Improving disease incidence estimates in primary care surveillance systems

Abstract

Background: In primary care surveillance systems based on voluntary participation, biased results may arise from the lack of representativeness of the monitored population and uncertainty regarding the population denominator, especially in health systems where patient registration is not required.

Methods: Based on the observation of a positive association between number of cases reported and number of consultations by the participating general practitioners (GPs), we define several weighted incidence estimators using external information on consultation volume in GPs. These estimators are applied to data reported in a French primary care surveillance system based on voluntary GPs (the Sentinelles network) for comparison.

Results: Depending on hypotheses for weight computations, relative changes in weekly national-level incidence estimates up to 3% for influenza, 6% for diarrhea, and 11% for varicella were observed. The use of consultation-weighted estimates led to bias reduction in the estimates. At the regional level (NUTS2 level - Nomenclature of Statistical Territorial Units Level 2), relative changes were even larger between incidence estimates, with changes between -40% and +55%. Using bias-reduced weights decreased variation in incidence between regions and increased spatial autocorrelation.

Conclusions: Post-stratification using external administrative data may improve incidence estimates in surveillance systems based on voluntary participation.

Keywords: Adjustment; General practitioners; Incidence estimation; Sentinel network; Surveillance; Volume of consultations.

Figure 1

Average number of cases reported versus average number of consultations per SGP and per week for ILI (left), AD (middle), and varicella (right) in French regions.

Figure 2

Relative number of consultations per GP compared to SGP (weight ratioW) according to region, week, and age group. Each bar shows the mean 2.5 and 97.5 quantiles of the distribution of weekly weight ratios over the two-year period (2009 week 32 to 2011 week 30).

Figure 3

Estimated regional incidence of GP consultations for ILI (top) and AD (bottom) using number of GPs (${\widehat{\mathbf{I}}}_{\mathbf{n}}\left(\mathbf{t}\right)$, dashed line) or number of consultations (${\widehat{\mathbf{I}}}_{\mathbf{c}}\left(\mathbf{t}\right)$, normal line) in the two regions with the most extreme changes.

Figure 4

Estimated French national incidence of GP consultations for ILI (top), AD (middle), and varicella (bottom) using number of GPs (${\widehat{\mathit{I}}}_{\mathit{n}}\left(\mathit{t}\right)$, dashed line) or number of consultations (${\widehat{\mathit{I}}}_{\mathit{c}}\left(\mathit{t}\right)$, normal line).

Figure 5

Autocorrelation in regional ILI incidence during an epidemic period. (top) National ILI epidemic profile. Week 0 is defined as the epidemic peak. (bottom) Moran’s index computed from regional incidence post-stratified on the number of GPs (${\widehat{I}}_n\left(r,t\right)$, dashed) or on the number of consultations (${\widehat{I}}_c\left(r,t\right)$, plain). The horizontal grey line shows the expected value without spatial autocorrelation.