Understanding Confidence Intervals Simply
A confidence interval puts a range around an estimate to show the uncertainty created by sampling. Learn how to read the bounds, margin of error, confidence level, and the limits that a narrow interval cannot solve.
What question does a confidence interval answer?
A confidence interval answers a practical question: what range of population values is reasonably compatible with the data, using a stated statistical method? It adds needed context to a point estimate. A point estimate is one number calculated from a sample, such as 52% of surveyed customers saying they would recommend a service. The unknown population value is the percentage that would answer that way if every relevant customer could be surveyed.Reporting only 52% can make the estimate look more exact than it is. Another random sample of the same size would usually give a slightly different result. A confidence interval expresses that sampling uncertainty as a lower and upper limit. For example, an interval from 49% to 55% says that, under the method and its assumptions, values in that range are plausible population percentages.The interval is not a guaranteed minimum and maximum for individual people, future observations, or every possible source of error. It is an interval for a population parameter, such as a population mean or proportion. The parameter, sampling plan, and calculation method all matter.
Read the estimate, margin of error, and limits
A simple symmetric confidence interval has three visible parts:Point estimate: the sample-based central estimate.Margin of error: the distance from the estimate to either endpoint.Confidence limits: the lower and upper endpoints of the interval.When a report already provides a margin of error, the arithmetic is straightforward:Lower limit = estimate − margin of errorUpper limit = estimate + margin of errorSuppose a survey estimate is 52%, with a margin of error of 3 percentage points. The lower limit is 52 − 3 = 49%, and the upper limit is 52 + 3 = 55%. The interval is 49% to 55%.Percentage points and percent change are different. Moving from 52% to 55% is an increase of 3 percentage points. It is also about a 5.8% relative increase because 3 divided by 52 is about 0.058. In confidence-interval reporting for a percentage, the margin of error is commonly stated in percentage points.For a measured quantity rather than a percentage, retain the original unit. A sample mean of 4.75 units with a margin of error of 0.36 units gives 4.39 to 5.11 units. Rounding should not create false precision: use enough decimal places to keep the endpoints meaningful, while matching the measurement’s practical resolution.
What a 95% confidence level really means
The most common misunderstanding is to say, “There is a 95% probability that the true value is inside this interval.” In the usual frequentist interpretation, that is not quite right. Once the data have been collected and the interval calculated, the population parameter is fixed and the particular interval either contains it or does not.The 95% refers to the procedure before the sample is observed. If researchers repeatedly drew comparable samples and used the same valid 95% method each time, about 95% of the resulting intervals would contain the true parameter in the long run. The uncertainty is in the sample and the interval-building process, not in the already-fixed population value. The NIST/SEMATECH Engineering Statistics Handbook explains this repeated-sampling interpretation for confidence limits.A confidence level is not a quality grade. A 99% interval is not automatically based on better data than a 95% interval; it is normally wider because its method is designed to capture the parameter more often over repeated sampling. Likewise, 95% is a convention in many fields, not a universal rule. The suitable level depends on the decision, the consequences of error, and the reporting standard being used.
Worked example: a survey percentage
Imagine a properly designed survey of a defined customer population. The survey estimates that 52% would recommend a service. Its chosen interval method reports a 95% margin of error of 3 percentage points.Start with the estimate: 52%.Subtract the margin of error: 52% − 3 percentage points = 49%.Add the margin of error: 52% + 3 percentage points = 55%.Report the result: 52%, with a 95% confidence interval from 49% to 55%.A careful interpretation is: assuming the sampling design, response process, and interval method are appropriate, the procedure used to create this interval has 95% long-run coverage. The result communicates sampling uncertainty around the estimate. It does not show that exactly 49% to 55% of every subgroup will recommend the service, nor does it prove the survey has no bias.If you need to check percentage arithmetic separately, use the site’s Percentage Calculator. It can help with the endpoint calculations, but it cannot choose the correct confidence-interval method or judge whether a survey sample represents its target population.
Why some intervals are wider or narrower
Formal confidence intervals often follow this general structure: point estimate ± critical value × standard error. The standard error describes how much a statistic would tend to vary from sample to sample under the model. Penn State’s confidence-interval overview presents this general form and explains that variability, sample size, and confidence level affect interval width.First, larger samples often reduce the standard error when the design and population conditions are otherwise comparable. For a simple mean, the standard error has a square-root relationship with sample size, so multiplying the sample size by four can roughly halve that component of uncertainty. It does not usually make an interval four times narrower.Second, more variable observations produce wider intervals. Estimating an average from values that differ greatly from one another is less precise than estimating it from otherwise similar values, all else equal.Third, a higher confidence level usually requires a larger critical value. Holding the same data and method fixed, a 99% interval is generally wider than a 95% interval, which is generally wider than a 90% interval. The higher level asks the procedure to cover the parameter more often across repeated samples.Finally, study design matters. Clustering, weighting, unequal selection probabilities, missing responses, and complex surveys can change the standard error and require design-aware methods. A simple formula intended for an independent random sample may be inappropriate for data collected in schools, households, clinics, or online panels.
Do not compare two intervals only by eye
Two estimates with overlapping confidence intervals may still differ under a formal comparison, and two non-overlapping intervals do not tell the full story of the comparison method. The answer depends on what is being compared and whether the estimates are independent, paired, weighted, or derived from a shared model.For example, comparing two groups is usually best handled by constructing a confidence interval for the difference between the groups, rather than inspecting two separate intervals. If a difference interval includes zero, that has a specific interpretation under the chosen method and confidence level. It is still not a substitute for considering the size and practical importance of the difference.Method choice also matters near boundaries. A simple symmetric interval for a proportion can extend below 0% or above 100%, which is impossible for a percentage. That is a signal to check whether the method is suitable, particularly when the sample is small or the observed proportion is near 0% or 100%. A valid interval should be selected for the statistic, distributional conditions, sample size, and sampling design—not merely because its arithmetic is easy.
What confidence intervals cannot fix
A narrow interval describes limited sampling variability under a method. It does not prove that the estimate is accurate. If the sample leaves out important people, question wording systematically pushes responses, measurements are faulty, or nonresponse is related to the outcome, the estimate can be biased. More data collected with the same systematic flaw can create a very narrow interval around the wrong value.This distinction between precision and accuracy is essential when reading polls, health studies, business reports, and news summaries. The CDC notes that sampling bias can occur when members of a population have unequal chances of inclusion and that a small margin of error does not necessarily mean the estimate is accurate. See the CDC’s sampling-bias guidance.Before relying on an interval, ask: What population does it describe? How was the sample selected? How many observations were missing? What quantity was estimated? Which interval method and confidence level were used? Were the model assumptions checked? These questions are often more useful than focusing only on whether the interval is narrow.Confidence intervals are a disciplined way to show uncertainty, not a promise of certainty. Read the estimate, the bounds, the confidence level, and the study design together.
Enable JavaScript to use interactive calculators.