Common Standardization Errors in CLT Problems

Why Standardization Errors Matter

The central limit theorem (CLT) turns a probability question about a sum or sample mean into a standard normal probability. The calculation is often short, but a small choice error can change the result substantially: using a population standard deviation where a standard error is needed, scaling a sum as if it were an average, subtracting the wrong center, or looking up the opposite tail.

A dependable method is to identify the random quantity in the question first, find its mean and standard deviation, and only then standardize it. This tutorial focuses on those decisions. It assumes the CLT is an appropriate approximation; it does not make a normal approximation exact unless the underlying distribution already gives one.

Definition: If a random quantity $W$ has mean $\mu_W$ and standard deviation $\sigma_W>0$, its standardized value is $Z_W=(W-\mu_W)/\sigma_W$. Under a suitable CLT approximation, $Z_W$ is approximated by a standard normal random variable $Z\sim N(0,1)$. The subtraction centers the quantity, and division expresses its distance from the center in standard-deviation units.

Set the Center and Scale for the Quantity Asked About

Suppose $X_1,\ldots,X_n$ are independent and identically distributed (iid) random variables with finite mean $\mu$ and finite, positive variance $\sigma^2$. Write $S_n=\sum_{i=1}^n X_i$ for their sum and $\bar X=S_n/n$ for their sample mean. Their centers and scales are different:

E[S_n]=n\mu,\qquad \operatorname{SD}(S_n)=\sigma\sqrt n; \qquad E[\bar X]=\mu,\qquad \operatorname{SD}(\bar X)=\frac{\sigma}{\sqrt n}.

Here $\sigma$ is the population standard deviation of one observation. The standard deviation of the sample mean, $\sigma/\sqrt n$, is its standard error when $\sigma$ is known. It is not the same as $\sigma$. Similarly, the sum's standard deviation is $\sigma\sqrt n$, not $\sigma/\sqrt n$.

Theorem: Under the iid assumptions above, as $n$ grows, the CLT gives the approximations $$ \frac{S_n-n\mu}{\sigma\sqrt n}\ \approx\ N(0,1), \qquad \frac{\bar X-\mu}{\sigma/\sqrt n}\ \approx\ N(0,1). $$ These are two forms of the same standardization because $\bar X=S_n/n$. For finite $n$, the accuracy depends on the population distribution and sample size. If the population is normal, the standardized quantities are exactly standard normal for every $n$.

Why the Scales Have These Forms

Independence explains the square-root scaling. For independent observations with common variance $\sigma^2$, variances add when observations are summed. Dividing the sum by $n$ divides its variance by $n^2$. Taking square roots then gives the standard deviations.

\operatorname{Var}(S_n) =\sum_{i=1}^n\operatorname{Var}(X_i) =n\sigma^2, \qquad \operatorname{SD}(S_n)=\sqrt{n\sigma^2}=\sigma\sqrt n,

\operatorname{Var}(\bar X) =\operatorname{Var}\left(\frac{S_n}{n}\right) =\frac{1}{n^2}\operatorname{Var}(S_n) =\frac{\sigma^2}{n}, \qquad \operatorname{SD}(\bar X)=\frac{\sigma}{\sqrt n}.

The CLT uses these standard deviations to put the centered sum or mean on a scale that approaches a standard normal distribution. This variance calculation justifies the scaling; it does not by itself establish that the distribution is approximately normal.

Name the random quantity.
Decide whether the event concerns a sum, an average, or another statistic. Do not choose a standard deviation before doing this.

Find its center and scale.
For an iid sum use $n\mu$ and $\sigma\sqrt n$; for an iid sample mean use $\mu$ and $\sigma/\sqrt n$.

Standardize the event.
Subtract the appropriate center from every boundary and divide by the positive standard deviation. Preserve the direction of each inequality.

Read the normal probability.
Use the standard normal cumulative distribution function $\Phi(z)=P(Z\le z)$. A lower tail is $\Phi(z)$; an upper tail is $1-\Phi(z)$.

Worked Examples: Scale, Center, and Tail

Worked Example: Using the Standard Error for a Sample Mean

A hypothetical portfolio has 64 independent policies. The annual claim amount $X_i$ for each policy has population mean $\mu=\$50$ and population standard deviation $\sigma=\$12$. Assume the policy amounts are identically distributed with finite variance and that a CLT approximation is reasonable. Approximate the probability that the average claim amount is between $\$47.75$ and $\$52.25$.

Method. The event is about the sample mean $\bar X$, so its center is $\mu=\$50$. Its standard error is not $\$12$; it is

\operatorname{SD}(\bar X)=\frac{\sigma}{\sqrt n} =\frac{12}{\sqrt{64}}=\frac{12}{8}=\$1.50.

Standardize each endpoint using this standard error:

z_L=\frac{47.75-50}{1.50}=\frac{-2.25}{1.50}=-1.50, \qquad z_U=\frac{52.25-50}{1.50}=\frac{2.25}{1.50}=1.50.

Thus the probability is approximated by the area between these standard normal values:

P(47.75<\bar X<52.25) \approx \Phi(1.50)-\Phi(-1.50) =0.9332-0.0668 =0.8664.

The calculation is checked: the interval is centered at the population mean, and its two standardized endpoints are equally far from zero. The approximate probability is therefore a central area. The common mistake is to divide by $\sigma=12$, which would give endpoints $-0.1875$ and $0.1875$ and an incorrectly small central probability. The population standard deviation describes one policy; the standard error describes the average of 64 policies.

Worked Example: Standardizing a Sum Instead of an Average

Suppose the number of service requests received on each of 100 independent days is identically distributed, with mean $\mu=8$ requests and standard deviation $\sigma=3$ requests per day. Assume the CLT is suitable. Approximate the probability that the total number of requests over the 100 days exceeds 830.

Method. The quantity is the total $S_{100}$, not the daily average. Its mean and standard deviation are

E[S_{100}]=100(8)=800, \qquad \operatorname{SD}(S_{100})=3\sqrt{100}=3(10)=30.

The standardized threshold is the threshold minus the sum's mean, divided by the sum's standard deviation:

z=\frac{830-800}{30}=\frac{30}{30}=1.

The event is an upper tail, so

P(S_{100}>830) \approx P(Z>1) =1-\Phi(1) =1-0.8413 =0.1587.

A check using the daily average gives the same result. The event $S_{100}>830$ is equivalent to $\bar X>8.3$. The mean's standard error is $3/\sqrt{100}=0.3$, so $(8.3-8)/0.3=1$. Both formulations produce the same upper-tail probability.

The common centering error is to calculate $830/30$, omitting the sum's expected value of 800. That does not measure the threshold's distance from the center and is not the CLT standardization. A second common error is to use 0.3 as the scale for the sum; 0.3 is the standard error of the average, not the standard deviation of the total.

Worked Example: Choosing the Lower Tail

A hypothetical insurer models claim amounts $X_1,\ldots,X_{36}$ from independent policies as identically distributed, with mean $\mu=\$40$ and standard deviation $\sigma=\$5$. Assume a CLT approximation is reasonable. Approximate the probability that aggregate claims are less than $\$1{,}410$. The claim amounts are treated as continuous, so no continuity correction is involved.

Method. Define $S_{36}=\sum_{i=1}^{36}X_i$. Its mean is $36(40)=\$1{,}440$, and its standard deviation is $5\sqrt{36}=5(6)=\$30$. The standardized threshold is

z=\frac{1410-1440}{30}=\frac{-30}{30}=-1.

The event is $S_{36}<1410$, a lower-tail event. Therefore,

P(S_{36}<1410) \approx P(Z<-1) =\Phi(-1) =0.1587.

The sign is a useful check: $\$1{,}410$ is below the expected total, so the standardized value must be negative and the requested probability should be below one-half. Using $1-\Phi(-1)=0.8413$ would compute the probability of exceeding the threshold, the opposite event. The approximate lower-tail probability is 0.1587.

Common Mistakes and Their Corrections

Before using a normal table or calculator, compare the expression you have written with the event in the problem. The following checks catch many errors quickly.

Using the population standard deviation for a sample mean. For $\bar X$, use $\sigma/\sqrt n$, not $\sigma$. For $S_n$, use $\sigma\sqrt n$, not $\sigma$.
Mixing the sum's center with the mean's scale. A sum is centered at $n\mu$; an average is centered at $\mu$. The chosen center and scale must belong to the same random quantity.
Forgetting to center. Standardization is $(\text{boundary}-\text{mean})/(\text{standard deviation})$, not boundary divided by standard deviation. A $z$-value measures distance from the expected value.
Reversing a tail. If the standardized boundary is $z$, then $P(W\le c)$ is approximated by $\Phi(z)$, while $P(W>c)$ is approximated by $1-\Phi(z)$. A negative $z$ does not, by itself, mean “take the complement.”
Treating a CLT approximation as exact. The mean and variance determine the standardization, but a normal approximation also requires suitable CLT conditions and enough accuracy for the population shape and sample size.

The same logic extends to independent, nonidentically distributed observations when their means and variances can be added and the relevant CLT conditions hold: center a sum at the sum of its means and use the square root of the sum of its variances as its standard deviation. Do not use the iid formulas $n\mu$ and $\sigma\sqrt n$ unless the common-mean and common-variance assumptions apply.

Key takeaway: Write the random quantity first, then write its mean and standard deviation, then standardize the event, and finally identify the requested tail. For an iid sample mean, the scale is $\sigma/\sqrt n$; for an iid sum, it is $\sigma\sqrt n$.

Check Your Understanding

For each problem, identify the random quantity and use the stated CLT approximation where appropriate. Give probabilities to four decimal places.

Independent measurements are identically distributed with mean 20 and standard deviation 5. For a sample of 25, approximate $P(\bar X>22)$. (Answer: 0.0228)
Independent daily counts are identically distributed with mean 4 and standard deviation 3. For 36 days, approximate the probability that the total is less than 150. (Answer: 0.6306)
Independent observations are identically distributed with mean 12 and variance 16. For $n=64$, approximate $P(11.5<\bar X<12.5)$. (Answer: 0.6827)
Independent daily totals are identically distributed with mean 3 and variance 4. Over 100 days, approximate the probability that the sum exceeds 320. (Answer: 0.1587)
Independent policy payments are identically distributed with mean $\$100$ and variance $\$400$. For 25 policies, approximate the probability that aggregate payments are below $\$2{,}400$. (Answer: 0.1587)

What You'll Learn