Confidence Level vs Confidence Interval (Definition)
The confidence level is chosen before any data is collected. It describes how the interval-construction method performs under repeated sampling — not the probability that any one computed interval captures the parameter. The interval itself is calculated from the sample and changes every time a new sample is drawn.
Writing "my confidence level is 46–54" mixes the two up. Writing "my confidence interval is 95%" does the same in the other direction. The correct form is: 95% CI [46.08, 53.92].
- Confidence level: 90%, 95%, or 99% — the procedure's long-run coverage target, set by the analyst before analysis
- Confidence interval: The sample-dependent range [lower, upper] — calculated from data, changes with each new sample
- Higher confidence level: Requires a larger critical value, which produces a wider interval (for fixed data and method)
- The 95% is not: A probability that the fixed parameter is inside this one specific computed interval
- Confidence level = 1 − α: For 95% CI, α = 0.05; for 99% CI, α = 0.01
Confidence Level vs Confidence Interval at a Glance
| Feature | Confidence Level | Confidence Interval |
|---|---|---|
| What is it? | Long-run coverage rate of the procedure | Range calculated from one sample |
| Example | 95% | [46.08, 53.92] |
| Format | Percentage | Lower and upper numerical bounds |
| Chosen or calculated? | Chosen before analysis | Calculated from the sample |
| Changes with each sample? | No — fixed by the analyst | Yes — depends on sample data |
| Related to α? | Confidence level = 1 − α | Built using the critical value for that α |
| Drives interval width? | Yes — higher level → wider interval | Width is the result, not the driver |
| Gives probability for this exact range? | No — not in frequentist statistics | The interval either contains the parameter or it does not |
What Is a Confidence Level?
Common values are 90%, 95%, and 99%. The analyst selects the confidence level before collecting or analyzing data — it is not calculated from the sample. Its value reflects a deliberate decision about the tradeoff between how often the procedure captures the parameter (coverage) and how wide the resulting intervals are (precision).
Choosing 95% is conventional in many fields, but there is nothing mathematically special about it. A 90% level produces narrower intervals that miss the true parameter more often under repeated sampling. A 99% level produces wider intervals that miss less often. The right choice depends on field conventions and the consequences of missing the true value.
What Is a Confidence Interval?
When a study reports "95% CI [1.2, 4.8] for the treatment effect," the [1.2, 4.8] part is the confidence interval. Calculated from sample data using a point estimate, a standard error, and a critical value, it changes every time a new sample is drawn. The 95% — the confidence level — stays fixed.
Once an interval has been computed from a specific dataset, the true population parameter either lies inside that range or it does not. There is no probability to assign to the specific computed interval. The 95% property belongs to the procedure over hypothetical repetitions, not to the specific numbers [46.08, 53.92].
The Main Difference Between Confidence Level and Confidence Interval
The confidence level describes how the method performs across many repetitions. The confidence interval is the concrete numerical output from one dataset. One is a property of the procedure; the other is a product of the data.
Writing "my confidence level is 48–54" is wrong — that's the interval. Writing "my confidence interval is 95%" is equally wrong — that's the level. Always report both together: 95% CI [48, 54].
Example — 95% Confidence Level vs a 95% Confidence Interval
A pollster surveys 400 registered voters. 52% say they prefer Candidate A. The standard error of this proportion is approximately 2.5 percentage points. Here is how the two terms apply:
Point estimate = 52% | Standard error ≈ 2.5 percentage points
Confidence level chosen: 95% — the analyst selected this before seeing the data. It is the procedure's long-run coverage target, not a property of this specific poll.
Critical value: For a two-sided 95% normal interval, z* ≈ 1.960.
Margin of error: 1.960 × 2.5 ≈ 4.9 percentage points (≈ 5 pp rounded).
Confidence interval: 52% ± 5% = [47%, 57%] — the numerical range computed from this specific sample of 400 voters.
✅ Correct reporting: "Candidate A is supported by 52% of voters, 95% CI [47%, 57%]."
The 95% is the confidence level. The [47%, 57%] is the confidence interval.
How Confidence Level Changes Confidence Interval Width
This is the most practically important relationship between the two concepts. When the sample, the point estimate, the standard error, and the interval method are all held constant, changing only the confidence level changes the width of the resulting interval.
z* comes from the confidence level
SE comes from the data
Higher confidence → larger z* → wider interval
The symmetric estimate ± MOE form applies to many common intervals (means, proportions with normal approximation). Some intervals are asymmetric — for odds ratios, hazard ratios, or certain proportion methods — and a single symmetric margin of error does not fully describe them. For t-based intervals the critical value also depends on degrees of freedom, not just the confidence level.
90% Confidence Interval
Point estimate = 50, standard error = 2. For a two-sided 95% normal procedure at 90% confidence, z* = 1.645:
50 ± 1.645 × 2 = 50 ± 3.29 = [46.71, 53.29] Width = 6.58
Narrowest of the three intervals. The tradeoff: lower long-run coverage (90% vs 95% or 99%).
95% Confidence Interval
Same data, z* = 1.960 at the 95% level:
50 ± 1.960 × 2 = 50 ± 3.92 = [46.08, 53.92] Width = 7.84
The conventional choice in most fields. Wider than the 90% CI because z* is larger.
99% Confidence Interval
Same data, z* = 2.576 at the 99% level:
50 ± 2.576 × 2 = 50 ± 5.152 = [44.85, 55.15] Width = 10.30
Widest of the three. Higher target coverage, but less numerical precision for the same dataset.
The bars below show all three intervals on the same scale. Point estimate and standard error are identical in each case — only the confidence level changes.
Confidence Level, Alpha, and Critical Values
For a standard two-sided confidence procedure, the confidence level and the significance level α are linked directly:
95% confidence → α = 0.05
90% confidence → α = 0.10
99% confidence → α = 0.01
| Confidence Level | α | Each tail (two-sided equal-tail normal) | z* (normal) |
|---|---|---|---|
| 90% | 0.10 | 0.05 | 1.645 |
| 95% | 0.05 | 0.025 | 1.960 |
| 99% | 0.01 | 0.005 | 2.576 |
For t-based intervals, the critical value also depends on degrees of freedom — there is no single universal "95% critical value" for all t procedures. See the t interval vs z interval guide for when each applies.
How Sample Size Affects Confidence Interval Width
Sample size and confidence level both influence interval width, but through entirely different mechanisms. The confidence level controls the critical value. Sample size controls the standard error via SE = σ/√n (or s/√n).
A larger sample reduces the standard error, which narrows the confidence interval at any given confidence level. You can achieve a narrower interval while keeping 95% confidence by collecting more data — you do not have to sacrifice coverage by dropping to 90%.
| Change | Confidence Level | Interval Width |
|---|---|---|
| 90% → 95% (same data, same method) | Higher | Wider |
| 95% → 99% (same data, same method) | Higher | Wider |
| Larger sample, same 95% level | Unchanged | Usually narrower |
| Lower population variability, same 95% | Unchanged | Usually narrower |
What Does 95% Confidence Really Mean?
The 95% confidence level describes how the procedure performs over repeated sampling, not the probability that any one computed interval contains the parameter. Run the experiment many times, each time drawing a fresh sample and computing a 95% CI. Over a large number of repetitions, about 95% of those intervals will contain the true population parameter. Any one particular interval either does or it does not.
Interactive: Repeated Sampling Simulation
Each horizontal bar is a 95% CI computed from a simulated sample (population mean μ = 50, σ = 10, n = 30). Green bars contain μ; red bars miss it. Roughly 95% cover the true value across many batches — but never exactly 95 in any single run.
The true coverage in any one batch will not be exactly 95%. That is expected, not a failure of the procedure. Over many batches the average converges toward the nominal 95%.
Why 95% Does Not Mean a 95% Chance for This Specific Interval
In standard frequentist statistics, a population parameter is treated as a fixed unknown constant, not a random variable. Because the parameter is fixed, a probability statement about where it sits cannot be assigned once the sample is observed.
Before sampling: "The procedure I will use produces intervals that contain the true mean 95% of the time." That describes the method. After computing [46.08, 53.92] from one sample: the true mean is either in that range or it is not.
It does not mean there is a 95% probability that the true parameter lies inside the specific interval [46.08, 53.92]. Under the frequentist framework, the parameter is fixed. The computed interval either contains it or it does not. The 95% refers to the procedure's long-run performance over hypothetical repetitions.
The phrase "we are 95% confident the interval [46.08, 53.92] contains the population mean" is commonly used in practice. It is acceptable shorthand, provided the reader understands the 95% refers to the procedure's performance property, not a posterior probability about the specific computed range.
A Bayesian credible interval supports a direct probability statement: "there is a 95% posterior probability the parameter lies between X and Y" — given a specified prior and model. Frequentist confidence intervals and Bayesian credible intervals answer different questions and should not be used interchangeably.
Common Misconceptions
| Mistake | Why it is wrong | Correct version |
|---|---|---|
| "My confidence interval is 95%." | 95% is the confidence level, not an interval. | 95% CI [46.08, 53.92] — report both. |
| "My confidence level is 42 to 58." | 42–58 is the confidence interval (a range). | The confidence interval is [42, 58]; the level is 95%. |
| "There is a 95% probability μ lies in [46.08, 53.92]." | Under frequentist reasoning, the fixed parameter either is or is not inside the computed range. | The 95% CI procedure produces intervals covering μ about 95% of the time under repeated sampling. |
| "95% of sample observations fall inside the CI." | A CI for the mean describes uncertainty about the mean, not where individual data points fall. | Use a prediction interval if you need a range for individual future observations. |
| "Higher confidence means higher precision." | Higher confidence typically widens the interval, which reduces numerical precision. | Higher confidence = wider interval = less precision, for fixed data and method. |
| "A 99% CI is always better than a 95% CI." | "Better" depends on purpose. Higher coverage comes at the cost of a wider, less precise range. | The right confidence level depends on field conventions and decision stakes. |
| "The point estimate is always the midpoint of the CI." | True for symmetric intervals, not for asymmetric ones such as some proportion or ratio intervals. | The midpoint equals the estimate only for symmetric interval constructions. |
| "MOE is always half the interval width." | True for symmetric ± MOE intervals; not universal for asymmetric methods. | MOE = (upper − lower) / 2 only when the interval is symmetric around the estimate. |
Interactive: Compare CI Width by Confidence Level
Enter a point estimate and standard error to see how changing the confidence level changes the interval at all three common levels simultaneously. Uses normal (z) critical values.
Confidence Interval Width Calculator
For small samples with unknown σ, use the full Confidence Interval Calculator, which applies t critical values with the correct degrees of freedom.
Frequently Asked Questions
Sources
- NIST/SEMATECH — e-Handbook of Statistical Methods, Section 7.2.1, "Confidence Intervals." itl.nist.gov
- Freedman, D., Pisani, R. & Purves, R. (2007) — Statistics (4th ed.). W.W. Norton. Standard reference for the frequentist interpretation of confidence intervals.
- Neyman, J. (1937) — "Outline of a Theory of Statistical Estimation Based on the Classical Theory of Probability." Philosophical Transactions of the Royal Society A, 236, 333–380. Original development of confidence intervals.