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


ISCB 2014 Vienna, Austria • Abstracts - Oral Presentations 39Tuesday, 26th August 2014 • 9:00-10:30 Monday25thAugustTuesday26thAugustThursday28thAugustAuthorIndexPostersWednesday27thAugustSunday24thAugust Tuesday, 26th August 2014 – 9:00-10:30 Invited session I4 New methods to control for unmeasured confounding Organizer: Michal Abrahamowicz I4.1 Bias sensitivity analysis of unmeasured confounding RH Groenwold1 1 UMC Utrecht, Utrecht, The Netherlands Observational epidemiologic research is prone to confounding bias, par- ticularly due to unmeasured confounding. Bias sensitivity analysis can help to guide discussions on the possible direction, magnitude, and impact of unmeasured confounding. Bias sensitivity analysis of unmeasured con- founding often focuses on a single unmeasured confounder, which may summarize the information of multiple weak unmeasured confounders. Although easy to evaluate, the downside of such a summary confounder is that it is not very intuitive and may be hard to conceptualize. Alternatively, the impact of multiple (possibly weak) confounders can be evaluated, in which case the correlation between those confounders is a key driver of the magnitude of unmeasured confounding. In this presentation, we will review the literature on bias sensitivity analy- sis of unmeasured confounding. Multiple bias sensitivity analyses will be illustrated with a study of ascorbic acid intake. The substantial mortality reduction associated with high ascorbic acid intake found in a observa- tional study was not replicated in an RCT, which may be the result of un- measured confounding in the observational study. Using this example, we will focus on the distinction between sensitivity analysis of unmeasured confounding due to a single (summary) confounder and multiple con- founders. I4.2 A Bayesian perspective on unmeasured confounding in large administrative databases LC McCandless1 , P Gustafson2 , JM Somers1 1 Simon Fraser University, Burnaby, Canada, 2 University of British Columbia, Vancouver, Canada Confounding creates terrible problems in observational studies using large administrative databases. The massive sample size crushes p-values and standard errors to zero that are calculated from standard analytic ad- justment. While this may delight health researchers who discover that ev- erything is significant, it obscures the role of bias, including unmeasured confounding. The Bayesian approach to statistics provides an appealing way forward because uncertainty about bias can be funneled into the analysis using prior distributions. The posterior distribution for model parameters incor- porates uncertainty from bias in addition to the usual random sampling error. In this talk I will discuss Bayesian approaches to adjustment for un- measured confounding in large administrative database studies. I will focus on the example of causal mediation analysis with confounding in the mediator-outcome relationship. The Bayesian method is illustrated in a mediation analysis of mortality among offenders with mental illness in British Columbia. I4.3 New statistical methods for using validation subsamples to adjust for unmeasured confounders in survival analysis M Abrahamowicz1 , R Burne1 1 McGill University, Montreal, Canada Observational studies of the effects of treatments on clinical outcomes typically rely on large administrative databases, which often lack informa- tion on important confounders such as clinical and lifestyle characteristics. However, such confounders are often recorded in smaller clinical ‘valida- tion’ datasets. Recently, a few methods have been proposed which use validation data to control for unmeasured confounding, however only Propensity Score Calibration (PSC) (Stürmer et al., Am. J. Epi. 2005) can be easily implemented in survival analysis. We propose a new method specifically designed for application in time- to-event analyses which makes use of such validation datasets in order to impute values for the missing confounders in the large databases. Our ap- proach uses martingale residuals as an indication of lack of fit in the mul- tivariable Cox proportional hazards (PH) model due to the unmeasured confounders. First, from a Cox model that includes only the measured confounders we obtain the martingale residuals, which can then be used within an impu- tation model for the unmeasured confounders. We expect the martingale residual to be informative about the value of the missing confounders. Thus, including the martingale residual in an imputation of the missing confounder(s) may improve the accuracy of the imputation. We assess this method in simulations under a variety of assumptions, altering the strength and direction of confounding and the censoring mechanism. The results are compared to (i) a conventional method which does not adjust for unmeasured confounders, (ii) PSC and (iii) standard imputation.  

Pages Overview