Skip to main content

Welcome to a Grasp lesson

Create a free account to keep studying and create your own study path.

Lesson 3

Interpreting Confidence Intervals: Evidence, Uncertainty, and Rare Harms

Recommended resources

The previous lesson turned relative risk changes into events per 1,000 people. Those calculations used a single effect estimate, but a study cannot measure the exact treatment effect in everyone. This lesson adds the missing piece: a confidence interval shows how much uncertainty surrounds an estimate.

You will use intervals to distinguish evidence of an effect from an inconclusive result. This matters particularly for rare statin harms: a study that finds no clear increase may still be too imprecise to rule one out.


What a confidence interval tells you

Imagine that a study estimates a treatment prevents 20 cardiovascular events per 1,000 people over ten years. That is its point estimate: one number calculated from the people studied. A different, comparable sample might yield a somewhat different number. A 95% confidence interval (95% CI) expresses the range of effect sizes reasonably compatible with the study’s data under its statistical method.

Read Examine’s explanation of why an interval can be narrow or wide, and its contrasting examples of precise and inconclusive results.

Confidence interval - Examine - Examine

Examine’s “Confidence interval” explains what the range adds to a study’s point estimate. Its examples help separate the size of an effect from the precision with which it was measured.

Under “Summary” and “What is a confidence interval?”, read the explanation of interval width and the meaning of 95%. Then, under “Examples of interpreting confidence intervals,” read the first example from the risk-ratio interpretation. Also read the blood-pressure supplement example through its conclusion. Focus on how the endpoints of each interval change what can be concluded from the point estimate.

The “95%” has a specific statistical meaning: if researchers repeatedly used the same method in comparable samples, approximately 95% of the intervals they calculated would contain the true population value. It does not mean there is a 95% probability that the true value lies inside this particular, already calculated interval.

A narrow interval generally indicates greater precision; a wide one leaves more possibilities open. Neither tells you by itself whether a study was well designed. Randomization, blinding, complete outcome collection, and relevance to the people you care about remain important.



First find the value for “no difference”

To read an interval, identify the measure before inspecting its numbers. The value representing no difference between groups depends on that measure:

Effect measure No-difference value Why
Risk difference, such as additional events per 1,000 Zero additional events means equal risks.
Risk ratio (RR), odds ratio (OR), or hazard ratio (HR) A ratio of one means equal values on that measure.

Watch a short visual explanation of confidence-interval bars and the vertical line representing no difference. The video’s plotted measure uses zero as its no-difference value; for the ratios in the statin figure later, use one instead.

Forest Plot Interpretation - Clearly Explained

In “Forest Plot Interpretation - Clearly Explained,” Steven Bradburn shows how a point estimate, horizontal confidence-interval bar, and no-difference line work together.

Watch interval width for the horizontal bars, then the reference line for how crossing it changes the statistical conclusion and how the direction of an effect depends on what is being compared. The video uses a measure whose no-difference value is zero, so transfer the visual method—not that particular number—to a plot of odds ratios.

The BMJ Best Practice summary supplies the corresponding rule for ratios and introduces a second distinction: a detectable effect is not necessarily an important one.

Understanding statistics: risk | BMJ Best Practice

BMJ Best Practice’s “Understanding statistics: risk” provides a compact guide to the no-difference values for ratios and differences, alongside the idea of a minimally important effect.

In “Significant difference,” read the bullets on RR, OR, and HR and the following “NOTE” about absolute differences. The paragraph beginning if the CI includes one is the key rule. Continue through “Minimal Clinically Important Difference (MCID),” including the importance threshold; stop before “Useful tools.”

For a conventionally calculated 95% CI, an interval entirely on one side of the no-difference value provides statistical evidence of an effect in that direction, subject to the study’s design and assumptions. If the interval includes the no-difference value, the result is statistically inconclusive at the conventional 5% threshold. Inconclusive means the data remain compatible with no effect—not that the study has proved no effect.

Now inspect the whole interval, not just whether it crosses a line. A range that includes no effect and a substantial increase in harm gives a different answer from one tightly clustered around no effect. Likewise, an interval excluding no effect may describe a change too small to matter much in practice.



Connect the interval to people affected

Return to the previous lesson’s hypothetical group: without treatment, 80 per 1,000 people experience a defined cardiovascular event over ten years. Suppose a suitable study estimates an RR of , with a 95% CI from to . Ratios below favor treatment for this undesirable outcome, and the entire interval is below , so the data provide statistical evidence of fewer events.

Using the same baseline risk and ten-year period, the point estimate implies events per 1,000 with treatment: 20 fewer. The interval’s endpoints imply:

Thus, on these assumptions, the data are compatible with approximately 4 to 28 fewer events per 1,000 over ten years. Reporting only “20 prevented” would hide that uncertainty.

Suppose instead that the same point estimate, , came with a wider 95% CI of to . At the upper endpoint, events per 1,000—8 more, rather than fewer. This interval is compatible with outcomes ranging from 28 fewer to 8 more per 1,000. It includes , so the apparent benefit is inconclusive, not disproved.

These translations are illustrations, not statin trial findings. They also treat the baseline risk of 80 per 1,000 as fixed; a real baseline estimate can have uncertainty of its own. Even an interval wholly below does not, by itself, establish causation if a study has substantial bias or confounding.



Read a statin safety figure without overreading it

The forest plot below reports odds ratios for several adverse-event categories by statin type, each compared with a control reference group. Each marker is an estimate; its horizontal line is a 95% CI. The vertical reference line is at an OR of , meaning no difference in odds.

Forest plot showing odds-ratio estimates and 95% confidence intervals for adverse-event categories by statin type versus a control reference group. Horizontal intervals that cross the vertical line at an odds ratio of 1 remain compatible with no difference.

Consider two rows, keeping their outcome labels distinct:

  • Under muscle symptoms, the rosuvastatin row reports an OR of , with a 95% CI of to . The interval lies just above : it provides statistical evidence of higher odds of reported muscle symptoms in that comparison. It does not tell us the additional number of people affected without a suitable baseline risk, and an odds ratio must not automatically be treated as a risk ratio.
  • Under muscle disorders, the simvastatin row reports an OR of , with a 95% CI of to . Its point estimate is below , but the interval spans values consistent with substantially lower odds, no difference, or substantially higher odds. This row cannot establish that simvastatin prevents muscle disorders or that it causes them.

“Muscle disorders” in this figure must not be silently relabelled “rhabdomyolysis” or another specific serious condition. Nor should one row be generalized to every statin, dose, or population. The exercise here is to read what each reported outcome and interval can support; assessing causation also requires knowing how the underlying comparison was conducted.



Why rare harms are especially hard to exclude

A rare event can leave researchers with very few cases even when many people participate. With little event information, the estimated effect may have a wide interval. No statistically clear increase can therefore coexist with uncertainty about a potentially meaningful increase.

For scale, imagine an entirely hypothetical serious outcome that occurs in 1 per 10,000 untreated people over five years. Suppose a study estimates an RR of , with a 95% CI of to . Applied to that assumed baseline, the interval is compatible with approximately to events per 10,000 treated people over five years. Its upper endpoint permits 6 additional events per 10,000, despite the interval also including no increase. The figures are illustrative, not estimates of any statin harm.

To judge whether a study offers reassurance about a particular rare harm, examine its upper confidence limit, the absolute event frequency that limit would allow, and whether the outcome was captured reliably for long enough. A tightly estimated result for common muscle symptoms cannot rule out a rare, different muscle injury. A confidence interval chiefly represents statistical sampling uncertainty; incomplete adverse-event collection or participants unlike those seen in practice pose additional limitations. Later lessons will examine those safety questions more closely.



Key takeaways

A point estimate gives one best estimate from a study; its confidence interval shows the range of effects compatible with the data under the method used. For risk differences, compare the interval with ; for RR, OR, and HR, compare it with . Crossing the no-difference value means inconclusive, not proof of no effect. Then examine the interval’s endpoints and, when appropriate, translate them into absolute frequencies over a stated period.

This provides a way to read the uncertainty in cardiovascular benefit estimates as well as harm estimates. The next lesson uses worked primary- and secondary-prevention cases to compare absolute cardiovascular benefits while keeping events and deaths separate.