Understanding Normal Approximation to Binomial

The binomial distribution describes the probability of observing exactly k successes in N independent trials, each with success probability p. For small N, the binomial probabilities follow a stepped, discrete pattern. However, as N grows, this stepped distribution converges to a smooth, continuous curve that closely matches a normal (Gaussian) distribution centered at μ = Np.

This convergence is what the central limit theorem guarantees: the sum of many independent random events tends toward normality, regardless of the underlying distribution. By treating the binomial as normal, you avoid computing factorial-heavy binomial coefficients and instead use standard normal tables or software to find probabilities in seconds.

The approximation works best when both N × p ≥ 5 and N × (1 − p) ≥ 5, ensuring the binomial isn't too skewed toward either tail.

Key Formulas

Given N trials with success probability p, calculate the mean, variance, and standard deviation of the equivalent normal distribution. Then standardize your observation using the z-score formula to look up probabilities in the standard normal table.

μ = N × p

σ² = N × p × (1 − p)

σ = √(σ²)

z = (x − μ) / σ

  • N — Number of independent trials or occurrences
  • p — Probability of success on each trial (0 < p < 1)
  • q — Probability of failure, equal to 1 − p
  • μ — Mean of the normal distribution (expected number of successes)
  • σ² — Variance of the binomial distribution
  • σ — Standard deviation of the normal distribution
  • x — Number of observed successes (the value you want probability for)
  • z — Standardized score; the distance from the mean in units of standard deviation

When to Apply the Normal Approximation

Not every binomial problem benefits from normal approximation. The rule of thumb is straightforward: both N × p and N × (1 − p) must exceed 5. This ensures the binomial distribution is reasonably symmetric and close enough to normal shape.

For example:

  • Fair coin, 100 tosses: N × p = 100 × 0.5 = 50 ✓ Use normal approximation
  • Rare event (p = 0.01), 50 trials: N × p = 50 × 0.01 = 0.5 ✗ Sample too small; approximation fails
  • Rare event (p = 0.01), 1000 trials: N × p = 1000 × 0.01 = 10 ✓ Large enough now

When p is very close to 0 or 1, or when N is small, stick with exact binomial calculations or Poisson approximation instead.

The Continuity Correction

Because the binomial is discrete (0, 1, 2, ... successes) and the normal is continuous, a small adjustment improves accuracy. Add or subtract 0.5 from your observed count before calculating the z-score:

  • For P(X = k): Use P(k − 0.5 < X < k + 0.5)
  • For P(X ≤ k): Use P(X < k + 0.5)
  • For P(X < k): Use P(X < k − 0.5)
  • For P(X ≥ k): Use P(X > k − 0.5)

For large N (say, N > 100), the 0.5 correction matters less, but including it is good practice and ensures consistency with statistical software defaults.

Common Pitfalls and Practical Advice

Avoid these frequent mistakes when using normal approximation to binomial probabilities.

  1. Forgetting the sample size check — Many skip the <em>N</em> × <em>p</em> ≥ 5 test and apply the approximation blindly. Always verify both conditions before trusting your result. If either is false, use exact binomial or Poisson instead.
  2. Omitting continuity correction — The 0.5 adjustment is small but systematic. Ignoring it introduces consistent bias, especially when <em>N</em> is moderate (10–50). Most statistical references and software apply it by default, so do the same for consistency.
  3. Confusing z-score direction — The formula is (x − μ) / σ, not (μ − x) / σ. A negative z-score means your observation is below the mean. Double-check the sign before looking up probabilities in the standard normal table.
  4. Using the approximation for extreme proportions — When <em>p</em> ≈ 0 or <em>p</em> ≈ 1 (e.g., defect rates of 1% or 99%), the binomial skews heavily. Large <em>N</em> helps, but the normal approximation still may underperform compared to exact methods or Poisson approximation for rare events.

Frequently Asked Questions

What is the minimum sample size needed for normal approximation?

There is no universal minimum <em>N</em>; instead, both <em>N</em> × <em>p</em> and <em>N</em> × (1 − <em>p</em>) must be at least 5. For <em>p</em> = 0.5 (balanced), you need <em>N</em> ≥ 10. For more extreme <em>p</em> values, <em>N</em> must be larger. For instance, with <em>p</em> = 0.01, you need <em>N</em> ≥ 500 so that <em>N</em> × <em>p</em> = 5. Always calculate both products before applying the approximation.

Why do we add or subtract 0.5 in the continuity correction?

The binomial distribution is discrete (integer counts only), while the normal is continuous. The 0.5 adjustment bridges this gap by representing a whole number as a range. For example, "exactly 50 successes" becomes the interval [49.5, 50.5) on the normal curve. This correction typically improves accuracy by 1–3 percentage points, especially for smaller samples.

Can I use normal approximation if one product falls below 5?

No. If either <em>N</em> × <em>p</em> < 5 or <em>N</em> × (1 − <em>p</em>) < 5, the binomial distribution is too skewed or sparse for the normal curve to fit well. Instead, compute binomial probabilities exactly using the binomial formula, or consider Poisson approximation if <em>p</em> is very small and <em>N</em> is large.

How does sample size affect the accuracy of the approximation?

Larger samples produce better approximations. As <em>N</em> grows, the binomial distribution becomes more symmetric and closer in shape to a normal curve. A sample of 100 trials typically yields better results than 20, even if both meet the threshold of 5. For <em>N</em> > 100 with moderate <em>p</em>, the approximation is usually excellent (error < 1%).

What is a z-score, and how do I interpret it?

A z-score measures how many standard deviations a value lies from the mean: z = (x − μ) / σ. A z-score of 0 means your observation equals the mean. Positive z indicates values above the mean; negative z indicates below. Standard normal tables convert z to a cumulative probability. For example, z = 1.96 corresponds to roughly 97.5% of the distribution, and z = −1.96 to about 2.5%.

When should I prefer Poisson over normal approximation?

Poisson approximation suits rare events with small <em>p</em> and large <em>N</em>. If <em>N</em> × <em>p</em> is small (< 5) but well-defined, Poisson often outperforms normal. For example, with <em>N</em> = 1000, <em>p</em> = 0.001, normal approximation may fail, but Poisson (with λ = 1) works well. When <em>N</em> is moderate and <em>p</em> is not extreme, normal approximation is usually the fastest and most convenient choice.

More statistics calculators (see all)