Central Limit Theorem: Worked Examples and Simulation

The Central Limit Theorem says that as sample size n grows, the sampling distribution of the sample mean x̄ approaches a normal distribution with mean μ (the population mean) and standard deviation σ/√n (the standard error), regardless of the shape of the underlying population. CLT is the bridge that lets us use normal-based inferential procedures (z-tests, t-tests, confidence intervals) even when the underlying data is not normal. The conventional rule of thumb is n ≥ 30 for most distributions, but severely skewed populations may need n closer to 100; only the underlying shape determines exactly how fast convergence happens.

Consider an exponential population (heavily right-skewed) with rate parameter λ = 0.5, so μ = 2 and σ = 2. We draw samples of various sizes and compute the sample mean. At n = 5: Sampling distribution of x̄ has mean = 2 and SE = 2/√5 = 0.894. The shape is still noticeably right-skewed at this small sample size — CLT has not yet kicked in fully. At n = 30: Mean = 2, SE = 2/√30 = 0.365. Distribution is now much closer to normal but with slight residual right-skew. Most CLT-based procedures will work reasonably here. At n = 100: Mean = 2, SE = 2/√100 = 0.20. Distribution is essentially normal — exponential origin completely "forgotten" at the sample-mean level. Probability calculation. What is P(x̄ > 2.5) for n = 50? SE = 2/√50 = 0.283. Z = (2.5 − 2)/0.283 = 1.768. P(Z > 1.768) = 1 − 0.9615 = 0.039. About 3.9% chance of observing a sample mean above 2.5, even though individual values from the exponential distribution can easily exceed 2.5.

A uniform distribution between 0 and 1 has μ = 0.5 and σ² = 1/12, so σ ≈ 0.289. Uniform is symmetric (no skew) so CLT convergence is fast. At n = 5: SE = 0.289/√5 = 0.129. Sampling distribution is already approximately normal — the uniform parent has no skew to overcome. At n = 30: SE = 0.289/√30 = 0.053. At n = 100: SE = 0.289/√100 = 0.029. For symmetric populations, n = 5 to 10 is often sufficient for the CLT approximation to work well in practice. The fact that "n = 30" became the textbook rule of thumb reflects worst-case skewed populations, not all populations. Probability calculation. What is P(x̄ < 0.45) for n = 100? SE = 0.029. Z = (0.45 − 0.5)/0.029 = −1.724. P(Z < −1.724) = 0.0424. About 4.2% chance of observing a sample mean below 0.45, even though about 45% of individual uniform draws are below 0.45.

The CLT is what justifies the standard error formulas used in inferential statistics. For sample mean: SE(x̄) = σ/√n (CLT directly). For sample proportion p̂ from binary data: SE(p̂) = √(p(1−p)/n). This comes from CLT applied to the binomial: the sample proportion is a sample mean of 0/1 indicator variables. Approximation works when np > 10 and n(1−p) > 10 (the "success-failure condition"). For difference of two means: SE(x̄_1 − x̄_2) = √(σ²_1/n_1 + σ²_2/n_2). This is the formula behind two-sample t-tests; it follows from independence of the two samples and CLT applied to each. For regression coefficients: the CLT applied to linear combinations of residuals gives normal-approximate distributions for slope estimates, justifying t-tests on slopes. Without the CLT, none of these formulas would work for non-normal data. The reason a sample of n = 50 from a heavily skewed population still produces a valid t-test on the mean is precisely that CLT has converted the sample-mean distribution to approximately normal.

The bootstrap is a resampling technique for estimating sampling distributions when analytical formulas are unavailable or assumptions are violated. The procedure: resample n observations WITH replacement from the original sample, compute the statistic of interest, repeat thousands of times, and use the empirical distribution of bootstrap statistics as the sampling distribution. Why the bootstrap works: CLT applies to the bootstrap distribution of x̄ (and many other statistics) under mild conditions. The bootstrap distribution converges to the sampling distribution of the statistic, which by CLT is approximately normal for large n. This is why bootstrap confidence intervals match analytical confidence intervals when the analytical ones are valid. The bootstrap is especially useful when: - The statistic has no closed-form sampling distribution (e.g., median, trimmed mean, ratio of medians). - Sample size is small but the statistic is asymptotically normal. - Data has unusual structure (clustered, censored) that standard formulas do not handle. Limitation: the bootstrap does NOT rescue you from severely skewed small samples where CLT has not yet kicked in. If n is too small for analytical CLT-based inference, n is too small for bootstrap inference too.

CLT problems span every introductory and intermediate statistics course, the AP Statistics exam, and most applied research design discussions. The key skills are recognizing when CLT applies, computing the sampling-distribution parameters (mean μ, SE = σ/√n), and using them for probability or interval calculations. Snap a photo of any CLT problem and StatsIQ identifies the population parameters, computes SE, evaluates the n requirement, and produces the requested probability with the area under the sampling-distribution curve visualized. For multi-part problems involving sample-size planning or bootstrap inference, StatsIQ chains the steps. This content is for educational purposes only and does not constitute statistical advice.