Objectives: An accurate knowledge of the population at risk is a fundamental requirement for determining rates and making comparisons in epidemiological research. The major obstacle of studying the epidemiology of sentinel practice networks is the determination of population at risk, in this case, the reference population of medical practices. This article is intended to give a brief overview of major denominator approaches used in practice based epidemiology today, to discuss their underlying assumptions, their strengths and limitations.
Design: The literature used in this paper was searched from Medline databases of 1970-1997 using the logical expression "denominator and practice". More literature was identified from the references cited in those articles and from research reports that were available to the authors.
Main results: There are various approaches to the denominator at different levels of complexity, which were presented akin to the well known "iceberg phenomenon": with only a small portion of the iceberg visible above the surface, inference as to the size of the invisible part may still be made under certain assumptions. Crude numbers of cases may still reflect trends in the true epidemiology of disease and may be useful for time-series analyses. Differences in the number of network participants over time and across region may be controlled for by using the number of sentinel practices as a denominator. The number of consultations is a first step towards a population-based denominator, reflecting characteristics of both patients and the network. The yearly or quarterly contact group is a true person-based denominator, yet disregarding the population not consulting. The population in practices' catchment areas can be either determined from patient lists or estimated using mathematical models. The ideal denominator is the total population in a geographically defined area, though this information can be directly related to medical practices only in very few countries.
Conclusions: Although a person, or ideally a population-based denominator is desirable, even "lower-level" denominators may be suitable for certain research topics. In countries without patient registration, the estimation of incidences and prevalences has many methodological uncertainties that limit the use of sentinel practice systems. Assuming representativeness, valid analytical or time-series studies, however, can still be carried out even if there is very little information on the population at risk covered by particular medical practices.