Continuity Correction Calculator — Binomial

The binomial lives on integers; the normal curve lives everywhere. Bridging them costs half a unit — the continuity correction — and this calculator shows exactly what that half buys. Enter n, p, the count k and the event you want, and it returns the normal approximation with and without the correction, beside the z-scores each one used, so the size of the correction’s effect is never a matter of faith.

Half a Unit, Honestly Spent

Picture the binomial as a histogram: a bar of width 1 centred on each integer, its area the exact probability. The normal curve approximating it must swallow whole bars to price an event like X ≤ 10 — the bar at 10 spans 9.5 to 10.5 — so the curve should be cut at 10.5, not at 10. That half-unit shift, applied toward whichever integers the event includes, is the entire correction, and its direction follows one rule: expand the event’s territory by half a unit at every boundary. “At most k” reaches up to k + 0.5; “less than k” stops at k − 0.5; “exactly k” claims the full strip from k − 0.5 to k + 0.5 — the only event with two boundaries, and the one where skipping the correction is not approximate but meaningless, as the calculator’s exact-k mode demonstrates by printing no uncorrected figure at all.

Worked Example: Ten or Fewer Heads in Twenty Tosses

The defaults are the classroom’s home turf: n = 20 fair tosses, P(X ≤ 10). The binomial’s mean is 10 and its SD √5 ≈ 2.236. Without the correction the boundary sits exactly at the mean and the answer is a suspiciously clean 0.5000 — the curve knows nothing of the bar at 10. With it, z = (10.5 − 10) ÷ 2.236 ≈ 0.224 and the answer is 0.5885. The exact binomial probability is 0.5881. The correction has carried the approximation from an error of 0.088 to an error of 0.0004 — a two-hundred-fold improvement for half a unit of honesty, which is why the technique survives in the curriculum long after software made exact binomials trivial: it teaches what an approximation is doing, geometrically, in a way a black-box exact answer never will.

When the Approximation Itself Deserves Trust

The correction refines the bridge; it does not extend it. The normal approximation is respectable when the binomial is plump in both tails — the working rule np ≥ 10 and n(1−p) ≥ 10, printed with your numbers in the calculator’s note — and it degrades exactly where p crouches near 0 or 1 and the distribution goes lopsided, leaning on one wall of its cage. In those territories the Poisson (for rare events) or the exact binomial is the right instrument, and no half-unit of good intention rescues a curve fitted to the wrong shape. Yates built the same correction into his chi-square adjustment for 2×2 tables, where the debate about its conservatism still runs; in the present, purely computational setting its value is unambiguous, and the calculator’s side-by-side printing is the whole argument in one line.

A Second Example, Off-Centre

The fair-coin example flatters the correction by symmetry, so run a lopsided one through the same machinery: a quality process with defect probability p = 0.1, n = 50 items, and the event X ≤ 8 defectives. The mean is 5, the SD √(50 × 0.1 × 0.9) ≈ 2.121. Uncorrected, the boundary 8 sits at z = 1.414, giving about 0.9213 — an error of about 0.021 against the exact binomial value of 0.9421. Corrected, the boundary moves to 8.5, z = 1.650, giving about 0.9505 — the error cut to roughly 0.008, nearly threefold better, with the skew of p = 0.1 (the approximation’s harder direction) responsible for the residue that remains. The pattern generalises beyond the numbers: the uncorrected error concentrates exactly at the boundary bar the event cares most about, and the half-unit repair is worth most where the distribution’s mass is thickest relative to its spread. Small np products are where the correction earns its name; it is no accident the classroom examples all live there.

The Correction in the Wild: Polls and Quality Lines

Two working habitats keep the half-unit employed outside classrooms. Election polling reports proportions from samples, and the margin-of-error arithmetic is this page’s machinery run on counts: the normal curve fitted to a binomial of decided voters, with careful shops applying the correction when translating a candidate’s count lead into a probability of leading at all — at n in the hundreds its effect is small but real, and at the small subgroup sizes where headlines breed (“voters under 30 in one state”) it is the difference between a defensible and an embarrassing probability. Acceptance sampling in manufacturing is the second habitat: lot-acceptance probabilities are binomial tail events, historically read off normal approximations with the correction applied, and the operating-characteristic curves in older quality manuals are this calculator’s outputs plotted by the hundred. Software has retired the necessity in both fields; literacy has not retired the understanding.

The de Moivre Lineage

The approximation this page refines is historically the first central limit theorem: Abraham de Moivre derived the normal curve in 1738 precisely as the limit of the binomial for fair coins, computing by hand (with Stirling’s help) the probabilities this calculator now renders in a blink. Laplace generalised it, and the continuity correction accreted over the following centuries as practitioners noticed the systematic half-bar error at the boundaries. There is a pleasing honesty in the lineage: the correction is not a modern patch on an old formula but the oldest formula in probability being read at the resolution its inventor’s own diagrams implied — bars first, curve second, and the knife falling where the bars say it should.

Related Tools

More from the Statistics Calculators hub: the Normal Approximation Calculator for the uncorrected parent method, and the False Positive Calculator for binomial thinking in the diagnostic direction.

Frequently Asked Questions

Why exactly 0.5?

Because each integer’s histogram bar is exactly one unit wide and centred on the integer, so its edges lie a half unit to either side. The correction is not a fudge factor fitted by experiment; it is the bar’s literal geometry — the curve is being asked to cover bars, so its knife must fall on bar edges. Any other constant would be an approximation of the approximation, which is a game with no floor.

Does the correction matter for large n?

Proportionally less each year of n: the half unit is measured against an SD growing like √n, so at n = 20 it moves z by 0.22 SD while at n = 2,000 it moves it by 0.022 SD — a rounding decision. This is why working statisticians computing by software omit it happily at scale, and why examiners insist on it at classroom scale: the correction’s importance is inversely proportional to how much the user needs a calculator at all, a small irony the page happily preserves.

What about “between” events, like 8 ≤ X ≤ 12?

Two boundaries, two corrections, both outward: the curve runs from 7.5 to 12.5 — the lower edge retreats by half a unit, the upper advances by half a unit, and the event swallows the full bars of 8 and 12 alike. The calculator prices single-boundary events and the exact case; a between probability is the difference of two of its “at most” answers at the corrected boundaries (12.5’s answer minus 7.5’s), a composition the worked example’s logic makes routine.

Is this the same as Yates’ correction in chi-square tests?

The same geometry wearing a different uniform: Yates adjusted the 2×2 chi-square by shrinking each |observed − expected| by 0.5 before squaring, for the identical reason — a discrete count distribution being read off a continuous curve. There the correction has critics (it can overcorrect, costing power), and modern practice often prefers exact tests at small counts. Here, in the direct normal approximation, the correction is uncontroversially an improvement, as the worked example’s error figures testify.

Why not always use the exact binomial?

Today, in software, you should — the exact answer is a function call away and this page says so. The approximation endures for three honest reasons: it is computable by hand and table (the historical classroom constraint), it reveals the binomial’s normal skeleton (the central limit theorem made visible, which the exact sum conceals), and in genuinely large-n work — thousands of trials, simulation pipelines — the normal form is the practical instrument. Learn the correction to understand the bridge; cross by the exact bridge when one is standing.

What if p is very small, like 0.01?

Then the rule in the note fires its warning: at n = 100 and p = 0.01, np = 1 and the binomial is a skewed spike the normal curve cannot impersonate, correction or no. The Poisson distribution with λ = np is the classical approximation there, and exact computation the modern one. The calculator will still print its figures — arithmetic never refuses — but the note’s np figures are its conscience, and at values like those they instruct the user to walk to a different instrument.