Biology Reference
In-Depth Information
Statistical considerations
At the discovery-phase level, recommendations for biomarker identifica-
tion and qualification in clinical proteomics have recently been proposed
to avoid overinterpretation of results, the use of inappropriate technologies
or statistics, and inefficiency in the construction of a multimarker panel
[34]
. For reporting on observational study design and diagnostic accuracy,
two guidelines have been established by the STROBE and STARD initiatives
[35,36]
.
Sample sizes
The number of specimens that should be tested depends on the objective
of the study and the extent of biomarker variability in the study. When the
objective is to select a subset of biomarkers from a pool, the following fac-
tors contribute to variability: the prevalence of the disease and the subtypes
of disease (i.e., skin GVHD, GI GVHD) among the study samples, the capac-
ity of the biomarkers to discriminate among the various disease subtypes,
the number of biomarkers under study, the number of case and control
subjects, and the statistical algorithm used to select promising biomark-
ers. Thus, as suggested by Pepe et al.
[33]
, there are no simple methods for
recommending sample sizes. In particular, traditional sample-size calcula-
tions that are based on statistical tests of hypotheses are not relevant. Pepe
et al.
[33]
propose that computer simulations guide the choice of sample
sizes, meaning that simulations should be performed with the guidance of
investigators on biologically plausible models to generate data. By varying
the numbers of cases and controls, one can assess at what sample size a
reasonable proportion of promising biomarkers is likely to be selected for
further study
[33]
.
457
Receiver operating characteristics (ROC) curves for
diagnosis
Several statistical methods can be used to estimate the diagnostic likeli-
hood ratio of a continuous biomarker, but the ROC curve, an entity used to
evaluate diagnostic biomarkers, is primarily used because the diagnostic
likelihood ratio is mathematically related to the slope of the ROC curve
[37]
.
The ROC curve is a plot of the true-positive rate (sensitivity = 1 − false-nega-
tive error rate) versus the false-positive rate (1 − specificity), which is associ-
ated with rules that classify an individual as “positive” if the marker value is
above a threshold
C
for all possible thresholds
[38]
. In addition, combina-
tions of multiple markers are often required; combining the ROC curves of
all biomarkers is an optimal way of estimating the risk score, defined as the
probability of disease given data on multiple markers, as the ROC curve is
maximized at every cutpoint
[39]
.
Figure 19.3
shows the ROC for aGVHD
diagnostics.
Single-versus-multiple-marker panels in GVHD evaluation
As mentioned above, while several biomarkers exist, none is sufficiently
sensitive or specific on its own for either a diagnostic or a predictive test.
Thus, the simultaneous use of several markers may increase specificity or
Search WWH ::
Custom Search