Odds Ratio vs Risk Ratio vs Risk Difference: Which Effect Measure Should Researchers Report?
Risk difference, risk ratio, and odds ratio describe different aspects of binary outcomes and should not be treated as interchangeable. This Resource explains how the estimand, study design, model, baseline risk, and communication goal should guide effect-measure selection and reporting.
For a binary outcome, researchers can often describe the same 2 × 2 data using a risk ratio (RR), odds ratio (OR), or risk difference (RD). These measures are mathematically related, but they do not answer the same scientific question and should not be treated as interchangeable.
The recurring problem is especially important for the odds ratio vs risk ratio distinction. Logistic regression routinely produces odds ratios, but an odds ratio is not a risk ratio simply because both are relative measures. Likewise, an impressive-looking relative effect may correspond to a small absolute change when the underlying event risk is low.
Practical principle: Choose the effect measure from the estimand, study design, and communication goal—not simply from the statistic that software reports.
Risk differences describe absolute changes in risk. Risk ratios describe proportional changes in risk. Odds ratios describe proportional changes in odds and have particular roles in logistic regression and case-control research. None is universally the best measure (Agresti, 2013; Borenstein et al., 2021; Lash et al., 2021).
Quick Comparison: Risk Difference vs Risk Ratio vs Odds Ratio
Let p1 denote the probability, or risk, of the event in one group and p0 the corresponding risk in the comparison group.
| Measure | Calculation concept | Interpretation | Null value | Common context | Major interpretation trap |
|---|---|---|---|---|---|
| Risk difference (RD) | p1 − p0 | Absolute difference in event probabilities between groups | 0 | Trials, cohort/follow-up data, absolute-effect reporting, meta-analysis | Ignoring its dependence on baseline risk or presenting it as though it were a relative effect |
| Risk ratio (RR) | p1 / p0 | Risk in one group divided by risk in the comparison group | 1 | Trials, cohort/follow-up data, relative-effect reporting, meta-analysis | Reporting a large relative change without the underlying risks or absolute impact |
| Odds ratio (OR) | Ratio of p / (1 − p) between groups | Odds of the event in one group divided by the odds in the comparison group | 1 | Logistic regression, 2 × 2 tables, case-control studies, meta-analysis | Interpreting the OR as though it were a risk ratio or saying “times as likely” when the quantity is actually odds |
Agresti develops the difference of proportions, relative risk, and odds ratio as distinct association measures for categorical data, while Borenstein et al. distinguish the same three measures as alternative effect-size indices for binary outcomes (Agresti, 2013; Borenstein et al., 2021).
Risk Is Not the Same as Odds
Understanding relative risk vs odds ratio starts with distinguishing probability from odds.
If the probability of an event is p, its odds are:
odds = p / (1 − p)
Probability expresses the proportion expected to experience the event. Odds compare the probability of the event with the probability of its complement.
These quantities become increasingly different as the event probability increases. An odds ratio therefore compares odds, not probabilities. Logistic regression uses this distinction directly: it models the log odds of a binary outcome, so exponentiating a regression coefficient produces an odds ratio for the specified predictor contrast (Agresti, 2013).
Interpretation boundary: “The odds were twice as high” and “the risk was twice as high” are different statements. Interpreting an odds ratio as though it were a risk ratio can therefore be misleading.
Risk Ratio: A Relative Comparison of Probabilities
The risk ratio, also commonly called the relative risk, compares two event risks:
RR = p1 / p0
An RR of 1 is the null value: the two groups have the same risk. An RR above 1 indicates higher risk in the numerator group, while an RR below 1 indicates lower risk, assuming the event and group orientation have been defined consistently.
Agresti describes the relative risk as the ratio of two group probabilities and develops confidence intervals on the log-RR scale. Altman likewise treats the relative risk as a ratio-based comparison of two proportions (Agresti, 2013; Altman, 1991).
What question does the risk ratio answer?
Risk-ratio question: How large is the event risk in one group relative to the event risk in the comparison group?
That makes it a natural candidate when the scientific estimand itself is relative and the study design permits risks to be estimated.
But relative magnitude does not communicate absolute impact by itself. A proportional change has different practical consequences depending on the underlying event probabilities.
Risk Difference: The Absolute Effect
The risk difference is:
RD = p1 − p0
Its null value is 0 because no effect on this scale means no difference between the two risks.
The RD answers an absolute question:
How many percentage points higher or lower is the event probability in one group?
For example, an RD of −0.05, under a treatment-minus-control convention, represents an absolute reduction in event risk of 0.05, or 5 percentage points.
Borenstein et al. explicitly distinguish the risk difference from the two ratio measures: RR and OR are relative measures, whereas RD is an absolute measure. They also note that risk differences are analyzed in their raw units rather than on the logarithmic scale used for RR and OR (Borenstein et al., 2021).
Why absolute risk difference matters
Absolute and relative measures can present different aspects of exactly the same data.
Borenstein et al. emphasize that the risk difference is sensitive to baseline risk, while risk and odds ratios tend to be less sensitive to baseline event rates. That sensitivity is not simply a statistical defect: when the objective is to communicate clinical impact, an absolute effect can be particularly informative (Borenstein et al., 2021).
Thus, the question should not be whether the absolute risk difference or relative measure gives the larger-looking number. The question is which quantity represents the effect the researcher wants readers to understand.
Odds Ratio: A Relative Comparison of Odds
For each group:
odds = p / (1 − p)
The odds ratio compares those two odds:
OR = [p1 / (1 − p1)] / [p0 / (1 − p0)]
The null value is 1.
An OR above 1 means higher odds of the modeled event in the numerator group; an OR below 1 means lower odds. Reversing the comparison can invert the odds ratio, so researchers should always identify the event, numerator group, and reference group.
Essential interpretation: An OR of 2 means twice the odds—not automatically twice the probability or twice the risk.
When Do the Odds Ratio and Risk Ratio Differ Materially?
The odds ratio can resemble the risk ratio when the event probabilities involved are small. Altman's discussion of case-control data shows the basis of this approximation: when the relevant outcome is rare, probabilities and corresponding odds are numerically close enough for the odds ratio to approximate a relative risk under the setting he describes (Altman, 1991).
The approximation should not be generalized indiscriminately.
As event risk increases, probability p and odds p / (1 − p) increasingly diverge. Consequently, the numerical difference between an OR and an RR can become important. Modern Epidemiology notes that, in comparisons where the measures depart from their nulls, the risk ratio would generally be expected to lie closer to the null than the corresponding odds ratio (Lash et al., 2021).
Major reporting trap: An OR substantially above 1 should not be translated directly into the same proportional increase in risk unless such a conversion is actually justified.
Study Design Changes What You Can Estimate
Effect-measure choice is not only about interpretability. The sampling design determines which quantities can be estimated directly.
Randomized trials and cohort studies
In prospective designs where participants are followed from defined groups and event risks can be estimated, risk differences and risk ratios can be calculated directly.
Odds ratios can also be calculated, but their availability does not make them automatically preferable.
Decision: Ask whether the scientific target is an absolute contrast, a relative contrast, or an odds-based contrast.
Case-control studies
Ordinary case-control sampling selects subjects according to outcome status. The observed proportion of cases is therefore determined partly by the investigator's sampling scheme and cannot be interpreted as the population risk of disease among exposed or unexposed subjects (Altman, 1991).
A conventional risk ratio cannot simply be calculated from those sample disease proportions as though the study were a cohort.
Decision: The odds ratio has a special role because it can be calculated from cross-classified exposure and outcome data despite outcome-based sampling.
Altman further explains the rare-outcome approximation under which the odds ratio can approximate the relative risk (Altman, 1991).
Practical rule: Do not choose RR merely because it is easier for readers to understand when the sampling design does not permit the required risks to be estimated directly.
Why Logistic Regression Produces Odds Ratios
A major reason odds ratios appear so often in medical literature is logistic regression.
For a binary outcome with probability p, logistic regression models:
log[p / (1 − p)]
as a function of predictors. A regression coefficient therefore represents a change on the log-odds scale. Exponentiating that coefficient gives an odds ratio.
The resulting OR should be interpreted on the odds scale. In a multivariable model, it is also conditional on the other covariates represented in the model (Agresti, 2013).
Methodological distinction: “My software fitted logistic regression and reported an odds ratio” is an explanation of the model output. It is not, by itself, a scientific justification for choosing the odds ratio as the primary effect measure.
If the scientific target is an absolute risk difference or another probability-based contrast, researchers should distinguish that target from the default coefficient scale of the fitted model.
Relative and Absolute Effects Answer Different Questions
The contrast between relative and absolute measures is central to choosing binary outcome effect measures.
A relative measure asks about proportional change. An absolute measure asks about the change in probability itself.
Borenstein et al. note that RR and OR tend to be relatively insensitive to differences in baseline event risk, whereas RD is highly sensitive to baseline risk. They therefore describe situations where a ratio measure can be useful for synthesis while an absolute measure can better convey clinical impact (Borenstein et al., 2021).
Modern Epidemiology adds an important conceptual point: effect-measure behavior itself depends on scale. Risk differences and risk ratios can show different patterns across strata of the same population. An effect that is uniform on one scale need not be uniform on another (Lash et al., 2021).
There is no scale-free answer to “How large is the effect?” Researchers must specify how effect is being measured.
A Decision Framework for Choosing the Effect Measure
1. Define the target estimand
Ask what the scientific question actually requires before opening the software.
- For an absolute change in event probability, consider RD.
- For a proportional comparison of event risks, consider RR.
- For a comparison of odds, OR is the relevant measure.
Decision: Do not substitute one merely because another is easier to obtain computationally.
2. Check the study design
Risk-based measures require meaningful estimation of event probabilities in the compared groups.
In conventional case-control sampling, sample disease proportions do not directly estimate population disease risks, so the OR has a design-specific role that RR and RD calculated from those sample proportions do not have (Altman, 1991).
3. Separate scientific and model scales
Logistic regression naturally produces odds ratios. That does not make an OR synonymous with the desired scientific effect.
Decision: Report and interpret the quantity the model actually estimates, and do not silently rename an odds ratio a relative risk.
4. Consider absolute impact
A relative effect without baseline risk can conceal the practical magnitude of an effect.
Where absolute consequences are important, the risk difference or underlying group risks can provide information that a ratio alone cannot communicate (Borenstein et al., 2021).
5. Consider synthesis across studies
Meta-analysis requires each study to contribute an effect on a sufficiently coherent scale for the synthesis.
For prospective binary data, RR, OR, and RD are alternative effect-size indices, and the choice should consider both substantive and technical factors (Borenstein et al., 2021).
Odds Ratio vs Risk Ratio in Meta-Analysis
For meta-analysis, effect-measure choice is part of defining what the pooled effect means.
Borenstein et al. analyze risk ratios and odds ratios on logarithmic scales. The study effects and their variances are handled on the log scale, after which pooled results and confidence limits are transformed back to the ratio scale for presentation. Risk differences, in contrast, are handled in raw units (Borenstein et al., 2021).
The choice should not be made by asking which measure happens to be reported by the largest number of papers.
Meta-analysis question: What common effect does the meta-analysis intend to summarize?
If the synthesis concerns proportional differences in risk, RR may align with the target. If odds are the intended scale, OR is appropriate. If the objective is an absolute change in event probability, RD addresses that question.
Baseline risk also matters. Because absolute risk differences are sensitive to baseline event risk, variation in baseline risk across studies can produce variation in RD even when a relative effect is more stable. Conversely, that same baseline dependence may make the absolute measure valuable when the substantive objective is clinical impact rather than relative constancy (Borenstein et al., 2021).
Confidence Intervals Are Part of the Effect Estimate
Reporting an OR, RR, or RD without uncertainty is incomplete.
For RR and OR, conventional large-sample inference commonly works on the logarithmic scale and transforms confidence limits back to the ratio scale. Agresti and Altman both describe this approach for relative risks and odds ratios (Agresti, 2013; Altman, 1991).
| Effect measure | Null value | Meaning at the null |
|---|---|---|
| Risk ratio (RR) | 1 | Equal risks |
| Odds ratio (OR) | 1 | Equal odds |
| Risk difference (RD) | 0 | No absolute difference in risk |
But the confidence interval should not be reduced to a yes/no question about whether it contains the null. Altman's broader emphasis on confidence intervals is that they convey information about both the estimated magnitude and its precision.
Practical interpretation: What range of scientifically important effects remains compatible with the data?
A wide interval may encompass materially different conclusions even when its point estimate looks impressive.
Reporting Recommendations for Binary Outcomes
A clear binary-outcome report should make the estimand and scale unmistakable.
Where the design and analysis permit, report the event risks in the comparison groups alongside the chosen effect measure. Then state explicitly whether the effect is an RD, RR, or OR, give its confidence interval, and identify the event and reference group.
Avoid
“Participants were 2.4 times more likely to experience the outcome”
when 2.4 is actually an odds ratio.
Prefer
“The estimated odds of the outcome were 2.4 times those in the comparison group.”
Similarly, do not let statistical significance substitute for magnitude. A binary-outcome effect should be interpreted in terms of its scale, estimated size, uncertainty, baseline risks where available, and scientific or clinical relevance.
Common Mistakes
Calling an odds ratio a risk ratio
OR and RR are different mathematical quantities. An OR should remain an odds-based interpretation unless a justified transformation or approximation is being made (Agresti, 2013; Altman, 1991).
Reporting “times more likely” for an odds ratio
Likelihood language suggests probability or risk. An OR directly compares odds.
Choosing OR because logistic regression produced it
Model output does not define the scientific estimand. Logistic regression has an odds-based coefficient scale; the research question still needs to be defined independently.
Assuming the OR is always close to the RR
The approximation is most defensible in settings where the relevant event probabilities are small. The measures can diverge materially as risks increase (Altman, 1991; Lash et al., 2021).
Reporting only a relative measure
A large relative effect can coexist with a small absolute difference when baseline risk is low. Relative and absolute effects answer different questions (Borenstein et al., 2021).
Calculating risks directly from ordinary case-control samples
Outcome-based sampling means the observed case proportion is not a population disease risk. Study design must be incorporated into effect-measure interpretation (Altman, 1991).
Selecting the most dramatic number
Effect measures are not alternative ways to optimize presentation. Choose the scale that represents the prespecified scientific question.
Practical Reporting Checklist
Before reporting an effect for a binary outcome, check:
- Have I clearly defined the event?
- Have I identified the target population and comparison groups?
- Is my scientific question about an absolute risk contrast, relative risk, or odds?
- Does the study design permit the required risks to be estimated?
- If this is a case-control study, have I accounted for outcome-based sampling?
- If I used logistic regression, am I interpreting exponentiated coefficients as odds ratios rather than risk ratios?
- Have I avoided “times more likely” wording when reporting an OR?
- Have I reported the underlying group risks where they are estimable and useful?
- Have I reported a confidence interval with the effect estimate?
- Have I interpreted the interval in terms of magnitude and precision rather than only statistical significance?
- If baseline risk affects practical interpretation, have I made that clear?
- For meta-analysis, do all included estimates represent a coherent effect on the chosen scale?
- Does my conclusion distinguish association from causal effect when the study design requires that distinction?
Bottom Line
The question odds ratio vs risk ratio cannot be answered by declaring one universally superior.
Use the risk ratio when the estimand is a proportional comparison of risks and the design permits those risks to be estimated. Use the risk difference when the scientific or communication goal is an absolute change in event probability. Use the odds ratio when odds are the relevant target, when working on the logistic-regression scale, or in designs such as conventional case-control studies where its sampling properties give it a particular role (Agresti, 2013; Altman, 1991; Borenstein et al., 2021).
Most importantly, report the measure as what it actually is. An odds ratio is not a risk ratio, an absolute effect is not interchangeable with a relative effect, and software output should not be allowed to define the scientific question.
Reporting workflow: define the estimand → respect the study design → select the corresponding effect measure → report its confidence interval → provide the underlying risks or absolute context when available and scientifically useful.
FAQs
What is the main difference between odds ratio and risk ratio?
A risk ratio compares two probabilities or risks. An odds ratio compares two odds, where odds are p / (1 − p). They are therefore different quantities, even though both have a null value of 1 (Agresti, 2013).
Can I interpret an odds ratio as a relative risk?
Not automatically. When event risks are small, an odds ratio can approximate a risk ratio in appropriate settings, but the approximation becomes less reliable as event probabilities increase. An OR should otherwise be interpreted as a ratio of odds (Altman, 1991; Lash et al., 2021).
What is the null value for an odds ratio and risk ratio?
Both OR and RR have a null value of 1, representing equal odds or equal risks, respectively. The risk difference has a null value of 0, representing no absolute difference in risk.
When should researchers report risk difference?
Consider risk difference when the target is an absolute effect: the difference in event probability between two groups. It can be particularly useful for communicating practical or clinical impact, although its magnitude depends strongly on baseline risk (Borenstein et al., 2021).
Why does logistic regression report odds ratios?
Binary logistic regression models the log odds of the event. Exponentiating a regression coefficient therefore produces an odds ratio for the corresponding predictor contrast. That OR should not automatically be renamed a risk ratio (Agresti, 2013).
Why are odds ratios commonly used in case-control studies?
In conventional case-control studies, participants are sampled according to outcome status, so sample disease proportions do not directly estimate population risks. The odds ratio can nevertheless be estimated from the case-control cross-classification, giving it a special role in this design (Altman, 1991).
Should researchers report both relative and absolute effects?
The approved sources support the value of distinguishing the two because they answer different questions. A ratio describes proportional effect, whereas the risk difference describes absolute impact. Where both quantities are estimable and relevant to the research and communication goals, presenting absolute context alongside a relative effect can prevent readers from mistaking relative magnitude for absolute impact (Borenstein et al., 2021).
Which effect measure should be used in a meta-analysis of binary outcomes?
There is no universally preferred measure. Borenstein et al. identify RR, OR, and RD as legitimate binary-outcome effect-size indices and recommend considering substantive as well as technical factors. The chosen measure should represent a coherent effect across the studies being synthesized (Borenstein et al., 2021).
References
Agresti, A. (2013). Categorical data analysis (3rd ed.). Wiley.
Altman, D. G. (1991). Practical statistics for medical research. Chapman & Hall.
Borenstein, M., Hedges, L. V., Higgins, J. P. T., & Rothstein, H. R. (2021). Introduction to meta-analysis (2nd ed.). Wiley.
Lash, T. L., VanderWeele, T. J., Haneuse, S., & Rothman, K. J. (2021). Modern epidemiology (4th ed.). Wolters Kluwer.
Need help with a similar research question?
Share a short, non-confidential summary of your study and the decision you need to make.