Portal de Periódicos da CAPES

Causation, bias and confounding: a hitchhiker's guide to the epidemiological galaxy.: Part 1. Principles of causality in epidemiological research: time order, specification of the study base and specificity

2008; BMJ; Volume: 34; Issue: 2 Linguagem: Inglês

10.1783/jfp.34.2.83

2045-2098

S. Shapiro,

Cancer Risks and Factors

Definitions and methodsIn this series of articles the focus is on epidemiology as a tool in the exploration of causation.Other functions (e.g.administrative functions such as monitoring birth or death rates, or life expectancy, or the use of epidemiological data for guidance in health policy) are not considered.In epidemiological research possible causation (or prevention) is explored by determining and comparing the incidence ('occurrence') of diseases ('outcomes', 'events'), as these occur among exposed and non-exposed persons in defined populations, over specified time intervals.Incidence per unit time is known as an incidence rate.A defined or implied population/time experience is known as a study base, and it can be analysed in two ways: in a follow-up ('cohort') study or in a case-control ('casereferent') study. Follow-up studiesIn a follow-up study, incidence rates of disease are compared among groups of exposed and non-exposed persons.Quite commonly, but not always, the rates are based on units of person-time (e.g.person-years).If the rates are similar (and unbiased and unconfounded: see Parts 1d, 1e and 2a 1 ) the evidence does not support causation; if the rate is higher among exposed persons it does; if it is lower it supports protection.The incidence rate in the exposed, divided by the rate in the non-exposed, is the estimated relative risk (RR) (the terms 'rate ratio' and 'incidence rate ratio' are synonyms; the terms 'hazard ratio' and 'hazard rate ratio' are slightly different, but for practical purposes the differences can be disregarded).Any RR estimate is only a rough approximation, and the practice of representing RRs to two decimal places is pseudo-precision; 2,3 it has nevertheless become popular.A RR estimate of 1.0 denotes no association ('the null'), and hence no evidence of causation or protection; >1.0 denotes a positive association (which may be causal) and <1.0 denotes a negative (and possibly protective) association.If an association is causal, the incidence rate in the non-exposed group, subtracted from the rate in the exposed group, roughly represents the excess incidence attributable to the exposure, or the absolute risk ('attributable risk', 'risk difference', 'rate difference' or, if an association is protective, 'risk reduction', 'risk difference', 'rate difference).To confuse matters, however, the term 'attributable risk' is sometimes used, erroneously, Causation, bias and confounding: a hitchhiker's guide to the epidemiological galaxy Part 1. Principles of causality in epidemiological research: time order, specification of the study base and specificity