Estimating Standard Errors of Parameters Obtained by the EM-Algorithm

In this thesis, two different methods of standard error estimation when using the EM-algorithm will be compared. One of these methods is a long existing and commonly used method called Louis' method. The main idea of this method is to use the first and second derivatives of the complete-data log likelihood to obtain an estimate for the observed information matrix. The second method estimates the observed information matrix by taking the negative of the second derivative matrix of the incomplete-data log likelihood and has to my knowledge not been proposed in earlier research. A simulation experiment shows that the standard error estimates by the second method are generally closer to the simulated standard error estimates when the EM-algorithm is used to estimate parameters of a aussian mixture model with two components.