For sample sizes greater than one, we must take account of the covariance structure between different observations on the collection D being used for the adjustment. These adjustments then exploit exchangeability. Even when the sample size is one, if you have available all the relevant beliefs, you can force the adjustment to be exchangeable by switching on the exchangeable control. You might do this, for example, if you are concerned with design issues involving a range of plausible sample sizes. In particular, output pertaining to the maximal resolution transform (crudely, what happens when you take an extremely large sample size) are available only for exchangeable adjustments. You may also wish to use special algorithms that are available for fast computation for pure exchangeable (but not general exchangeable) adjustments, that are enabled by switching on the matchcd control. (We recommend that you do so: see §16.3 for further details.)
We suppose that the quantities involved in the collection being used for
the adjustment are of the form described in §16.1, that
is a sequence of vectors 
. In the notation
of that section, it remains to give the source of the belief
specifications G and G+U. The latter is the variance matrix over
each such vector 
, and the former is the covariance matrix
between a pair of vectors. That is,
The infovar control should point to the belief store containing G+U, and the modelvar control should point to the belief store containing G. As above, the prior variances supplied, together with any observations over D, must be coherent. The COHERENCE: command can be used to perform coherence checking.