Imputing data below detection limit

Yes, Jessica, the practice -- of which I also have been and continue
to be guilty -- does not really make a lot of sense. It usually
doesn't affect estimation all that much, but it can certainly mess up
inference. The proper approach is to use the proper approach: model it
as left-censored data. The problem with that is:

1. It's complicated, and is beyond the statistical background of most
folks who deal with such data -- it's a ubiquitous issue in science
and engineering after all.

2. Typically, the LOD isn't: that is, there often is not a well
defined value and that which is chosen is both arbitrary and
inaccurate. What one often sees is an increasing loss of relative
precision as one "approaches" the designated value. Modeling this
effectively gets even more complicated. David Rocke and colleagues has
published methodology on this, mostly in TECHNOMETRICS if memory
serves.

3. So, as in other situations, we muddle along with rather crude
statistical approaches and hope that they are adequate. Probably in
most circumstances they are, but ...

Cheers,
Bert

On Mon, Aug 13, 2012 at 1:15 AM, Jessica Streicher

Imputing data below detection limit

Thread (5 messages)