Biology Reference
In-Depth Information
40
µ
0
µ
1
NNBA
30
20
10
0
2007
2008
2008
2008
IV
II
III
IV
Time (weeks)
Figure 12.8
Negative binomial CUSUM for the <1 group. The interpretation of the lines shown is as in
Figure 12.6.
Based on 1000 realizations of
I
(
T
A
≤ 6 5|τ = ∞), the probability is estimated to be
0.57, which is surprisingly high compared to the rough estimate of 0.28 in Section
12.4. However, the two numbers are not completely comparable as the simula-
tion uses a negative binomial model and observations are not independent. If
the above Monte Carlo estimated false-alarm probability of the Farrington algo-
rithm should be near 10%, we would have to choose a much smaller α. Instead,
we use the Markov chain approximation to determine that a threshold of
g
≈
2.2 gives a similar probability for the negative binomial CUSUM. Figure 12.8
contains the result of the CUSUM monitoring with this threshold.
The CUSUM behaves slightly different than the Farrington algorithm in
Figure 12.3. In the last weeks of 2007, an increased number of cases above
the baseline is accumulated leading to a steady decrease of NNBA. In week
01, the threshold is nearly reached, but as for the Farrington procedure, an
alarm is first generated for week 02 in 2008. However, the sustained excess
above baseline leads to a further alarm in week 08, which was not detected
by the Farrington algorithm, as here, the excess alone in that week is not
enough to get beyond the threshold.
12.6 Conclusions
In this chapter, we have given an introduction to the capabilities of the
open-source R package
surveillance
for epidemiological biosurveillance.
Further advantages of choosing R to conduct such analyses exist: R pro-
duces high-quality graphics in a variety of formats, including TIFF, PNG,
EPS, and PDF which, combined with Sweave or odfWeave Leisch (2002),
