🚀 go-pugleaf

> Suppose you are handed a data set with a variable number of multiple obs
> per subject.  Short of eliminating data on a random basis so that each
> ... the data
> do not represent independent observations  ...

Multiple observations seem to me to be a field where the underlying ideas
are often obscured by a desire to impose jargon. Having multiple observations
on the same subjects does not necessarily make those observations not
independent. There are two alternatives: the variate measured has some
"true value" and "distribution" each time it is measured, and each
observation does not affect the value or precision of subsequent ones ---
in this case each subject acts as a "blocking factor", and you can define
a "measurement error" for each subject are for between subjects; OR, the
value having been measured, that value affects subsequent measurements ---
the true "repeated measure" design. There is, I think, a third alternative
which is the experimental scientist who "refines his technique", so that
each measurement affects subsequent ones by providing background information.
This may be susceptible to a formal treatment by a sort of binary-chop
model, but in practice you have complications like the scientist who
discards "bad" measurements (cf. Milliken oil-drop notebooks).

In this example, if each observation within each person might (for practical
or theoretical reasons) be treated as independent, might I ask (as an Aunt
Sally) why you cannot use a weighting for each value of 1/(no of obs for that
person) and fit a line to the full data?

Does true "repeated measures", as defined above, necessarily imply a
"memory" in the system?