Continuity correction is a small adjustment (adding or subtracting 0.5) made when you use a smooth, continuous distribution like the normal curve to estimate probabilities for data that can only take whole numbers. It bridges the gap between discrete counts and the continuous curves used to approximate them, producing more accurate probability estimates.
Why Whole Numbers Need an Adjustment
Many real-world outcomes are counted in whole numbers: coin flips, survey responses, defective items in a batch. The binomial distribution describes these perfectly, but calculating exact binomial probabilities by hand gets tedious as sample sizes grow. Statisticians long ago discovered that the bell-shaped normal distribution provides a good shortcut, letting you swap a messy sum for a quick calculation. The problem is that the normal distribution is continuous. It assigns probabilities to every value along a smooth curve, while the actual data can only land on integers like 10, 11, or 12.
Picture a bar chart of binomial probabilities. Each integer value sits at the center of a bar that stretches half a unit in each direction. The probability of getting exactly 10 heads, for instance, is represented by a bar running from 9.5 to 10.5. When you lay a smooth normal curve over that bar chart, matching the probability of “exactly 10” means capturing the area under the curve from 9.5 to 10.5, not just the single point at 10 (which would have zero area under a continuous curve). That half-unit expansion on each side is the continuity correction.
How the 0.5 Adjustment Works
The direction you shift depends on the inequality in your probability statement. The core idea is to widen the range by 0.5 so the continuous curve covers the same integers as the original discrete question:
- P(X = a) becomes P(a − 0.5 < X < a + 0.5)
- P(X ≥ a) becomes P(X ≥ a − 0.5)
- P(X ≤ a) becomes P(X ≤ a + 0.5)
- P(X > a) becomes P(X > a + 0.5)
- P(X < a) becomes P(X < a − 0.5)
The logic is consistent: you always nudge the boundary so that the integer value you want to include stays inside the range, or the integer value you want to exclude stays outside it.
A Coin-Flip Example
Suppose you flip a fair coin 50 times and get 29 heads. You want to know the probability of getting 29 or more heads under a fair coin. Using the normal approximation without any correction, you’d calculate a probability of about 0.129. The exact answer from the binomial distribution is 0.161. That’s a noticeable gap. The normal approximation underestimates the true probability by about 20%.
With continuity correction, you calculate P(X ≥ 28.5) instead of P(X ≥ 29), which shifts the boundary half a unit to the left and captures more area under the curve. This brings the estimate closer to the exact binomial answer. The difference matters most when sample sizes are moderate. With very large samples, the normal curve fits the discrete bars so tightly that the half-unit adjustment barely changes anything.
When You Can Use the Normal Approximation
Continuity correction only makes sense when the normal approximation is already reasonable. For a binomial distribution with n trials and probability of success p, the standard rule of thumb is that both np and n(1 − p) should be at least 10. If you’re flipping a fair coin, that means you need at least 20 flips. If the probability of success is 0.05, you’d need at least 200 trials before the normal curve fits well enough to approximate the binomial.
Below those thresholds, the binomial distribution is too lopsided for a symmetric bell curve to approximate it well, and no half-unit correction will fix that. You’d be better off calculating exact binomial probabilities directly, which modern software handles instantly regardless of sample size.
Yates’ Correction for Chi-Square Tests
The same principle shows up in a different context: the chi-square test for 2×2 tables. When you compare observed and expected counts in a contingency table, the test statistic follows a continuous chi-square distribution, but the observed counts are discrete. In 1934, Frank Yates proposed subtracting 0.5 from the absolute difference between each observed and expected value before squaring. This reduces the test statistic slightly, making it harder to reach statistical significance.
Yates’ correction is built into many software tools. In R, for instance, the chisq.test() function applies it by default for 2×2 tables through a parameter called correct, which is set to TRUE unless you change it. Python’s scipy library behaves similarly. If you’ve ever run a chi-square test on a 2×2 table and gotten a slightly different result than you expected from a hand calculation, the default continuity correction is likely the reason.
The Overcorrection Problem
Yates’ correction has been controversial for decades. Research has consistently shown that it tends to be overly conservative, meaning it makes you less likely to detect a real effect. The uncorrected Pearson chi-square statistic generally provides adequate control over false positive rates on its own, without the additional adjustment. Despite this, medical researchers in particular have continued to use the corrected version routinely, partly because software defaults to it and partly out of habit.
The practical consequence is that Yates’ correction can push a borderline result from statistically significant to not significant. If your 2×2 chi-square test just barely misses significance, it’s worth checking whether the correction was applied and whether the uncorrected version tells a different story. Many statisticians now recommend skipping Yates’ correction entirely for 2×2 tables, or using Fisher’s exact test when sample sizes are small enough to warrant concern about the chi-square approximation.
When Continuity Correction Matters Most
The correction has the biggest impact with small to moderate sample sizes, roughly in the range of 20 to 100 observations. In that zone, the discrete steps between integers are large relative to the spread of the distribution, so the half-unit adjustment represents a meaningful shift. As sample sizes grow into the hundreds or thousands, each integer step becomes tiny compared to the overall distribution width, and the correction makes almost no practical difference.
For the normal approximation to the binomial, continuity correction consistently brings estimates closer to the true binomial probability. For chi-square tests on 2×2 tables, the picture is murkier, since the correction can overcorrect and reduce your ability to detect real associations. The safest approach in modern practice is to use exact tests when sample sizes are small and skip the correction for chi-square tests on larger samples, where it’s unnecessary anyway.