 3 responses × 3 values of RNASTART = 9 possible transition probabilities at any setting of X can be uniquely specified using only 4 parameters (A1–A4), modeled as
ln(Ai) = bij*Zij + bij*Zij*T + bij*Zij*N + bij*T*N + hi,
i,j=1,4, where, Zij = 1 + aij*ADH, ADH = l*MEMS + (1-l)*AQ, the b, a, and l are parameters to be estimated, and the hi are normally distributed random individual effects. NONMEM is used for estimation, which is stabilized by penalizing all fixed-effect parameters, except for the baseline effect (b11), for deviation from zero, the “null” value. Multiple imputation is performed for missing MEMS and AQ. Standard-errors are estimated by bootstrap.
Results: New treatment duration (T) and prior exposure to NNRTI are significant: the objective function (OF) decreases by 86. The best cut-off point for T (early vs. late) is 2 months. ADH = average (l = 0.537) of MEMS and AQ significantly affects RNA (OF decrease = 11.5). For MEMS, the simple fraction of compliant days is as good as any other measure tested.
Conclusion: We have defined a Markov model for the direction of viral RNA change at bi-monthly intervals in extensively treated AIDS patients. The model recognizes the influences of prior RNA (Markov property), drug adherence, time since start of (new) treatment, and prior exposure to NNRTIs.
References:
[1] Vrijens B. Analyzing time-varying patterns of human exposure to xenobiotics and their biomedical impact. PhD Thesis. University of Ghent, 2002. Ghent, Belgium.
Reference: PAGE 12 () Abstr 450 [www.page-meeting.org/?abstract=450]
Poster: poster