Central Limit Theorem

Central Limit Theorem (CSSGB BoK III.B.2 — Understand)

Quick Answer: The Central Limit Theorem (CLT) says that averages of random samples tend toward a normal distribution as the sample size increases, even when individual observations come from a skewed or non-normal population (with finite variance). That is why Green Belts can use normal-theory confidence intervals, many mean tests, and X-bar control limits on subgroup averages.

Statement of the CLT

Consider independent observations X₁, X₂, …, Xₙ from a population with mean μ and finite variance σ². Let the sample mean be

X̄ = (X₁ + … + Xₙ) / n

Then, as n → ∞, the distribution of X̄ approaches a Normal(μ, σ²/n) distribution. Equivalently, the standardized mean

Z = (X̄ − μ) / (σ/√n)

approaches a standard normal distribution.

What the CLT does not say:

• It does not require the raw data to be normal for the theorem to apply (normality of X̄ is the conclusion for large n).
• It does not fix a single magic n for every population; highly skewed or heavy-tailed processes need larger n before the normal approximation is comfortable.
• It does not make a single sample’s histogram look normal; it describes the sampling distribution of the mean across many hypothetical samples of size n.

Rule-of-thumb (practical, not a law): many quality texts treat n ≥ 30 as a common threshold for relying on CLT for means when the population is not extremely irregular; if the population is already roughly normal, smaller n can suffice. Always pair the rule of thumb with residual checks and process knowledge.

Mean and Standard Error

Regardless of normality, for random samples:

• Expected value: E(X̄) = μ (X̄ is unbiased for μ)
• Standard error of the mean: σ_X̄ = σ / √n

When σ is unknown, replace σ with the sample standard deviation s to get s/√n. The √n in the denominator is the statistical reason larger samples give tighter estimates: quadrupling n halves the standard error.

Worked example — standard error: A filling process has σ = 2.0 g. Samples of size n = 16 have

• σ_X̄ = 2.0 / √16 = 2.0 / 4 = 0.50 g

Samples of size n = 64 have σ_X̄ = 2.0 / 8 = 0.25 g. Same process variation; the mean of a larger sample is less variable.

Worked example — CLT probability: Suppose fill weights have μ = 500 g and σ = 2 g (not necessarily normal). For n = 36, X̄ is approximately Normal(500, 2/√36) = Normal(500, 1/3).

Approximate P(X̄ > 500.5):

• Z = (500.5 − 500) / (1/3) = 0.5 × 3 = 1.5
• From standard normal tables, P(Z > 1.5) ≈ 0.0668

So there is roughly a 6.7% chance a 36-unit sample mean exceeds 500.5 g under this model—even without assuming each bottle is normal—because the CLT approximates the distribution of the mean.

CLT and Confidence Intervals

A large-sample (1 − α)100% confidence interval for μ when σ is known is

X̄ ± z_{α/2} · (σ/√n)

When σ is unknown and data are approximately normal (or n is large enough for CLT + t approximation),

X̄ ± t_{α/2, n−1} · (s/√n)

Why CLT matters: the interval’s coverage probability is derived from the sampling distribution of X̄ (or of the t-statistic). If means were wildly non-normal for your n, the stated 95% confidence would not be trustworthy. Green Belts should remember:

• Confidence attaches to the method over repeated samples, not to a mystical probability that “μ sits inside this one interval” after seeing the data in Bayesian language.
• Narrower intervals need larger n or smaller σ—not wishful z-values.

Worked CI sketch: n = 40, X̄ = 12.0, s = 2.5, approximate 95% interval using z ≈ 1.96 (large-n):

• SE ≈ 2.5 / √40 ≈ 2.5 / 6.325 ≈ 0.395
• Margin ≈ 1.96 × 0.395 ≈ 0.77
• Interval ≈ (11.23, 12.77)

(For exam precision with smaller n, use t critical values; the CLT still motivates treating the mean as roughly normal.)

CLT and Hypothesis Testing

Many CSSGB-relevant tests for means (one-sample z or t, two-sample t under standard assumptions) use test statistics of the form

(statistic − hypothesized value) / standard error

Under H₀, that ratio is referred to a normal or t reference distribution. The CLT (plus measurement assumptions: independence, constant variance, adequate n) is the bridge from “I collected a sample mean” to “I can compute a p-value.”

Implications for Green Belts:

• Tests on means are more robust to non-normal individuals than tests that assume normality of every observation without averaging—especially as n grows.
• Tests on variances or highly skewed individuals (or rare defectives with tiny n) do not get a free pass from the CLT for free; choose methods matched to the data type (later BoK topics).
• Sample size planning uses the same SE = σ/√n logic: detecting a small shift in μ requires larger n because the sampling distribution of X̄ is tighter only when n is large enough relative to the shift.

CLT and Control Charts

X-bar charts plot subgroup averages. Even if individual measurements are skewed, subgroup means of size n (often n = 4 or 5 in manufacturing) are closer to normal than the individuals—this is the CLT in SPC clothing. Control limits for an X-bar chart are commonly

CL = X̄̄
UCL / LCL = X̄̄ ± A₂R̄ (or using estimated σ_X̄)

which assume approximate normality of X̄ for false-alarm probabilities to match the usual 3-sigma design.

Individuals (I-MR) charts plot single observations. They rely more heavily on approximate normality of the individuals (or on careful interpretation), because there is no within-subgroup average for the CLT to smooth. That is one reason practitioners prefer X-bar charts when rational subgrouping is possible.

Practical SPC link:

Putting CLT to Work in DMAIC

Measure: Use averages and standard errors when reporting baseline means; do not treat one observation as if it had the same precision as a mean of 50.

Analyze: When testing whether a process change shifted μ, the test’s legitimacy often rests on CLT + design assumptions—not on a perfectly bell-shaped histogram of n = 8 points.

Control: Prefer rational subgroups and X-bar logic when you need stable false-alarm behavior; investigate capability of individuals separately from control of means.

Summary Pitfalls

1. Claiming “CLT says my population is normal.”
2. Using n = 2 averages and assuming perfect normality without process knowledge.
3. Confusing the standard deviation of individuals (σ) with the standard error of the mean (σ/√n).
4. Applying mean-based CLT comfort to every statistic (medians, variances, and rare events need their own care).
5. Forgetting independence: automated oversampling of autocorrelated process data can violate the i.i.d. setup behind textbook CLT applications.

If you remember only one sentence for BoK III.B.2: averages become approximately normal and less variable as n increases, and that single idea underwrites confidence intervals for means, many hypothesis tests, and X-bar control charts.