What Does "50% Lower Risk" Mean in People? — Reading Relative Risk, Absolute Risk, and NNT

Even when bad cholesterol is lowered by the exact same amount, how much the cardiovascular risk drops varies across different groups. Looking across 17 cholesterol drug trials, every time 1 mmol/L was lowered, the fraction by which risk decreased was larger where cardiovascular events were rare and smaller where they were frequent. At a level where cardiovascular events occurred at a rate of 1% per year among people who received a placebo, it was 36%, and at a level where they occurred at a rate of 3%, it was 13% (PMID 42404863).

This result clashes with the assumption set by the primer article in the series, "If the Risk Was Cut in Half, How Many Fewer People Got Sick?". That piece ran its calculations on the premise that the relative risk reduction is identically 50% whether in a group with an original risk of 20% or in one with 0.2%. Relative risk reduction is a value indicating by what percentage the risk has shrunk relative to its original size. This article follows up on two things left behind by the primer. One is the place where that assumption wobbles in actual research. The other is how to read how much the NNT—which indicates "how many people need to be treated for one person to benefit"—can wobble.

⚠️ The names of drugs and training appearing below are all examples just to show how to read numbers. They are not judgments on whether or not to use a certain treatment, and each number has meaning only within that study's participants, baseline risk, and observation period.

The Numbers Set in the Primer

The rate at which that event actually occurred in each group is the absolute risk. The rate at which that event occurs in the untreated group is the baseline risk, and it is the starting point for comparison. Three numbers are made from the absolute risks of the two groups. Dividing the absolute risk of the treated group by the absolute risk of the untreated group gives the relative risk. Subtracting the relative risk from 1 gives the relative risk reduction. Subtracting the two absolute risks gives the absolute risk reduction, and since it is a value obtained by subtracting percentages from each other, the unit is written as percentage points.

Multiplying the relative risk reduction by the baseline risk yields the absolute risk reduction. Therefore, to cross from relative numbers to "how many people", you must necessarily go through the baseline risk. NNT is the reciprocal of the absolute risk reduction, and is a figure indicating how many people must be treated for one person to benefit. Because of this calculation, whenever you encounter a headline line saying "% reduction", you ask "out of how many people?" and "for how long?".

In this article, ranges are frequently attached in parentheses behind numbers. The confidence interval is a range that records how much an estimate can wobble. The two numbers inside the parentheses are both ends of that range. What that range means is covered by "How Much Does a Number Counted Once Wobble? — The Two Numbers in the Parentheses, the Confidence Interval" from Numbers from Scratch.

The primer placed three groups with originally different risks on a single table, and set the relative risk reduction to 50% for all three rows. Whether that set value holds true in reality is the first question of this article.

Is 50% 50% in Any Group?

The cholesterol number at the beginning came from a meta-analysis. A meta-analysis is an analysis that gathers the results of several trials and combines them into one. This meta-analysis gathered 17 randomized trials comparing cholesterol-lowering medication with a placebo, involving 105,879 people. The subjects were a population with few or no cardiovascular disease patients (less than 20%). The counted events totaled 6,076, combining cardiovascular death, non-fatal myocardial infarction, and non-fatal stroke, and the median follow-up period was 4.4 years.

The key to the analysis was that the original event rates differed from trial to trial. The rate of events occurring in the group receiving the placebo (the placebo group) ranged from 0.52% to 3.78% per year depending on the trial. The authors paired this event rate and the relative risk reduction per 1 mmol/L for each trial, and drew the relationship between the two as a line (regression). On that line, the relative risk reduction per 1 mmol/L decreased from 36% at a level where the event rate was 1% per year to 13% at a level of 3% (p < 0.0001, an indication that a relationship this strong is unlikely to happen by chance alone). This means that the group with more frequent original events had a smaller share of risk reduction when lowered by the same magnitude.

The authors' conclusions are twofold. First, among the primary prevention population, in other words, people who do not yet have the disease, the lower-risk group gains a significantly greater relative benefit per 1 mmol/L. Second, the higher-risk population has a smaller benefit per 1 mmol/L, so to obtain a similar relative benefit, cholesterol must be lowered much more (PMID 42404863).

When translating these results into the number of people, the original event rate must be looked at together. This is because the side where the relative risk reduction became smaller is the side where the original events occurred more frequently. Multiplying the original event rate by the relative risk reduction according to the calculation seen earlier yields 0.36 percentage points per year (for the side where events are rare) and 0.39 percentage points (for the frequent side) per 1 mmol/L. Both are around 0.4 percentage points. This means that although the relative shares differed greatly, translating them into the number of people yields a similar magnitude.

Three conditions are attached to this multiplication. First, because the actual magnitude lowered differed from trial to trial, this value is a value per 1 mmol/L and is not the actual number of people who got sick less often. Second, this multiplication is arithmetic done by this article using the two numbers from the abstract. Because they are two point estimates (a single number obtained from data) written in the abstract without confidence intervals, it cannot be determined which of the two is larger. Third, the "smaller benefit" mentioned by the authors is the relative benefit per 1 mmol/L, so it does not conflict with this multiplication.

The relative risk reduction does not always fluctuate greatly when moving between groups. An example is a trial that randomly compared a combination pill containing aspirin, a statin, and two blood pressure medications with a placebo. The subjects were 378 people who had no reason (indication) to use any component of this drug separately, but whose 5-year cardiovascular risk was estimated to be 7.5% or higher. In an analysis that re-divided the data after the trial ended (post-hoc analysis), the magnitude by which risk factors were reduced varied widely from group to group. The group with high "bad" cholesterol at the start of the trial went down 1.1 mmol/L more than the placebo, and the group with low cholesterol went down 0.6 mmol/L more. The relative risk reduction calculated from that reduction amount averaged 48% (95% confidence interval 43–52). The "bad" cholesterol at the start of the trial and the estimated cardiovascular risk were factors that significantly changed this value. The authors conclude that the effect of the combination pill on risk factors varies depending on the level of those risk factors.

However, among groups divided by cholesterol or blood pressure at the start of the trial, while the magnitude of risk factor reduction differed greatly, the relative risk reduction calculated from that reduction differed only moderately. The authors write that this suggests that even if risk factors are only slightly elevated, people whose overall cardiovascular risk has risen can benefit from the combination pill (PMID 26743587).

Both results tell us one thing together. "The relative risk reduction remains the same even when moving between groups" is an assumption that must be checked for each set of data. In the cholesterol meta-analysis, which directly counted events, it fluctuated greatly, and in the combination pill analysis, calculated from the magnitude of risk factor reduction, it fluctuated only moderately.

The same problem follows the NNT. The NNT is inversely proportional to both the relative risk reduction and the baseline risk. This means that if either of the two becomes smaller, the NNT grows larger. Therefore, even if the average NNT looks good, when the power to predict who will benefit is weak, the response of some groups (subgroups) may be very poor or cause harm. Summarizing this point by comparing it to a diagnostic test, the author of the critique concludes that in such cases, the NNT can be misleading, so its value must be weighed cautiously. They also add that the upper limit of the confidence interval can reflect the NNT better than the average (PMID 36170611). The upper limit of the NNT is the end that says more people must be treated for one person to benefit, in other words, the less optimistic end.

How to Read the Range Attached to an NNT

The effect of reducing knee anterior cruciate ligament injuries in female athletes through neuromuscular training is described with the same data presented in two side-by-side scales (a review of 12 prospective controlled studies, PMID 22745221). Looking only at injuries that occurred without colliding with another player, the relative risk reduction is 73.4% (95% confidence interval 62.5–81.1). Expressing the same effect as an NNT, to prevent one such injury, 108 players (95% confidence interval 86–150) must be trained over the course of a season. Expanding to total injuries including those caused by collisions, the relative risk reduction is 43.8% (28.9–55.5), and the NNT is 120 players (74–316).

The same effect sounds large as a relative risk reduction and modest as an NNT. The authors also summarize that while relative risk reduction points to the preventive effect, a relatively large NNT means many players are needed to prevent a single event. They then suggest creating a screening system to identify athletes at high risk in order to lower the NNT (PMID 22745221).

The range in parentheses is the confidence interval. Because this NNT includes both the duration answering "For how long?" (one season) and a range, the material to read is in place.

The range for an NNT is usually not measured separately, but calculated from the range of the absolute risk reduction. The confidence interval of an NNT can be obtained by inverting each end of the absolute risk reduction confidence interval into reciprocals and swapping their positions (PMID 11306148). Let us look at this with an example created for this article. Suppose the absolute risk reduction is 2 percentage points and its confidence interval is 1 to 3 percentage points. Then the NNT is 1 ÷ 0.02 = 50. The range is from 1 ÷ 0.03 ≈ 33 to 1 ÷ 0.01 = 100. The upper end of the absolute risk reduction becoming the lower end of the NNT is what "swapping positions" means.

Once inverted, the range is uneven from left to right. Even if the range of the absolute risk reduction has an equal width on both sides centered on the point estimate, the range of the NNT stretches longer toward the larger numbers. In the example above, 33 is 17 away from 50, whereas 100 is 50 away. The total injury NNT from the previous review also places 120 close to the lower end between 74 and 316. However, how that review obtained its range is not in the abstract.

There is also a trap in the calculation method for finding the range of the absolute risk reduction. According to a 2001 paper summarizing this method, it appears that only the simple Wald method is used in practice, and this method often makes the interval too narrow. When the interval is narrow, the estimate looks less shaky than it actually is. The author shows that using a different calculation method called the Wilson score method improves the calculation and presentation of NNT confidence intervals (PMID 11306148).

The hardest case to read is when the range of the absolute risk reduction includes 0. That the range includes 0 means the data cannot rule out the possibility that there is no effect. Let us take a second example created for this article, where the absolute risk reduction is 1 percentage point and the confidence interval is −1 to 3 percentage points. The NNT is 100. The upper end of the range, 3 percentage points, becomes an NNT of 33. The problem lies in the path toward the lower end. As the absolute risk reduction approaches 0, the NNT grows endlessly, and once it passes 0 and becomes negative, it crosses over to the side where treatment causes more harm.

The numbers on the side of causing harm have a separate name. The NNH is the number of people who end up being treated until one more harmful event occurs. A clinician-oriented commentary that advises considering baseline risk, time, and confidence intervals together when interpreting the NNT also introduces this NNH as the counterpart to the NNT (PMID 40839106). The lower end of −1 percentage point in the example above is an NNH of 100. The NNT scale starts at 1, passes through infinity, and continues to −1 (PMID 11306148). This means that once the number for the side that benefits grows endlessly and crosses infinity, it connects to the number for the side that suffers harm. Therefore, the range in this example splits into two pieces: "from 33 people to infinity toward the side that benefits" and "from 100 people to infinity toward the side that suffers harm."

If the range of the absolute risk reduction includes 0, the range of the NNT passes through infinity and breaks into two pieces. One proposed paper lists this breakage as the first limitation of the NNT. The NNT is an easy-to-understand absolute effect index, but it only makes sense when the absolute risk reduction (risk difference) is statistically significant. If not significant, the confidence interval of the NNT becomes a broken interval that includes implausible values. The second limitation is that trading off benefits and harms on the NNT scale becomes complicated. This paper suggests abandoning the NNT in decision-making contexts such as clinical guidelines, policies, and health technology assessments (PMID 40782977). An earlier review also noted that NNT confidence intervals can be difficult to interpret (PMID 16174840).

The Range-Free NNT

Ranges are frequently missing from the NNTs published in papers. Looking at 51 drug intervention studies reporting NNTs in the top 25 general and internal medicine journals, only 32 studies (62.7%) included confidence intervals for the NNT (PMID 28571585). Ingredients answering "Out of how many?" and "For how long?" were not always attached either. Baseline risk was noted in 72.5% and duration in 74.5%.

There were also cases where the calculation method was incorrect. Studies that did not follow basic calculation recommendations numbered 15 (29.4%), and in meta-analyses, it was especially high at 13 out of 23 studies (56.5%). The authors conclude that while NNTs are useful for helping clinical decisions, they cannot be interpreted if reported incompletely, and can mislead if the calculation method does not fit the study design and variables. They add that these results need to be confirmed by further research (PMID 28571585).

When calculating NNTs through a meta-analysis, there is one more value that must be decided. To calculate an NNT from relative measures that split two groups, such as odds ratios or risk ratios (relative risk), the rate at which events occur in the control group—that is, the baseline risk—must be determined as a single value. Depending on whether the average of the studies, the value combined via meta-analysis, or expert opinion is used as that value, the NNT changes. A methodological paper using Alzheimer's cognitive enhancer studies as an example concludes that that choice can affect the NNT and decision-making, so interpretation requires caution (PMID 30905696). This means that behind a single NNT produced by a meta-analysis lies this choice.

What an NNT missing its range and baseline risk actually looks like can be seen in the abstract of a meta-analysis. This meta-analysis compared a class of drugs called GLP-1 receptor agonists with placebos using 25 randomized trials and 109,846 patients, with a mean follow-up period of 3.48 ± 1.51 years. The relative risks for each outcome were similar to one another. Myocardial infarction was 0.86, cardiovascular death 0.87, major adverse cardiovascular event 0.87, and stroke 0.88, all with p < 0.01. The NNTs from the same data were 207 people, 170 people, 67 people, and 335 people in order. Gastrointestinal adverse events had a relative risk of 1.55 and an NNH of 9. The authors' conclusion is that this drug class lowered the risk of the above outcomes across a wide range of patients spanning the presence or absence of type 2 diabetes and prior cardiovascular disease, supports its role in cardiovascular disease prevention, and does so especially in groups with high body mass index (PMID 40652242).

Confidence intervals are not attached to the NNTs in this abstract. Nor is it written in the abstract which baseline risk they were calculated with. In the previous survey, 37.3% of NNTs were reported without confidence intervals, and this is precisely that kind of NNT. From these numbers alone, we can neither know how much they might wobble nor what group they were counted against.

The relative risks are nearly the same, yet the NNTs differ by fivefold. Since the relative risks are nearly the same, what made the difference is mainly how often each outcome originally happens—that is, the baseline risk. Because the abstract does not state those incidence rates, this is a reading derived from the relationships among the numbers. Also, one must not place them side by side and compare them as "benefit to 1 in 67 people, harm to 1 in 9 people." What the benefit side measures are events like myocardial infarction and cardiovascular death, whereas what the harm side measures are gastrointestinal symptoms, so the scale's graduations are different. To place them on the same scale, you must first place them on the same outcome and the same duration.

Common Misconception 1 — "Relative Risk is a Bad Metric"

It is easy to jump to this conclusion when you hear that relative numbers make an effect sound bigger. However, relative metrics also have their uses. Relative metrics are advantageous when comparing benefits across populations with different baseline risks. A review paper noting this point also wrote that describing treatments using absolute metrics can make decisions more conservative and potentially more useful (PMID 16174840). This means the two metrics are strong at different things.

Relative risk reduction is also used as a standard scale where scattered results need to be pooled together. A research team trying to estimate the health effects of HIV testing and counseling extracted 473 effect metrics from 55 papers, and these metrics were scattered in five different formats. The team created an algorithm to convert all of them into relative risk reduction, and presented conceptual arguments for why relative risk reduction is better than other metrics. And after reviewing the literature, they concluded that although relative risk reduction is easy to understand and useful, it appears to be underused in original studies and review studies (PMID 25726522). This is not a study that measured the superiority of metrics through an experiment.

There are also strong arguments in the opposite direction. One commentary editorial concludes that relative risk reduction is a misinformative metric that masks the magnitude of risk reduction, and should not be used to measure the efficacy of treatments and vaccines (PMID 36785641). This editorial did not measure new data, but built its argument using clinical trials and COVID-19 vaccine efficacy reports as examples. What this editorial disputes is which scale the same data should be written in. This article does not judge the efficacy of any vaccine or drug based on this editorial.

There is no consensus on which metric is better. The argument to exclude relative risk reduction from efficacy reporting came from an editorial without new data (PMID 36785641), but it has also been pointed out that people who receive only relative expressions tend to overestimate effects (PMID 28339913). What is close to a consensus is a single recommendation. It is to report both relative and absolute metrics, along with their respective confidence intervals (PMID 28339913, PMID 12201860).

Common Misconception Two — "A High Relative Risk Means a Big Problem"

Even if the multiplier is large, the number of people can be small. The data of 18- to 35-year-olds in Denmark seen in the primer article "If the Risk Was Cut in Half, How Many Fewer People Got Sick?" is an example viewed from the risk of each and every person. Sudden cardiac death in people who suffered a myocardial infarction at a young age was more than fifty times that of those who did not, measured by a model. Still, assuming 100,000 people were tracked for a year each, it was at a rate of 305 people (PMID 22753866). The main point of this data is that the relative risk is high, but the absolute risk is low.

The same property also changes where an effort will be effective across a whole group as well. Data on smoking during pregnancy in Victoria, Australia, is an example. The rate ratio appearing here is the multiplier obtained by dividing the smoking rates of two groups, but what it is a multiplier compared against is not written in the abstract. The rate ratio of smoking during pregnancy for teenage mothers was distinctly high at 3.26 (95% confidence interval 3.00–3.54), and for low-income mothers at 4.67 (4.17–5.22). Yet each of these high-risk groups made up a small share of all mothers. Therefore, the majority of smokers during pregnancy were 25 to 34 years old, educated, lived in cities, were not indigenous, and were not impoverished. The authors conclude that the majority of smokers during pregnancy do not belong to traditional high-risk groups.

The authors write implications in two directions here. If smoking rates are lowered at the whole-population level, the absolute smoking rate of high-risk groups can be lowered even if their high relative risk remains as is. Conversely, focusing only on high-risk groups may fail to lower the overall smoking rate (PMID 20649777).

There is no guarantee that the group with the large multiplier accounts for the majority of that event. How many people there are can be found only by looking at the multiplier and the size of that group together.

Common Misconception Three — "You Just Need to Look at Absolute Risk"

The simplification in the opposite direction does not hold either. Depending on whether you base your decision on absolute risk or absolute risk reduction, the target for treatment changes. A prospective study in British primary care (neighborhood clinics) reported in 2003 measured this directly. The British recommendation at the time was to prescribe lipid-lowering drugs for preventive purposes only to people whose 5-year absolute coronary risk exceeded 15%. This study compared that threshold with a criterion that selected people whose absolute risk reduction from the drug would exceed 4.45% (the absolute risk reduction criterion).

Out of 2,351 people, the two criteria disagreed on 111 people (4.7%). Eighty-two people were eligible for treatment only by the absolute risk criterion, and 29 only by the absolute risk reduction criterion. The two groups were different people. The group selected only by the absolute risk criterion was much older (average 68.1 years versus 49.1 years), and the group selected only by the absolute risk reduction criterion had more smokers (72% versus 35%) (PMID 12624557).

The authors clearly concluded in favor of one side. Relying solely on absolute risk as a criterion leaves out young people with marked hyperlipidemia and high relative risk, whereas the absolute risk reduction criterion ensures that drugs go to those who will see the greatest benefit. They also added that delaying treatment until absolute risk crosses an arbitrary threshold allows coronary risk to accumulate in young people in the meantime (PMID 12624557).

Because that judgment was made within the 2003 British recommendation and this study population, this article does not carry it over as a prescribing standard. What this article takes away is the fact that precedes the judgment. For the majority, the two criteria made the same determination (91% were ineligible for both, 4.3% were eligible for both). Where the two criteria diverge, which metric you use as your baseline is who you select as your target.

Common Misconception Four — "NNT Is Intuitive, So You Can Just Use It"

NNT is read properly only when the baseline risk, duration, and scope are attached. These are the conditions established in the preceding two sections. Even with those conditions met, the literature is divided on whether to continue using NNT. On one side, there is a proposal to step away from it in decision-making contexts (PMID 40782977). The advocating side also has substantial arguments.

According to a paper reviewing the use of NNT in the psychiatric literature, results can be translated from different units like scale scores or weight changes into a common unit called "per patient." Then, clinicians can anticipate how often they will see a real difference in practice. The same paper also notes limitations. NNT must be calculated in well-designed studies that treat people similar to actual clinic patients with realistic dosages, and direct calculation is possible only for outcomes divided into two branches: "yes or no." The authors' conclusion is that NNT helps predict treatment response in both efficacy and tolerability (how well one tolerates the medicine) (PMID 18479317). On the other hand, a review covering NNT concludes that while NNT is a useful index, it depends on multiple variables and patients may not easily understand it (PMID 16174840).

This article takes neither side. However, this article sees that either way, the condition that baseline risk, time, and confidence intervals must be attached remains the same to read NNT. One commentary for clinicians also recommends considering those three together (PMID 40839106).

To Go Deeper

When you encounter "% reduction" or NNT in an abstract, behind the two questions from the primer—"Out of how many?" and "For how long?"—you can ask three more. First, has it been confirmed that the relative risk reduction measured in this group is the same in other groups? Second, is a range written next to the NNT? Third, does the range of absolute risk reduction that created that range include zero? If you want to review the five metrics—odds ratio, relative risk, relative risk reduction, absolute risk reduction, and NNT—with examples, a summary titled "Which One Should Be Used?" is the starting point (PMID 12201860).

"Same Result, Different Sentence — Where Format Changes Judgment, and Where It Doesn't" covers whether the format in which the same number is written actually changes the reader's judgment, and the chain where numbers are compressed once more as they move from a paper to a press release, and from a press release to a news article.

If only one thing remains, it is this: "50% reduction" and NNT finally tell you how many people only when that group, that period, and that range are written together.

Source Papers

• "CV event rate modifies response to LDL-C lowering in primary prevention." (2026) — PMID 42404863 — Meta-regression (17 papers). Relative risk reduction per 1 mmol/L was 36% (event rate 1%/year) → 13% (3%/year). Conclusion = low-risk populations have a larger relative benefit per 1 mmol/L, while high-risk populations have a smaller benefit per 1 mmol/L requiring a larger absolute reduction for a similar relative benefit. The multiplication in the text is our arithmetic.
• "Estimated CV relative risk reduction from a polypill." (2016) — PMID 26743587 — Post-hoc analysis (378 participants). While the magnitude of risk factor reduction varied widely, the difference in calculated relative risk reduction was moderate.
• "Accuracy of the NNT Compared With Diagnostic Testing." (2023) — PMID 36170611 — Critique. NNT is inversely proportional to relative risk reduction and baseline risk. Conclusion = scrutinize carefully and use the upper limit of the confidence interval rather than the average.
• "…neuromuscular training to reduce ACL injury in female athletes…" (2012) — PMID 22745221 — Review (12 papers). Relative risk reduction 73.4% · NNT 108 (86–150), one season. Conclusion = effective but NNT is large, proposes screening for at-risk athletes.
• "Calculating confidence intervals for the number needed to treat." (2001) — PMID 11306148 — The NNT interval is obtained by inverting and swapping the two ends of the absolute risk reduction interval (the scale runs 1 → infinity → −1). The Wald method tends to produce intervals that are too short, making the Wilson score method preferable.
• "How to interpret the number needed to treat for clinicians." (2026) — PMID 40839106 — Commentary. Introduces NNH, and discusses baseline risk, time, and confidence intervals in interpretation.
• "The NNT: it is time to bow out gracefully." (2025) — PMID 40782977 — Proposal paper. Intervals break when non-significant. Proposal = stop using in decision-making contexts.
• "NNT: a descriptor for weighing therapeutic options." (2005) — PMID 16174840 — Review. Relative measures are advantageous for comparisons between populations with different risks, absolute measures are potentially more useful, and the confidence interval of the NNT can be difficult to interpret. Conclusion = useful, but patients may not easily understand it.
• "NNT in clinical literature: an appraisal." (2017) — PMID 28571585 — Completeness survey (51 papers). Confidence intervals 62.7%. Conclusion = incompletely reported NNTs are uninterpretable and require confirmation through further studies.
• "The NNT in pairwise and network meta-analysis…" (2019) — PMID 30905696 — Methodology. The NNT changes depending on what is chosen as the control group event rate.
• "GLP-1 receptor agonist in MI and ASCVD risk reduction…" (2025) — PMID 40652242 — Meta-analysis (25 papers). NNT 207, 170, 67, 335, NNH 9, no confidence intervals for NNT. NNT 67 and NNH 9 cannot be directly compared.
• "RRR is useful metric to standardize effect size…" (2015) — PMID 25726522 — Methodological study. Converted 473 metrics to relative risk reduction. Conceptual rationale. Conclusion = appears to be underutilized.
• "RRR: Misinformative measure…COVID-19 vaccine efficacy." (2022) — PMID 36785641 — Commentary editorial (no new data). Do not extrapolate as a basis for judging efficacy.
• "Relative risk versus absolute risk: one cannot be interpreted without the other." (2017) — PMID 28339913 — Commentary. Recommendation = report both along with 95% confidence intervals.
• "Odds ratio, relative risk, ARR, and the NNT — which of these should we use?" (2002) — PMID 12201860 — Recommendation = report relative and absolute measures together with confidence intervals.
• "Prior MI in the young: high relative risk but low absolute risk of SCD." (2013) — PMID 22753866 — Denmark, ages 18–35. 4.1 per 100,000 person-years → 305.0, hazard ratio 55.5.
• "Balancing absolute and relative risk reduction in tobacco control policy." (2010) — PMID 20649777 — Descriptive epidemiology. While the rate ratio for high-risk groups is high (comparison group not in the abstract), the majority of smokers are outside the high-risk group. Implication = population-wide reduction can also bring about absolute reduction in that group.
• "A prospective study…absolute risk or absolute risk reduction." (2003) — PMID 12624557 — Prospective study (UK). The two criteria select different people. Author's conclusion (supporting the absolute risk reduction criterion) does not translate to prescribing criteria.
• "Compelling or irrelevant? Using NNT can help decide." (2008) — PMID 18479317 — Proponent side. Conclusion = helps anticipate responses in terms of both efficacy and tolerability.