The Structure of Objectivist Studies - Evaluation Methods in Biomedical Informatics - page 105

assert that these effects are causal rather than merely coincidental. The ran-
domized clinical trial, as described in the literature of clinical epidemiol-
ogy, 15 is the consummate demonstration study.
It is not always possible, however, to manipulate the environment to the
extent required to conduct such a pure experiment. In some study settings,
the investigator may be compelled to forego random assignment of subjects
to groups, and thus conduct what is called a “quasi-experiment.” For
example, health professionals who work in unionized organizations may
refuse to be randomized into studies, insisting instead that they be able to
choose the alternative that they believe in advance is best suited for them.
Even when the results of a quasi-experiment reveal a difference between
groups of subjects on the dependent measure, the source of the effect
cannot be isolated as the independent variable. If the subjects selected
themselves into groups instead of being randomly assigned, some unknown
difference between the groups' members may be the causal factor.
Correlational Studies
In other cases, investigators conduct correlational studies that explore the
hypothesized relationships among a set of variables the researcher mea-
sures but does not manipulate in any way. Correlational studies are guided
by the researcher's hypotheses, which direct the choice of variables included
in the study. The independent variables are the hypothesized predictors of
an outcome of interest, which is the dependent variable. Correlational
studies are linked most closely to the “comparison-based” and “decision-
facilitation” approaches to evaluation discussed in Chapter 2. Correlational
studies are also called observational, retrospective, or ex post facto studies.
So-called data-mining studies that seek to extract interesting relationships
from existing datasets are a form of correlational study. Data-mining studies
are becoming increasingly common in both clinical and biological applica-
tion domains. 16,17 Outcomes research and case-control studies in epidemi-
ology can be seen to fall into this category as well.
In informatics, an example of a correlational study is one in which the
researcher analyzes the extent of use of an information resource (the
dependent variable) as a function of the clinical workload and the senior-
ity (two independent variables) of each care provider (the subjects) in a
hospital. In this study, the values of all the variables are properties of the
subjects that cannot be manipulated and are studied retrospectively. As
another example of a correlational study, the proportion of physicians in a
hospital who enter their own medication orders may be followed over time.
If an increase in this proportion is observed soon after a new user interface
is implemented, the investigator might argue that the change in user inter-
face was the cause of the increase. Although such an uncontrolled study
cannot be the basis of a strong inference of cause and effect, correlational
studies can be highly persuasive when carefully conceived and conducted.