Rob van Hest Capture-recapture Methods in Surveillance - RePub ...
Rob van Hest Capture-recapture Methods in Surveillance - RePub ...
Rob van Hest Capture-recapture Methods in Surveillance - RePub ...
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
General discussion<br />
the screen<strong>in</strong>g programme. Often cross-validation of the population estimates from<br />
truncated models is not possible. However, we could compare our estimates with an<br />
<strong>in</strong>dependent assessment of the number of problematic illicit drug users <strong>in</strong> Rotterdam <strong>in</strong><br />
2003 established through two-source capture-<strong>recapture</strong> analysis that used a similar casedef<strong>in</strong>ition<br />
of the target group. These capture-<strong>recapture</strong> estimates were comparable to<br />
those of the truncated models. In the context of its ad<strong>van</strong>tages and limitations, we<br />
conclude that the use of truncated population estimation models is a feasible and valid<br />
method for estimat<strong>in</strong>g the coverage of a public health <strong>in</strong>tervention programme among<br />
hidden populations.<br />
A more detailed study of the validity of truncated population estimation models<br />
<strong>in</strong> <strong>in</strong>fectious disease surveillance compared the performance of some truncated<br />
population estimation models with three-source log-l<strong>in</strong>ear capture-<strong>recapture</strong> analysis<br />
us<strong>in</strong>g data from published and current capture-<strong>recapture</strong> studies on <strong>in</strong>fectious disease<br />
<strong>in</strong>cidence (chapter 10). This comparative research was triggered by the results of the<br />
studies described <strong>in</strong> chapter 6 and chapter 8, <strong>in</strong>dicat<strong>in</strong>g that conventional three-source<br />
log-l<strong>in</strong>ear capture-<strong>recapture</strong> models sometimes break down and produce unexpected and<br />
implausible results. Solely rely<strong>in</strong>g on three-source capture-<strong>recapture</strong> analysis seemed<br />
<strong>in</strong>appropriate and we perceived that a tool is needed to cross-validate capture-<strong>recapture</strong><br />
estimates for <strong>in</strong>fectious disease <strong>in</strong>cidence. Ideally, this would be a model robust to<br />
violation of all capture-<strong>recapture</strong> assumptions but such models do not exist. We used<br />
truncated models because they performed well when compared to log-l<strong>in</strong>ear capture<strong>recapture</strong><br />
analysis earlier. 30 The comparative research was feasible as truncated models are<br />
easy to apply and can be used on the data of three-source capture-<strong>recapture</strong> studies (but<br />
not vice-versa). A limitation of our approach is that the number of possible frequency<br />
classes is restricted to three, violat<strong>in</strong>g the “<strong>in</strong>f<strong>in</strong>ite number of sources” assumption for<br />
truncated Poisson models. The truncated models are at least as vulnerable to violation of<br />
the perfect positive predictive value, perfect record-l<strong>in</strong>kage, closed population and<br />
<strong>in</strong>dependent registers assumptions as capture-<strong>recapture</strong> models. It has been argued that<br />
truncated models are more robust to heterogeneity among the patients than capture<strong>recapture</strong><br />
studies. 25,31 In contrast, the equiprobability assumption of the truncated models<br />
is almost certa<strong>in</strong>ly violated when us<strong>in</strong>g multiple-source data from capture-<strong>recapture</strong><br />
studies. Therefore we <strong>in</strong>troduced the coefficient of variation, a measure of variability <strong>in</strong><br />
the coverages of the three data sources for capture-<strong>recapture</strong> studies, to estimate the error<br />
from the data. We conclude that, <strong>in</strong> the context of validity, for estimat<strong>in</strong>g <strong>in</strong>fectious<br />
disease <strong>in</strong>cidence and completeness of notification <strong>in</strong>dependent and parsimonious threesource<br />
log-l<strong>in</strong>ear capture-<strong>recapture</strong> models are preferable compared to the truncated<br />
models exam<strong>in</strong>ed. When saturated models are selected as best-fitt<strong>in</strong>g model and the<br />
estimates are unexpectedly high and seem implausible, the data should be re-exam<strong>in</strong>ed<br />
with truncated models as a heuristic tool, <strong>in</strong> the absence of a gold standard. Possible<br />
failure <strong>in</strong> the saturated log-l<strong>in</strong>ear model or unidentified violation of the underly<strong>in</strong>g<br />
assumptions should be suspected when the truncated models produce a (considerably)<br />
lower estimated number of <strong>in</strong>fectious disease patients.<br />
169