All Times EDT
Keywords: model misspecification, covariance structure, simulation study
This project is a simulation study of model misspecification when analyzing cardiovascular MRI data observed post myocardial infarction. A covariance structure was designed specifically for left ventricular (LV) data based on clinical observations. Using the American Heart Association's segmentation model, LV data can be segmented into three levels, based on the location of the segments. Observations from the same level were assumed to have correlation depending on the distance between segments while observations from different levels share a common covariance. Rotation of the LV was simulated using the multivariate normal; a fixed-effects model was constructed to compare the rotation of diabetics and non-diabetics after adjusting for the level of LV. We modeled the simulated data using a simple fixed-effects model specifying the LV level and diabetic status as predictors. We examined bias, relative efficiency, power, and choice of working covariance structure via fit indices. Briefly, bias is close to zero for all structures, even when the covariance structure is misspecified and relative efficiency showed that our proposed structure resulted in the smallest standard error in most cases. However, type I error and power are inflated for the exponential and unstructured working structures, thus should not be specified when analyzing data of this type. When examining model fit indices, the proposed structure was chosen 99.90% of the time by both the AIC and BIC while the other structures where chosen less than 1% of the time. Specifying the proposed structure as the working structure, unsurprisingly, give the best results in terms of type I error and power and is chosen as the best fit by fit indices.