Geoscience Reference
In-Depth Information
can be performed during a graduate student's normal tenure. Research stud-
ies usually address an immediate problem or crisis or inform an impend-
ing management decision. When that problem, crisis, or decision is resolved,
motivation to continue the study drops. For this reason, a spatial design that
provides the highest possible statistical precision to estimate a single param-
eter in the shortest amount of time is usually paramount.
In contrast, a key characteristic of most monitoring studies is that they
are not designed to answer a specific question. The objectives of monitoring
studies are usually summarized as watching an environmental resource.
For example, objective statements for monitoring commonly contain phrases
like “to estimate current status and detect trends.” As such, and especially
when compared to research studies, the objectives of monitoring studies
can seem vague. Monitoring studies tend to be long in duration, typically
10 to 30 years, because annual variation of parameters is usually large and
trend detection requires an extended period. Monitoring studies also tend
to be large scale in their geographic scope, where large means that the study
incurs significant travel between sample sites. Monitoring studies also tend
to be funded by agencies with management authority over the resource.
Spatial designs used in monitoring studies generally cannot focus on a
single variable or objective, while designs used in research studies can be
optimized to provide the highest possible precision for a single parameter.
Real-world monitoring studies almost always have multiple objectives, and
it is difficult to optimize the placement of locations for estimation of mul-
tiple parameters. For a monitoring study to survive one to three decades,
its spatial design needs to be robust enough to provide high quality data on
multiple parameters. It must be easy to implement and maintain, and it must
be able to provide data on unforeseen issues when they arise.
Unfortunately, much of the statistical literature on spatial design is rel-
evant only when interest lies in a single parameter. Experimental design
topics such as those by Steel and Torrie (1980) and Quinn and Keough (2002)
discuss techniques like blocking and nesting as ways to maximize the ability
to estimate a treatment affect. Stratification, discussed in Chapter 2, is usu-
ally thought of as a technique to improve precision, but it generally works
only for a single parameter. Maximum entropy (Shewry and Wynn, 1987;
Sebastiani and Wynn, 2000), spatial prediction (Müller, 2007), and Bayesian
sampling methods (Chaloner and Verdinelli, 1995) are designed to locate
sites in a way that maximizes a criterion (such as information gain). But, the
criteria used by these designs are functions of a single variable; thus, these
procedures optimize for one parameter at a time. Focusing on one parameter
is generally undesirable in monitoring studies because near-optimal designs
for one parameter are suboptimal for other variables unless spatial correla-
tion is strong.
