CLT Probability Calculations for General Sample Averages

Why Probabilities for Averages Are Often Easier

An individual measurement may have a skewed or otherwise unfamiliar distribution, while the average of many independent measurements can have a distribution that is close to normal. The central limit theorem (CLT) makes this useful: it lets us approximate probabilities about an average using its mean, its standard error, and the standard normal distribution.

This tutorial focuses on turning an event about a sample average into a standard normal probability. The observations need not be Bernoulli, and their original distribution need not itself be normal. The approximation does depend on appropriate assumptions, including independence and a sufficiently large sample for the distribution of the average to be adequately close to normal.

Definition: Let $X_1,\ldots,X_n$ be independent and identically distributed observations with finite mean $\mu$ and finite, positive variance $\sigma^2$. Their sample average is $$ \bar{X}=\frac{X_1+\cdots+X_n}{n}. $$ The CLT says that, for sufficiently large $n$, the distribution of $\bar{X}$ is approximately normal with mean $\mu$ and variance $\sigma^2/n$. Equivalently, $(\bar{X}-\mu)/(\sigma/\sqrt{n})$ is approximately standard normal.

Write $\Phi(z)=P(Z\le z)$ for the cumulative distribution function of a standard normal random variable $Z\sim N(0,1)$. The standard error of $\bar{X}$, meaning the standard deviation of its sampling distribution, is $\sigma/\sqrt{n}$. It is not the standard deviation $\sigma$ of an individual observation.

From an Average Event to a Standard Normal Probability

For a threshold $c$, subtract the mean $\mu$ from the threshold and divide by the standard error. This produces the standardized value $z=(c-\mu)/(\sigma/\sqrt{n})$. Since $\bar{X}$ increases as its standardized value increases, upper and lower tails keep their direction: a below-threshold event uses a lower tail, and an above-threshold event uses an upper tail.

P(\bar{X}\le c)\approx \Phi\!\left(\frac{c-\mu}{\sigma/\sqrt{n}}\right), \qquad P(\bar{X}>c)\approx 1-\Phi\!\left(\frac{c-\mu}{\sigma/\sqrt{n}}\right).

For an interval, standardize both endpoints and subtract the lower cumulative probability from the upper one. The same calculation works for strict or non-strict inequalities when the approximating normal distribution is continuous; the probability of an exact endpoint under that approximation is zero.

P(a\le \bar{X}\le b)\approx \Phi\!\left(\frac{b-\mu}{\sigma/\sqrt{n}}\right) - \Phi\!\left(\frac{a-\mu}{\sigma/\sqrt{n}}\right).

These expressions use the CLT approximation, not necessarily an exact distribution for $\bar{X}$. If the individual observations are normally distributed, their average is exactly normal for every $n$. Otherwise, the accuracy of the approximation depends on the distribution and sample size. Use the assumptions and condition checks appropriate to the problem; a large sample is not a substitute for independence or finite variance.

Why the Standard Error Has This Form

The mean and variance of the average explain why its standardization uses $\sigma/\sqrt{n}$. For independent observations with common mean $\mu$ and variance $\sigma^2$, linearity of expectation gives $E[\bar{X}]=\mu$. Independence makes the variance of a sum equal to the sum of the variances.

Find the mean.
$E[\bar{X}]=E[(X_1+\cdots+X_n)/n]=(n\mu)/n=\mu.$

Find the variance.
$\operatorname{Var}(\bar{X})=\operatorname{Var}(X_1+\cdots+X_n)/n^2=n\sigma^2/n^2=\sigma^2/n.$

Take the square root.
The standard deviation of $\bar{X}$ is $\sqrt{\sigma^2/n}=\sigma/\sqrt{n}$, which is its standard error.

The CLT supplies the approximate normal shape; the mean and variance calculation supplies its center and spread. If the observations are independent but have different means $\mu_i$ and variances $\sigma_i^2$, then the average has mean $\sum_{i=1}^n\mu_i/n$ and variance $\sum_{i=1}^n\sigma_i^2/n^2$. When conditions for a CLT approximation hold, standardize with those quantities rather than assuming one common $\mu$ and $\sigma$.

Worked Probability Calculations

Worked Example: An Interval for a Sample Average

A hypothetical laboratory records 49 independent processing times. Suppose the times have a common mean of 200 minutes and a common standard deviation of 21 minutes, with a distribution for which a CLT approximation is appropriate at this sample size. Approximate the probability that the average processing time is between 196 and 203 minutes.

Method. Use the CLT for the average, first calculating its standard error. The observation distribution is not specified as normal, so this is an approximation.

The sample size is $n=49$, so the standard error is $21/\sqrt{49}=21/7=3$ minutes. The lower endpoint standardizes to $(196-200)/3=-4/3$, and the upper endpoint standardizes to $(203-200)/3=1$.

P(196\le \bar{X}\le 203) \approx \Phi(1)-\Phi(-4/3) =0.841344746-0.091211220 =0.750133526.

Rounded only at the final step, the probability is 0.7501 to four decimal places. As a check, the lower endpoint is below the mean and has a negative standardized value, while the upper endpoint is above the mean and has a positive one; the resulting interval probability is between 0 and 1.

Common mistake avoided: Rounding $-4/3$ to $-1.33$ before looking up the probability gives approximately $0.8413-0.0918=0.7495$. That is not the requested four-decimal answer: retain the exact endpoint $-4/3$ through the normal-probability calculation, then round the final result.

Worked Example: A Two-Sided Range for Average Claim Amount

In a hypothetical portfolio study, 64 independent claim amounts have a common mean of $80 and standard deviation of $24. Assume the CLT approximation is suitable. Approximate the probability that the average claim amount is between $77 and $84.50.

Method. Standardize both endpoints using the standard error of the average. The claim amounts are not assumed to be Bernoulli or normal.

Here $n=64$, so the standard error is $24/\sqrt{64}=24/8=\$3$. The lower endpoint has standardized value $(77-80)/3=-1$. The upper endpoint has standardized value $(84.50-80)/3=4.50/3=1.5$.

P(77\le \bar{X}\le 84.50) \approx \Phi(1.5)-\Phi(-1) =0.933192799-0.158655254 =0.774537545.

The probability is 0.7745 to four decimal places. As a check, the interval extends one standard error below the mean and 1.5 standard errors above it, so it should contain a substantial portion of the distribution.

A common error is to use the individual standard deviation, $24, as the denominator. That would produce standardized endpoints $-3/24$ and $4.5/24$, which describe neither the sampling distribution of the average nor the requested event. The relevant spread is the standard error, $3.

Worked Example: Applying the CLT to Uniform Observations

A hypothetical sensor’s error $X$, in a single reading, is uniformly distributed on the interval from $-6$ to $6$. For 36 independent readings, approximate the probability that their average error exceeds 1. The individual errors are not normally distributed.

Method. Use the mean and variance of a uniform distribution, then standardize the average. For a uniform distribution on $[a,b]$, the mean is $(a+b)/2$ and the variance is $(b-a)^2/12$.

Here $a=-6$ and $b=6$, so the mean is $(-6+6)/2=0$. The variance is $(6-(-6))^2/12=12^2/12=12$, and the standard deviation is $\sqrt{12}$. With $n=36$, the standard error is $\sqrt{12}/6=1/\sqrt{3}$.

The threshold 1 standardizes to $(1-0)/(1/\sqrt{3})=\sqrt{3}$, approximately 1.7321. Therefore,

P(\bar{X}>1) \approx 1-\Phi(\sqrt{3}) =1-0.958367742 =0.041632258.

The approximate probability is 0.0416 to four decimal places. The check is consistent with the setup: the threshold is about 1.73 standard errors above the mean, so the upper-tail probability should be small. The individual readings are uniform, but the CLT supports a normal approximation for their average.

Common Traps When Standardizing Averages

Using the individual standard deviation. For an average of $n$ independent, identically distributed observations, use $\sigma/\sqrt{n}$, not $\sigma$.
Reversing a tail. For an upper-tail event, subtract the cumulative probability from 1. For an interval, subtract the lower cumulative probability from the upper one.
Using the wrong center. For identically distributed observations, center at their common mean $\mu$, not at zero unless $\mu=0$.
Rounding standardized values too early. Keep exact fractions or several guard digits until the probability is calculated, then round the final probability as requested.
Treating an approximation as exact. Unless the observations are normally distributed, the CLT probability is generally approximate, and its quality depends on the assumptions and sample size.

Key takeaway: For a suitable CLT approximation, standardize an average by subtracting its mean and dividing by its standard error. Then evaluate the corresponding standard normal tail or interval probability, keeping full precision until the final rounding.

Check Your Understanding

For each problem, identify the mean and standard error of the average before standardizing. Give probabilities to four decimal places.

A hypothetical production process yields independent measurements with common mean 50 and standard deviation 10. For $n=25$, approximate $P(\bar{X}<54)$. (Answer: 0.9772)
Independent readings have common mean 12 and standard deviation 5. For $n=100$, approximate $P(11.5\le\bar{X}\le13)$. (Answer: 0.8186)
Independent observations have common mean 30 and standard deviation 14. For $n=49$, approximate the probability that their average is at most 27. (Answer: 0.0668)
Twenty-five independent observations have mean 10 and variance 9, and another 25 independent observations have mean 14 and variance 16. Assume independence across all observations and that a CLT approximation is suitable. For the average of all 50 observations, approximate the probability it exceeds 13. (Answer: 0.0228)
Independent hypothetical waiting times are exponentially distributed with rate $0.2$ per minute. For $n=100$, approximate the probability that the average waiting time exceeds 6 minutes. (Answer: 0.0228)

What You'll Learn