Biology Reference
In-Depth Information
Max Kuhn, and Steve Weaston (2009)
,
allows for automatic report genera-
tion using LaTeX/OpenOffice in literate programming fashion. Also HTML
pages containing text, graphics, and tables of the results can automatically be
generated from R using, for example, the package
R2HTML
(Lecoutre (2003)
or
hwriter
Pau (2009)). Altogether, using the command
R> demo(biosurvbook)
the analyses of this chapter can be reproduced after the package has been
loaded.
We introduced the time-varying negative binomial CUSUM as an alterna-
tive to the Farrington aberration detection method, as it is better embedded
within the framework of statistical process control. A shortcoming of the
suggested GLM modeling to determine in-control and out-of-control values
is that any uncertainty of the estimation was ignored when plugging in the
estimators for
μ
0,
t
and
μ
1,
t
into the CUSUM. Furthermore, no auto-correlation
between observations was taken into account—neither in the GLM model
nor in the likelihood ratio-based CUSUM. However, if trend and seasonality
are adequately modeled, little auto-correlation is expected to remain as, for
example, shown in the simulation study by Farrington et al. (1996). If auto-
correlation is a concern, different modeling strategies can be applied: gen-
eralized estimating equations (used for mortality modeling in, for example,
Fouillet et al. (2008); integer auto-regressive models (Freeland and McCabe
2004, Held et al. 2005, Wei 2007), or pairwise likelihood models (Varin and
Vidoni (2006)). An auto-regressive approach is, for example, implemented in
the function
glrnb
by using the control argument
change = “epi”
(see
Höhle and Paul, 2008 for details on the methodology). The same reference
also discusses how to estimate the out-of-control state κ at each time point
using generalized likelihood ratio CUSUMs instead of a fixed prior specifi-
cation. An alternative to the independence assuming likelihood ratio-based
CUSUM is the Shiryaev-Roberts detector, which also works for auto-corre-
lated observations (see for example Frisén (2003) for details). As an example,
the spatio-temporal cluster detection of Assunçáo and Correa (2009)—imple-
mented as function
stcd
in
surveillance
—uses this detector. Further
package developments are the extension to categorical time series, for exam-
ple, the monitoring of binomial and multinomial data.
With respect to the Danish mortality monitoring, the presented analyses
illustrated the potential of using
surveillance
and R for this task since
they provide methods for the visualization, modeling, and aberration detec-
tion. A big advantage of the regression-based models in the CUSUM detection
is their flexibility for extending them with additional covariates as illustrated
by population size. Such covariates could for example, be the number of influ-
enza-like illness cases or temperature. A limitation of the current methods
is that mortality reporting is governed by a delay between the day of death
and the reporting to health authorities. Quantification and handling of such
