EBM Tools for Practice: Risk Reductions

Clinicians are constantly presented with data and results intended to influence the treatment they render to patients. However, in spite of the gravity of the consequences of treatment, we are often not supplied with appropriate and adequate data on which to base our judgment of effectiveness. Such situations become clearer when we see media reports with "relative risk reduction" in treatments rather than statistics on "absolute risk reduction." The difference in data presentation often makes the treatments appear much better than they actually are. Evidence-based medicine compels us to integrate the best research evidence with our clinical expertise and patient values.

In order for clinicians to make an informed decision in evaluating relative and absolute risk reduction data, they must understand the difference between the two forms of risk and the data upon which it is based. There are two types of risk assessment. Absolute risk reduction (ARR) is risk stated without any context. For example, you have a 50 % chance when flipping a coin of coming up with heads. The 50% probability is independent of other factors and is independent of prior tests. It is not compared to any other risk. Relative risk reduction (RRR) is a comparison between different risk levels and between treatment effect and the risk without treatment. For example, the relative risk for lung cancer is (approximately) 10 times greater for a smoker compared to a nonsmoker. An important feature of relative risk is that it gives no information about the actual baseline risk but it can be important in evaluating how significant a relative increase might be.

A small increase in risk in a large population can skew the effect and make the result seem larger than it actually is. Simply stated, we often lack a standard by which to judge the superior clinical decision. Whether RRR "overestimates" the effect or ARR "underestimates" the effect is a value judgment because we often lack a gold standard for what is the "best" decision.

Let’s now apply these criteria to some specific situations to gain a better insight. The baseline risk is critical for determining changes in absolute risk. For example, thyroid cancer is diagnosed in slightly less than 1 in 100,000 persons per year, whereas first myocardial infarction is diagnosed in about 400 in 100,000 people per year. These numbers are based upon absolute risk. If relative risk were applied and a 10% increase occurred, then for thyroid cancer there would be 0.10 x 1 = 0.1 new cases per 100,000 people. On the other hand, a 10% increase in myocardial infarction affects an additional 40 per 100,000 people. Accordingly, if we assume the population of the United States is 300 million (which is 3,000 times 100,000), the small increase in thyroid cancer would result in 0.1 x 3,000 = 300 new cases. By contrast, the same increase in rate using RRR in first myocardial infarction would result in 40 x 3,000 = 120,000 new cases.

To further understand the application of the data presented, we must be informed of the average number of patients that need to be treated for one to benefit as compared with a control in a clinical trial within a given period of time. This is called Number Needed to Treat (NNT). NNT is a more informative statistic for comparison because it describes the number of patients who must be treated over a set period of time to prevent one person from suffering an event or for seeing a benefit of the treatment. Number needed to treat is the inverse of absolute risk. NNT is derived from absolute risk and does not rely upon relative risk for its calculation. The higher the NNT, the less effective the treatment.

The NNT value is time-specific. For example, if a study ran for five years and the NNT was 100, in one year the NNT would be multiplied by five to estimate a one-year NNT of 500. However this method of calculation is controversial because an effect may not be constant over the full course of measurement. If the slope of the curve from the placebo group in a study remains fairly constant throughout a trial, this method is an acceptable and quick method of adjustment for the study’s duration when evaluating the treatment effect as calculated from NNT. In addition to concerns about extrapolating over time, there is the potential limitation when one randomized controlled trial is compared to another when there is a different baseline risk.

Despite criticism of relative risk, it must be considered that relative risk is a comparator against absolute risk, much like clinical trials examine therapies relative to conventional therapy. Relative risk must be interpreted alongside absolute risk to tell you if the therapy is worth pursuing. Does framing the data in RRR alter the perception of therapeutic effectiveness in physicians? In the Helsinki Heart Study, after five years of treatment with gemfibrozil, 2.73% of patients in the treatment arm experienced a cardiac event comparing to 4.14% in the placebo arm. Without mentioning the name of the trial or the medication, the results of the Helsinki Heart Study in various formats were distributed among 148 physicians. Physicians' willingness to prescribe the drug was 77% when the data was presented in terms of RRR while 24% were willing to prescribe the drug when data was expressed in terms of ARR.¹ Influence of RRR on physician’s perception of treatment benefits has been reported in several other trials.^2,3

Applying the criteria of relative risk and absolute risk to some other lipid trials, we find the following: The ASCOT-LLA Trial was a primary prevention study that examined the benefit of atorvastatin 10mg in patients with hypertension but with no previous cardiovascular disease. Over more than three years, the relative risk of a cardiovascular event was reduced by 36%.⁴ The absolute risk reduction, however, was much smaller. This study determined that taking atorvastatin for 3.3 years would lead to an absolute risk reduction of only 1.02%. The number needed to treat would then be 99.7 for the 3.3 years period to prevent one cardiovascular event.

Absolute event rates in both the rosuvastatin and placebo group arms of the JUPITER trial were low, but the relative effect was very large.⁵ The relative risk reduction from the use of rosuvastatin in this population was 44% and even higher if there was a family history of premature CHD. When primary endpoints of MI, stroke, CV death, angina requiring hospitalization and revascularization were defined, the NNT was found to be 25 patients over a 5-year period to prevent one of these endpoints. The absolute risk reduction over the two years of the study is 1.2%, reflecting a significantly lower baseline risk in this population. From the relative risk reduction and from the calculated NNT, statin therapy in patients with elevated high-sensitivity C-reactive protein and low LDL cholesterol seem to be comparable to many other interventions for primary cardiovascular prevention, but significantly higher than that seen for statin interventions for secondary prevention.* Again, we cannot compare NNT across these populations with very different baseline risks.

Exploitation of "information framing" is a well-recognized strategy in marketing and mass media. Perception of probabilities and outcomes can predictably shift when the same problem is framed in a different way. As for medical interventions when the results are presented in RRR rather than ARR, it appears that the enthusiasm for the intervention increases as both physicians and patients downplay other attributes of the treatment such as side effects.⁷ Clinicians must be vigilant for this type of data presentation and seek out absolute risk reduction figures. In so doing, we can determine if the absolute benefit of the treatment is large enough to justify its potential risks and costs.

*Another set of data to emerge from JUPITER was a high relative risk for new onset diabetes. Meta-analysis of 13 statin trials shows a 9% increase in the relative risk of new-onset diabetes.⁶ This meta-analysis shows that 255 patients have to be treated for four years before one statin-induced case of incident diabetes is seen. However, a composite of nine vascular events (including death, myocardial infarction, stroke, and coronary revascularization) would be avoided by treating 255 patients over the same time period. It is intuitively clear that these risks and benefits are independently derived and each population of 255 patients is an individual set.

Disclosure statement: Dr. Kroll has received honoraria related to speaking from Abbott Laboratories, GlaxoSmithKline and AstraZeneca.