Central Limit Theorem vs Sampling Distribution

The central limit theorem is a statement about one object: the sampling distribution of $\bar{x}$ (x-bar), the distribution you would get by taking every random sample of size $n$ from a population and recording each sample's mean. For independent observations from a population with a finite mean $\mu$ (mu) and a finite standard deviation $\sigma$ (sigma), that distribution gets closer to normal as $n$ grows, centered at $\mu$ with standard deviation $\frac{\sigma}{\sqrt{n}}$.

Two sentences come up constantly and both are wrong: "with a large enough sample, the data become normal," and "a large sample makes the population normal." Nothing about the population or the sample changes shape. A right-skewed population stays right-skewed however much of it you collect, and a histogram of your 200 observations is a picture of that population, not of $\bar{x}$. The normal shape belongs to a distribution you never see in a single study.

The familiar $n \ge 30$ is a rule of thumb, not part of the theorem, and it guarantees nothing. Draw from a strongly right-skewed population (an exponential, $\mu = 1$, $\sigma = 1$) and build the usual 95% t-interval for the mean. Across two million simulated samples the interval captured $\mu$ 88.3% of the time at $n = 5$, 91.2% at $n = 15$, 92.7% at $n = 30$, 94.2% at $n = 100$, and 94.8% at $n = 500$.

The misses are lopsided too. At $n = 30$, 6.5 of those 7.3 missed percentage points were intervals sitting entirely below $\mu$. A sample that draws none of the long right tail has both a low mean and a small $s$, so it produces a short interval in the wrong place. Skew does not only widen the error, it aims it.

Two edges. The population needs a finite standard deviation. And if the population is already normal, the sampling distribution of $\bar{x}$ is exactly normal at every $n$, so the theorem is not needed. Topic 2.12 is Sampling Distributions and the Central Limit Theorem.

Three different distributions get called "the distribution" and only one of them is this one. The population distribution holds a value for every individual. The distribution of a sample holds the values in the one sample you actually collected. A sampling distribution is neither: its individuals are entire samples of a fixed size $n$, and the number recorded for each is a statistic computed from it, such as $\bar{x}$ (x-bar) or $\hat{p}$ (p-hat).

Small cases can be written out in full. Take the population 2, 4, 6, 8, with $\mu = 5$ (mu) and $\sigma = 2.2361$ (sigma), and draw samples of size 2 with replacement. There are exactly 16 such samples, and their means come out as 2 once, 3 twice, 4 three times, 5 four times, 6 three times, 7 twice, and 8 once. That list is the sampling distribution of $\bar{x}$. Its mean is 5, matching $\mu$, and its standard deviation is 1.5811, which is $\sigma/\sqrt{2}$. The population is flat; the sampling distribution is already triangular at $n = 2$.

The misreading is blunt: "the sampling distribution is the distribution of my sample." A histogram of the 40 numbers you collected is a picture of the sample. As $n$ grows it comes to look more like the population, skew and all, and it does not narrow. The sampling distribution is the thing that narrows, and $\sigma/\sqrt{n}$ describes it rather than your data.

Outside enumerable toy cases you never actually build one. Theory stands in for it, the central limit theorem for means and the binomial for counts, or a simulation approximates it: 10,000 simulated sample means draw a close picture of the sampling distribution without being it.

Topic 2.12 introduces sampling distributions. Every confidence interval and every p-value later in the course is a statement read off one.

One sampling distribution gets reported on the news every month. The US unemployment rate is not a census, it is an estimate from a household survey of about sixty thousand homes, which is why it is revised and why it carries a margin of error at all: the unemployment rate.