Practical and theoretical considerations of the application of marginal structural models to estimate causal effects of treatment in HIV infection.
Doctoral thesis, UCL (University College London).
Standard marginal structural models (MSMs) are commonly applied to estimate causal effects in the presence of time-dependent confounding; these may be extended to history-adjusted MSMs to estimate effects conditional on time-updated covariates, and dynamic MSMs to estimate e¤ects of pre-speci�ed dynamic regimes (Cain et al., 2010). We address methods to assess the optimal time for treatment initiation with respect to CD4 count in HIV-infected persons, and apply these to CASCADE cohort data. We advocate the application of all three types of MSM to address such causal questions and investigate gaps in the literature concerning their application. Of importance is the construction of suitable inverse probability weights. We have structured this process as four key decisions, de�fining a range of strategies; all demonstrated a bene�ficial effect of ART in CASCADE. We found a trend towards greater treatment bene�fit at lower CD4 across a range of models. Via large simulated randomised trials based on CASCADE data, longer grace periods (permitted delay in treatment initiation) and in particular less-frequently observed CD4 indicated higher optimal regimes (earlier treatment initiation at higher CD4), although similar AIDS-free survival rates may be achieved at these higher optimal regimes. In realistically-sized observational simulations, the optimal regime estimates lacked precision, mainly due to broadly constant AIDS-free survival rates at higher CD4. Optimal regimes estimated from dynamic MSMs should be interpreted with regard to the shape of the outcome-by-regime curve and the precision. In our clinical setting, we found that allowing a 3-month grace period may increase precision with little bias under the interpretation of no grace period; under longer grace periods, the bias outweighed the efficiency gain. In our CASCADE population, immediate treatment was preferable to delay, although estimation was limited by relatively short follow-up. Comparison across the MSM approaches offers additional insights into the methodology and clinical results.
|Title:||Practical and theoretical considerations of the application of marginal structural models to estimate causal effects of treatment in HIV infection|
|Open access status:||An open access version is available from UCL Discovery|
|UCL classification:||UCL > School of Life and Medical Sciences > Faculty of Population Health Sciences > Institute of Epidemiology and Health Care > Infection and Population Health|
Archive Staff Only