Standardizing Sums and Sample Means

Why Standardization Matters

A sum of many independent observations can have a distribution that is difficult to calculate directly. The central limit theorem (CLT) makes many such probability questions manageable: after centering and scaling, the sum or sample mean is approximately standard normal. The key is to use the correct mean and standard deviation for the quantity in the question.

A frequent source of error is switching between a sum and an average without changing the scale. A sum becomes more variable as the sample size grows, while the sample mean becomes less variable. Both use the same underlying observations, but their standard deviations differ by a factor of $n$.

Definition: Let $X_1,\ldots,X_n$ be independent and identically distributed (IID) random variables with $E[X_i]=\mu$ and $0<\operatorname{Var}(X_i)=\sigma^2<\infty$. Define the sum $S_n=\sum_{i=1}^n X_i$ and the sample mean $\bar X=S_n/n$. Their means and standard deviations are $$ E[S_n]=n\mu,\qquad \operatorname{SD}(S_n)=\sigma\sqrt n, $$ $$ E[\bar X]=\mu,\qquad \operatorname{SD}(\bar X)=\frac{\sigma}{\sqrt n}. $$ For sufficiently large $n$, the CLT approximates the standardized quantities by a standard normal variable $Z\sim N(0,1)$: $$ \frac{S_n-n\mu}{\sigma\sqrt n}\ \approx\ Z,\qquad \frac{\bar X-\mu}{\sigma/\sqrt n}\ \approx\ Z. $$

The hypotheses matter. The formulas for the mean and variance above follow from independence and identical distribution, and the CLT approximation requires a finite, positive variance under these IID conditions. If the observations are dependent, have different distributions, or have infinite variance, these formulas or this CLT justification may not apply as stated. If each $X_i$ is itself normally distributed, then the standardized quantities have standard normal distributions exactly for every $n$; otherwise the normal result is generally an approximation.

Choosing the Correct Standardization

Use the version that matches the random quantity in the event. For a sum, subtract $n\mu$ and divide by $\sigma\sqrt n$. For an average, subtract $\mu$ and divide by $\sigma/\sqrt n$. In both cases, a boundary value is converted to a standard normal score, often called a $z$-score.

z_{\text{sum}}=\frac{s-n\mu}{\sigma\sqrt n}, \qquad z_{\text{mean}}=\frac{\bar x-\mu}{\sigma/\sqrt n}.

Once the event boundaries have been standardized, use the standard normal cumulative distribution function $\Phi(z)=P(Z\le z)$. For example, an interval probability is found by subtracting the lower-tail probability at its lower $z$-score from the lower-tail probability at its upper $z$-score. A strict or non-strict inequality gives the same result for a continuous normal approximation.

Worked Example: Standardizing an Aggregate Sum

A hypothetical service operation records a daily workload $X_i$, measured in workload units. Assume the daily values are IID with mean $\mu=160$ units and variance $\sigma^2=3{,}600$ units squared. Approximate the probability that the total workload over $n=36$ days is between 5,400 and 6,120 units. No interest rate or mortality assumption is involved; the model specifies the distribution through its IID assumption, mean, and variance.

Method and model parameters. The quantity is a sum, so use its mean $n\mu$ and standard deviation $\sigma\sqrt n$. The individual standard deviation is $\sigma=\sqrt{3{,}600}=60$ units.

Sum mean and standard deviation. $$ E[S_{36}]=36(160)=5{,}760,\qquad \operatorname{SD}(S_{36})=60\sqrt{36}=60(6)=360. $$

Standardize both endpoints. $$ z_{\text{lower}}=\frac{5{,}400-5{,}760}{360} =\frac{-360}{360}=-1, \qquad z_{\text{upper}}=\frac{6{,}120-5{,}760}{360} =\frac{360}{360}=1. $$

Find the probability. By the CLT, $$ P(5{,}400<S_{36}<6{,}120) \approx P(-1<Z<1) =\Phi(1)-\Phi(-1) =0.8413447-0.1586553 =0.6826894. $$ To four decimal places, the approximate probability is 0.6827.

Check and trap avoided. The endpoints are one sum standard deviation below and above the sum mean, so a probability near 0.68 is reasonable. Dividing by $60$, the standard deviation of one day, would be the wrong scaling for a 36-day sum.

Worked Example: Standardizing a Sample Mean

Suppose a hypothetical group of $n=49$ independent account records has IID values $X_i$, with mean $\mu=18$ units and variance $\sigma^2=49$ units squared. Approximate the probability that the sample mean is between 16.5 and 19 units. There is no interest or mortality assumption in this statistical model.

Method and model parameters. The event concerns $\bar X$, not $S_n$. The observation standard deviation is $\sigma=\sqrt{49}=7$, so the standard error of the sample mean is $$ \operatorname{SD}(\bar X)=\frac{7}{\sqrt{49}}=\frac{7}{7}=1. $$

Standardize the lower endpoint. $$ z_{\text{lower}}=\frac{16.5-18}{1}=-1.5. $$ Standardize the upper endpoint. $$ z_{\text{upper}}=\frac{19-18}{1}=1. $$

Use the standard normal distribution. $$ P(16.5<\bar X<19) \approx \Phi(1)-\Phi(-1.5) =0.8413447-0.0668072 =0.7745375. $$ Thus the probability is approximately 0.7745 to four decimal places.

Check and trap avoided. The mean is 18 and the standard error is 1, so the standardized endpoints must be $-1.5$ and $1$, respectively. Using $7\sqrt{49}=49$ as the standard deviation would confuse the sample mean with the sum and produce scores that are much too close to zero.

Why the Two Formulas Agree

The sum and mean standardizations are not separate rules. They are equivalent because $\bar X=S_n/n$. The variance calculation explains the sample-size scaling: independence makes the variance of a sum equal to the sum of the individual variances.

Find the sum variance.
Independence gives $\operatorname{Var}(S_n)=\sum_{i=1}^n\operatorname{Var}(X_i)=n\sigma^2$, so $\operatorname{SD}(S_n)=\sigma\sqrt n$.

Scale to obtain the mean.
Since $\bar X=S_n/n$, its mean is $n\mu/n=\mu$, and its variance is $n\sigma^2/n^2=\sigma^2/n$. Hence its standard deviation is $\sigma/\sqrt n$.

Compare the standardized forms.
For an endpoint $\bar x$, the sum endpoint is $n\bar x$. Therefore $\frac{n\bar x-n\mu}{\sigma\sqrt n}=\frac{\bar x-\mu}{\sigma/\sqrt n}$. The two methods give the same $z$-score.

This equivalence is also a useful check on a calculation. If the event is stated in terms of an average, it can be standardized directly using the standard error, or it can be rewritten as an event about the sum and standardized there. The resulting normal scores must match.

Worked Example: An Asymmetric Interval for a Sample Mean

A hypothetical measurement process produces IID observations with mean $\mu=30$ units and variance $\sigma^2=64$ units squared. For $n=64$, approximate the probability that the sample mean lies between 28.5 and 32 units. No interest rate or mortality assumption is relevant to this model.

Method and model parameters. Since the target is the sample mean, first calculate its standard error: $$ \sigma=\sqrt{64}=8,\qquad \operatorname{SD}(\bar X)=\frac{8}{\sqrt{64}}=\frac{8}{8}=1. $$

Convert each boundary to a $z$-score. $$ z_{\text{lower}}=\frac{28.5-30}{1}=-1.5,\qquad z_{\text{upper}}=\frac{32-30}{1}=2. $$ The interval is not symmetric about 30: the lower endpoint is 1.5 units below the mean and the upper endpoint is 2 units above it.

Calculate the interval probability. $$ P(28.5<\bar X<32) \approx \Phi(2)-\Phi(-1.5) =0.9772499-0.0668072 =0.9104427. $$ To four decimal places, the approximation is 0.9104.

Check and trap avoided. The lower and upper scores are unequal in magnitude, so this is not a symmetric interval probability such as $P(-z<Z<z)$. Subtracting the two CDF values gives the required interval probability directly.

Common Standardization Errors

The most reliable way to avoid a scaling error is to name the target random variable before writing a $z$-score. Then write its mean and standard deviation explicitly. For IID observations, these are $n\mu$ and $\sigma\sqrt n$ for the sum, and $\mu$ and $\sigma/\sqrt n$ for the sample mean.

Using $\sigma$ for a sum. The sum contains $n$ independent contributions; its standard deviation is $\sigma\sqrt n$, not $\sigma$.
Using $\sigma\sqrt n$ for a mean. Averaging divides the sum by $n$, so the sample mean standard deviation is $\sigma/\sqrt n$.
Confusing variance and standard deviation. If the problem gives $\sigma^2$, take its square root to obtain $\sigma$ before applying either standard deviation formula.
Centering at the wrong mean. A sum is centered at $n\mu$; a sample mean is centered at $\mu$.
Subtracting rounded normal probabilities too early. Keep sufficient digits in the CDF values, subtract them, and round the final probability. For instance, $0.9772499-0.1586553=0.8185946$, which rounds to 0.8186, not 0.8185.

Key takeaway: For an IID sample with mean $\mu$ and variance $\sigma^2$, standardize a sum with center $n\mu$ and scale $\sigma\sqrt n$; standardize a sample mean with center $\mu$ and scale $\sigma/\sqrt n$. Then use the standard normal CDF to approximate the requested probability.

These translations are useful whenever a question describes a total accumulated across observations or an average across a sample. They keep the role of sample size explicit and make it easier to check whether a probability calculation is on the right scale. The normal approximation is aligned with the SOA VEE Mathematical Statistics syllabus; its accuracy depends on the model and sample size, rather than on standardization alone.

Check Your Understanding

For each problem, identify whether the target is a sum or a sample mean, standardize using the appropriate scale, and then check the stated answer.

A hypothetical batch of $n=49$ IID processing times has mean 20 minutes and variance 49 minutes squared. Use the CLT to approximate the probability that their total is between 931 and 1,029 minutes. (Answer: 0.6827.)
IID observations have mean 8 and variance 36. For $n=36$, approximate the probability that the sample mean is between 7 and 9. (Answer: 0.6827.)
IID account amounts have mean 4 units and variance 9 units squared. For $n=100$, approximate the probability that their sum exceeds 430 units. (Answer: 0.1587.)
A hypothetical measurement sample has $n=64$, with IID observations of mean 15 and variance 16. Approximate the probability that the sample mean is between 14.25 and 15.5. (Answer: 0.7745.)
IID workload values have mean 100 units and variance 400 units squared. For $n=25$, approximate the probability that the total is between 2,400 and 2,700 units. (Answer: 0.8186.)

What You'll Learn