Significance testing of fixed slopes in a pattern submodel missing data strategy

Blume and Mercaldo (2020) proposed an approach to handle missing data: pattern submodels (PS). It is a set of submodels for every missing data pattern that is fit using data from that pattern. The present study focused on comparing the predicted data using PS with the actual data that were generated from a simulation. Three dimensions were manipulated: the correlation between predictor and outcome, the missing data mechanism (the relationship between the variable and the missing data), and the missing data rate. The difference between the predicted slope and the actual slope, the difference between the standard deviation of the sample data and the standard deviation of the population data, and the rejection rates (the rate of rejecting the preset claim) under different manipulated conditions were examined. This study revealed that PS did not work when missingness rates were too low or too high. The difference between the predicted fixed slope of X₁ (with missing data) and the actual fixed slope of X₁ was very small. But the difference between the predicted fixed slope of X₂ and the actual fixed slope of X₂ (without missing data) was large. The study also examined the difference between the predicted outcome and the actual outcome using PS and found that the results were better when the missing data was unrelated to other variables and with less missingness. Overall, this study does not recommend using PS to deal with missing data.