182 / 274
8
RECOMMENDATIONS
1. Use of ‘robust’ estimators for measures of center location is non-controversial, as the measure
of centrality is based on central data.
2. Use of ‘robust’ estimators for measures of variation or spread is deprecated, as they will be
substantially biased low.
3. A circumstance in which ‘robust’ estimators of variation might be recommended is when:
a. The underlying distribution is
known a priori
to be normally distributed or substantial
additional evidence (other than the actual data in question) supports this assertion.
b. The observed data are
known a priori
to be contaminated with data a foreign distribution, and
this contamination is exclusively found in the tails of the empirical distribution.
c. The outliers present are
known a priori
to not be identifiable for assigned cause.
This circumstance might arise, for example, in PT data where it can be supposed that
substantially different variants of the method in question may be in use, and these variants cannot
be identified from the information collected in the study. Inclusion of all data in such a study may
result in an estimate of reproducibility standard deviation that is
known a priori
to be much too
large.
4. In all other circumstance, reproducibility standard deviation should be estimated in the usual
way after removal of outliers for assignable cause.
129