Heard at JSM
On dropping missing data
I'm at JSM, which is always a fantastically rich event. One is awed by the variety and breadth of topics over hundreds of sessions of talks.
Each year, someone will mention a pet peeve. The other day, a speaker explains why "practitioners" often decide to drop missing data instead of imputing them.
By his observation, most real-world datasets have tons of missing values. Many variables may be up to 40% missing. Given this practical fact, he thinks it obscene to impute that amount of missing data. You'd be making a lot of assumptions.
The flip side of the argument is that when you don't impute, and simply drop the units with missing data, you do not make any assumptions, and most importantly you do not "fabricate" any data.
Sure, imputation requires making some assumptions, like any statistical procedure does.
But dropping missing data also makes one assumption.
And this one assumption is usually the worst possible assumption one could make!
The assumption is that the units with missing data are exactly like the units without missing data. In other words, when you drop the missings, you assume that the missings are the same as the non-missings.
In most practical datasets, data are missing for reasons. Thus, the assumption that they are absent for no special reason is bad.
P.S. In case it's not clear, let's dissect why dropping missing data makes this assumption. After you drop the missing data, you proceed to analyze the remaining data as if they are your full dataset. You then take the learning from that dataset and generalize them to your entire population (which includes the units with missing data). So you have just assumed that those with missing are the same as those without.
