Normal Approximation To The Binomial

Why Approximate a Binomial Count?

A binomial model describes a fixed number of independent trials, each with the same probability of success. When the number of trials is large enough and both outcomes are sufficiently represented, the distribution of the total number of successes can be approximated by a normal distribution. This can turn a sum of many binomial probabilities into a standardized normal probability.

The approximation is useful only when its assumptions and limitations are understood. In particular, a binomial count is integer-valued, while a normal random variable is continuous. The formulas below use raw count boundaries, with no continuity correction; that choice makes them quick to apply, but can affect accuracy near a boundary.

Definition: If $X\sim\operatorname{Bin}(n,p)$, then $X$ is the number of successes in $n$ independent trials, each having success probability $p$. Write $q=1-p$. Its mean and variance are $\mu=np$ and $\sigma^2=npq$, so the normal approximation uses $Y\sim N(np,npq)$, with standard deviation $\sigma=\sqrt{npq}$.

The binomial model requires a fixed trial count, independent trials, and a common success probability. For example, a hypothetical insurer could model the number of policyholders who file at least one claim during a year as binomial only if the model treats each policyholder’s outcome as independent and assigns the same claim probability to each. Whether those assumptions are plausible is a separate question from whether the normal approximation is accurate.

Standardizing the Count

For a count $x$, its standardized value is $z=(x-\mu)/\sigma$. The standard normal distribution function $\Phi(z)$ gives the probability that a standard normal random variable is at most $z$. Applying this standardization gives the following raw-boundary approximations.

P(X\le k)\approx\Phi\!\left(\frac{k-\mu}{\sigma}\right), \qquad P(X\ge k)\approx 1-\Phi\!\left(\frac{k-\mu}{\sigma}\right).

For an inclusive integer interval, the corresponding direct approximation is $P(a\le X\le b)\approx \Phi((b-\mu)/\sigma)-\Phi((a-\mu)/\sigma)$. These formulas substitute the stated integer boundaries directly. They should not be combined with a different boundary convention partway through a calculation. A continuity correction adjusts the boundaries to reflect the discrete count; it is a separate refinement.

Point-probability approximation: A continuous normal variable has probability zero at any single point, so $P(X=k)$ is not approximated by a normal point probability. A local normal approximation instead uses the normal density at $k$, multiplied by one count unit: $$ P(X=k)\approx \frac{1}{\sigma}\,\phi\!\left(\frac{k-\mu}{\sigma}\right), \qquad \phi(z)=\frac{1}{\sqrt{2\pi}}e^{-z^2/2}. $$ This is most useful when $npq$ is reasonably large and $k$ is not far into a tail.

Why the Normal Model Can Work

A binomial count can be written as a sum of independent indicator variables, one for each trial. Each indicator is 1 for a success and 0 otherwise, with mean $p$ and variance $pq$. Adding the $n$ indicators gives the binomial mean and variance, and the central limit theorem explains why the standardized sum is approximately normal when many independent contributions are combined.

X=\sum_{j=1}^{n} I_j, \qquad E[X]=\sum_{j=1}^{n}E[I_j]=np, \qquad \operatorname{Var}(X)=\sum_{j=1}^{n}\operatorname{Var}(I_j)=npq.

The variance sum uses independence: the covariance between distinct indicators is zero. The central limit theorem motivates the approximation, but does not make it exact for a finite $n$. A useful initial check is whether both $np$ and $nq$ are at least about 5; requiring both to be at least 10 is a more cautious rule of thumb. These are practical guidelines, not guarantees. Approximation quality also depends on how far the event lies into a tail and on whether the binomial distribution is strongly skewed.

Check the model.
Confirm that a fixed number of independent, equal-probability trials is being modeled.

Calculate the normal parameters.
Set $\mu=np$, $q=1-p$, and $\sigma=\sqrt{npq}$.

Check suitability.
Compare $np$ and $nq$ with a practical threshold, and be cautious with strongly skewed or extreme-tail events.

Standardize and evaluate.
Use the appropriate raw-boundary formula or local point formula, and round only the final probability.

Worked Applications

Worked Example: Estimating an Upper-Tail Probability

A hypothetical portfolio has 400 independent policies. For each policy, the probability of at least one claim during the year is 0.20. Estimate the probability that at least 90 policies have a claim, using the raw-boundary normal approximation.

Model and suitability. Let $X$ be the number of policies with a claim. Under the stated assumptions, $X\sim\operatorname{Bin}(400,0.20)$. Here $n=400$, $p=0.20$, and $q=0.80$. The expected success and failure counts are $np=80$ and $nq=320$, so both are large enough for a normal approximation to be plausible.

Find the normal parameters. $\mu=np=400(0.20)=80$, and $\sigma^2=npq=400(0.20)(0.80)=64$. Thus $\sigma=\sqrt{64}=8$, and the approximating distribution is $N(80,64)$.

Standardize the stated threshold. For $k=90$, $z=(90-80)/8=10/8=1.25$. Using the raw-boundary upper-tail formula, $P(X\ge90)\approx1-\Phi(1.25)$.

Evaluate and check. Since $\Phi(1.25)\approx0.8944$, the estimate is $1-0.8944=0.1056$, or 10.56%. This is a plausible upper-tail probability: the threshold is 10 above a mean of 80, or 1.25 standard deviations above it. The method avoids summing 311 binomial probabilities, but the result remains an approximation.

Worked Example: Estimating an Inclusive Interval Probability

A hypothetical quality audit samples 200 independent components. Each component has probability 0.50 of passing a specified test. Estimate the probability that between 90 and 110 components, inclusive, pass, using raw interval boundaries.

Model and suitability. Let $X$ count the components that pass. Then $X\sim\operatorname{Bin}(200,0.50)$. The expected numbers of passes and failures are both $200(0.50)=100$, so the normal approximation is suitable as a working estimate.

Calculate the parameters. The mean is $\mu=200(0.50)=100$, and the variance is $\sigma^2=200(0.50)(0.50)=50$. Therefore, $\sigma=\sqrt{50}\approx7.0711$.

Standardize both endpoints. The lower endpoint gives $z_L=(90-100)/\sqrt{50}\approx-1.4142$. The upper endpoint gives $z_U=(110-100)/\sqrt{50}\approx1.4142$. Thus the raw-boundary interval estimate is $\Phi(1.4142)-\Phi(-1.4142)$.

Evaluate and check. By symmetry, $\Phi(-1.4142)=1-\Phi(1.4142)$, and $\Phi(1.4142)\approx0.92135$. Keeping the CDF values unrounded until subtraction gives $0.92135-(1-0.92135)\approx0.8427$. The estimated probability is 0.8427, or 84.27%. It is sensible for an interval extending about 1.41 standard deviations on each side of the mean to contain a substantial majority of the normal probability.

Worked Example: Estimating a Single Count Probability

A hypothetical system monitors 400 independent devices, each of which has probability 0.25 of triggering an alert during a day. Use a local normal approximation to estimate the probability of exactly 100 alerts.

Model and suitability. Let $X$ be the alert count. Then $X\sim\operatorname{Bin}(400,0.25)$, with $q=0.75$. The expected numbers of alerts and non-alerts are $100$ and $300$, so the normal approximation is reasonable for this central point.

Calculate the parameters and standardized value. The mean is $\mu=400(0.25)=100$, and the variance is $\sigma^2=400(0.25)(0.75)=75$. Hence $\sigma=\sqrt{75}\approx8.6603$. For $k=100$, $z=(100-100)/8.6603=0$.

Apply the local formula. The standard normal density at zero is $\phi(0)=1/\sqrt{2\pi}\approx0.3989423$. Therefore, $P(X=100)\approx\phi(0)/\sqrt{75}\approx0.3989423/8.6603\approx0.0461$.

Check the interpretation. The estimate is 0.0461, or 4.61%. It is a probability assigned to one integer count by approximating the binomial mass near the mean with the normal density over a one-unit count scale. It is not obtained by asking for the probability that a continuous normal variable equals 100, which would be zero.

Common Mistakes and Practical Limits

The approximation should not be treated as a substitute for checking the model or the event. When $np$ or $nq$ is small, the binomial distribution is often skewed and the normal model may assign misleading probability near the edge of the possible count range. The binomial count must lie between 0 and $n$, but the normal distribution extends without bound in both directions. An approximation that puts noticeable mass below 0 or above $n$ is a warning sign.

Using the wrong variance. For a binomial count the variance is $npq$, not $np$ unless a separate approximation is being considered. Use $\sqrt{npq}$ as the standard deviation.
Using a point CDF difference for a point mass. A continuous distribution assigns zero probability to a single point. For $P(X=k)$, use the local density approximation, not $P(Y=k)$.
Mixing boundary conventions. Raw-boundary formulas and continuity-corrected formulas use different standardized endpoints. Apply one convention consistently.
Assuming a large $n$ is sufficient. If $p$ is very small or very close to 1, one of $np$ and $nq$ may remain small even when $n$ is large.
Rounding intermediate values. Retain the unrounded standard deviation, z-scores, and CDF values through the calculation. Round the final probability to four decimal places.

Key takeaway: For $X\sim\operatorname{Bin}(n,p)$, use a normal model with mean $np$ and variance $np(1-p)$ when both expected outcome counts are sufficiently large. Standardize carefully, distinguish point masses from tail or interval probabilities, and treat the result as an approximation whose accuracy depends on the model and the event.

Check Your Understanding

Use raw count boundaries for tail and interval estimates, and retain full precision until the final probability, reported to four decimal places.

A hypothetical group contains 250 independent members, each with probability 0.40 of renewing a subscription. Estimate the probability that at least 110 renew. (Answer: 0.0984)
Among 160 independent inspected items, each has probability 0.50 of meeting a standard. Estimate the probability that between 74 and 86 items, inclusive, meet it. (Answer: 0.6572)
A hypothetical monitoring network has 300 independent sensors, each with probability 0.20 of sending an alert. Use the local normal approximation to estimate the probability of exactly 64 alerts. (Answer: 0.0488)
A binomial model has $n=40$ and $p=0.08$. Assess whether a normal approximation is a good routine choice by checking the expected numbers of successes and failures. (Answer: $np=3.2$ and $nq=36.8$; the success count is too small for the usual rule-of-thumb check, so prefer the exact binomial calculation.)
A hypothetical portfolio has 500 independent policies, each with probability 0.12 of a claim. Using raw interval boundaries, estimate the probability of between 50 and 70 claims, inclusive. (Answer: 0.8312)

What You'll Learn