Health Atlas
Menu
Evidence-led reference
What does this actually mean?Explanation 62

“Raises risk by 20%”: 20% of what?

How different are relative risk, absolute risk and percentage points?

8 sources · Latest source check 2026-10-15 UTC
On this page

Relative effects need a baseline risk to describe absolute benefit or harm. The same relative reduction can mean very different numbers of events in different populations.[1]

What does this actually mean?

The meaning

Risk is a chance of a specified outcome over a specified period. Absolute risk describes that chance in a group: for example, 10 out of 1,000 people over five years. Relative risk compares two such chances. Both can be accurate, yet they answer different questions and can create very different impressions.[7]

The worked example

Imagine the five-year risk rises from 10 to 12 people out of 1,000. The increase is two extra people per 1,000, or 0.2 percentage points. Relative to the original 10, two extra cases are a 20% increase. “Raises risk by 20%” therefore does not mean that 20 extra people out of every 100 will be affected.[7]

Why the starting chance matters

A relative comparison needs a baseline. A 20% increase from 100 cases per 1,000 means 120 cases: 20 additional cases, rather than two. These are invented arithmetic examples, not estimates for a food or medicine. Without baseline risk, a headline cannot show the absolute difference.[8]

Keep time and people aligned

Comparing five-year risk in one group with lifetime risk in another is not a fair comparison. The outcome, starting population and follow-up all need to match. A useful presentation puts the treated and comparison groups on the same denominator and states the time period before asking whether the difference matters.[7]

Where number needed to treat fits

Number needed to treat translates an absolute risk reduction into how many people, on average, receive treatment for one additional outcome to be prevented. It belongs to a particular comparison and follow-up period. It is not a guarantee that exactly one named person benefits.[6]

What to check

Ask for the event counts, group sizes, timeframe and absolute difference, alongside uncertainty. Put benefits and harms into comparable units. Then check whether the study population resembles the people facing the decision. A small absolute change can matter for a serious outcome, but an impressive relative percentage cannot make that value judgment for you.[8]

Read the effect with its context

Compare the same outcome, population, and follow-up. Check whether the number is a risk ratio, hazard ratio, rate, or mean difference; they are not interchangeable. Consider harms and certainty as well.[1]

Risk, odds and hazard answer different questions

Risk is the proportion experiencing an event over a stated period. Odds compare events with non-events; odds ratios differ from risk ratios, particularly for common events. Hazard ratios summarize relative event rates, rather than percentage-point differences. Specify the event, population and time window.[2]

A hypothetical example of baseline risk

Imagine a treatment halves a one-year event risk. Starting from 4%, that changes 40 events per 1,000 people to 20: 20 fewer per 1,000. Starting from 0.4%, it changes four to two: two fewer per 1,000. Both are 50% relative reductions, but the absolute differences are two percentage points and 0.2 percentage points. These are teaching numbers, not a prediction for any treatment.[2]

What a p-value does not tell you

ASA emphasizes that a p-value does not measure the probability that a hypothesis is true or quantify the importance of a result. A threshold such as 0.05 does not replace study design, effect size, transparency and context. Trying many analyses and highlighting only favorable results can undermine interpretation. Statistical significance and clinical importance are different judgments.[4]

Reading a pooled estimate

A pooled estimate weights study results rather than averaging headlines. Precision can improve while differences in methods, populations and outcomes still limit interpretation. Combining biased studies does not remove their biases. Check whether the component studies answer a sufficiently similar question before treating one combined number as universally applicable.[3]

Counts, percentages and person-time describe different quantities

Counts, percentages and person-time describe different quantities
MeasureNumerator and denominatorWhat it answers
Incidence proportion / riskNew cases divided by the initially at-risk population.How many developed the outcome during the stated period?[5]
Incidence rateNew cases divided by accumulated person-time.How quickly did new cases occur?[5]
PrevalenceExisting cases relative to the specified population and time.How common is the condition, rather than how many new cases appeared?[5]

A rate is not automatically a percentage

CDC’s archived methods text distinguishes incidence proportion from incidence rate. A denominator of person-years is time accumulated across people, not just a head count. A rate per 1,000 person-years cannot simply be read as the same percentage risk over an unstated follow-up.[5]

Number needed to treat uses an absolute difference

Invented illustration: over one year, suppose an outcome occurs in 20 of 1,000 people in one group and 10 of 1,000 in the comparison group. The absolute reduction is 10 per 1,000, or 0.01. Using CEBM’s inverse-absolute-risk-reduction definition gives an NNT of 100 for that outcome and timeframe. These are teaching numbers, not a treatment estimate.[6]

ASA emphasizes that a p-value does not measure effect size or importance. A result can meet a statistical threshold while its absolute effect remains small. Effect size, uncertainty, relevance and harms therefore need explanation alongside a significance label.[4]

Evidence & uncertainty

A confidence interval describes statistical precision, not every source of uncertainty. A result compatible with no difference does not establish that treatments are equivalent.[1]

Selected-source overview · Independent review pending

Published questions

Is a relative risk reduction a percentage-point reduction?

Selected-source overview · Independent review pending. Read the evidence context.

No. Absolute change depends on the baseline risk.[1]

Evidence behind this answer

This answer summarizes selected sources. An independent appraisal of the complete evidence has not been completed.

Does a confidence interval settle every uncertainty?

Selected-source overview · Independent review pending. Read the evidence context.

No. Bias, indirectness, and other evidence limitations still matter.[1]

Evidence behind this answer

This answer summarizes selected sources. An independent appraisal of the complete evidence has not been completed.

Does a non-significant difference prove equivalence?

Selected-source overview · Independent review pending. Read the evidence context.

No. Inspect the effect estimate and interval rather than treating a threshold as proof.[1]

Evidence behind this answer

This answer summarizes selected sources. An independent appraisal of the complete evidence has not been completed.

What context does a number needed to treat require?

Selected-source overview · Independent review pending. Read the evidence context.

NNT comes from an absolute risk difference for a defined outcome and follow-up; it changes with baseline risk and comparison.[2]

Evidence behind this answer

This answer summarizes selected sources. An independent appraisal of the complete evidence has not been completed.

Can I read an odds ratio directly as a risk ratio?

Selected-source overview · Independent review pending. Read the evidence context.

Generally no. They compare different quantities and can differ substantially when the event is common.[2]

Evidence behind this answer

This answer summarizes selected sources. An independent appraisal of the complete evidence has not been completed.

Does a non-significant result prove there is no effect?

Selected-source overview · Independent review pending. Read the evidence context.

No. Its estimate and confidence interval may still allow important benefit or harm; design and bias matter too.[1]

Evidence behind this answer

This answer summarizes selected sources. An independent appraisal of the complete evidence has not been completed.

Does p = 0.03 mean a claim is 97% likely to be true?

Selected-source overview · Independent review pending. Read the evidence context.

No. ASA explicitly distinguishes p-values from the probability that a scientific hypothesis is true.[4]

Evidence behind this answer

This answer summarizes selected sources. An independent appraisal of the complete evidence has not been completed.

Does a precise pooled estimate apply to every population?

Selected-source overview · Independent review pending. Read the evidence context.

No. Differences in study populations, outcomes and methods still affect what the combined result can answer.[3]

Evidence behind this answer

This answer summarizes selected sources. An independent appraisal of the complete evidence has not been completed.

What is the difference between a risk being doubled and a risk rising by two percentage points?

Selected-source overview · Independent review pending. Read the evidence context.

Doubling is a relative change; percentage points describe an absolute difference. In a teaching example, increasing risk from 1% to 2% doubles it but adds one percentage point. Increasing 1% to 3% adds two percentage points and triples it. Specify the baseline, outcome and timeframe before interpreting either statement.[2]

Evidence behind this answer

This answer summarizes selected sources. An independent appraisal of the complete evidence has not been completed.

Is a rate per 1000 person-years automatically the same percentage risk?

Selected-source overview · Independent review pending. Read the evidence context.

No. Person-time and a population head count are different denominators. The follow-up and definition must be specified.[5]

Evidence behind this answer

This answer summarizes selected sources. An independent appraisal of the complete evidence has not been completed.

Are incidence and prevalence the same measure?

Selected-source overview · Independent review pending. Read the evidence context.

No. Incidence concerns new cases; prevalence concerns existing cases in a specified population and time.[5]

Evidence behind this answer

This answer summarizes selected sources. An independent appraisal of the complete evidence has not been completed.

Does a one-percentage-point absolute reduction give an NNT of 100 in the illustrative example?

Selected-source overview · Independent review pending. Read the evidence context.

Yes. One percentage point is 0.01, and its inverse is 100. This applies to the invented outcome and one-year comparison, not a named medicine.[6]

Evidence behind this answer
  • Number Needed to Treat (NNT)

    institutional method definition · Date not stated
    Indexed original method definition and formula inspected; direct page retrieval unavailable · NNT as the inverse of absolute risk reduction

This answer summarizes selected sources. An independent appraisal of the complete evidence has not been completed.

Does statistically significant necessarily mean a large or clinically important effect?

Selected-source overview · Independent review pending. Read the evidence context.

No. ASA states that a p-value does not measure effect size or importance. Absolute magnitude and clinical relevance need separate interpretation.[4]

Evidence behind this answer

This answer summarizes selected sources. An independent appraisal of the complete evidence has not been completed.

What information is missing from a headline that only says risk rises by 20%?

Selected-source overview · Independent review pending. Read the evidence context.

You still need the starting risk, the outcome being counted, the relevant population and the time period. Ask for both groups as cases out of the same number of people. A relative percentage alone cannot show how many additional people might be affected.[8]

Evidence behind this answer

This answer summarizes selected sources. An independent appraisal of the complete evidence has not been completed.

Can I compare a relative benefit with an absolute harm percentage?

Selected-source overview · Independent review pending. Read the evidence context.

Those formats can make the benefit look larger and the harm look smaller. Convert both to the same denominator and relevant time period first, while keeping the outcomes distinct. The comparison also needs uncertainty and context: preventing one severe event is not equivalent to avoiding one minor side effect.[8]

Evidence behind this answer

This answer summarizes selected sources. An independent appraisal of the complete evidence has not been completed.

A confidence interval is not a complete credibility score

An estimate and its confidence interval describe statistical uncertainty under the analysis assumptions. The interval does not automatically account for bias, a poorly chosen outcome or indirectness for another population. A result that is not statistically significant can still be compatible with worthwhile benefit or harm. It is necessary to read the range of plausible effects and the evidence limitations together.[1]

Questions still to cover

Further work: predictive values and worked test-accuracy examples; competing risks and detailed time-to-event interpretation; complete outcome-specific uncertainty and harms appraisal.

Try the absolute-risk example

Hypothetical values for 1,000 people over the same follow-up period. This uses a risk ratio; it is not a study result or a personal risk estimate.

Assumes no other differences between groups. This example does not model uncertainty, treatment harms, or an individual’s risk.

References

  1. Chapter 15: Interpreting results and drawing conclusions

    Evidence methods

    Cochrane Handbook · Content/publication date: 2024
    Accessed: 2026-10-06 UTC · Selected page sections
    Cited sections: 15.3; 15.4

  2. Choosing effect measures and computing estimates of effect

    Methods handbook

    Cochrane · Content/publication date: 2024
    Accessed: 2026-10-06 UTC · Selected handbook sections; not a full independent methods review
    Cited sections: Sections 6.4.1 and 6.7: risk, odds, differences and time-to-event effects

  3. Analysing data and undertaking meta-analyses

    Methods handbook

    Cochrane · Content/publication date: 2024
    Accessed: 2026-10-06 UTC · Selected handbook sections; not a full independent methods review
    Cited sections: Sections 10.1–10.2: pooling estimates and differing studies

  4. Statement on Statistical Significance and P-Values: official release

    Agency explanation

    American Statistical Association · Content/publication date: 2016
    Accessed: 2026-10-06 UTC · Selected pages of official three-page statement release
    Cited sections: Six principles and interpretation of statistical significance

    Methods guidance, not a clinical intervention study.

  5. Principles of Epidemiology, Lesson 3, Section 2: Morbidity Frequency Measures

    Agency methodological definitions

    CDC · Content/publication date: 2012
    Accessed: 2026-10-07 UTC · Selected archived original methods text inspected
    Cited sections: Incidence proportion, incidence rate, person-time and prevalence

    Archived textbook definitions, not a current clinical recommendation or new disease-incidence estimate.

  6. Number Needed to Treat (NNT)

    Institutional method definition

    Centre for Evidence-Based Medicine, University of Oxford · Content/publication date: not stated
    Accessed: 2026-10-07 UTC · Indexed original method definition and formula inspected; direct page retrieval unavailable
    Cited sections: NNT as the inverse of absolute risk reduction

    Worked counts in the article are explicitly invented editorial arithmetic, not published treatment data.

  7. Simple tools for understanding risks: from innumeracy to insight

    Methodological review

    Gigerenzer and Edwards · Content/publication date: 2003
    Accessed: 2026-10-15 UTC · Abstract and indexed full-text passages; PMC direct access challenged
    Cited sections: Relative risks; reference classes; natural-frequency examples

    Original methodological explanation; arithmetic examples here are editorial.

  8. What is my risk, doctor? How to convey disease risk and treatment effects

    Methodological review

    Bretthauer and Kalager · Content/publication date: 2023
    Accessed: 2026-10-15 UTC · Indexed original journal passages; direct access blocked
    Cited sections: Absolute and relative treatment effects; baseline risk and decision thresholds

    Indexed selected passages; direct HTML/PDF retrieval blocked.

Scope & review

Population
Readers of study results
Exposure or intervention
Effect interpretation
Comparison
Same outcome and follow-up
Outcomes
Absolute and relative effects
Timeframe
Specified study follow-up
Jurisdiction
Cochrane methods guidance

Selected-source educational explainer. Cited page sections or indexed abstracts were inspected as specified for each source. This is not a systematic search, a formal evidence grade, or an independent clinical expert review; topic and applicability judgments are editorial. The edition 3.8 check applies to selected additions and named reused passages; older references retain their individual access dates.

Sources checked on their individual access dates. Selected sources were checked; ongoing surveillance and qualified review have separate statuses on the coverage page.

Revision history
  • 2026-10-06 UTC: Bounded foundation added with source locators, scope, and unresolved expansion areas.
  • 2026-10-06 UTC: Expanded definitions, categories, evidence boundaries and locally sourced questions; selected sources checked with access limits recorded. Independent expert review remains incomplete.
  • 2026-10-07 UTC: Edition 3.8: added definitions, worked distinctions and sourced questions about evidence, testing, medicines and exposure. Qualified review remains pending.
  • 2026-10-15 UTC: Added the plain-English series explanation with definitions, distinctions, practical questions and selected primary sources; qualified review remains pending.

What this article still leaves open

  • predictive values and worked test-accuracy examples
  • competing risks and detailed time-to-event interpretation
  • complete outcome-specific uncertainty and harms appraisal
  • Outcome-specific absolute effects and appropriate interval estimates in the populations a reader is comparing.
Suggest a correction

Submit a content correction to the editorial queue, or download a local note. Leave out personal medical details. Submitted text is stored privately with a browser receipt session and is visible to this browser and the editor. Clearing cookies removes access to your submission history. This is not an account or a medical service.

More in the plain-English series

Keep making sense of it.

Explore Prevention & care explanations See every explanation

Explore the topic map

Browse the health domains connected to this article. The map also identifies topics whose articles remain unwritten.