| dc.description.abstract | Relational Event History (REH) data is a valuable type of social network data due to
its increasing availability and precise time resolution. However REH data is especially vulnerable to
missing data, a small proportion of missing values lead to biased estimates which makes valid
inference impossible. Traditionally missing values in REH data were handled using complete case
analysis (CCA), which does not solve these issues. A potential solution can be using multiple
imputation (MI), but research incorporating this method is scarce. This paper aims to continue on
the work of an earlier simulation study by testing the effectiveness of MI in handling more complex
missing values patterns in REH data.
True estimates are produced by performing relational event model (REM) analysis on a
fully observed REH dataset. REH datasets with missing values are simulated from the full dataset
by introducing missingness according to a variety of MAR simulation settings. These datasets are
then analyzed using MI and CCA. The estimates produced separately by the two methods are
compared to the true estimates, assessing whether they produced valid inferences.
MI produced unbiased estimates in most simulations, but in higher proportions of
missingness some of the estimates were biased. Variance of the estimates was underestimated,
resulting in low coverage rates and false significant p-values. CCA produced biased estimates in all
simulations.
The study provided further evidence that MI can be an effective solution to the missing
data problem in REH data. Valid inference was achieved in most MAR simulations, but did show
reduced effectiveness for certain analyzed statistics and in higher proportions of missingness.
Future simulation research should focus on investigating their effects on MI results further. | |
