If you understand one statistics concept for psychiatry board exams, make it this one: a confidence interval tells you both how big an effect probably is and how precisely the study measured it. Everything a bare p value hides — direction, magnitude, clinical importance, the possibility that the study was simply too small — is visible in the interval.
This article explains what a 95% confidence interval means without chest-beating, why interval width is a precision statement, how it differs from a p value, and how to read a real trial result end to end. By the end you should be able to take any odds ratio or mean difference in a vignette and say, in one sentence, what the study did and did not show.
What a confidence interval actually is
Start with the estimation problem. A trial measures an effect in a sample — say a 12 percentage-point difference in response rates — and you want a statement about the underlying population effect. The confidence interval is the honest version of that statement: a range of plausible values for the true effect, computed from the data, using a procedure that has a stated long-run coverage property.
Three things follow: a point estimate at its centre, bounds that are usually asymmetric once effects are ratios, and a statement about plausible values of the parameter — not the probability that a particular patient benefits.
What "95%" really means (and what it does not)
The correct interpretation is procedural, not probabilistic about the data: if you repeated the study many times and built a 95% interval each time, about 95% of those intervals would contain the true population effect. You have one study and one interval; either it captured the true effect or it is in the unlucky 5%, and you have no way of knowing which from the output alone.
The two interpretations candidates often give in vivas are both wrong:
- "There is a 95% probability the true value lies between these bounds." — no; the interval is fixed and the parameter is either in it or not. Only Bayesian credible intervals license a probability statement about the parameter.
- "95% of patients fall within this range." — no; that describes individual observations, a different object entirely. The confusion is common when reading mean depression-score differences.
The 95% figure itself is a convention. 90% intervals are narrower, 99% wider; nothing biological changes at 0.05 — the level is a decision about how much error you are willing to build into the procedure, and it is the level at which the interval and the two-sided p value at 0.05 will always agree on significance.
Width equals precision
Once you accept that the interval is a range of plausible values, its width becomes the most useful number on the page. A narrow interval says: the effect is pinned down, and it is small. A wide interval says: the study cannot yet distinguish between a worthwhile benefit, no effect, and harm.
Width is driven by three levers you can reason about at the table:
- Sample size and event count. More participants — and, for binary outcomes, more events — shrink the standard error. Doubling n narrows the interval by a factor of roughly √2, which is why underpowered trials produce comically wide ranges.
- Variability of the outcome. Continuous outcomes with large standard deviations (most psychometric scales) need bigger samples for the same precision than tightly distributed biomarkers.
- The scale itself. Ratios (odds ratios, risk ratios) are read on a log scale, so their intervals are asymmetric — an OR of 1.5 (1.05–2.13) and an OR of 1.5 (1.10–2.05) have the same point estimate but very different evidential weight.
A useful habit: ask "what is the most benefit, and the most harm, compatible with these bounds?" That range is what you counsel a patient about — the discipline Preppsy's concept library drills into each statistics block.
Confidence interval vs p value
The two come from the same calculation and answer different questions. The p value asks: if there were exactly no effect, how surprising would these data be? It is a single number with no information about magnitude. The confidence interval asks: which effect sizes are compatible with these data? It carries magnitude, direction and precision at once — which is why the best answer to "summarise this trial" is the interval, with the p value if pressed.
This is also where the two most-cited statistics aphorisms land. Altman and Bland's famous BMJ note is one line long: absence of evidence is not evidence of absence. A non-significant p value in an underpowered study tells you almost nothing; a confidence interval that spans from a large benefit to a large harm tells you precisely why you cannot conclude anything. For the companion treatment of hypothesis testing, thresholds and the 0.05 trap, see our guide to the p value in psychiatry exams.
Exam trap: "no significant difference" does not mean the treatments are equivalent
A p value above 0.05 means only that the data did not exclude zero difference. If the confidence interval is wide — for example a risk difference of −8 to +14 percentage points — the study remains compatible with meaningful benefit and meaningful harm, and the correct verdict is "inconclusive", not "no difference". Equivalence or non-inferiority requires a pre-specified margin and an interval that sits inside it. Exam stems routinely reward candidates who say "the study was underpowered; look at the width of the confidence interval" over those who simply report "not significant".
A worked example: reading a trial result
Suppose a randomised trial of a novel antidepressant augmentation strategy reports response in 84 of 200 patients (42%) versus 60 of 200 (30%) on placebo. Here is the same result expressed four ways, with plausible intervals:
| Measure | Point estimate | 95% confidence interval | How to read it |
|---|---|---|---|
| Risk ratio | 1.40 | 1.07 to 1.83 | 40% greater response, plausibly 7% to 83% greater; excludes 1, so significant at 0.05 |
| Odds ratio | 1.69 | 1.12 to 2.55 | Benefit ranges from 12% to 155% higher odds — note the asymmetric bounds |
| Absolute risk difference | +12 percentage points | +3 to +21 | Between 3 and 21 extra responders per 100 patients treated |
| Number needed to treat | 8 | 5 to 37 | Same result as the risk difference, restated: treat 5–37 patients for one extra response |
Notice what the p value alone hides: the absolute difference ranges from a marginal 3-point gain to a 21-point gain, and the lower bound may be below what justifies the drug's side effects. Width matters most here — "NNT 8" sounds precise, but "NNT 5 to 37" says the study cannot yet promise efficiency.
How to phrase results the way examiners want
A defensible answer follows the same three steps every time. State the estimate with its interval ("risk ratio 1.40, 95% CI 1.07 to 1.83"). Translate the bounds into clinical language ("between 7% and 83% greater odds of response"). Name what remains uncertain ("the lower bound represents a small absolute gain, so the result is statistically significant but of uncertain clinical importance"). That last sentence is where most candidates leave marks on the table.
One technical footnote for the exam room: when an interval touches the null value exactly (an upper bound of 1.00), the two-sided p value sits on 0.05 — the two presentations always tell the same story.
Frequently asked questions
How do you interpret a 95% confidence interval in simple terms?
A 95% confidence interval is a range of values that is plausibly compatible with the data you collected. Formally, if you repeated the same study many times, 95% of the intervals you would compute would contain the true population effect. So when a trial reports a risk difference of +12 percentage points with a 95% CI of +3 to +21, the honest summary is that the true effect is somewhere in that span, with the study design — not the p value — determining how tightly it is pinned down.
What does it mean when a confidence interval crosses 1 (or 0)?
It means the null value is among the results compatible with your data, so the corresponding two-sided test at the 5% level is not significant. But crossing 1 by a hair (1.00 to 1.90) is very different from straddling it widely (0.60 to 1.55). In the first case the point estimate still suggests benefit with reasonable precision; in the second the study is uninformative, and "no significant difference" must be reported as "inconclusive" rather than "the treatments are the same".
What is the difference between a confidence interval and a p value?
They come from the same calculation but carry different information. The p value is a single probability: how extreme your data would be if the true effect were exactly zero. It says nothing about the size of the effect. The confidence interval gives the range of effect sizes compatible with the data, so it conveys magnitude, direction and precision simultaneously — which is why reporting the interval (with the p value if required) is the preferred practice in Cochrane reviews and most journals.
Why does confidence interval width tell you about the sample size of a study?
Because interval width is essentially 1.96 times the standard error, and the standard error shrinks as the square root of the sample size (or number of events) grows. Wide intervals therefore signal few participants, few events or highly variable data. If you see a 95% CI for an odds ratio running from 0.7 to 3.4, you can correctly infer that the trial was underpowered long before anyone tells you its sample size.
Key takeaways
- A confidence interval is the range of true effect sizes compatible with the data — magnitude and precision in one statement.
- "95%" describes the long-run behaviour of the procedure, not the probability that this interval is correct.
- Width is a precision signal: wide intervals mean the study cannot distinguish benefit from harm, regardless of the p value.
- Null value 1 for ratios, 0 for differences — crossing it means non-significant at that level, not equivalence.
- Always report point estimate + interval + a clinical translation of the bounds; that three-part answer is what examiners mark.
Confidence intervals are one of several reasoning tools you need to read evidence fluently — if statistics still feels shaky, the MRCPsych pathway lays these concepts out in the order examiners test them.
Cover image: “Confidence Intervals Explained (cover)” by Preppsy Editorial Team (Original artwork) via Preppsy.
