Mixed CLT Problems for Means and Proportions

Start by Identifying the Target

A central limit theorem (CLT) problem can look different from one question to the next: it may ask about an average measurement, a combined total, or a fraction of individuals with a particular outcome. The first task is to identify which quantity the probability statement describes. That choice determines the sampling distribution’s center and standard error.

The normal approximation describes the variation of a statistic across repeated samples, not the spread of individual observations. In particular, a sample mean is usually less variable than a single observation, and a total is usually more variable than a single observation. Keeping those distinctions clear prevents a common error: using the population standard deviation in place of the statistic’s standard error.

CLT models: Suppose $X_1,\ldots,X_n$ are independent observations from the same population, with mean $\mu$ and finite standard deviation $\sigma$. For a sufficiently large sample, the CLT gives the approximations $$ \bar{x}\ \text{approximately follows}\ N\left(\mu,\frac{\sigma^2}{n}\right), \qquad T=\sum_{i=1}^{n}X_i\ \text{approximately follows}\ N(n\mu,n\sigma^2). $$ For independent binary outcomes, let $X$ be the number of successes in $n$ trials, each with success probability $p$, and let $\hat{p}=X/n$. When the success-failure counts are sufficiently large, $$ X\ \text{approximately follows}\ N\bigl(np,np(1-p)\bigr), \qquad \hat{p}\ \text{approximately follows}\ N\left(p,\frac{p(1-p)}{n}\right). $$

The normal model for a sample mean or total requires independent observations from a population with finite variance; the CLT supports the approximation as $n$ grows. If the population itself is normal, the sampling distribution of the mean is normal for any sample size. Otherwise, a larger sample is generally needed, and no single sample-size rule guarantees accuracy for every population. A random sample and an appropriate independence condition are important: for sampling without replacement, observations from a small finite population are not exactly independent, so the relevant finite-population adjustment must be considered.

For the proportion model, each observation must be a success or a failure, with the same probability $p$ and independent outcomes under the model. A useful check is that both expected counts, $np$ and $n(1-p)$, are sufficiently large for the normal approximation. The course’s earlier material on success-failure counts explains this condition and the accuracy limitations in more detail.

A Short Justification for the Centers and Standard Errors

The formulas follow from the expected value and variance of sums. Independence makes the variance of a sum equal to the sum of the individual variances. Dividing a sum by $n$ divides its variance by $n^2$. Thus:

E(T)=n\mu,\quad \operatorname{Var}(T)=n\sigma^2, \qquad E(\bar{x})=\mu,\quad \operatorname{Var}(\bar{x})=\frac{n\sigma^2}{n^2}=\frac{\sigma^2}{n}.

The standard error is the standard deviation of a statistic’s sampling distribution. Therefore the standard errors are $\sigma\sqrt{n}$ for a total and $\sigma/\sqrt{n}$ for a mean. For a binary observation, the mean is $p$ and the variance is $p(1-p)$. Applying the same sum-and-divide reasoning gives standard errors $\sqrt{np(1-p)}$ for the success count and $\sqrt{p(1-p)/n}$ for the sample proportion. The CLT is what justifies approximating these standardized statistics with a standard normal distribution under its conditions.

Once the model is chosen, standardize the event using its center and standard error. If $Z$ denotes a standard normal random variable, a value $z$ measures how many standard errors the target lies from its center. For a continuous normal approximation, probabilities can then be found from the standard normal distribution.

Worked Examples

Worked Example: A Probability About a Sample Mean

Suppose a hypothetical production process has a mean tablet dissolution time of 52 minutes and a standard deviation of 12 minutes. A random sample of 64 tablets is selected independently. Approximate the probability that the sample mean dissolution time exceeds 55 minutes.

Check the model. The sample is assumed random and independent, and the population has a finite standard deviation. The population’s shape has not been specified, so we rely on the CLT for a sample of 64; the normal approximation is reasonable if the population is not so extremely skewed or heavy-tailed that this sample size is inadequate.

The target is a sample mean, so its approximate center is 52 minutes and its standard error is $$ SE(\bar{x})=\frac{\sigma}{\sqrt{n}}=\frac{12}{\sqrt{64}}=1.5\text{ minutes}. $$ The standardized boundary is $$ z=\frac{55-52}{1.5}=2. $$ Therefore, $$ P(\bar{x}>55)\approx P(Z>2)=1-\Phi(2)\approx 0.0228. $$ Under the stated assumptions and approximation, about 2.28% of samples of 64 tablets would have a mean dissolution time above 55 minutes. This is a probability about sample means, not the percentage of individual tablets whose dissolution times exceed 55 minutes.

Worked Example: A Probability About a Total

In a hypothetical bottling process, an individual bottle’s fill volume has mean 500 mL and standard deviation 8 mL. Suppose 100 bottles are selected independently from the process. Approximate the probability that their combined fill volume is less than 49,800 mL.

Check the model. We assume the selected bottles are representative, their fill volumes are independent, and the individual volume distribution has finite variance. Since the individual distribution is not specified, the CLT is used for the total of 100 measurements; unusually heavy tails could make the approximation less reliable.

The target is the total $T$, not the average. Its approximate mean and standard error are $$ E(T)=n\mu=100(500)=50{,}000\text{ mL}, \qquad SE(T)=\sigma\sqrt{n}=8\sqrt{100}=80\text{ mL}. $$ The standardized boundary is $$ z=\frac{49{,}800-50{,}000}{80}=-2.5. $$ Thus, $$ P(T<49{,}800)\approx P(Z<-2.5)=\Phi(-2.5)\approx 0.00621. $$ The model estimates that about 0.621% of groups of 100 bottles would have a combined fill volume below 49,800 mL. The units in the boundary and standard error are both millilitres, so their ratio is dimensionless, as a $z$-score must be.

Worked Example: A Probability About a Sample Proportion

Imagine a hypothetical process in which 8% of items have a particular defect. A random sample of 200 items is inspected, with independent outcomes. Approximate the probability that the sample proportion of defective items is at most 0.0525.

Check the model. Each inspected item has a binary outcome, and the model assumes independent trials with the same defect probability $p=0.08$. The expected success and failure counts are $$ np=200(0.08)=16,\qquad n(1-p)=200(0.92)=184. $$ Both are sufficiently large for a normal approximation in this setting. The sample proportion is based on the defective-item count $X$, which has approximate mean 16 and variance $$ np(1-p)=200(0.08)(0.92)=14.72. $$

Because the proportion changes in increments of $1/200=0.005$, translate the event into a count event before applying the continuity correction: $$ \hat{p}\leq 0.0525 \quad\Longleftrightarrow\quad \frac{X}{200}\leq 0.0525 \quad\Longleftrightarrow\quad X\leq 10.5 \quad\Longleftrightarrow\quad X\leq 10. $$ For the normal approximation to the count, the continuity-corrected boundary is 10.5. Its $z$-score is $$ z=\frac{10.5-16}{\sqrt{14.72}}\approx -1.43354. $$ Therefore, $$ P(\hat{p}\leq 0.0525)\approx P(Z\leq -1.43354)\approx 0.0759. $$ The estimated probability is about 7.59%. The continuity correction is applied to the discrete count boundary; using the uncorrected boundary 10 would not give the same approximation.

A Reliable Problem-Solving Sequence

The same sequence works across many applied CLT questions. Writing down the target and the model before calculating helps reveal whether a mean, total, or proportion formula belongs in the solution.

Name the target.
Decide whether the event concerns a sample mean, a sum or total, a count, or a sample proportion. Translate between a count and a proportion when necessary.

Check the assumptions.
Establish that the sampling process is appropriate and that observations are independent or that a suitable finite-population adjustment is used. For a proportion, check the expected success and failure counts.

Write the approximate distribution.
State its center and standard error, including units. Do not confuse the spread of individual observations with the standard error of the statistic.

Standardize and interpret.
Convert the event boundary to a $z$-score, use the appropriate tail or interval probability, and describe the result as a probability about repeated samples in the stated context.

Common Mistakes and Why They Matter

Using the wrong standard error. For a mean, use $\sigma/\sqrt{n}$, not $\sigma$. For a total, use $\sigma\sqrt{n}$, not $\sigma/\sqrt{n}$. For a proportion, use $\sqrt{p(1-p)/n}$.
Treating individual variation as sampling variation. A population with a large standard deviation can still have a much less variable sample mean when $n$ is large. The probability calculation must use the statistic’s sampling distribution.
Skipping the conditions. A large sample does not automatically make every normal approximation accurate. Dependence, a small expected count, or an extreme population shape may matter.
Forgetting the count’s discreteness. A sample proportion comes from a discrete count. When approximating a count probability by a continuous normal curve, use the appropriate continuity correction, as in the third example.
Giving the calculation without its meaning. A probability such as 0.0759 is not a claim about the chance that the assumed value of $p$ is true. It describes the chance of the specified sample outcome under the model and assumptions.

These distinctions matter because a probability calculation is only as useful as the model behind it. Identifying the target correctly connects the question to the right standard error; checking conditions explains whether the approximation is credible; and interpretation in context tells a reader exactly what repeated-sampling outcome the probability describes.

Check Your Understanding

For each question, identify the target and explain the model choice before calculating or describing the required probability.

A random independent sample of 81 measurements comes from a population with mean 30 cm and standard deviation 9 cm. What are the approximate center and standard error of the sample mean?
For 50 independent observations with individual mean 4.2 units and standard deviation 1.1 units, state the approximate mean and standard error of the total.
A binary outcome has probability $p=0.25$ in each of 120 independent trials. What expected counts should be checked before using a normal approximation for the sample proportion?
Explain why the standard deviation of individual observations is not the standard error of a sample mean.
A question asks for the probability that at least 18 successes occur in 100 independent trials. Which boundary would receive the continuity correction in a normal approximation to the count?