Is the correction always necessary?

No. Exact discrete calculations do not need this approximation adjustment, and some methods use different approximations.

Does it always improve accuracy?

No universal guarantee applies; its usefulness depends on the distribution and event.

Does a normal distribution need a continuity correction for its own probabilities?

No. The adjustment arises when representing a discrete distribution through a continuous approximation.

AI Glossary / Statistics and probability / Intermediate

Continuity correction

Definition

A continuity correction adjusts a boundary when a continuous distribution approximates a discrete one. For integer-valued counts, the usual normal approximation moves the relevant boundary by half a unit to represent the full probability mass associated with the included integers. It can improve an approximation, but it does not make the underlying continuous model exact or repair unsuitable distributional assumptions.

What it means

A count jumps from one integer to the next, whereas a normal curve spreads probability smoothly across a line. When calculating the probability of twelve or fewer successes, stopping the curve at twelve cuts through the area representing twelve. Moving the boundary to twelve and a half includes that entire notional count interval.
The correction concerns how an event is translated between two mathematical representations. It does not change the observed number of successes, add half a person to a dataset, or modify the actual event being studied.

How it works

First express the discrete event precisely. For an integer-valued X approximated by a continuous Y, P(X <= k) is commonly approximated by P(Y < k + 0.5). Conversely, P(X >= k) uses P(Y > k - 0.5). For a bounded event a <= X <= b, use a - 0.5 and b + 0.5. A single value X = k corresponds to the interval from k - 0.5 to k + 0.5.
Choose the continuous approximating distribution and calculate its parameters. For a binomial count with n independent trials and constant success probability p, the normal approximation has mean np and variance np(1-p). Apply the boundary correction before standardizing by that mean and standard deviation.
Check whether the approximation is appropriate for the parameters and event. Highly skewed distributions, extreme tails, and limited counts can produce poor results despite a correction. Direct computation of the discrete probability is preferable when practical and when accuracy matters.
The half-unit adjustment assumes unit-spaced values on the count scale. Different spacing requires corresponding care. Yates's correction in a two-by-two chi-square analysis is a related application of continuity correction, with its own formula and limitations; it should not be confused with a universal instruction to add 0.5 everywhere.

Three practical examples

An operations analyst approximates the probability that no more than a specified number of orders need rework.
A teaching tool compares exact binomial probabilities with normal approximations to show why boundary placement matters.
An evaluation team reviews whether a reported chi-square result used a continuity correction before comparing it with another software output.

Think of it this way

Think of counting whole paving stones and then measuring their combined length with a ruler. If each stone is centered on an integer mark, the outer edge lies halfway to the next mark. Measuring only to the center of the last included stone leaves out half of its represented width.

Worked example

Suppose X counts successes in twenty independent trials, each with success probability 0.5. Find P(X <= 12). The mean is 20 x 0.5 = 10, and the variance is 20 x 0.5 x 0.5 = 5, giving standard deviation sqrt(5), approximately 2.236.
With a continuity correction, the normal boundary is 12.5. Its standardized value is (12.5 - 10)/sqrt(5), approximately 1.118. The standard normal cumulative probability is approximately 0.868224.
Without the correction, the boundary twelve gives a standardized value approximately 0.894 and probability 0.814453. Directly summing the binomial probabilities from zero through twelve gives 0.868412. Here, the corrected approximation is much closer: its absolute error is about 0.000188, compared with 0.053959 without correction.
These calculations concern a deliberately simple probability model. The close agreement in this example does not establish that every corrected approximation will be accurate, particularly in an extreme tail or with a success probability near zero or one.

Illustrative case study

Fictional case: Harbor Desk builds an educational dashboard for quality supervisors. Its first version estimates count probabilities with a normal curve ending exactly at the requested integer. A reviewer notices systematic discrepancies when comparing several moderate-size examples with exact binomial results.
The team adds event-aware continuity correction and displays the exact calculation alongside the approximation. It also labels the independent-trial and constant-probability assumptions. When batches contain related items from the same machine setting, the team investigates the dependence instead of assuming a boundary adjustment solves the modeling problem. The result is a clearer teaching dashboard, not a guarantee of real-world defect probabilities.

Common misconceptions and mistakes

A common mistake is always adding 0.5, even for a lower boundary that should move downward. Another is confusing fewer than twelve with twelve or fewer: the first event ends at eleven and uses 11.5 as its upper corrected boundary. Applying the adjustment after standardization also uses the wrong scale unless it is transformed appropriately. Finally, a corrected approximation is still an approximation; a large sample rule of thumb cannot guarantee accuracy for every tail probability.

Business use

Businesses encounter this concept when reviewing approximate probabilities, quality analyses, and statistical software settings. Document the discrete event, model, approximation, and correction so another analyst can reproduce the result. Where a decision turns on a small probability difference, compare against an exact calculation or another justified numerical method rather than relying on an unlabeled normal approximation.

Check your understanding

For an integer count, what normal boundary approximates P(X >= 8)?

Answer: Use 7.5 as the lower boundary.

What interval represents P(X = 8)?

Answer: From 7.5 to 8.5.

Can a correction remove dependence between trials?

Answer: No; it changes approximation boundaries, not the data-generating assumptions.

Related terms

Learn next

Compare corrected and uncorrected normal approximations against exact binomial probabilities for several success probabilities. Then examine how the error changes near a distribution's center versus its tails, keeping the event definition identical across methods.

Sources

Philip B. Stark, SticiGui, The Normal Curve, the Central Limit Theorem, and Markov's and Chebychev's Inequalities for Random Variables, continuity correction section: https://www.stat.berkeley.edu/~stark/SticiGui/Text/clt.htm SciPy documentation, chi2_contingency, correction parameter: https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.contingency.chi2_contingency.html

Previous: Continuer Next: Continuity in probability