Depends on Measurement Framework:
Basic distinction: controlled versus measured
Sinton's (1978) framework for spatial information time,
space and theme, one must be fixed, one "controlled"
in order to measure the other.
Error in control is different from error in a measurement. Different
frameworks have different expectations (about completeness, etc.)
Isolated Objects:
Well-defined points (as in nautical chart), the name of the point
(identity as a "feature") is easy to determine, but
is it indeed the correct one? The position is measured for the
object "Point No Point Lighthouse"
Some of the extra attributes (not the ones used to determine
identity) may operate in a reversed mode where the spatial object
is the control with a measurement attached (as in choropleth maps).
The frequency of the radio beacon at Point No Point may be an
attached attribute.
Testing well-defined points: one to one correspondence, yes/no
answers for categories, divergence (difference for continuous
attributes)
Extends (more or less) to line features and area features, but
there are geometric issues that creep in (yes they match, but
incomplete coverage one way or other)
Choropleth maps (collection zones): spatial object came first
(control), attribute attached to object. The spatial averaging
of the collection zone affects accuracy (modifiable units problem).
Testing choropleth maps: zone by zone? as above
Surfaces
continuous attribute measured as general intent, but modifications
are used in practice.
Contours: attribute CONTROLLED, position measured. Note that
surveying standards mix it up with positional accuracy (since
it can be measured in same metric.)
Sample points can be compared with the value recorded on the
surface. The difference can be recorded and treated like a "distance"
(as with position).
Coverages
attribute controlled, position measured.
Point sampling method: Fitzpatrick-Lins, Rosenfield, remote sensing
reported as misclassification matrix for sample points and %
correct
(Sub topic: what does %correct really tell you...)
Polygon overlay alternative
reported as misclassification matrix of areas (not samples)
see Chrisman and Lester for further details