Continuity Correction for Discrete Counts

Why a Continuous Curve Needs an Adjustment

A binomial random variable can take only whole-number values: 0, 1, 2, and so on. A normal random variable, by contrast, can take any value on a continuous scale. When a normal curve is used to approximate probabilities for a binomial count, the boundary between included and excluded integer counts matters. The continuity correction accounts for this difference by moving each count boundary half a unit.

This adjustment is useful because the normal curve represents probability as area over intervals, while a discrete count assigns probability to individual integers. A count such as 35 is represented approximately by the area in a narrow interval centered at 35, rather than by the area at the single point 35 (which is zero under a continuous distribution).

Definition: Suppose \(X\sim\operatorname{Bin}(n,p)\) is a binomial count, and let \(Y\) be its normal approximation, \(Y\sim N(np,np(1-p))\). The continuity correction moves an integer boundary by \(0.5\) so the normal area includes the appropriate whole-number count or counts. The binomial count \(X\) remains discrete; \(Y\) is a continuous approximation to its distribution.

Here \(n\) is the fixed number of trials and \(p\) is the probability of success on each trial. The binomial model assumes independent trials with the same success probability \(p\). Its mean is \(np\), and its variance is \(np(1-p)\), so the normal approximation has standard deviation \(\sqrt{np(1-p)}\). The approximation is useful only when the binomial distribution is adequately represented by a normal curve; shifting boundaries cannot by itself make a poor approximation reliable.

How the Half-Unit Adjustment Arises

Imagine placing a unit-wide interval around each integer count. The interval centered at 10 runs from 9.5 to 10.5; the one centered at 11 runs from 10.5 to 11.5. These intervals meet without gaps. Thus, the probability assigned by a continuous curve to the interval from 9.5 to 10.5 can stand in for the probability that the count equals 10. Adding adjacent intervals approximates the probability of a range of integer counts.

This gives the correction its direction: to include an endpoint count, extend the continuous interval halfway beyond that integer. To exclude that count, stop halfway before it. Since a continuous random variable has probability zero at any single point, it does not matter whether the corrected normal interval uses a strict or inclusive inequality at its endpoint.

Translate the count event.
Identify the smallest and largest integer counts included by the event, paying attention to whether the inequality is strict or inclusive.

Adjust the boundaries.
Move the lower boundary down by 0.5 and the upper boundary up by 0.5 so the normal area covers the included integer counts.

Calculate the normal area.
Use the normal approximation with mean \(np\) and standard deviation \(\sqrt{np(1-p)}\), standardize the corrected boundary or boundaries, and evaluate the probability.

The main translations are summarized below. In each row, \(k\), \(a\), and \(b\) are integers, and \(Y\) has the normal approximation to \(X\).

Binomial count event	Approximate normal event
\(X\le k\)	\(Y\le k+0.5\)
\(X<k\)	\(Y\le k-0.5\)
\(X\ge k\)	\(Y\ge k-0.5\)
\(X>k\)	\(Y\ge k+0.5\)
\(a\le X\le b\)	\(a-0.5\le Y\le b+0.5\)

For example, \(X<k\) includes counts only up to \(k-1\). The upper edge of the interval representing \(k-1\) is \((k-1)+0.5=k-0.5\), which explains that row. Similarly, \(X\ge k\) begins with the interval representing \(k\), whose lower edge is \(k-0.5\).

Worked Examples

Worked Example: A Lower-Tail Probability

Suppose a hypothetical process produces 100 independent items, each with probability 0.40 of meeting a specified criterion. Let \(X\) be the number that meet it. Approximate the probability that at most 35 items meet the criterion.

The count has the binomial model \(X\sim\operatorname{Bin}(100,0.40)\). Its normal approximation has mean \(\mu=np=100(0.40)=40\) items and standard deviation \(\sigma=\sqrt{np(1-p)}=\sqrt{100(0.40)(0.60)}=\sqrt{24}\approx4.899\) items. We use this approximation on the understanding that it is suitable for this binomial distribution.

The event is \(X\le35\), so the corrected normal boundary is \(35.5\), not 35. Therefore,

P(X\le35)\approx P(Y\le35.5) =\Phi\!\left(\frac{35.5-40}{\sqrt{24}}\right) =\Phi(-0.9186) \approx0.1792.

The approximate probability that at most 35 of the 100 items meet the criterion is 0.1792, or about 17.92%. The correction includes the area representing count 35 as well as the areas for all smaller counts.

Worked Example: A Probability Between Two Counts

In a hypothetical set of 80 independent trials, each trial succeeds with probability 0.25. Let \(X\) count the successes. Approximate the probability of getting from 18 through 22 successes, including both endpoints.

The model is \(X\sim\operatorname{Bin}(80,0.25)\). The normal approximation has mean \(np=80(0.25)=20\) successes and standard deviation \(\sqrt{np(1-p)}=\sqrt{80(0.25)(0.75)}=\sqrt{15}\approx3.873\) successes. We assume the normal approximation is adequate for this example.

The integer values included are 18, 19, 20, 21, and 22. Extending half a unit beyond the endpoints gives the corrected interval from 17.5 to 22.5. Thus,

\begin{aligned} P(18\le X\le22) &\approx P(17.5\le Y\le22.5)\\ &=\Phi\!\left(\frac{22.5-20}{\sqrt{15}}\right) -\Phi\!\left(\frac{17.5-20}{\sqrt{15}}\right)\\ &=\Phi(0.6455)-\Phi(-0.6455)\\ &\approx0.7407-0.2593\\ &\approx0.4814. \end{aligned}

The probability of obtaining between 18 and 22 successes, inclusive, is approximately 0.4814, or 48.14%. Both boundaries move outward because both endpoint counts are included.

Worked Example: An Upper-Tail Probability

Suppose a hypothetical machine makes 200 independent parts, each with probability 0.30 of having a particular feature. Let \(X\) be the number of parts with that feature. Approximate the probability that at least 70 parts have it.

Here \(X\sim\operatorname{Bin}(200,0.30)\). The normal approximation has mean \(np=200(0.30)=60\) parts and standard deviation \(\sqrt{np(1-p)}=\sqrt{200(0.30)(0.70)}=\sqrt{42}\approx6.481\) parts. We use the normal model as an approximation to this count distribution.

“At least 70” means 70 or more, so the count of 70 is included. The normal area must begin halfway below 70, at 69.5:

\begin{aligned} P(X\ge70) &\approx P(Y\ge69.5)\\ &=1-\Phi\!\left(\frac{69.5-60}{\sqrt{42}}\right)\\ &=1-\Phi(1.4659)\\ &\approx1-0.9286\\ &\approx0.0714. \end{aligned}

The approximate probability of at least 70 parts having the feature is 0.0714, or 7.14%. Using 70 as the normal boundary would omit roughly half of the continuous area corresponding to count 70.

Why the Rule Works—and What It Does Not Do

The half-unit shift is justified by how the normal approximation represents individual count probabilities. For an integer \(j\), the unit-width interval centered at \(j\) has endpoints \(j-0.5\) and \(j+0.5\). If the normal density changes gradually across that interval, its area there is a reasonable approximation to the binomial probability at \(j\). For consecutive counts, these intervals join together.

\begin{aligned} P(a\le X\le b) &=\sum_{j=a}^{b}P(X=j)\\ &\approx\sum_{j=a}^{b}P(j-0.5\le Y\le j+0.5)\\ &=P(a-0.5\le Y\le b+0.5). \end{aligned}

The last equality follows because the intervals for consecutive integers meet at their endpoints, and a continuous normal variable assigns probability zero to those shared endpoints. This argument explains the correction; it does not make the approximation exact. The binomial probabilities and the normal areas can still differ, especially when the normal curve does not closely resemble the count distribution.

The correction also does not change the binomial model, the values of \(n\) and \(p\), or the normal approximation’s mean and standard deviation. It adjusts only the area boundaries. Exact binomial probabilities remain a separate way to calculate count probabilities when needed.

Common Mistakes

Using the integer itself as the boundary. For \(P(X\le35)\), using \(P(Y\le35)\) misses part of the area representing count 35. Use 35.5 instead.
Moving every boundary in the same direction. For a lower-tail event, the upper boundary moves up by 0.5. For an upper-tail event, the lower boundary moves down by 0.5. For an interval, the lower endpoint moves down and the upper endpoint moves up.
Ignoring strict inequalities. \(X>70\) starts at count 71, so its corrected boundary is 70.5. In contrast, \(X\ge70\) includes count 70 and uses 69.5.
Assuming the correction guarantees accuracy. It improves the match between discrete count ranges and continuous areas, but it cannot ensure that a normal curve is a good approximation to a particular binomial distribution.

Key takeaway: First identify exactly which integer counts an event includes. Then extend the normal interval by half a unit to cover those counts. Standardize the corrected boundary or boundaries and calculate the resulting normal area.

Check Your Understanding

For each question, decide which integer counts are included before choosing corrected boundaries.

What corrected normal event approximates \(P(X\le k)\), where \(k\) is an integer?
For \(P(X<24)\), what is the largest included count, and what is the corresponding corrected upper boundary?
Write the corrected normal interval for \(P(12\le X\le19)\).
Why does an event \(X\ge40\) use a lower normal boundary of 39.5 rather than 40.5?
In your own words, explain why the continuity correction cannot make an unsuitable normal approximation reliable.