Using Continuity Correction With a Binomial Test

Why a Discrete Count Needs a Correction

A binomial test asks how unusual an observed number of successes would be if a specified null success probability were true. The count can take only whole-number values. A normal distribution, by contrast, is continuous: it assigns probability to intervals, and the probability of any single exact point is zero. If we approximate a binomial tail by simply placing a normal cutoff at the observed count, the two distributions do not line up at their boundaries. A continuity correction shifts the normal cutoff by half a count to better match the integer values included in the binomial event.

The correction is a refinement of the normal approximation, not a replacement for the binomial model. When the null probability and sample size are known, the exact binomial tail can be calculated directly. Comparing the two is useful for understanding what the correction changes and how closely the approximation matches the exact result.

The Binomial Model and Its Normal Approximation

Definition: Under a one-proportion null hypothesis \(H_0:p=p_0\), let \(X\) be the number of successes in \(n\) independent trials, each with success probability \(p_0\). Then \(X\sim\operatorname{Bin}(n,p_0)\), with mean \(\mu_X=np_0\) and standard deviation \(\sigma_X=\sqrt{np_0(1-p_0)}\). When the normal approximation is appropriate, approximate \(X\) by a normal random variable \(Y\sim N(np_0,np_0(1-p_0))\), whose second parameter is its variance.

For a one-sided test, the p-value is the probability, calculated under \(H_0\), of a count at least as extreme as the observed count in the direction specified by \(H_a\). If \(H_a:p>p_0\) and the observed count is \(x\), the relevant event is \(X\ge x\). If \(H_a:p<p_0\), it is \(X\le x\). The binomial events include the observed count; this inclusion determines which way the normal cutoff must move.

Before using a normal approximation, check the conditions for it. The trials should be independent, and the null expected numbers of successes and failures, \(np_0\) and \(n(1-p_0)\), should be sufficiently large under the course’s approximation check. Meeting a rule of thumb makes the approximation reasonable, not exact. The continuity correction addresses the discrete-versus-continuous boundary; it does not fix poor normal approximation conditions.

How to Place the Half-Count Boundary

A useful way to see the correction is to associate each integer count \(j\) with a continuous interval extending half a unit on either side: from \(j-0.5\) to \(j+0.5\). For example, the integer count 6 corresponds approximately to the interval from 5.5 to 6.5 on the normal scale. Consequently, the integer tail \(X\ge x\) is represented by the normal area to the right of \(x-0.5\); the integer tail \(X\le x\) is represented by the area to the left of \(x+0.5\).

Continuity-correction rule: For an upper-tail event \(X\ge x\), use \(Y>x-0.5\). For a lower-tail event \(X\le x\), use \(Y<x+0.5\). In both cases the strict or non-strict inequality gives the same normal probability, because a continuous normal variable has probability zero at any single boundary point.

Writing \(SE_0=\sqrt{np_0(1-p_0)}\), the corresponding corrected normal statistics are

z_{\text{upper, cc}}=\frac{x-0.5-np_0}{SE_0}, \qquad z_{\text{lower, cc}}=\frac{x+0.5-np_0}{SE_0}.

The corrected upper-tail approximation is the standard normal probability to the right of \(z_{\text{upper, cc}}\); the corrected lower-tail approximation is the probability to the left of \(z_{\text{lower, cc}}\). The uncorrected method instead uses \(x\) as the count-scale boundary in either direction. That can place the normal boundary half a count away from the edge of the binomial tail being approximated.

Worked Upper-Tail Example

Worked Example: At Least 60 Successes

Suppose a hypothetical test has \(n=100\) independent trials and tests \(H_0:p=0.50\) against \(H_a:p>0.50\). The observed count is \(x=60\). The null expected successes and failures are both 50, so the normal approximation check is satisfied. Find the corrected normal approximation and the exact one-sided p-value.

Under the null, the count has mean \(np_0=100(0.50)=50\) and standard deviation \[ SE_0=\sqrt{100(0.50)(0.50)}=\sqrt{25}=5. \] The upper-tail event is \(X\ge60\), so the corrected count-scale boundary is \(60-0.5=59.5\). Thus \[ z_{\text{upper, cc}}=\frac{59.5-50}{5}=1.900. \] The corrected approximation is \(P(Z\ge1.900)\approx0.02872\), where \(Z\sim N(0,1)\).

The exact binomial p-value uses the whole-number tail, including 60: \[ P(X\ge60)=\sum_{j=60}^{100}\binom{100}{j}(0.50)^j(0.50)^{100-j} \approx0.02844. \] The corrected approximation is close to the exact value. Without correction, the cutoff would be \(z=(60-50)/5=2.000\), giving an upper-tail probability of about 0.02275, which is noticeably smaller than the exact tail.

In this hypothetical test, the exact tail probability is about 0.02844, and the corrected approximation is about 0.02872. Both quantify how often a count of at least 60 would occur if the null model were true; neither is the probability that \(H_0\) is true.

Worked Lower-Tail Example

Worked Example: At Most 5 Successes

Suppose a hypothetical process has \(n=20\) independent trials. Test \(H_0:p=0.50\) against \(H_a:p<0.50\), and observe \(x=5\) successes. Under the null, the expected numbers of successes and failures are both 10, meeting the stated normal-approximation check. Calculate the lower-tail probability using both methods.

The null mean count is \(20(0.50)=10\), and the null standard deviation is \[ SE_0=\sqrt{20(0.50)(0.50)}=\sqrt{5}\approx2.236. \] The event is \(X\le5\). For a lower tail, add one half to the observed count, so the normal boundary is \(5.5\). Then \[ z_{\text{lower, cc}}=\frac{5.5-10}{\sqrt{5}}\approx-2.012. \] The corrected normal approximation is \(P(Z\le-2.012)\approx0.0221\).

The exact probability sums the binomial probabilities for counts from 0 through 5. Because \(p_0=0.50\), each count has a factor of \((0.50)^{20}=1/1{,}048{,}576\): \[ P(X\le5)=\frac{\binom{20}{0}+\binom{20}{1}+\binom{20}{2}+\binom{20}{3}+\binom{20}{4}+\binom{20}{5}}{2^{20}} =\frac{21{,}700}{1{,}048{,}576}\approx0.02069. \] The corrected approximation, about 0.0221, is close to the exact value. The uncorrected boundary would be \(z=(5-10)/\sqrt{5}\approx-2.236\), yielding a lower-tail probability of about 0.0127. In this case, omitting the correction understates the exact tail substantially.

Worked Example: An Upper Tail With the Same Null Model

Worked Example: At Least 14 Successes

Keep \(n=20\) and \(H_0:p=0.50\), but suppose a separate hypothetical test is upper-tailed, with \(H_a:p>0.50\), and observes \(x=14\). The expected successes and failures remain 10 each. Find the corrected and exact upper-tail probabilities.

The null mean is 10 and the standard deviation is \(\sqrt{5}\approx2.236\). The event is \(X\ge14\), so the corrected boundary is \(13.5\): \[ z_{\text{upper, cc}}=\frac{13.5-10}{\sqrt{5}}\approx1.565. \] The corrected normal approximation is \(P(Z\ge1.565)\approx0.0587\).

The exact tail is \[ P(X\ge14)=\sum_{j=14}^{20}\binom{20}{j}(0.50)^{20}. \] By symmetry of a \(\operatorname{Bin}(20,0.50)\) distribution, this equals \(P(X\le6)\). Therefore, \[ P(X\ge14)=\frac{1+20+190+1{,}140+4{,}845+15{,}504+38{,}760}{1{,}048{,}576} =\frac{60{,}460}{1{,}048{,}576}\approx0.05766. \] The corrected approximation is again close to the exact probability. The uncorrected boundary gives \(z=(14-10)/\sqrt{5}\approx1.789\), whose upper tail is about 0.0368. That value is much smaller than the exact probability because it omits part of the continuous area corresponding to the count 14.

Why the Correction Works—and What It Does Not Do

The half-count rule follows from matching whole-number outcomes to regions on a continuous scale. For the upper-tail event, the integer outcomes \(x,x+1,x+2,\ldots\) are represented by adjoining intervals whose combined region begins at \(x-0.5\). For the lower-tail event, the outcomes \(\ldots,x-2,x-1,x\) are represented by intervals whose combined region ends at \(x+0.5\). The correction therefore aligns the normal area with the boundary of the binomial event, rather than with its central integer value.

This is a better boundary match, but the normal curve still only approximates the binomial distribution. A corrected probability can differ from the exact result, especially when the null expected successes or failures are small or when the distribution is strongly skewed. When feasible, the exact binomial tail is the direct calculation under the specified binomial model. The corrected approximation is useful when a normal calculation is being used and a more faithful tail boundary is needed.

For a two-sided test, do not automatically apply a one-sided correction and double the result. The exact two-sided p-value depends on the definition of “at least as extreme” used for that test, and the approximation must reflect that same definition. The rules here apply directly to a specified upper- or lower-tail event.

Common Pitfalls

Shifting in the wrong direction. Using \(x+0.5\) for an upper tail moves the boundary farther into the tail. For \(X\ge x\), use \(x-0.5\); for \(X\le x\), use \(x+0.5\).
Correcting the observed proportion instead of the boundary. The correction changes the count-scale cutoff used for the normal approximation. It does not alter the observed count, the null proportion, or the data.
Treating a corrected result as exact. The half-count adjustment improves the boundary match; it does not eliminate approximation error. Compare with the exact binomial tail when it is available.
Ignoring the test direction. The tail is determined by \(H_a\), not by which result looks more surprising after observing the data. Establish the null model and the appropriate tail before calculating the p-value.

Key takeaway: For a binomial upper tail \(P(X\ge x)\), place the normal boundary at \(x-0.5\). For a lower tail \(P(X\le x)\), place it at \(x+0.5\). Check the normal-approximation conditions, then interpret the corrected area as an approximation to the binomial tail—not as an exact probability.

Check Your Understanding

For each question, identify the relevant tail before deciding where the corrected normal boundary belongs.

A test uses \(H_a:p>p_0\) and observes \(x=42\) successes. What count-scale boundary should be used for the corrected normal approximation?
A test uses \(H_a:p<p_0\) and observes \(x=18\) successes. What count-scale boundary should be used?
For \(n=40\) and \(p_0=0.25\), calculate the null expected numbers of successes and failures. What approximation condition should be checked before using a normal tail?
Explain why using the uncorrected boundary \(x\) may give a different tail probability from the continuity-corrected boundary.
Does a corrected normal p-value equal the exact binomial p-value by definition? Explain the distinction.