Statisticians learned long ago that "missing" needs to be treated special. It should be either an entirely separate signal or a "sentinel value" riding on the existing signal that everyone knows and couldn't possibly be a normal operation value. Which is why sentinel values are usually a huge bunch of 9's or the maximum possible value of the field or something.

Interesting that the sentinel value is zero in this case. In data analysis that's usually a terrible sentinel value, but here it's the most practical one.