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


64 ISCB 2014 Vienna, Austria • Abstracts - Oral PresentationsWednesday, 27th August 2014 • 9:00-10:48 Monday25thAugustTuesday26thAugustThursday28thAugustAuthorIndexPostersWednesday27thAugustSunday24thAugust C34.3 Case-wise diagnostics for the multinomial logit- link regression model L Blizzard1 , DW Hosmer2 , S Quinn3 , JD Canary1 1 Menzies Research Institute Tasmania, Hobart, Australia, 2 University of Massachusetts, Amherst, United States, 3 Flinders University, Adelaide, Australia   For nominal outcomes with more than two attributes, odds ratio estimates are obtained by fitting a multinomial logistic regression model. Several summary measures of goodness-of-fit provide a global test of the adequa- cy of a fitted multinomial model, and a variety of diagnostic quantities can be used to identify observations that influence the estimated coefficients of the fitted model and/or its predicted probabilities. These case-wise di- agnostics are a natural adaption of those proposed by Pregibon (1981) for the binary logistic model, and were extended to the multinomial logistic model by Lesaffre andAlbert (1989). The multinomial diagnostics have not been implemented in statistical packages. Hosmer and Lemeshow (Hosmer, Lemeshow and Sturdivant, 2013) continue to recommend that until they are, the fit of a multinomial logistic regression model can be investigated by assessing the fit of sepa- rate binary models fitted to the data. Whilst this approach due to Begg and Gray (1984) is generally sound, we show by demonstration and data simulations that it may fail to detect a lack of fit that would otherwise be revealed by the multinomial diagnostics. The use of the multinomial diagnostics in combination with graphical plots is demonstrated, some troublesome cases in which different diag- nostics provide conflicting results are highlighted, and guidance in the interpretation of their values is offered with tentative guidelines for iden- tifying outlying and influential observations. C34.4 Interpreting small differences in mean z-scores in sick populations: does dichotomisation help? J Peacock1 , O Sauzet2 , J Lo1 1 King’s College London, London, United Kingdom, 2 Universität Bielefeld, Bielefeld, Germany   Background: A recent study in ex-preterm children, ie ‘sick’ individuals, observed a small difference in lung function mean z-scores which though statistically significant was of uncertain clinical importance. However, the corresponding difference in proportion at high risk was substantial. We explore this issue by comparing effects of a small shift in mean z-score in ‘normal’and‘sick’(Rose IJE 1985) populations. Methods: We assumed a comparison of sample mean z-scores in two ‘normal’ (mean z near 0) and two ‘sick’ populations (mean z near-1) simi- lar to data observed. We defined high risk as z-score<-1.96, <-1.64, <-1.28 (2.5th ,5th ,10th centiles respectively in a ‘normal’ population). We used a distributional approach to calculate proportions at high risk assuming z- score was Gaussian (Peacock StatMed 2012). This approach provides dif- ferences in proportions with the same precision as differences in means. Results: For ‘normal’ populations a difference of 0.25 in mean z-score equates to a difference of 1.9 percentage points in individuals with z- score<2.5th centile. This contrasts with a difference of 7 percentage points for ‘sick’ populations. Results for percentage of individuals <10th centile show a small mean difference of 0.25 equates to a large difference of 10 percentage points in‘sick’populations. Conclusions: Small differences in mean z-scores equate to larger differ- ences in proportions at high risk in‘sick’compared to‘normal’populations. Hence reporting means alone is potentially misleading; we recommend a dual approach reporting differences in means and proportions at high risk calculated using the distributional approach.   C34.5 Confidence bounds for monotone dose-response relationships C Baayen1 , P Hougaard1 1 H. Lundbeck A/S, Valby, Denmark   An important aim of drug trials is to characterize the dose-response rela- tionship of a new compound. Such a relationship can often be described by a parametric (non-linear) function that is monotone in dose. To estab- lish proof of concept, or find the minimal effective dose, it is of interest to know the uncertainty of the estimated dose-response curve. It is well known that Wald confidence intervals are based on linear approximations and may be unsatisfactory in nonlinear models.They can be unreasonable in the sense that the lower confidence limit of the difference to placebo can be negative even when the overall test shows significant positive ef- fect under a monotonicity assumption. In nonlinear models, profile like- lihood based confidence intervals for the parameters have been shown to have better coverage. In this work we use a similar approach to com- pute confidence intervals for the dose-response curve. These confidence bounds have a more reasonable shape (as function of dose) than Wald confidence intervals. Finally, the method is robust when there is poor in- formation (few doses, or irregular choice of doses) for estimating the dose response curve. C34.6 Simpler is better: a comparison of methods for construction of fetal reference charts D Nevo1 , M Mandel1 , E Ein-Mor2 , O Chen2 , E Daniel-Spiegel3,4 , S Yagel2 1 The Hebrew University of Jerusalem, Jerusalem, Israel, 2 Hadassah University Hospital-Mount Scopus, Jerusalem, Israel, 3 Ha’Emek Medical Center, Afula, Israel, 4 Technion Israel Institute of Technology, Haifa, Israel   Reference charts for fetal measures have been developed over the years in order to estimate gestational age and fetal weight. These reference charts have also been used for early detection of pregnancies that should be monitored closely, since values in the tail of measure´s distribution are as- sociated with fetal defects and disorders. Construction of reference charts is essentially an estimation of quantiles of a distribution as function of the gestational age. Existing methods were developed under various modelling assumptions, typically by fitting a polynomial regression for certain functionals of the distributions (e.g., mean, standard deviation, quantiles). We relax the as- sumptions of a parametric polynomial link between the distribution pa- rameters and the age. We consider nonparametric regression and discreti- zation of the age in order to allow more flexible models. We use a large cross sectional data with repeated measures to compare between the various existing and suggested methods. The question of homogeneity of reference charts is also of interest. Curves built using the same method but using data from different hospitals, located 100km from each other, are compared. We conclude that simple methods should be preferred, provided enough data is available, and that reference charts should be constructed sepa- rately for different subpopulations.  

Pages Overview