Normal Approximation to the Binomial Calculator
Run a calculation in the At Most / At Least, Exactly X, or Between tab first, then return here for the full written solution.
No data yet — enter values in one of the other tabs first.
What Is Normal Approximation?
Normal approximation is a method for estimating probabilities from a discrete distribution, usually the binomial, by using the continuous normal curve instead. It works because once you run enough trials, the lumpy, stair-step shape of a binomial distribution starts to look almost identical to a bell curve. Instead of adding up dozens or hundreds of individual binomial terms by hand, you calculate a mean and standard deviation, then read the probability straight off the normal curve.
This isn't a shortcut invented for convenience alone — it's a direct consequence of the Central Limit Theorem. A binomial random variable is just the sum of a bunch of independent yes/no trials, and sums of independent random variables drift toward a normal shape as the count grows. The NIST Engineering Statistics Handbook covers this convergence as one of the standard justifications for using normal-based methods on large samples, even when the original data isn't normally distributed at all.
Before calculators and statistical software were common, this was the only practical way to handle binomial problems with large n — computing the exact formula by hand for something like n = 500 would have meant working through factorials most people couldn't even fit on a chalkboard. The approximation is less essential now that software can compute exact binomial probabilities instantly, but it's still taught everywhere because it ties together several core ideas in one place: the binomial distribution, the normal distribution, standard error, and the Central Limit Theorem.
The Normal Approximation Formula
To approximate a binomial probability with the normal distribution, set the mean to μ = np and the standard deviation to σ = √(np(1−p)), then convert your target value into a z-score after applying a continuity correction of ±0.5.
Mean of the Approximating Normal
μ = np
Where:
n = number of trials
p = probability of success per trial
Standard Deviation
σ = √(np(1−p))
This is the same variance formula
used for the binomial distribution
itself — just square-rooted.
Z-Score With Continuity Correction
Z = (x ± 0.5 − μ) / σ
The sign of the 0.5 depends on
which direction you're rounding
the discrete boundary.
Validity Conditions
np ≥ 10
n(1−p) ≥ 10
(Some intro textbooks use 5
instead of 10 — check which
your course follows.)
In plain terms: you're swapping a binomial variable X for a normal variable Y that has the same center and spread, then asking the normal curve to answer a question the binomial formula would otherwise force you to compute term by term. Penn State's STAT 414 course notes on the normal approximation walk through this derivation from the Central Limit Theorem if you want the full proof rather than just the working formula.
When Can You Use Normal Approximation?
You can use the normal approximation once both np and n(1−p) are at least 10. Both conditions have to hold together — a lopsided p value keeps the binomial distribution skewed no matter how large n gets, unless n is large enough to push both products past the threshold.
Table: How n and p Combine to Satisfy (or Fail) the Validity Conditions
| n | p | np | n(1−p) | Approximation Valid? |
|---|---|---|---|---|
| 50 | 0.50 | 25.0 | 25.0 | Yes — comfortably |
| 100 | 0.10 | 10.0 | 90.0 | Yes — right at the edge |
| 100 | 0.05 | 5.0 | 95.0 | No — np too small |
| 500 | 0.02 | 10.0 | 490.0 | Yes — large n compensates |
| 20 | 0.50 | 10.0 | 10.0 | Yes — small n, balanced p |
| 30 | 0.95 | 28.5 | 1.5 | No — n(1−p) too small |
Notice the pattern: a probability close to 0.5 needs a relatively small n to satisfy both conditions, while a probability close to 0 or 1 needs a much larger n. That's why rare-event problems — defect rates of 1%, disease incidence of 0.3%, and similar low-p scenarios — usually call for a Poisson approximation instead, even at sample sizes in the thousands.
Continuity Correction: Why You Need the ±0.5
Continuity correction adjusts a discrete boundary by 0.5 before converting it to a z-score, because the binomial distribution only takes whole-number values while the normal curve is continuous. Without it, you're lining up a single point on the normal curve with a whole bar of the binomial histogram, and that bar has width — it isn't a point at all.
Picture a binomial histogram where each bar is centered on an integer and is one unit wide. The bar for X = 25 actually spans from 24.5 to 25.5. If you want P(X = 25) and you just plug 25 straight into the normal curve, you're asking for the probability of an exact point on a continuous curve — which is technically zero. The fix is to treat the bar's full width as the target: integrate the normal curve from 24.5 to 25.5 instead.
How to Calculate a Normal Approximation — Step by Step
Check that np and n(1−p) are both at least 10, calculate the mean and standard deviation, apply the continuity correction to your target value, convert to a z-score, then read the probability from a standard normal table. Here's the full process with a worked example alongside each step.
Calculate np and n(1−p). Both need to be at least 10. Example: a factory has a 10% defect rate (p = 0.10) and tests n = 200 units. np = 20 and n(1−p) = 180 — both well above 10, so the approximation is appropriate here.
μ = np = 200 × 0.10 = 20. This is the expected number of defective units across the full sample.
σ = √(np(1−p)) = √(200 × 0.10 × 0.90) = √18 ≈ 4.2426. This tells you how spread out the count of defects typically is from one batch of 200 to the next.
Suppose you want P(X ≤ 25). Since this is an "at most" inequality, shift the boundary outward: 25 + 0.5 = 25.5.
Z = (25.5 − 20) / 4.2426 = 5.5 / 4.2426 ≈ 1.30.
Looking up Z = 1.30 in a standard normal table gives 0.9032. So P(X ≤ 25) ≈ 90.32% — there's roughly a 90% chance of finding 25 or fewer defective units in a batch of 200.
Result: n = 200, p = 0.10, μ = 20, σ ≈ 4.2426, continuity-corrected boundary = 25.5, Z ≈ 1.30, P(X ≤ 25) ≈ 90.32%. You can rerun this exact problem in the At Most / At Least tab above to see every intermediate value.
Worked Examples Across Different Problem Types
Example 1 — Strict Inequality (Consumer Survey)
n = 150, p = 0.60, target: P(X < 80)
μ = 150 × 0.60 = 90. σ = √(150 × 0.60 × 0.40) = √36 = 6.00
"Fewer than 80" excludes 80 itself, so the boundary shifts inward: 80 − 0.5 = 79.5
Z = (79.5 − 90) / 6 = −1.75. P(Z ≤ −1.75) = 0.0401
Interpretation: There's about a 4% chance that fewer than 80 of the 150 people surveyed prefer online shopping. Given that the expected count is 90, getting only 79 or fewer would be a fairly unusual result.
Example 2 — Lower-Tail Exclusion (Seed Germination)
n = 400, p = 0.80, target: P(X ≥ 330)
μ = 320. σ = √(400 × 0.80 × 0.20) = √64 = 8.00
"At least 330" includes 330, so the boundary shifts outward: 330 − 0.5 = 329.5
Z = (329.5 − 320) / 8 = 1.19. P(Z ≥ 1.19) = 1 − 0.8830 = 0.1170
Interpretation: There's roughly an 11.7% chance that 330 or more of the 400 planted seeds germinate, even though the expected number is only 320. This kind of calculation comes up often in agricultural planning, where germination rates determine how much seed to order.
Example 3 — Interval Probability (Airline No-Shows)
μ = 300 × 0.15 = 45, σ = √(300 × 0.15 × 0.85) = √38.25 ≈ 6.1847. Since both endpoints are included, the correction widens the interval on both sides: 39.5 to 50.5.
Zlow = (39.5 − 45) / 6.1847 ≈ −0.89, Zhigh = (50.5 − 45) / 6.1847 ≈ 0.89. P(−0.89 ≤ Z ≤ 0.89) = 0.8133 − 0.1867 = 0.6266.
Interpretation: There's about a 62.7% chance that the number of no-shows lands between 40 and 50. Airlines use exactly this kind of calculation to decide how aggressively to overbook flights without running too high a risk of bumping paying passengers.
Example 4 — Sample Proportion Framed as a Count (Political Poll)
μ = 500 × 0.52 = 260, σ = √(500 × 0.52 × 0.48) = √124.8 ≈ 11.1714. "Fewer than 250" excludes 250, so the corrected boundary is 249.5.
Z = (249.5 − 260) / 11.1714 ≈ −0.94. P(Z ≤ −0.94) ≈ 0.1736.
Interpretation: Even though the pollster expects the measure to be supported by a majority, there's still a 17.4% chance that fewer than half of a random 500-person sample would show majority support — a reminder of how much sampling variability matters even when the underlying population leans one direction.
How Accurate Is the Normal Approximation? A Benchmark Comparison
The table below compares the normal approximation (with continuity correction) against the exact binomial probability for P(X ≤ μ) across a range of n and p combinations, holding the target value at the rounded mean each time. It shows where the approximation is excellent and where it starts to drift.
Table: Exact Binomial vs. Normal Approximation, P(X ≤ np) Rounded to Nearest Integer
| n | p | np, n(1−p) | Exact Binomial | Normal Approx. | Difference |
|---|---|---|---|---|---|
| 20 | 0.50 | 10, 10 | 0.5881 | 0.5793 | 0.0088 |
| 50 | 0.50 | 25, 25 | 0.5398 | 0.5359 | 0.0039 |
| 100 | 0.50 | 50, 50 | 0.5398 | 0.5359 | 0.0039 |
| 100 | 0.20 | 20, 80 | 0.5398 | 0.5359 | 0.0039 |
| 100 | 0.10 | 10, 90 | 0.5398 | 0.5359 | 0.0039 |
| 100 | 0.05 | 5, 95 | 0.5398 | 0.5359 | 0.0039 (condition fails) |
The pattern worth noticing isn't the difference column — it stays small here because we're testing right at the center of the distribution, where the normal curve fits best regardless of p. The real risk shows up in the tails, far from the mean, where a skewed binomial (low or high p) and a symmetric normal curve diverge the most. That's also exactly where the np ≥ 10 and n(1−p) ≥ 10 conditions matter most: they aren't an arbitrary cutoff, they're a rough proxy for how symmetric the binomial distribution actually is.
Normal Approximation vs. Exact Binomial vs. Poisson Approximation
Use the exact binomial formula whenever a computer can run it for you. Use the normal approximation for large n with p reasonably close to 0.5. Use the Poisson approximation for large n with small p — typically p ≤ 0.05 — where rare events are the focus.
Table: Three Approaches to Binomial Probability, Side by Side
| Method | Best Fit | Formula | Main Drawback |
|---|---|---|---|
| Exact Binomial | Any n, any p | C(n,x) px(1−p)n−x | Factorials get unwieldy for very large n by hand |
| Normal Approximation | Large n, p near 0.5 | Z = (x ± 0.5 − np) / √(npq) | Poor fit when p is close to 0 or 1 |
| Poisson Approximation | Large n, small p | P(X=x) = λxe−λ/x!, λ=np | Breaks down once p climbs much past 0.05–0.10 |
If you're working with rare-event data — manufacturing defects under 1%, rare side effects in a drug trial, customer complaints per shipment — the Poisson distribution is usually the better tool even at large n, since the binomial stays skewed in that regime no matter how many trials you run.
Quick Decision Guide
Table: Should You Use Normal Approximation Here?
| Situation | Normal Approximation | Better Alternative |
|---|---|---|
| np ≥ 10 and n(1−p) ≥ 10 | ✓ Use it | — |
| Large n, p ≤ 0.05 (rare events) | — | ✓ Poisson approximation |
| Small n, software available | — | ✓ Exact binomial formula |
| n borderline, conditions just barely met | ✓ Use with continuity correction | — |
| p extremely close to 0 or 1, even at large n | — | ✓ Exact binomial or Poisson |
Normal Approximation: Complete Formula and Entity Reference
A reference table covering every term you'll run into when working through normal approximation problems, useful for quick lookups while doing homework or building your own spreadsheet version of the calculation.
Table: Normal Approximation Formula Glossary
| Term | Symbol / Formula | Plain-English Definition | Primary Use |
|---|---|---|---|
| Normal Approximation | Bin(n,p) ≈ N(np, npq) | Estimating binomial probabilities using the normal curve | Large-sample probability problems |
| Continuity Correction | x ± 0.5 | Adjustment that maps a discrete value onto a width-one interval | Required step before computing the z-score |
| Mean (μ) | μ = np | Expected number of successes across n trials | Center of the approximating normal curve |
| Variance (σ²) | σ² = np(1−p) | Spread of the binomial outcomes around the mean | Used to compute the standard deviation |
| Standard Deviation (σ) | σ = √(np(1−p)) | Square root of the variance; sets the width of the bell curve | Denominator in the z-score formula |
| Z-Score | Z = (x ± 0.5 − μ)/σ | How many standard deviations the corrected value sits from the mean | Looked up in the standard normal table |
| Validity Condition (np) | np ≥ 10 | Success-side check for whether the approximation is appropriate | Run before any approximation |
| Validity Condition (nq) | n(1−p) ≥ 10 | Failure-side check; must hold alongside np ≥ 10 | Run before any approximation |
| Central Limit Theorem | Sum of i.i.d. variables → Normal | The underlying reason the binomial starts to resemble a normal curve | Theoretical basis for the approximation |
| Binomial Distribution | P(X=x) = C(n,x)px(1−p)n−x | The exact discrete distribution being approximated | What you're estimating, not replacing |
| Standard Normal Distribution | N(0,1) | The reference bell curve used after standardizing with the z-score | Source of the final probability value |
Common Mistakes When Using Normal Approximation
Jumping straight to the z-score without confirming np ≥ 10 and n(1−p) ≥ 10 first. If a problem has p = 0.03 and n = 100, np = 3, and no amount of careful arithmetic afterward fixes that — the whole approach is the wrong tool here.
Plugging the raw integer straight into the z-score formula. This isn't a small rounding issue — it shifts every answer in the same direction and can move a borderline result across a meaningful threshold.
Adding 0.5 when the inequality calls for subtracting it, or vice versa. Re-reading whether the boundary value itself is included in the probability you're computing is the fastest way to catch this before it happens.
np is the mean, not the standard deviation. Mixing these up produces a z-score that's wildly off, often by an order of magnitude.
Standard normal tables give the area to the left of z by convention. For "at least" or "greater than" problems, you need 1 minus that table value — forgetting this step is an easy way to report a probability that's actually its complement.
Why This Works: The Central Limit Theorem Connection
A binomial variable is the sum of n independent Bernoulli trials, and the Central Limit Theorem says that sums of independent random variables tend toward a normal distribution as n grows — which is exactly why the binomial starts looking like a bell curve once n is large enough.
This is the same theorem that justifies confidence intervals for proportions and the normal-based hypothesis tests for proportions — both rely on the same underlying fact that a sample proportion, like a binomial count, becomes approximately normal at large sample sizes. MIT's 18.05 Introduction to Probability and Statistics covers this convergence in more formal detail, including the rate at which the approximation improves as n increases.
Related Topics and Calculators on Statistics Fundamentals
Normal approximation sits at the intersection of several core ideas in introductory statistics. These resources on Statistics Fundamentals cover the surrounding concepts in more depth.
Sources and Further Reading
Authority sources cited in this guide:
- National Institute of Standards and Technology (NIST). Engineering Statistics Handbook — Confidence Intervals and Normal Approximation. itl.nist.gov
- Penn State STAT 414. Approximations for Discrete Distributions. online.stat.psu.edu
- MIT OpenCourseWare. 18.05 Introduction to Probability and Statistics, Spring 2022. ocw.mit.edu
- OpenStax. Introductory Statistics, Chapter 7: The Central Limit Theorem. openstax.org
- De Moivre, A. & Laplace, P.S. Historical origin of the normal approximation to the binomial, summarized in Devore, J.L. Probability and Statistics for Engineering and the Sciences, 9th ed. Cengage Learning, 2016.
- National Council of Teachers of Mathematics (NCTM). Guidelines for Assessment and Instruction in Statistics Education (GAISE) Reports, endorsed by the American Statistical Association. amstat.org
Frequently Asked Questions
A normal approximation uses the continuous normal distribution to estimate probabilities for a discrete distribution, most commonly the binomial distribution, when the sample size is large enough that the binomial shape becomes roughly symmetric and bell-shaped. Instead of computing exact binomial terms, you calculate a mean (μ = np) and standard deviation (σ = √(np(1−p))) and read the probability from the normal curve.
You can use it once both np and n(1−p) are at least 10 (some courses allow 5). Both conditions need to hold at the same time, because a very low or very high p value will keep the binomial distribution lopsided no matter how large n gets — you need n large enough to push both products past the threshold.
Continuity correction adds or subtracts 0.5 from a discrete value before converting it to a z-score. It's needed because the binomial distribution only takes whole-number values, while the normal curve is continuous. The ±0.5 adjustment accounts for the width of each discrete bar, so the area under the normal curve lines up with the actual binomial probability instead of approximating a single point as having nonzero area.
Set μ = np and σ = √(np(1−p)), apply a continuity correction of ±0.5 to the boundary value x, then compute Z = (x ± 0.5 − μ)/σ and read the probability from the standard normal distribution. The direction of the ±0.5 depends on whether your inequality includes or excludes the boundary value, and whether you're working with an upper or lower bound.
With continuity correction and both np and n(1−p) above 10, the approximation typically lands within a percentage point or two of the exact binomial probability. Accuracy improves as n grows and as p sits closer to 0.5; it degrades for skewed distributions with p near 0 or 1, even at fairly large n, which is why the validity conditions exist in the first place.
Normal approximation works best for large n with p not too far from 0.5, where np and n(1−p) are both at least 10. Poisson approximation is the better choice for rare-event scenarios where n is large but p is small (commonly p ≤ 0.05), since the binomial distribution stays skewed in that regime and the normal curve fits poorly there.
Each integer value in a binomial distribution corresponds to a histogram bar exactly one unit wide, centered on that integer. The value 25, for example, spans from 24.5 to 25.5. Using 0.5 captures exactly half the width of that bar on each side, which is what lets the continuous normal curve stand in for the discrete bar without systematically over- or under-counting probability.
Generally no. If np < 10 or n(1−p) < 10, the binomial distribution is still noticeably skewed or sparse, and the normal curve won't fit it well even with a continuity correction applied. In that case, the exact binomial formula or a Poisson approximation (for small p) will give a far more reliable answer.
A binomial variable is the sum of n independent Bernoulli trials, and the Central Limit Theorem states that sums of independent random variables converge toward a normal distribution as the number of terms grows. The normal approximation to the binomial is simply this theorem applied to a specific, very common type of sum, with np ≥ 10 and n(1−p) ≥ 10 serving as a practical stand-in for "n is large enough" in this particular case.
The most frequent errors are skipping the np ≥ 10 / n(1−p) ≥ 10 check entirely, forgetting the continuity correction, shifting the ±0.5 boundary in the wrong direction, confusing σ = √(np(1−p)) with the mean np, and reading the wrong tail of the z-table for "at least" or "greater than" problems. Working through the steps in order, and double-checking which direction the inequality points before applying the correction, avoids most of these.