(24) Comparison of Methods to Quantify Uncertainties in Averaging Incidence Estimates: An Illustration of Approximating Lyme Borreliosis (LB) Incidence in High-Risk Regions in Europe

Publication Number: 1069

Background: Many countries in Europe conduct public health surveillance to quantify the burden of disease, trends, identify populations at risk, and guide the planning of intervention and prevention programs. This epidemiological evidence is often made available in annual surveillance reports, where average estimates across a multi-year period are usually not available but necessary for decision-making. Appropriate methods should be employed to obtain unbiased estimates (no systematic difference between the estimated incidence and its true value) of the average incidence, with improved coverage probability of its confidence interval (higher probability that this interval contains the true average incidence).

Objectives: To investigate and compare methods for the construction of confidence intervals (CIs) for the average estimates of incidence.

Methods: We used Lyme Borreliosis (LB) surveillance data available from 24 European countries to demonstrate the performance of different methods. The average number of cases or incidence over a multi-year period was calculated as the weighted estimate of the annual number of cases, or yearly incidence with weights equal to the yearly denominators. We used the combination of stratum-specific CIs, based on the F-distribution, to calculate the 95% CI for the average incidence. A simulation study was performed to compare the performance of this method with the conventional imputation method. In the imputation method, the mean annual population was used as the denominator, and the mean annual number of cases was the numerator. We considered various scenarios where we varied the population size, the true proportion, and the number of data points to be averaged. Finally, these methods were applied to quantify the LB average incidence in 24 European countries.

Results: The simulation study showed that our method produced higher coverage probability and equivalently (very) small bias when compared to the imputation method. More specifically, the coverage probability using the imputation method decreased substantially (from 94% in the best scenario to below 20% in the worst case). The method we used, in contrast, preserved good coverage in all scenarios (above 95%). To our data application, average LB incidence varies considerably in Europe (from 0.2 in Portugal to 81.3 in France and 263 cases per 100,000 person-years in Estonia).

Conclusions: The proposed method (weighted average with 95 % stratum-specific CI based on F-distribution) is straightforward to implement and provided better estimates compared to the conventional imputation method.
Concerning LB incidence, the interpretation of disease burden is complicated by heterogeneous surveillance systems that utilize various case definitions.