Assajos TFM's de Susana Pérez i de Juan Vicente Torres

Seminari GRASS - 22 de juny de 2010

PROGRAM:

10:00-10:45 Assaig TFM Susana Pérez

> TÍTOL: Tests of Genetic Association for Quantitative Traits Based on Both Mean and Variance
> RESUM: (Detallat al final del e-mail)

10:45-11:30 Assaig TFM Juanvi

> TÍTOL: Efficiency of stroke clinical trials with ordinal outcomes: a simulation study
> RESUM: (Detallat al final del e-mail)

11:30-14:00 SEGUIMENT PROJECTE.

Revisió del full Excel: estat de les tasques del projecte. Presentació de'n Carles i el Jorge segons els avenços pel que fa referència a la tasca G.

14:00-15:00 DINAR GRASS

Resums dels projectes

Assaig TFM Susana Perez

> TÍTOL: Tests of Genetic Association for Quantitative Traits Based on Both Mean and Variance

> RESUM: In this master thesis we have described several statistical methods to asses the association between a quantitative trait and genotypes. We have focus on comparing some common approaches that only use trait mean with a new proposed bayesian approach to test this association using both trait mean and trait variance. In summary, we compared three different tests: linear regression, Bayes factors for linear models (BFlm) and a Bayes factor for nested models (BFdv). BFdv allows the mean and the variance to change over genotypes, whereas linear regression and BFlm assume a linear trend in trait mean and constant variance. Both direct and indirect association were evaluated, in addition to perform simulation studies to set the most suitable values for Bayes Factors parameters. We found that BFdv, which uses both trait mean and variance, improves the power to detect association between a quantitative trait and genotypes, but this not applies to all the situations.

Assaig TFM Juanvi (Juan Vicente Torres, Cap de Biometria, Recerca Clínica, SL)

> TÍTOL: Efficiency of stroke clinical trials with ordinal outcomes: a simulation study

RESUM:

Introduction: Stroke is the second most common cause of death and a major cause of disability worldwide. Although at least 178 randomized clinical trials enrolling > 73 000 patients were conducted for 75 promising agents, only 3 trials reported positive findings and only 1 agent has been approved by the FDA.

As a result, several new methods to improve the efficiency of Stroke clinical trials have been developed. These can be grouped in four groups: (1) End point analysis; (2) Shift Analysis; (3) Responder analysis; and (4) Global analysis.

The objective of this study is to assess the power of the most common statistical methods as well as some newer proposals by means of a simulation.

Patients and methods: 571 patients from the placebo arms of 4 real clinical trials were randomly split by study in two groups of equal sizes. These were measured on three common ordinal outcome scales in stroke trials: the Barthel Index (BI), the modified Rankin Scale (mRS) and the National Institute of Health Stroke Scale (NIHSS). A treatment effect size equivalent to ORs of 1.20, 1.25, 1.30 and 1.35 was added following two different approaches: (1) OAST based on the generation of new scores following an ordinal logistic model adjusted by important prognostic variables; and (2) CHOI based on directly improving the outcome a certain number of subjects. The power of different statistical methods was evaluated.

Results: The utilization of the ordinal characteristic of the data proved to be the best way to improve the power of finding a positive result. Then, the utilization of prognostic variables and the incorporation of other scales to assess a global outcome showed also relevant power improvements. The most powerful methods were the ordinal logistic regression adjusted by prognostic variables and the responder analysis based on quartiles.

Discussion: The most powerful method combined the benefit of considering all ordinal scale information together with the adjustment of baseline characteristics. The third improvement was to incorporate the information of other scales to reduce the inter-subject variability. Although this last method was only applied to the dichotomized outcomes, it could be interesting to study the way of conducting such an analysis with overall ordinal outcomes. In order to check the stability of the results, performing both replications in other real databases and simulations under alternative treatment effect scenarios are desirable. A full parameterized simulation, independent of real data, may help to better understand the hierarchy of methods.