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


48 ISCB 2014 Vienna, Austria • Abstracts - Oral PresentationsTuesday, 26th August 2014 • 11:00-12:30 Monday25thAugustTuesday26thAugustThursday28thAugustAuthorIndexPostersWednesday27thAugustSunday24thAugust Tuesday, 26th August 2014 – 11:00-12:30 Invited session I5 Prediction to support clinical decision making Organizer: Alessandra Nardi I5.1 What price Cox regression? Ranking predictions from semiparametric and parametric hazard regression models via focused information criteria NL Hjort1 1 University of Oslo, Oslo, Norway   Consider survival data of the familiar kind, involving possibly censored sur- vival times along with a covariate vector for each individual. The standard methods for analysing such data are based on Cox´s proportional hazard regression model, involving an unspecified baseline hazard function mul- tiplied with a log-linear component, featuring regression coefficients etc. I shall explore fully parametric alternatives to Cox´s semiparametric model, where also the baseline hazard is modelled parametrically. I give results detailing how much the semiparametric methods may lose in terms of precision of estimates of the most relevant parameters, like survival curves and median survival time for given covariates, and also develop model and variable selection methods via suitable focused information crite- ria (FIC). This is different from the FIC apparatus developed in Hjort and Claeskens (JASA, 2006), in that these previous methods concentrated on selecting covariates inside the semiparametric regression framework; the present task also involves comparing and ranking models of both semipa- rametric and parametric types. The methods are being illustrated on real survival data. I5.2 Individualized predictions of event times using joint longitudinal-survival models JM Taylor1 1 University of Michigan, Ann Arbor, United States   Following radiation therapy treatment for prostate cancer patients are monitored by regular measurements of prostate specific antigen (PSA), a simple blood test. Increasing trends in PSA are suggestive that the cancer may be regrowing and that clinical recurrence of a detectable tumor may be imminent. The patient may choose to start hormone therapy if the risk is perceived to be high. Thus for each patient it would be useful to be able to calculate the risk of clinical recurrence in the next short period of time under two conditions, either do start hormone therapy or don’t start hor- mone therapy. Using a large training dataset we build a joint longitudinal model for the PSA values and survival model for the clinical recurrences. The longitudinal model involves random effects and the survival model involves a proportional hazards model with PSA as a time-dependent co- variate. Markov chain Monte Carlo methods are used for estimation. To provide individualized predictions for a new patient, the posterior distri- bution from the training dataset is used as a prior for the data from the new patient. Calculation of the probability of recurrence in the next three years for the new patient involves a second MCMC algorithm. In this talk I will discuss estimation from the training data and of the predictions, how validation might be performed and the interpretation of a treatment of effect hormonal therapy in this setting. I5.3 Model selection and ensemble predictive performance R Henderson1 1 Newcastle University, Newcastle, United Kingdom   With the increase in size of data sets now routinely available for analysis, traditional concepts such as significance tests or confidence intervals have become less central. Tiny differences are statistically significant and con- fidence intervals can be very narrow. Instead, as statisticians we need to think about practical differences between competing models say, which can only be done in close collaboration with subject specialists. It also means that the role of predictive performance is attracting more atten- tion. This talk considers predictive performance and model comparison in the area of event history analysis, including standard survival analysis as a spe- cial case. We look at measures of fit for complex repeated event data and examine the role and potential for ensemble prediction, as used in meteo- rology.There is no interest in model selection per se, but rather in combin- ing predictions from competing models, with an appropriate weighting and a carefully chosen performance measure. We will consider prediction in the context of dynamic models in which subject-specific non-predictable histories are informative for the future. A particular question in event history analysis is the length of follow-up needed before heterogeneity between patients can reliably be quantified.

Pages Overview