Comparison of Multiple Imputation models with two endpoints: Simultaneous imputation against separate imputation models
Specialeforsvar: Sylwia Agnieszka Piasecka
Titel: Comparison of Multiple Imputation models with two endpoints: Simultaneous imputation against separate imputation models.
Abstract: This thesis compares two approaches to Multiple Imputation (MI) in the case of two incomplete variables. One is to impute the values simultaneously with a two-dimensional MI model using Fully Conditional Specification (FCS). The other is to impute the two variables independently using a univariate MI model. Moreover, it compares the models depending on the correlation between the two incomplete variables.
In the thesis, we present data from a clinical trial conducted by Novo Nordisk. In the study, missing values were imputed using univariate MI. What we do, is additionally use a two-dimensional approach as an alternative method. Next, we fit a linear model to the data imputed by the two MI models. After that, we compare the estimates obtained by the two models. Since we do not know the underlying values of the missing data, we do not know which estimates are more accurate. Therefore, we execute a set of simulations to compare the two MI models in terms of the accuracy of the latter analysis.
What we find is that when the two incomplete variables are highly correlated then a two-dimensional MI model gives more accurate results. However, it has a slightly worse computational time than the other method.
Vejledere: Helle Sørensen,
Censor: Sören Möller, SDU