Please activate JavaScript!
Please install Adobe Flash Player, click here for download


ISCB 2014 Vienna, Austria • Abstracts - Oral Presentations 53Tuesday, 26th August 2014 • 11:00-12:30 Monday25thAugustTuesday26thAugustThursday28thAugustAuthorIndexPostersWednesday27thAugustSunday24thAugust by deriving their conditional hazard functions. We recursively simulate inter-event-times using the conditional cumulative hazards and adopt the univariate approach of Bender et. al. (StatMed 2003). We extend our methods to incorporate fixed and random covariates into a proportional hazards model. Methods are illustrated by simulating clinical trial data with recurrent events under different total time models (weibull, log-normal, step- function). This is used to obtain the empirical power of an Andersen-Gill analysis depending on censoring, shape of hazards and between-subject- variation. We find that censoring most affects the power of a study if haz- ards vary with total time and between subjects. This highlights the use- fulness of our simulation approach for planning clinical trials, as different potentially relevant scenarios, e.g. for progressive diseases, can easily be investigated.   C28 Marginal structural models C28.1 Addressing measurement error in time-varying covariates through the use of simex-adjusted marginal structural models RP Kyle1 , EEM Moodie1 , M Abrahamowicz1 1 McGill University, Montreal, Canada   Background: The assumption of no unmeasured confounding is funda- mental to unbiased estimation of causal parameters from marginal struc- tural models (MSMs). While substantial measurement error is known to in- duce residual confounding in unweighted regression models, few authors have addressed this issue in the context of the treatment model used to generate inverse probability weights (IPW) for use in MSMs. Objectives: Our primary aims were to validate a novel application of the simulation-extrapolation (SIMEX) procedure to reduce the impact of mea- surement error in MSMs, and to demonstrate its utility by application to an analysis of empirical data. Methods: The SIMEX is a simulation-based method for reducing measure- ment error when the measurement error variance of a given covariate is known precisely or may be well-estimated (Cook & Stefanski, JASA 1994). In a series of simulation studies, we examined varying degrees of measure- ment error in a single time-varying covariate in the treatment model from which IPW were obtained. We modified several simulation parameters, in- cluding sample size, treatment and covariate effect sizes, and assessed the robustness of error correction given differing assumptions of error vari- ance. Following analyses of simulated data, we fit SIMEX-adjusted MSMs to data from the Multicenter AIDS Cohort Study. Conclusions: Correcting measurement error via SIMEX in MSMs improves covariate balance, and is a useful tool for reducing bias and improving pre- cision in treatment estimates from MSMs.   C28.2 A caution on the use of stabilized weights in marginal structural models D Talbot1,2 , J Atherton1 , AM Rossi3 , SL Bacon3,4 , G Lefebvre1 1 Université du Québec à Montréal, Montréal, Canada, 2 Université Laval, Québec, Canada, 3 Concordia University, Montreal, Canada, 4 Montreal Behavioural Medicine Centre, Montreal, Canada   Marginal structural models (MSMs) are commonly used to estimate the causal effect of a time-varying treatment in presence of time-dependent confounding. When fitting a MSM to data, the analyst must specify both the treatment model for the inverse-probability-of-treatment weights and the marginal structural model for the outcome. With MSMs, the use of stabilized weights is recommended since they are generally less vari- able than the standard weights. In this work, we are concerned with the use of the common stabilized weights when the structural model is speci- fied to only consider partial treatment history, such as the current or most recent treatments. We present various examples of settings where these stabilized weights yield biased inferences while the standard weights do not. These issues are first investigated on the basis of simulated data and subsequently exemplified using data from the Honolulu Heart Program. In conclusion, we suggest replacing the common stabilized weights with basic stabilized weights that do not share the problems of the former.   C28.3 Dialysis, catheter use and mortality: challenges in applying marginal structural models to data from a clinical registry J Kasza1 , R Wolfe1 , K Polkinghorne1 1 Monash University, Melbourne, Australia   With aging populations and growing rates of obesity, the burden of kid- ney disease is increasing in many countries. Dialysis, the most frequently- used treatment for end-stage kidney disease, is undertaken using one of three modalities: haemodialysis, classified by location of delivery (home/ facility), or peritoneal dialysis. Haemodialysis requires a vascular access be in place, and there are three alternative access types: central venous cath- eter, arteriovenous graft, or arteriovenous fistula. Using data from the Australian and New Zealand Dialysis and Transplant Registry (ANZDATA), we seek to determine the effect of dialysis modality, with haemodialysis sub-classified by vascular access type and location, on mortality. Complicating the estimation of this effect is that throughout the course of a patient’s treatment, their dialysis modality and vascular access type may change. Further, co-morbid conditions such as coronary artery disease are time-dependent confounders affected by dialysis history. We use marginal structural models to estimate the effect of dialysis mo- dality on mortality. However there are several difficulties posed by the structure of ANZDATA that are typical of clinical registries. In particular, we will discuss approaches to accounting for the clustering of patients in dialysis treatment centres when not all dialysis modalities are available at all centres. We also explore the bias in the estimation of treatment effects when modality changes are recorded as they occur (as is optimal), but vas- cular access and comorbidities are only recorded annually. We also discuss methods for dealing with large stabilised inverse probability of treatment weights, which are an issue in this context.   C28.4 Non-specific effects of vaccines on child morbidity examined with a marginal structural model for recurrent events. AKG Jensen1,2 , H Ravn2 , PK Andersen1 1 University of Copenhagen, Copenhagen, Denmark, 2 Statens Serum Institut, Copenhagen, Denmark   Non-specific vaccination effects denote a possible boost or attenuation of the immune system following a vaccination. Observational studies from Guinea-Bissau where even ordinary infectious diseases can be fatal have shown a dramatic effect on child survival due to non-specific effects. To examine non-specific effects in a high-income setting based on Danish registries where infant morbidity (infectious hospitalizations) is the out- come and recurrent events are common we introduce a marginal structur- al model for recurrent events. In this model the recurrent outcome is asso- ciated with future risk of infectious hospitalization and future vaccination- status. In addition the recurrent outcome is affected by past vaccinations and hence also plays the role as time-dependent confounder affected by previous treatment in the marginal structural model.

Pages Overview