Consider weighing a gold nugget 100 times on a pan balance, a prototypical
example of repeated measurement. It almost goes without saying that the purpose of
weighing the nugget is to determine its weight. But how does one deal with the fact
that the observed weight varies from trial to trial? We assume that statisticians and
nonstatisticians alike would regard these fluctuations as resulting from errors in the
measurement process. But given this variation, how should we use the 100
measurements to arrive at the object’s weight? Should all the measurements be
used? Perhaps not, if they are all not equally accurate. A novice might attempt to
deal with this question by trying to separate the 100 measurements into two classes:
those that are truly accurate versus those that are not. The problem then becomes
how to tell which observations are truly accurate, because the actual weight is not
known.