I just think that this MAR and MCAR are much more common than we normally think. The patterns that you would normally expect to see in the bad kind of missingness, the MNAR [Missing Not at Random] kind of missingness, are a lot more complex than we think. They're not nearly as common as what we might imagine, and certainly when people say, 'Well, what if the cause of missingness is MNAR?' I would argue that that statement is never true. There might be a study out there somewhere for which it is true, but I've never seen a study for which that statement is ever true, that the cause of missingness is MNAR. Maybe a cause of missingness is MNAR, but never the cause. There are a lot of causes of missingness. There are a lot of reasons why if you're looking at a particular variable and you have different missing values on that particular variable, there are a lot of reasons why that variable will be missing, why different data points will be missing. In fact if you look at a single data point, a single subject that doesn't give you data for a particular variable, even that one point may have multiple causes of missingness. There may be several reasons why that particular data point is missing. It's not really an easy thing to say that what if the cause of missingness is MNAR.

« Back to Article