Statistics › Section G · The central limit theorem › Topic G.4

Approximating the binomial with the normal

Three chapters ago, the binomial histogram turned bell-shaped once np and n(1−p) grew past about 10. This topic cashes that observation in: replacing awkward discrete calculations with one normal lookup, when the trade is safe, and the half-unit fix that makes it accurate.

The bell returns

Three chapters ago, the binomial histogram turned bell-shaped once \(np\) and \(n(1-p)\) grew past about 10. The Poisson's bar chart does the same when \(\lambda\) grows. This topic cashes both observations in: how to replace awkward discrete calculations with one normal lookup, when the trade is safe, and the half-unit fix that makes it accurate.

Worked example: 400 fair flips

When \(np \ge 10\) and \(n(1-p) \ge 10\) - the success-failure condition - a binomial count \(X\) is approximately normal:

X \approx N\left(np, \sqrt{np(1-p)}\right)

which simply restates the binomial's own mean and SD. Try it on 400 fair flips: the head count is approximately \(N(200, 10)\), since \(np = 200\) and \(\sqrt{400 \times 0.25} = 10\). Getting more than 220 heads - two SDs above centre - has probability about \(P(Z > 2) = 0.0228\), near the empirical rule's rougher 2.5%. The next card's correction refines this to about 0.020, matching the exact binomial value of 0.0201.

The continuity correction

The binomial is a bar chart; the normal is a smooth curve. The count \(k\)'s bar runs from \(k - 0.5\) to \(k + 0.5\), so the curve must cover the bar's full width: to approximate \(P(X \le 220)\), compute \(P(Y \le 220.5)\), extending half a unit up; to approximate \(P(X \ge 65)\), start the shading at 64.5. The picture beats memorising signs: draw the bars, shade the ones that count, and let the shading's edges - not the integers - set the curve's cutoffs. The correction matters most for moderate \(n\), where bars are wide relative to the SD.

Sample proportions go normal too

Divide a binomial count by \(n\) and the sample proportion \(\hat{p}\) emerges, with mean \(p\) and standard deviation

\sigma_{\hat{p}} = \sqrt{\frac{p(1-p)}{n}}

It, too, is approximately normal for large \(n\) - the proportion's version of the CLT. If 60% of voters support a measure, a poll of 1,000 lands within about 3 percentage points of 0.60 roughly 95% of the time: \(2\sqrt{0.24/1000} \approx 0.031\). That "about 3 points" is the familiar poll margin of error, derived here in one line.

Worked example: a defect spike

A factory's defect rate is 5%. In a batch of 1,000 items, what is the chance of more than 65 defects? 1. Condition: \(np = 50 \ge 10\) and \(n(1-p) = 950 \ge 10\) - approximation licensed. 2. Centre and spread: \(np = 50\), \(\sqrt{1000(0.05)(0.95)} \approx 6.9\). 3. Continuity correction: more than 65 means 66 and up, so shade from 65.5. 4. Standardise:

P(Y > 65.5) = P\left(Z > \frac{65.5-50}{6.9}\right) = P(Z > 2.25) \approx 0.012

5. Check: about a 1% event (the exact binomial gives about 0.015, as some right skew remains) - a batch that defective suggests the process changed, not bad luck.

Approximating the Poisson

A Poisson count with a large mean is also approximately normal: for \(\lambda\) of 20 or more,

X \approx N\left(\lambda, \sqrt{\lambda}\right)

A hospital averaging 36 births a day sees more than 48 births - \(2\sqrt{36} = 12\) above the mean, two SDs - on roughly 2.3% of days: \(P(Z > 2) = 0.0228\), close to the exact Poisson value of about 0.023. The same success-count logic governs both approximations: piles of small chances add up to a bell, whether the piles come from repeated trials or from intervals.

When not to approximate

The success-failure condition is not a formality. With \(n = 20\) and \(p = 0.05\), \(np = 1\): the binomial is a stubby right-skewed pile at 0 and 1, nothing like a bell, and the normal approximation would pronounce absurd probabilities. The honest routes are exact binomial calculation, or - for rare events among many trials - the Poisson approximation from the Poisson topic. Small samples and rare events are exactly where approximations fail; check before computing.

Software changes the arithmetic, not the idea

Modern software computes exact binomial and Poisson probabilities instantly, so approximating to save labour is rarely necessary now. The approximations still matter for understanding: they explain why sample proportions and averages behave predictably, they power the quick margins of error quoted in polls, and they underlie the confidence intervals and tests of the coming chapters. Learn the approximation to see the structure; use exact computation to get the numbers.

Common mistakes to avoid

First, approximating when the success-failure condition fails - check \(np\) and \(n(1-p)\) first, every time. Second, skipping the continuity correction on exact count questions, which shifts answers noticeably for moderate \(n\). Third, using \(\sqrt{p(1-p)}\) without the \(n\) for a proportion's SD: that is the single-trial spread. Fourth, mixing counts and proportions in one calculation: 50 defects and a 5% rate need consistent units. Fifth, applying the Poisson normal approximation at \(\lambda = 3\). Sixth, moving the corrected cutoff away from the bars that count - the correction covers bars, it does not abandon them.

Key takeaways

Large-sample binomial counts are approximately \(N\left(np, \sqrt{np(1-p)}\right)\) once \(np\) and \(n(1-p)\) both reach 10; the continuity correction - extend by half a unit to cover the counted bars - buys accuracy. Sample proportions inherit the bell: mean \(p\), SD \(\sqrt{p(1-p)/n}\), so a 1,000-voter poll lands within about 3 points of the truth 95% of the time. Poisson counts with \(\lambda \ge 20\) are approximately \(N\left(\lambda, \sqrt{\lambda}\right)\). When conditions fail, compute exactly or use the rare-event Poisson route. Software supplies exact numbers; the approximation supplies the understanding.

Flashcards

Flashcard

Success-failure condition

np ≥ 10 and n(1−p) ≥ 10

The licence for the normal approximation.

Test your understanding

Question 1 of 5

A fair coin is flipped 100 times. The head count is approximately normal with what mean and SD?

Question 2 of 5

A binomial has n = 20, p = 0.05, so np = 1. The best tool for its probabilities is:

Question 3 of 5

To approximate P(X ≤ 220) for a binomial count using the normal curve, compute:

Question 4 of 5

Defects: mean 50, SD about 6.9, batch of 1,000. With the continuity correction, P(X > 65) is about:

Question 5 of 5

With p = 0.6 and n = 1,000, what is the standard deviation of the sample proportion?

← Previous topicNext topic →