P2.61: An implementation of automated individual matching for observational studies

✍ Scribed by Markus Schröder; Johannes Hüsing; Karl-Heinz Jöckel

• Publisher: John Wiley and Sons
• Year: 2004
• Tongue: English
• Weight: 76 KB
• Volume: 46
• Category: Article
• ISSN: 0323-3847
• DOI: 10.1002/bimj.200490213

✦ Synopsis

AGM, the German nation-wide sentinel practice network on measles, has been operational since 1999. From more than 1200 medical practices, all clinical cases of measles along with their basic data are reported and population-based incidence rates are derived at the Robert Koch Institut. Sentinel doctors are entitled to have serum, throat swabs or urine from clinical cases analysed in a reference laboratory for confirmation. About 40% of all measles notifications were laboratory tested, and 60% of those were confirmed by serology and/or PCR. Questions remained, however, (a) whether the result of a laboratory test may be predicted from data of the case and the practice, (b) what goodness of prediction (sensitivity, specificity, PPV) can be achieved by that model in comparison to the true test results, and (c) what percentage of true cases the model may predict for notifications without laboratory tests. Data from 3086 measles notifications (12/1999 -07/2002) and from 1287 medical practices were available. By Chi2 tests and U tests, 10 qualitative and 8 quantitative parameters were identified, each of them significantly associated (p < 0.05) with laboratory test results. Multivariate logistig regression (stepwise foreward procedure with pin = 0.05) with internal cross-validation ("leave one out") selected the most important 7 predictors out of those 18 parameters. The derived probability function (cutpoint 0.5) predicted 61% of all tested cases as positive, yielding a sensitivity of 0.78 (specificity 0.65; PPV 0.76). Measles notifications not being tested by laboratory measures were predited to be positive ("true measles") in 92%. Furthermore, the average probability to be a true masles case was predicted to be 0.59 among the tested notifications and 0.73 among the non-tested notifications. Results from laboratory tests for measles can be predicted from data of cases and notifying practices with satisfactory goodness. Predictions are plausible, since doctors are more likely to use laboratory tests when the clinical diagnosis is less clear. The validity of the prediction model could only be assessed by additional data. With a valid prediction, however, incidence rates derived from those notifications could be adjusted for their probability of being "true" measles cases.