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


ISCB 2014 Vienna, Austria • Abstracts - Oral Presentations 69Wednesday, 27th August 2014 • 14:00-15:30 Monday25thAugustTuesday26thAugustThursday28thAugustAuthorIndexPostersWednesday27thAugustSunday24thAugust matched sets consisting of cases and controls with the same treatment, and a conditional logistic regression identified no significant treatment effect. From the weighted Cox regression, we estimated a significant risk associated with radiotherapy and evidence of effect modification, with a hazard ratio of 2.71 (1.37 - 5.36) in smokers. Conclusion: Inverse probability weighting provides a way to exploit nest- ed case-control data in real settings to overcome weaknesses in design and address new research questions with existing data.   C37.3 Performance of targeted maximum likelihood estimation in point-exposure studies using high-dimensional covariate data M Pang1 , T Schuster1,2 , KB Filion1,3 , M Eberg1 , RW Platt2,4 1 Lady Davis Institute, Centre For Clinical Epidemiology, Montreal, Canada, 2 McGill University, Dept of Epidemiology and Biostatistics, Montreal, Canada, 3 McGill University, Division of Clinical Epidemiology, Montreal, Canada, 4 Research Institute of the McGill University Health Centre, Montreal, Canada   Double robust targeted maximum likelihood estimation (TMLE) has been proposed for estimating marginal causal effects, allowing specification of both treatment and outcome models. While inverse probability weight- ing (IPW) methods are known to be sensitive to violation of the positivity assumption, the consequences of such violation in the TMLE framework have not been widely investigated. As non-positivity is frequently pres- ent in high-dimensional covariate settings, a better understanding of the mechanism of TMLE is of particular interest in pharmcoepidemiological studies using large databases. Using plasmode simulation, we evaluated the performance of TMLE com- pared to that of IPW estimator based on a typical point-exposure drug ef- fect cohort study of statin use post-myocardial infarction and the 1-year risk of all-cause mortality from the Clinical Practice Research Datalink. A variety of model specifications were considered inducing different de- grees of non-positivity. Our simulations showed that the performance of TMLE and IPW estima- tor was comparable when the dimension of the treatment model was modest; however, they diverged when a large number of covariates were considered. In some cases, we found irregular bias, large standard errors, and non-convergence results with TMLE even with a correctly specified but saturated treatment model. IPW estimator showed slightly better MSE performance with high-dimensional model specifications. In conclusion, TMLE and IPW estimator using the same modeling can per- form differently due to their different sensitivity to the positivity violation. AlthoughTMLE is appealing for its double robustness property, a near vio- lation of the positivity assumption in a high-dimensional covariate setting might be problematic.   C37.4 The impact of pCR after neoadjuvant chemotherapy in patients with large operable breast cancer on survival outcomes: a causation analysis A Efendi1 , L Slaets1 , S Vandenberghe2 , J Bogaerts1 1 EORTC Headquarters, Brussels, Belgium, 2 Dept. of Applied Mathematics, Universiteit Gent, Gent, Belgium   Purpose: Pathological complete response (pCR) is a well-used endpoint in neoadjuvant breast cancer trials. Although it is established as a predictive marker for long-term outcome (PFS/OS), surrogacy was not demonstrated (Cortazar et al., The Lancet 2014). Commonly used statistical analyses con- sidered pCR as one of the independent covariates in a multivariate model, along with baseline data. However, pCR is measured after treatment expo- sure and dependent on baseline characteristics. This study examines the causal effect of pCR on PFS/OS independent of baseline covariates, using causal modelling. Methods: Inverse Probability Weighting (IPW) enables us to account for potential confounding between pCR and baselines. IPW creates a pseudo- population, a re-weighted version of the original population, in which the measured association between baseline characteristics and pCR is re- moved. A causal effect is then obtained by fitting the weighted Cox regres- sion for PFS/OS with pCR as the sole covariate. Landmarking will be used to account for lead-time bias, that is, the analysis is restricted to patients that are event-free at time of pCR assessment. Results: The estimation of the effect of pCR on PFS and OS independent of baseline covariates will be presented and compared to the classical ap- proach (multivariate modeling) in a large neoadjuvant breast cancer trial (N=1856). Conclusion: In clinical trials, causal modeling can provide meaningful es- timates of the added predictive value of endpoints or markers, assessed during/after treatment exposure (e.g. pCR, toxicity), on long-term out- come. We observe that pCR indeed has a causal effect on PFS, indepen- dent of treatment and baseline. C37.5 Causal mediation analysis in a clinical survival trials - can statistics help to understand treatment mechanisms? S Strohmaier1 , K Røysland1 , T Lange2 , Ø Borgan1 , T Pedersen1 , O Aalen1 1 University of Oslo, Oslo, Norway, 2 University of Copenhagen, Copenhagen, Denmark   When it comes to clinical survival trials, regulatory restrictions usually require the application of methods that solely utilize baseline covariates covariates and the intention-to-treat principle. Thereby a lot of potentially useful information is lost, as collection of time-to-event data goes hand in hand with collection of information on other internal time-dependent covariates and patients deviate from assigned treatment plans. Whereas considerable effort has been put into developing methods for dealing with treatment deviations, less attention is paid to a secondary objective, to employ those repeated measurements to shed more light on underlying treatment mechanisms. We have data from a large-scale secondary prevention trial available, that compared how different statin treatment strategies would effect the risk of cardiovascular disease among patients with history of acute myocardial infarction, comprising repeated measures of internal makers. To gain a better understanding about how treatment effects evolve over time, we adopt the model of analysis on dynamic path analysis, a model that can be viewed as an extension of classical path analysis and the con- cept of directed acyclic graphs (DAGs) to settings that involve time-to- event outcomes and time-dependent covariates. Additionally, we suggest a definition of direct, indirect and total effects that allows a causal inter- pretation and discuss other causal aspects that arise in this particular set- ting, where we obtain treatment effect estimates on the mediator whilst conditioning on survival.

Pages Overview