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

ISCB2014_abstract_book

110 ISCB 2014 Vienna, Austria • Abstracts - Poster PresentationsTuesday, 26th August 2014 • 10:30-11:00 Monday25thAugustTuesday26thAugustThursday28thAugustAuthorIndexPostersWednesday27thAugustSunday24thAugust parameter estimates and statistical significance, indicating that the man- ner in which missing data is handled is essential to obtain correct results in regression analysis.   P2.3.140 Imputation of an ordinal exposure derived from a semi-continuous variable with missing data: a simulation study L Rodwell1,2 , H Romaniuk1,2 , KJ Lee1,2 , JB Carlin1,2 1 Murdoch Children’s Research Institute, Melbourne, Australia, 2 The University of Melbourne, Melbourne, Australia   Multiple imputation (MI) requires careful specification of the imputation model, with which there are often a number of possible methods. We fo- cus on a specific, albeit relatively common, scenario where the analysis model includes an ordinal exposure variable with categories derived from a semi-continuous variable with missing data. We based our simulations on weekly alcohol consumption data from a longitudinal study of adolescents. We varied the proportion of zeros in three different semi-continuous distributions (25, 50 or 75 per cent). The semi-continuous data were then categorised to represent levels of alcohol consumption: none, 1-10 units, 11-20 units and 21+ units. Finally, we cre- ated a binary outcome, with which higher alcohol consumption was as- sociated with higher odds. We generated 2000 sets of 1000 observations. Within each of these we set a random 33% of observations for the alcohol variable(s) to missing. We examined five imputation methods which involve either deriving the ordinal variable first: (1) projected distance based rounding , (2) ordinal logistic regression; or imputing the semi-continuous variable: (3) two-part model, (4) the ‘just another variable’ method, (5) predictive mean match- ing. We assessed the performance of the imputation approaches by compar- ing the average estimates across 2000 simulations with those from a pseu- do-population of 1 million observations. All of the imputation methods performed reasonably well when used to estimate the association with the binary outcome, with over-coverage more common than under-coverage. We intend to investigate these ap- proaches in more realistic conditions where the data are missing at ran- dom. P2.3.169 Regularized approach for missing data problem C-H Tseng1 , Y-H Chen2 1 UCLA, Los Angeles, United States, 2 Academia Sinica, Taipei, Republic of China (Taiwan)   We present a framework to analyze missing data when the missing data mechanism is unknown. The regularized approach is used to accommo- date both ignorable and non-ignorable missing data. We investigate the impact of missing data mechanism uncertainty based on simulation and show that proposed method can provide stable and reliable estimates for both ignorable and nonignorable missing data. We apply our method to a longitudinal clinical trial of hypertension where no- nignorably missing data were a concern.   P2.3.188 What if my doctor would be as receptive to innovations in therapies as to innovations in statistical methods? A Güttner1 , Y Gong2 , M Akacha1 , F Hornig1 , P Mesenbrink3 , S Witte1 1 Novartis, Basel, Switzerland, 2 Novartis, Shanghai, China, 3 Novartis, East Hanover, United States   For chronic auto-immune diseases such as psoriasis, the evaluation of long-term effects is important for physicians and patients with respect to treatment decisions. In long-term clinical trials, missing data due to treatment interruptions or discontinuation of studies have to be handled within the statistical analysis of treatment comparison or estimation of re- sponse rates or disease activity. Statistical methods for imputation of missing data vary and improve over time and health authorities may change the acceptance of imputation methods. Within the same indication, this leads to publication of long- term data where different imputation methods were used. Hence, comparison of different compounds, e.g. for the education of treat- ing physicians and patients, although not studied in the same study, are not always straight-forward. Non-statisticians may not understand different assumptions underlying the imputation methods or associated bias. For effective and safe treat- ments, the amount of missing data is moderate, but not negligible. For response variables in psoriasis, non-responder imputation is often used for short-term comparisons to controls, but leads to decreasing response rates over time. If the handling of drop-outs is not considered in the in- terpretation, a diminishing effect may be, incorrectly, attributed to the treatment. Next to non-responder imputation, last-observation-carried- forward, less stringent non-responder imputation, observed data only and multiple imputations are other options for dealing with missing data. Comparing these methods on the same dataset shows that decreases in response rates over time are driven by non-responder imputation whereas other methods are relatively indistinguishable, in particular show constant response rates over time.   P2.3.196 Anemia is a risk factor for poor cognitive outcome after ischemic stroke VD Zietemann1 , FA Wollenweber1 , M Dichgans1,2,3 1 Institute for Stroke and Dementia Research, Munich, Germany, 2 Interdisciplinary Stroke Centre, Munich, Germany, 3 Munich Cluster for Systems Neurology (SyNergy), Munich, Germany   Background: Anemia is common in patients with stroke and has recently been shown to be a risk factor for poor functional outcome and mortality after stroke. However, the impact of anemia on cognitive outcome after stroke remains unexplored. Methods: 252 consecutively recruited patients with acute ischemic stroke and without pre-stroke dementia were included in this prospective ob- servational study. Anemia was defined by the WHO criteria (hemoglobin concentration, 13 g/dl for men and 12 g/dl for women). Blood samples were taken on the first morning after admission between 6 and 9 AM. Cognitive outcome was assessed by theTelephone Interview for Cognitive Status (TICS) 3 months post-stroke. Ordinal logistic regression was used to adjust for confounders. Complete case (CC) analyses as well as sensitivity analyses (inverse probability weighting (IPW), multiple imputation (MI)) were performed to control for loss to follow up. Results: 20% of the patients had anemia. 36% of patients with anemia and 14% of patients without anemia had missing TICS data. The proportion of subjects with missing TICS data was higher in patients with poor function- al status and poor functional status was associated with poor cognitive status. Anemia was associated with poor cognitive outcome in CC analy- sis (n=206; OR=4.19, 95%CI: 1.40-12.55) and results were confirmed using

Pages Overview