Choosing the initial sample size

We show in Figure 16.1 what steps must be considered in choosing the sample size for a preliminary adjustment. In turn, these are:

1. The usedata  control. When switched off, actual observations are ignored, and so a fictitious sample size must be set for the adjustment. Otherwise the sample size used is the same as the number of actual observations.
2. The obs  control. This is used to set a fictitious sample size, whenever necessary and assuming the absence of (or ignoring) actual data.
3. The exchangeable  control. This is used to force the use of exchangeable beliefs when the sample size, actual or fictitious, is n=1. (By default, when n=1 [B/D] assumes non-exchangeability for efficiency.)
4. The matchcd  control. This is used to force the use of special algorithms which exploit analytical results to avoid much unnecessary calculation when the default algorithms are used to calculate the adjustments. The setting of the control is relevant only when a predictive or mean component analysis for a pure exchangeable system is taking place. In these circumstances, the adjustments can be arranged so that the canonical directions for the mean and predictive adjustments are the same (except for normalising constants) in the case of rank degeneracy. (Otherwise, the canonical directions are the same anyway, except for normalising constants.) However, note that the setting of this control can, in the case of rank degeneracy, give canonical directions for the mean component adjustment which are valid but different (because of non-uniqueness) to the directions computed by the default algorithm. In particular, if you select the matchcd  control, in the special case of mean component adjustment for an exchangeable system under rank degeneracy, the canonical directions reported for the resolution matrix are not generally the same as those reported when using the default algorithms. However, we recommend that you set the matchcd  control to ON.

These controls are discussed in greater detail in §20.3. As an example, suppose that you wish to generate a preliminary adjustment of B by D (assume that these are also the names of the respective [B/D] bases) where you wish to take initially a sample of size n=1, and where you wish to ignore the actual observations that you have on D. Then the following commands are appropriate:

BD>control  -usedata ,obs =1,exchangeable 

BD>adjust  [B/D]

 
Figure 16.1:  Controls affecting exchangeable adjustments