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

ISCB2014_abstract_book

30 ISCB 2014 Vienna, Austria • Abstracts - Oral PresentationsMonday, 25th August 2014 • 16:00-17:30 Monday25thAugustTuesday26thAugustThursday28thAugustAuthorIndexPostersWednesday27thAugustSunday24thAugust Monday, 25th August 2014 – 16:00-17:30 Invited session I3 Inverse probability weighting techniques Organizers: Ronald Geskus and Karen Leffondré I3.1 Iterative inverse probability weighting R Geskus1,2 1 Academic Medical Center of Amsterdam, Amsterdam, The Netherlands, 2 Public Health Service of Amsterdam, Amsterdam, The Netherlands Inverse probability weighting (IPW) is used in many areas of statistics in order to correct for systematic or random imbalance in observed data. Examples of systematic imbalance are 1) confounding in causal inference based on observational data, 2) selection bias due to informative censor- ing in survival analysis, 3) differential sampling fractions in surveys. The mechanism that generates imbalance can be visualized via arrows in a di- rected acyclic graph. The purpose of IPW is to remove these arrows in the analysis by weighting individuals according to the observed imbalance generating mechanism. It eliminates bias in case of systematic imbalance and increases efficiency in case of random imbalance. There is a bias-variance tradeoff in the choice of model for the weights. Rarely is a simple saturated model sufficient to eliminate imbalance. We may have continuous variables or a large amount of variables that we need to correct for. However, using a too flexible weight model may cause near or complete violation of the positivity assumption. This will generate heavy tailed weight distributions and increase bias and variance. Weight truncation has been suggested as a way to tradeoff bias and variance. We describe an algorithm that performs IPW iteratively (IIPW). We exam- ine performance of IIPW in a couple of point treatment simulation studies with confounding. Compared to IPW estimators, IIPW estimators 1) suffer substantially less from small sample bias, 2) are less variable, and 3) are more robust against model misspecification. We give heuristics why the method improves upon standard IPW. I3.2 Bias-variance trade off in IPCW: Is it possible to hear the curse of dimensionality in a random forest? TA Gerds1 1 University of Copenhagen, Department of Biostatistics, Copenhagen, Denmark Estimating the nuisance parameter of a semiparametric model is more ef- ficient than not to, even if the nuisance parameter is known. This some- what surprising result is the heart of inverse of the probability (of censor- ing) weighting technique. In the context of a clinical study with survival endpoint, it implies that even if the censoring mechanism is known to be independent of the covariates it is advantageous to use a working regres- sion model to estimate the conditional censoring distribution. But, a bias is introduced into the IPCW estimate if the working model is misspecified. This is particularly delicate if the nuisance parameter is a conditional dis- tribution function (given covariates) because the curse of dimensionality seems to prohibit a purely non-parametric approach to estimation of the weights. In this talk I will provide an intuitive understanding of the efficiency gain, illustrate the bias-variance trade off in real data applications and by us- ing simulations, and propose to estimate the weights based on a machine learning approach: We shall see to what extend a random forest can ab- sorb the curse of dimensionality. I3.3 Inverse probability weighting methods for biomarker evaluation with case cohort studies T Cai1 1 Harvard University, Boston, United States Identification of novel biomarkers for risk prediction is important for both disease prevention and optimal treatment selection. However, dis- covering which biomarkers are useful for prediction will require the use of stored biological samples from large assembled cohorts, and thus the depletion of a finite and precious resource. To preserve these samples, the case cohort (CCH) design provides a resource-efficient sampling design, especially when the outcome is rare. However, existing methods for CCH designs focus on efficient inference of the relative hazard parameters from the Cox regression model, or have considered measures of predictive accuracy of only a single biomarker. In this talk, we will discuss inverse probability weighted approaches to deriving robust risk prediction rules under general survival models. A major obstacle for making inference under two phase studies is due to the correlation induced by the finite population sampling which prevents standard inference procedures such as the bootstrap from being used for variance estimation. We propose a novel resampling procedure to obtain p-values and confidence intervals for parameters of interest. The proposed procedure will be applied to a Danish case-cohort study of novel lipid markers for prediction cardiovas- cular risks.

Pages Overview