Improving incidence estimation in practice-based sentinel surveillance networks using spatial variation in general practitioner density

Abstract

BackgroundIn surveillance networks based on voluntary participation of health-care professionals, there is little choice regarding the selection of participants’ characteristics. External information about participants, for example local physician density, can help reduce bias in incidence estimates reported by the surveillance network.MethodsThere is an inverse association between the number of reported influenza-like illness (ILI) cases and local general practitioners (GP) density. We formulated and compared estimates of ILI incidence using this relationship. To compare estimates, we simulated epidemics using a spatially explicit disease model and their observation by surveillance networks with different characteristics: random, maximum coverage, largest cities, etc.ResultsIn the French practice-based surveillance network – the “Sentinelles” network – GPs reported 3.6% (95% CI [3;4]) less ILI cases as local GP density increased by 1 GP per 10,000 inhabitants. Incidence estimates varied markedly depending on scenarios for participant selection in surveillance. Yet accounting for change in GP density for participants allowed reducing bias. Applied on data from the Sentinelles network, changes in overall incidence ranged between 1.6 and 9.9%.ConclusionsLocal GP density is a simple measure that provides a way to reduce bias in estimating disease incidence in general practice. It can contribute to improving disease monitoring when it is not possible to choose the characteristics of participants.Electronic supplementary materialThe online version of this article (doi:10.1186/s12874-016-0260-x) contains supplementary material, which is available to authorized users.

Comparison between influenza coded primary care consultations and national influenza incidence obtained by the General Practitioners Sentinel Network in Portugal from 2012 to 2017

“…In a surveillance system, it is imperative to get information as close as possible to that obtained from a random and representative sample of the population [ 12 ]. In this context, general practitioners' voluntary participation, in contrast to a random representative selection, has been seen as a challenge for reducing bias in estimates of disease incidence in surveillance networks [ 13 ]. Although the age and sex distribution of the population under observation in the Sentinel Network is similar to the Portuguese population no geographical representativeness was achieved.…”

“…Recognizing possible bias in surveillance systems is an important step for improving them [ 12 , 13 ], and in order to achieve more accurate surveillance systems, it is always relevant to compare different data sources. In Portugal, an alternative data source of influenza incidences is the primary care consultations coding system, which has been mandatory since 2001 with the implementation of electronic medical records software in the Portuguese National Health System [ 14 ].…”

Shifting patterns of seasonal influenza epidemics

“…Third, the surveillance network of sentinel general practitioners is not a representative sample distributed in the country. Recent work proposed different techniques to correct ILI data in France against this possible bias [54,55]. We checked that these corrections would leave the peak time invariant and would lead to a small change of the threshold value that would however not alter the identification of the onset time, as discussed above.…”

Spatial distribution and differences of stroke occurrence in the Rhone department of France (STROKE 69 cohort)

“…The density of the supply of care could also be considered, specifically with general practitioners to determine if there could be a relation between stroke occurrence risk and the density of general practitioners per inhabitant 37 . In addition, the spatial analysis could be supplemented by a temporal analysis of stroke incidence.…”