Portal de Periódicos da CAPES

Measures of effect: Relative risks, odds ratios, risk difference, and ‘number needed to treat’

2007; Elsevier BV; Volume: 72; Issue: 7 Linguagem: Inglês

10.1038/sj.ki.5002432

1523-1755

Giovanni Tripepi, Kitty J. Jager, Friedo W. Dekker, Christoph Wanner, Carmine Zoccali,

Advanced Causal Inference Techniques

Epidemiological studies aim at assessing the relationship between exposures and outcomes. Clinicians are interested in knowing not only whether a link between a given exposure (e.g. smoking) and a certain outcome (e.g. myocardial infarction) is statistically significant, but also the magnitude of this relationship. The 'measures of effect' are indexes that summarize the strength of the link between exposures and outcomes and can help the clinician in taking decisions in every day clinical practice. In epidemiological studies, the effect of exposure can be measured both in relative and absolute terms. The risk ratio, the incidence rate ratio, and the odds ratio are relative measures of effect. Risk difference is an absolute measure of effect and it is calculated by subtracting the risk of the outcome in exposed individuals from that of unexposed. Epidemiological studies aim at assessing the relationship between exposures and outcomes. Clinicians are interested in knowing not only whether a link between a given exposure (e.g. smoking) and a certain outcome (e.g. myocardial infarction) is statistically significant, but also the magnitude of this relationship. The 'measures of effect' are indexes that summarize the strength of the link between exposures and outcomes and can help the clinician in taking decisions in every day clinical practice. In epidemiological studies, the effect of exposure can be measured both in relative and absolute terms. The risk ratio, the incidence rate ratio, and the odds ratio are relative measures of effect. Risk difference is an absolute measure of effect and it is calculated by subtracting the risk of the outcome in exposed individuals from that of unexposed. This study addresses the measures of effect, that is, the measures that are used to compare the frequency of disease (or other outcome) between two groups. The measures of effect are generally expressed as relative risks and odds ratios (OR) (relative measures of effect) or as risk difference (absolute measure of effect). The 'number needed to treat' (NNT) is another absolute measure of effect that is frequently used in clinical trials. The relative risk can be calculated as ratio between two incidence proportions (risk ratio, see Example 1) or two incidence rates (incidence rate ratio, see Example 2). Example 1: In the randomized prospective, Heart Outcomes Prevention Evaluation (HOPE) study,1.The Heart Outcomes Prevention Evaluation Study Investigators Effects of an angiotensin-converting-enzyme inhibitor, ramipril, on cardiovascular events in high-risk patients.N Engl J Med. 2000; 342: 145-153Crossref PubMed Scopus (8157) Google Scholar the effect of ramipril on the risk of cardiovascular (CV) events, was investigated by calculating the ratio between the incidence proportions of CV events in ramipril-treated and in placebo-treated patients. A value of 0.78 indicates that patients treated with ramipril had a lower risk than untreated patients and that this drug induced a 22% decrease in the risk of CV events (relative risk reduction). The risk ratio (as well as other measures of effect) is generally accompanied by a measure of the precision of the estimate: the confidence interval (CI). In the HOPE study, the CI was 0.70–0.86. The concept of the CI can be explained as follows: if 100 samples of the same size considered in the HOPE study would be drawn from the population and if we would calculate for each sample the risk ratio associated with ramipril treatment, we would obtain a 100 (all slightly different) risk ratio estimates. The 95% CI is the interval that includes the 95% of risk ratios of these 100 population samples. Thus, the 95% CI is the interval of values in which the true risk ratio is likely to lie with a probability of 95%. To be statistically significant with a P 5% that this effect is due to random error, i.e. the error purely attributable to chance). The odds are a way of representing probability, familiar to gamblers. For example, the odds that a single throw of a die produces a six are 1–5, that is 1 chance of success and 5 chances of failure (see Example 3). In a case–control study, the odds of exposure in cases and controls are calculated as the number of exposed individuals divided by the number of unexposed individuals in each group. If we know the odds of exposure in cases and controls, we can calculate the OR, that is the ratio between the odds of exposure in diseased and in non-diseased individuals. As discussed for the risk ratio, an OR 1.0 means that the odds are higher in exposed individuals. Example 3: Knoll et al.4.Knoll G.A. Wells P.S. Young D. et al.Thrombophilia and the risk for hemodialysis vascular access thrombosis.J Am Soc Nephrol. 2005; 16: 1108-1114Crossref PubMed Scopus (87) Google Scholar investigated the association between vascular access thrombosis and thrombophilia. They considered 107 patients with access thrombosis (cases) and 312 patients without fistula thrombosis (controls). Overall, among the 107 patients with access thrombosis, 59 had evidence of thrombophilia and 48 did not, while among the 312 without access thrombosis 122 had thrombophilia and 190 did not. Odds of thrombophilia in patients with vascular access thrombosis: 59/48=1.229.Odds of thrombophilia in patients without vascular access thrombosis: 122/190=0.642 The OR is 1.229/0.642=1.91. An odds ratio of 1.91 means that the odds of exposure to thrombophilia were 91% higher in patients with vascular access thrombosis than in those without this complication. This OR was statistically significant because the 95% CI of this estimate (95% CI 1.23–2.98) did not include 1. The authors concluded that although large, multicenter, prospective cohort studies are needed to confirm this observation, thrombophilia seems to be associated with vascular access thrombosis in hemodialysis patients. To understand why the risk ratio is not appropriate in the Knoll's study, we consider the example reported in Figure 1, showing that, unlike the OR, the risk ratio among others depends on the number of controls taken for each case. As shown in Figure 1 (left panel), the risk ratio for arteriovenous fistula thrombosis results to be 65% higher (risk ratio=1.65) in patients with thrombophilia than in those without this complication. Now, as an example, we consider the following hypothetical situation. If, to increase the study power, the investigator had decided to double the number of controls, hypothesizing that the exposure in this larger control group remains the same of that of the original control group, the risk ratio would have increased from 1.65 to 1.77 (!) (Figure 1, right panel). In contrast, the OR did not change in respect to the risk ratio (Figure 1). This is because of the way the two measures of effect are calculated. In fact, the incidence proportion (by which we calculate the risk ratio) is dependent on the number of both cases and controls and exposed and unexposed individuals while the odds, being a within-group absolute risk indicator, is only affected by the distribution of exposure in each group. Since the allocation fraction of the exposure in the larger control group did not change in respect to the original control group, the OR remained unchanged. OR is a good estimate of risk ratio when the disease/event is rare (rare disease assumption).5.Szklo M. Nieto F.J. Epidemiology. Beyond the Basics. John and Bartlett publishers, Sudbury, MA, USA2004: 110-114Google Scholar The effect associated to a specific treatment can be also calculated in terms of absolute risk difference. The calculation is just the difference between the incidence proportion of a disease/event in the control group and the incidence proportion of the same outcome in the treated group. Accordingly, in the HOPE study, the risk difference for CV events between patients on placebo and on ramipril is: 0.18–0.14=0.04 (i.e. 4% absolute risk difference attributable to ramipril). On the basis of the risk difference it is possible to calculate the NNT to prevent one adverse event. In the HOPE study, the number of patients to be treated to prevent one CV event in 5 years can be calculated as the inverse of the absolute risk difference (NNT=1/absolute risk difference): 1/0.04 is 25. So, we ought to treat 25 patients with ramipril to prevent one CV event in 5 years. The ideal NNT for a given treatment is 1. In fact a NNT=1 means that all treated individuals had a favorable outcome (i.e. had no events: incidence proportion=0), while all untreated individuals had adverse outcome (i.e. experienced the event in question: incidence proportion=1). If the number of events in the HOPE study had been 65 in ramipril group and 83 in placebo group (instead of 650 and 826, respectively), the NNT would have been 1/(0.018–0.014)=250, while the RR would have remained unchanged (RR=0.014/0.018, i.e. 0.78). Thus, the NNT provides a good insight into the clinical relevance of the effect. To estimate the magnitude of the association between exposure and outcomes we can use relative and absolute measures of effect. Relative measures of effect are risk ratio (i.e. the ratio between two incidence proportions), incidence rate ratio (the ratio between two incidence rates), and OR (the ratio between two odds). The risk difference is an absolute measure of effect (i.e. the risk of the outcome in exposed individuals minus the risk of the same outcome in unexposed). The risk difference is frequently used in clinical trials to calculate the NNT, that is the number of individuals that is needed to treat to prevent one adverse event in a given time period. Download .ppt (.1 MB) Help with ppt files Supplementary Material