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


78 ISCB 2014 Vienna, Austria • Abstracts - Oral PresentationsWednesday, 27th August 2014 • 16:00-17:30 Monday25thAugustTuesday26thAugustThursday28thAugustAuthorIndexPostersWednesday27thAugustSunday24thAugust Contributed sessions C43 Causal inference from observational data II C43.1 Using different propensity score matching methods to construct comparable control groups for disease management program evaluation R Riedl1 , A Berghold1 1 Inst Med Info, Stat & Docu, Medical University of Graz, Graz, Austria   In observational studies, confounders, defined as variables associated with both, treatment and disease outcome may induce a bias in the estimates of association. Different matching methods are frequently used to reduce systematic differences of baseline characteristics between treated and un- treated individuals. However, if the number of confounding variables is large, matching on the variables itself becomes challenging. Propensity scores (PS), defined as conditional probability of treatment assignment given observed baseline covariates are used to overcome this dimen- sionality problem. In the literature, the performance of several matching methods, including optimal matching, nearest neighbour matching or matching within calipers, have recently been investigated by simulation studies for constructing matched pairs. It has been noted that for different matching methods and also for the order in which individuals are selected for matching can result in different qualities of the matches. We investigate the influence of different propensity score matching methods in combination with exact matching methods on the ability to induce balance on baseline covariates between the treatment groups of the matched samples in praxis. We apply these methods to data of a dis- ease management program in patients with type 2 diabetes. Furthermore, we investigate the impact on balance if more than one control per case is matched. C43.2 Double propensity-score adjustment: a solution to incomplete matching P Austin1 1 Institute for Clinical Evaluative Sciences, Toronto, Canada   Propensity-score matching allows for estimation of the average treatment effect in the treated (ATT). However, popular matching methods such as nearest neighbour caliper matching, often result in some treated subjects being excluded from the final matched sample. This can lead to loss of generalizability of the estimated treatment effect, since the estimand only applies to the matched treated subjects, and not to the entire population of treated subjects. Alternative matching methods such as nearest neigh- bour matching (NNM) and optimal matching result in the inclusion of all treated subjects in the matched sample, at the cost of the elimination of a lesser degree of bias due to confounding variables. We propose a method based on using covariate adjustment using the pro- pensity score within a sample constructed using NNM or optimal match- ing to address these two limitations. Using a series of Monte Carlo simula- tions, we compared the performance of double propensity-score adjust- ment to caliper matching, NNM, and optimal matching. The proposed method results in improved generalizability compared to caliper matching and greater bias reduction compared to NNM or opti- mal matching alone. We illustrate the application of this method using a sample of patients hospitalized with a heart attack. C43.3 A structural equation modelling approach to explore the role of interferon-α on chronic immune activation in successfully treated HIV-infected patients M-Q Picat1 , I Pellegrin2 , J Bitard2 , L Wittkop1 , C Proust-Lima1 , B Liquet3 , J-F Moreau4 , R Thiébaut1 1 Centre Inserm U897- Epidémiologie - Biostatistique, Bordeaux, France, 2 Laboratoire d’Immunologie-Immunogénétique, Bordeaux, France, 3 School of Mathematics and Physics, Saint Lucia, Australia, 4 CNRS, UMR 5164, Bordeaux, France   Background: Chronic Immune Activation (CIA) is a predictor of Human Immunodeficiency Virus (HIV) progression. In successfully treated pa- tients, the understanding of mechanisms by which CIA persists is still lim- ited. We hypothesized that cytomegalovirus (CMV) could be an important factor of CIA in these patients through persistent production of interferon α (IFN-α). Methods: Data from 191 HIV-1-infected patients were analyzed. Patients initiated antiretroviral therapy between 2005 and 2008, and were treated with sustained virological suppression for at least two years. CMV-induced immune response was measured by QuantiFeron-CMV test (positive if >0.2 IU/mL) and CIA was defined by HLA-DR+/CD38+CD8+T-cells. Structural equation modeling (SEM) was used to evaluate the mediating role of IFN-α related gene-transcription (defined as one or several latent variables using 21 IFN-α-induced genes) in the relationships between CMV and CIA. Several definitions of the latent IFN-α were explored, includ- ing IFN-stimulated-genes (n=5) and MYD88 dependent (n=6) or indepen- dent (n=10) pathways. Results: The hypothesized SEM model revealed a strong association be- tween IFN-α latent variable and CIA (p=0.00034).This association persisted in modeling IFN-α by IFN-stimulated-genes (p=0.00129). Modelling IFN-α through two types of latent variables: MYD88 dependent and indepen- dent, revealed a strong association between IFN-α MYD88 independent variable and CMV (p = 0.00501) and CIA (p = 0.00001). Conclusion: SEM provides a flexible framework to explore complex rela- tionships between variables and to understand mediation. In our applica- tion, a major role of IFN- α was demonstrated in the association between CMV and CIA. C43.4 Independent censoring in survival analysis: a causal approach K Røysland1 1 University of Oslo, Dep of Biostatistics, Oslo, Norway   “Independent censoring” is a common assumption when using the Kaplan-Meier estimator. This means that an individual who has not (by chance) experienced the event in question has the same risk of experiencing the event in an infini- tesimal period, regardless of any previous censoring. The formal definition relies on martingale theory and yields a dynamic concept that is much weaker than assuming the censoring and event in question are indepen- dent in the usual probabilistic sense. It is tempting to think that independent censoring would mean that Kaplan-Meier curves represent the survival as would be seen if the censor- ing had been prevented. This, however, is a claim about causation, and can be treated formally us- ing graphical models and techniques from causal inference. “Local independence graphs” and “local characteristics” provide an anal- ogy to causal Bayesian networks, where the nodes also may represent counting processes. Independent censoring is actually a special case of local independence, so these graphical models provide a natural framework for our purpose. Suppose a model is causal with respect to change of censoring regimes.

Pages Overview