Analysis of Baseline Measures and Time-to-Death Post Stroke; Accounting for Missing Data

 

Anna France, Lancaster University

 

Abstract:

Due to the severity and complexity of stroke, it is very difficult to collect all intended observations on stroke patients, and subsequently stroke data is rarely handled appropriately to account for the missing data. This presentation aims to use Cox proportional hazards modelling to examine the factors affecting survivability post-stroke, and will focus on the following areas:

It will explore the missing data and discuss how multiple imputation using chained equations (MICE) can be adjusted to be compatible with Cox proportional hazards modelling, to develop a multiple imputation framework using MICE to impute the missing data.

Techniques such as Rubin’s rules and the multivariate Wald test will be discussed in order to fit a pooled Cox proportional hazards model to the multiply imputed data, and diagnostics will be carried out to validate the model.

To account for the violation of the proportional hazards assumption, a piecewise-proportional hazards model will be constructed through interactions with a time dependent covariate within the Cox proportional hazards model. To conclude, we will consider how diagnostic techniques can be appropriately adjusted to carry out model validation on the piecewise-proportional hazards model. 

 

======================================================

Sequential Monte Carlo Methods for Epidemic Data

 

Jessica Welding, Lancaster University

 

Abstract:

Epidemics often occur rapidly, with new data being obtained daily. Due to the frequently severe social and economic consequences of an outbreak, this is a field of research that benefits greatly from on-line inference. This provides the motivation for developing a Sequential Monte Carlo algorithm that uses a combination of reweighting and resampling to update current particle values as new information is received. The algorithm constructed is found to be comparable in estimation capabilities to current Markov chain Monte Carlo methods, with the advantage of being easily parallelized thereby allowing for a reduction in computation time.

              

Add to my calendar

Back to listing