Exposicions Congrés ERCIM'11 (Londres)

Dijous 15 de desembre

De 10:30 a 14:00h

* 10:30 - 11:15: Klaus

Títol: Linear regression models with an interval–censored covariate. A comparison of different residuals.

Resum:

We consider the analysis of a linear regression model with an interval–censored covariate. For this model, we propose a new definition of residuals and compare their behavior with other proposed residuals for this setting.

1.Introduction: Interval–censored survival data refer to a situation where the event of interest is not exactly observed and the time T to this event is only known to lie between two time points, L and R, that define a censoring interval for T. This kind of censoring is quite usual in longitudinal studies where subjects are followed over time and the event of interest is observed within consecutive visits. An extensive literature for interval–censored data exists, see for instance the review by Gómez et al. (2009). The vast majority of methods for interval censoring consider the interval–censored variable as the response variable. In this work, we consider the situation where the interval–censored observation is a covariate in a linear regression model. In this setting, Gómez et al. (2003) developed a likelihood approach to jointly estimate the regression coefficients of the model as well as the marginal distribution of the covariate

2. Residuals for a linear regression model with an interval–censored covariate: Checking the underlying model assumptions is essential to assure the validity of the inferences and conclusions of a linear regression analysis. These assumptions include a correctly specified regression function as well as independent and identically distributed model errors. Gómez et al. (2003) and Topp et al. (2004) proposed three different kinds of residuals for linear models that incorporate an interval–censored covariate. In this work, we propose new residuals for this setting. The existence of an interval–censored covariate generates interval–censored residuals. The new residuals are defined as their expected value conditional on the observed censoring interval. We compare through simulation the behavior of the new proposed residuals with that of the residuals considered in Topp et al. (2004).

* 11:15 - 12:00: Moisés

Títol: Sample Size and Asymptotic Relative Efficiency (ARE) when using Composite Endpoints.

Resum:

In Randomized Clinical Trials, Composite Endpoints (CE) are often used to assess the efficacy of a new treatment. The decision on whether to expand the study relevant endpoint T1 with an additional endpoint T2 is controversial. To assess the difference in efficiency between using logrank test Z based on  T1  or logrank  test  Z* based on the composite endpoint T*=min{T1,T2}, Gomez and Lagakos  base  the  strategy on the behaviour of the asymptotic relative efficiency (ARE) of  Z*  versus  Z. Given that both tests Z and Z* are asymptotically N(0,1) under the null H0 and H0*, respectively, and  asymptotically normal N(mu,1) and N(mu*,1) under a sequence of contiguous alternatives to the null hypothesis, their ARE=(mu*/mu)^2 .

The goal of this work is to check if the usual interpretation of ARE as the reciprocal ratio of sample sizes needed to attain a given power for a significance level is fulfilled when the two null hypothesis are not the same. To do so, we simulate the joint distribution of (T1,T2) by means of Frank's copula for different marginal distributions,  association degrees, probabilities of observing T1  and T2  and anticipated treatment effects with respect to each endpoint.

* 12:00 - 12:45: Susana
 

Títol: FARMS: A new strategy for model selection

Resum:

Selecting a regression model, when based on large datasets with a big number of covariates, needs efficient methods to pick up the variables to be included in the final model. A variable selection method should find a subset of variables having the optimal prediction performance. Sometimes, this prediction is not optimized during the process of variable selection, and testing for all potential subsets of variables is not possible.

Consequently, suboptimal methods for variable selection are applied and the prediction performance of regression models is estimated separately. We propose a new method that combines Forward variable selection and All subsets Regression for Model Selection: FARMS. In order to explore its properties, we fit different models using common methods and FARMS. We tested also its robustness. We performed these comparisons on a dataset with host genetic and immunological information of over 800 individuals from Lima (Peru) and Durban (South Africa) with HIV infection. This dataset includes around 500 variables with information on HIV immune reactivity (around 400 predictive variables) and individual genetic characteristics (around 100 predictive variables). We compared all the models obtained and confirmed that the application of FARMS proved to be more time-efficient and have better statistical properties.