| |
After the investigator
has specified the study questions and hypotheses, the next step
is to identify the variables needed (dependent, independent, and
confounding variables) for the analysis.
Measures of interest
often are constructed by combining two or more variables. References
are available to help guide measure selection. (McDowell
and Newell, 1996). Large secondary survey databases allow
the researcher to examine patient-level, provider-level, and area-level
factors that influence health outcomes. Variable selection and
analytic strategies require attention to the survey's inherent
limitations as well as its unique advantages, and differ somewhat
from those used in clinical trials (Hornberger
and Wrone, 1997; MacMahon
and Collins, 2001).
Briefly consider the
differences in study design between clinical trials and observational
studies that result in the need for different analytic methods.
Clinical trials use an experimental design to assess the efficacy
of a therapeutic intervention. Trials design minimizes confounding
of the association of the intervention with the outcome of interest
by (1) randomization, (2) specified inclusion criteria for participation,
and (3) implementation of the study protocol in a controlled experimental
setting. In actual practice, however, many uncontrolled factors
may mediate the effect of an intervention on the outcome of interest.
Patients encountered in practice are more likely to have comorbidities
or other factors that would have led to their exclusion from a
clinical trial, but which need to be considered in clinical decision
making and outcome assessment.
Studies that assess
the outcomes of a given intervention in "real world settings"
are known as effectiveness studies. Effectiveness studies may
be experimental or observational. They can assess outcomes for
individuals with multiple chronic conditions, advanced age and
frailty, who are commonly excluded from clinical trials. In observational
studies, since patients are not randomly assigned to an intervention,
but may choose or receive it based on other factors that may also
be related to the outcome of interest, the research must identify
potential confounders and include these variables in the analyses.
So, while clinical trials primarily use experimental design to
minimize confounding, in observational studies statistical methods
become critical for this purpose. Statistical strategies to account
for unmeasured confounding in observational studies include multi-variate
analysis, instrumental variables, propensity scores and sensitivity
analysis (Hadley
et al., 2000; Rosenbaum,
1991). In addition, hierarchical or multi-level models can
provide insights into the relative contribution of individual
level, provider level, and area level factors (Normand
and Zou, 2002). Click here
for an example of a conceptual framework that can be used to develop
such a model.
|