Environmental Engineering Reference
In-Depth Information
treatment. Because successive measurements are relative to and correlated with the
initial measurement, distinctive statistical analyses are necessary to account for the
correlation among successive measurements of experimental units due to violation
of the assumption for most analyses of independence among experimental units.
The use of randomization and replication in a study design determines the target
population and extent of inference from conclusions. These tools also maintain
study integrity and scientific objectivity (Morrison et al.
2001
). Study designs can
be distinguished by the rules used to govern randomization and replication. In
addition, the application and extent of randomization and replication are critical
factors in judging the reliability of conclusions from studies.
Randomization
,
according to Fisher (
1935
), is at the heart of experimental design. Randomization
refers to both (1) random selection of representative study units for sampling and
(2) random assignment of treatments to experimental units in an experiment. Most
traditional experimental designs are based solely on the rules for randomization
(Kirk
1982
). The extent of inference is directly related to the degree of random
sampling, which ensures that the study units are representative of the target
population. Failure to randomly assign treatments to experimental units increases
the potential impacts of nuisance variables leading to spurious results and potential
bias. Underlying statistical theory demonstrates that randomization ensures that
estimates of treatment effects and experimental error are unbiased estimates of their
respective population parameters. Random assignment of treatments to experimen-
tal units is necessary to satisfy the assumption that experimental errors are indepen-
dent by minimizing the effects of correlation between experimental units on
statistical results. Cox (
1980
: 313) summarized the importance of randomization
to studies, in that “randomization provides a physical basis for the view that the
experimental outcome in a given study is simply one of a set of many possible
outcomes. The uniqueness of the outcome, its significance, is judged against the
reference set of all possible outcomes under an assumption about treatment effects,
as such effects are negligible. For the logic of this view to prevail, all outcomes
must be equally likely, and this is achieved only by randomization.” In wetland
studies, true randomization within a target population can be difficult. The reasons
for this complicatedness are numerous, but include denial of access to study sites,
environmental conditions (e.g., drying of wetland when studying aquatic
invertebrates), equipment placement requirements (e.g., insufficient water depth
to measure water quality, flooding potential of deployed equipment), lack of
defined boundaries identifying experimental units, sampling logistics (e.g., travel,
time to collect samples), and presence, or suspected presence, of organisms of
interest (e.g., certain amphibian species, habitats used by certain avian species).
Replication
is the necessary practice of using more than one experimental unit
for each treatment. Replication is required to minimize the uncertainty of conclud-
ing that differences among treatments are due to treatment effects rather than
inherent differences among experimental units or due to random chance. Replica-
tion is required to measure the variation among and within treatments to make
conclusions regarding treatment effects, which is the basis for drawing inference
about treatment effects using traditional, univariate statistical techniques such as
