Continuity Correction vs. Confidence Interval: What’s the Difference?

A continuity correction and a confidence interval are not competing versions of the same method. A continuity correction helps approximate the probability of a discrete random variable using a continuous distribution. A confidence interval uses sample data to estimate an unknown population parameter, such as a proportion or mean.

Both may involve the normal distribution, which can make them easy to confuse. The key question is whether you are calculating a probability for a count you can observe, or estimating a parameter you do not know.

What a continuity correction does

A discrete count takes integer values, while a normal random variable can take any real value. When a normal curve approximates a discrete distribution, a continuity correction shifts the cutoff by one-half to better match the integer values included in the event.

For an integer-valued variable $X$, an integer cutoff $k$, and a continuous approximation $Y$, common translations are:

Discrete event	Approximate continuous event
$X \le k$	$Y \le k+0.5$
$X < k$	$Y \le k-0.5$
$X \ge k$	$Y \ge k-0.5$
$X > k$	$Y \ge k+0.5$

The half-unit adjustment reflects the width of the interval represented by each integer. For example, the integer value 12 is represented approximately by the continuous range from 11.5 to 12.5.

A continuity correction is optional in the sense that you can calculate a normal approximation without it. But including it often improves the approximation, especially when the count is not very large. It does not make the normal approximation exact.

Worked example: a binomial probability

Suppose $X$ is the number of heads in 20 fair coin flips. Then $X$ has a binomial distribution with $n=20$ and $p=0.5$. Its mean and standard deviation are

$$\mu=np=10, \qquad \sigma=\sqrt{np(1-p)}=\sqrt{5}\approx 2.236.$$

To approximate $P(X\le 12)$ with a normal random variable $Y$ having mean 10 and standard deviation $\sqrt{5}$, use the cutoff 12.5:

$$P(X\le 12)\approx P(Y\le 12.5) =\Phi\left(\frac{12.5-10}{\sqrt{5}}\right) =\Phi(1.118)\approx 0.868.$$

Without the correction, the approximation would use 12 instead of 12.5 and give $\Phi((12-10)/\sqrt{5})\approx 0.814$. The exact binomial probability is about 0.868, so the corrected approximation is much closer in this example.

What a confidence interval does

A confidence interval is calculated from sample data to estimate an unknown parameter. For instance, if $p$ is the proportion of a population with a certain property, a confidence interval gives a range of plausible values for $p$ under a stated procedure.

As a simple illustration, suppose 12 of 20 sampled people have the property. The sample proportion is $\hat p=12/20=0.60$. The approximate 95% Wald confidence interval is

$$\hat p\pm 1.96\sqrt{\frac{\hat p(1-\hat p)}{n}} =0.60\pm 1.96\sqrt{\frac{0.60(0.40)}{20}} \approx (0.385,\,0.815).$$

This interval estimates the population proportion. It is not a correction to the probability calculation in the coin-flip example. The Wald interval is shown for illustration; for small samples or proportions near 0 or 1, methods such as the Wilson interval generally have better properties.

In frequentist terms, a 95% confidence level is the procedure’s nominal long-run coverage: under repeated sampling, the method is designed to produce intervals containing the fixed population parameter about 95% of the time, assuming the model and sampling conditions hold. Actual coverage can differ from 95%, especially for approximate methods such as the Wald interval. It does not mean there is a 95% probability that the already fixed parameter lies in this particular interval.

How they can appear together

A confidence interval may use a normal approximation to the sampling distribution of a statistic. That does not automatically mean a continuity correction is needed. For example, the Wald interval above uses a normal-based formula for a sample proportion, but it does not shift its endpoints by 0.5.

Continuity corrections are relevant when approximating probabilities for discrete values or events, such as binomial counts, with a continuous distribution. Some procedures for discrete data also account for discreteness when calculating tail probabilities or inverting tests to form intervals. The details depend on the interval method; there is no universal rule to add 0.5 to every confidence interval.

A useful distinction is:

• Probability question: “What is the chance that the count is at most 12?” Use the count’s distribution; if using a normal approximation, consider a continuity correction.
• Estimation question: “What range of population proportions is compatible with this sample?” Choose a confidence-interval method appropriate to the data and assumptions.

Common mistakes

• Using the correction for the wrong task. The half-unit shift belongs to a normal approximation for a discrete event, not to every calculation involving a normal distribution.
• Shifting in the wrong direction. For $X\le k$, use $k+0.5$; for $X\ge k$, use $k-0.5$. The direction follows which integer values the event includes.
• Calling a probability range a confidence interval. A range of count values containing, say, 95% of outcomes is a prediction or probability interval, not necessarily a confidence interval for a parameter.
• Assuming a continuity correction fixes a poor approximation. If the normal approximation is unsuitable, shifting the boundary cannot make it reliable. Consider an exact binomial calculation or an interval method designed for the sample size and data.