What a Confidence Interval Actually Tells You (and What It Doesn't)

A single number lies by omission. "The average order is 84 dollars." "45 percent support the proposal." Each of those came from a sample, and a different sample would have produced a slightly different figure. A confidence interval is the honest version of the same result: a range, plus a stated level of confidence, that admits how much wiggle room the sampling left behind. Once you read intervals fluently, a bare point estimate starts to look incomplete.

This guide walks through what a confidence interval is for a mean and for a proportion, what the 95 percent label actually promises, where the margin of error comes from, and why more data tightens the range. There is one formula, one fully worked example, and a calculator you can use to reproduce every number.

What 95% confidence actually means

Here is the sentence almost everyone gets wrong: "I'm 95 percent sure the true mean is between 95.84 and 104.16." That reading is intuitive and it is also incorrect. The true mean is a fixed, unknown number. It is either inside that particular range or it is not. There is no probability attached to one finished interval.

The 95 percent is a property of the method, not of any single result. If you repeated the whole exercise many times, drawing a fresh sample each time and building an interval the same way, about 95 percent of those intervals would contain the true value. You only ever see one of them, and you accept that the procedure that produced it has a 95 percent hit rate over the long run. That distinction sounds pedantic until someone makes a decision on the false reading, treats one interval as a 95 percent bet, and is surprised when reality lands in the 5 percent.

So the right way to write up a result is to describe the confidence as the reliability of the procedure: "a 95 percent confidence interval of [95.84, 104.16]," not "a 95 percent chance the truth is in this range."

The formula for a mean, and where z comes from

For a population mean, the confidence interval is:

CI = mean ± z × (sd / √n)

The mean is your sample average, sd is the sample standard deviation, and n is the sample size. The piece in parentheses, sd / √n, is the standard error of the mean: how much the sample average itself tends to bounce around from sample to sample. The z is the critical value that turns that standard error into a margin for your chosen confidence level.

Where does z come from? A 95 percent two-sided interval leaves 2.5 percent of the normal curve in each tail, so you need the z that puts 97.5 percent of the area to its left. That value is 1.95996, which everyone rounds to 1.96. By the same logic 90 percent confidence uses 1.645 and 99 percent uses 2.576. Higher confidence demands a larger z, which widens the interval; that is the price of being more sure.

The whole expression after the ± sign, z × (sd / √n), is the margin of error. It is the half-width of the interval, the number you tack onto a point estimate to quote it honestly.

A worked example you can reproduce

Suppose you measured 50 items and found a sample mean of 100 with a standard deviation of 15. Build a 95 percent interval.

Standard error: sd / √n = 15 / √50 = 15 / 7.071 = 2.121.
Margin of error: z × standard error = 1.96 × 2.121 = 4.16.
Interval: 100 ± 4.16, which runs from 95.84 to 104.16.

So you report the mean as 100 with a margin of error of 4.16 at 95 percent confidence. Drop the numbers into the confidence interval calculator and you will see exactly [95.84, 104.16], with the margin of error printed on its own line. The calculator uses the exact inverse-normal z (1.95996) rather than a rounded 1.96, so if you computed by hand with 1.96 you may see a difference in the last decimal. That is the calculator being more precise, not a mistake on your part.

I keep that example in my head as a reference point. The first time I had to defend a number in a review, I presented a flat average and got asked, reasonably, "how confident are you in that?" I had no answer. Now I reach for the interval first and the point estimate second, because the range is what makes the figure defensible. A single number invites the question; the interval answers it before anyone asks.

Proportions work the same way, with a different standard error

When your outcome is a yes/no rather than a measurement, you are estimating a proportion, and the formula swaps in a different standard error:

CI = p ± z × √(p(1−p) / n)

Here p is the sample fraction. Poll 100 people, 45 say yes, and p = 0.45. The margin is 1.96 × √(0.45 × 0.55 / 100) = 1.96 × 0.0497 = 0.0975, about 9.8 percentage points. The interval is roughly [35.2 percent, 54.8 percent]. That is why a poll is reported as "45 percent, ±10 points," and why a result inside the margin is effectively a tie.

Two quirks are worth knowing. First, the spread depends on p itself and is widest near 0.5, so a 50/50 split carries the largest margin for a given sample size. Second, a proportion cannot go below 0 or above 1, so the bounds get clamped to that range. One common slip: enter the proportion as a decimal (0.45), not a percentage (45). Typing 45 falls outside the valid 0-to-1 range, and the tool will reject it. Convert to a fraction first, then read the bounds back as percentages.

How sample size narrows the interval

Look again at both formulas. The margin always divides by √n. That square root is the whole story of how data buys precision, and it is less generous than people expect.

Because of the √n in the denominator, quadrupling the sample size only halves the margin. With sd = 15 at 95 percent confidence: n = 30 gives a margin of about 5.37, n = 120 cuts it to about 2.68, and n = 300 brings it to about 1.70. Each step toward a tighter interval costs disproportionately more data. To halve your error you need four times the sample. To cut it to a third you need nine times. This is why precise estimates are expensive and why "just collect more data" has sharply diminishing returns past a point.

That relationship is also a planning tool. If you need a mean estimate within ±2 units at 95 percent confidence, you can try a sample size, read the margin, and bump n upward until the margin drops under 2, all before you collect a single data point. The trial-and-error converges fast precisely because of the √n behavior.

One last caution on the mean: the z-based interval assumes the population standard deviation is known or the sample is large, with n ≥ 30 as the usual rule of thumb. For a small sample where you only have the sample standard deviation, reach for the t-distribution instead. Its fatter tails produce a wider, more cautious interval that is honest about the extra uncertainty a small n carries. If you are also working with standardized scores, the z-score calculator handles the companion question of how far a single value sits from the mean in standard-deviation units.

Read intervals this way and a point estimate never looks finished again. The number tells you the best guess; the interval tells you how much to trust it.

Made by Toolora · Updated 2026-06-13