Missing data and multiple imputation: A sampling study with the SAGA cohort

Background: Missing data in epidemiological research is a common occurrence where there is no method that conclusively performs best. The aim of this thesis is to compare selected methods on the Stress-And Gene-Analysis (SAGA) cohort. Methods: Using: Complete case analysis (CCA), Single imputation using predictive mean matching (SI-PMM), Multiple imputation using predictive mean matching (MI-PMM) and Multiple imputation using the default methods from MICE (MI-MICE) on the SAGA cohort, and fitting a Poisson model with robust error variance on a binary outcome, the estimates and confidence intervals were compared. Following this, a sampling study was conducted by drawing differently sized (1.000, 5.000, 10.000, 15.000 and 20.000) random samples from the complete cases of the data and imposing different levels of missingness, i.e. 5%, 10%, 25%, 50% and 75%. The same missingness pattern from the whole data was used to create the missingness on five covariates under the Missing at random (MAR) mechanism. Here the percentage bias and coverage was compared. A brief look into different strategies of choosing auxiliary variables was conducted as well as looking at how the results differ under Missing not at random (MNAR) mechanism. Results: The different methods performed similarly on the whole SAGA cohort. From the sampling study, CCA performed worse than the imputation methods with higher bias and standard errors. MI-PMM performed marginally better than MI-MICE where the bias on average stayed below the 5% benchmark in all sample sizes except 1.000. Conclusions: MI-PMM is recommended since it is user friendly, fast and resulted in the lowest amount of bias.

Author dc:creator

• Marín Dögg Bjarnadóttir 1995-

Contributors dc:contributor

• Háskóli Íslands

• Líftölfræði

Language dc:language.iso

en

Handle dc:identifier.uri

http://hdl.handle.net/1946/45867

OAI identifier oai:identifier

oai:skemman.is:1946/45867