STUDIES ON MODIFIED NESTED CASE-CONTROL DESIGNS
    
                   Roger Månsson
 

The nested case-control design in a cohort study has appeared
as a way to reduce the amount of covariate data collection,
which may be expensive. Instead of using the whole cohort to
investigate the impact of a certain exposure on the risk of
dying (or getting a disease) it is sufficient to gather
information about the cases, i.e. patients who die, and about a
proportion of those patients who are alive; the latter ones are
used as controls. In a nested case-control design we sample for
each case a number of controls randomly without replacement from
the persons at risk, not including the case.
 
In this paper we have studied modified nested case-control
designs where we have tried to re-use selected controls, as
controls to cases later on. The results indicate that it is
sometimes possible to get better designs, in terms of smaller
standard error of the logarithm of the relative risk, than the
conventional nested case-control design, where new controls are
sampled to each case. Compared with the 1:1 design, i.e. when one
control is sampled to each case, a more efficient design is seldom
obtained by re-using selected controls as additional controls to
cases later on. On the other hand, compared with the 1:2 design, i.e.
when two controls are sampled to each case, effiency can be
maintained although the number of different controls is reduced.
This is attained as follows: after an initial fraction of the
controls have been sampled as usual, one of the two controls to a
case is re-sampled from the previously selected controls;
re-sampling should not be initiated too early. These findings are
considered in an application where the design of a clinical trial
is addressed.