Why is the n = 30 rule of thumb for the Central Limit Theorem unreliable on skewed data?

## Where 'n = 30' came from The Central Limit Theorem is a limit statement. It says the standardised average converges to a standard normal as the sample size grows without bound; it contains no number 30, no number 100, and no threshold of any kind. The rule of thumb is a pedagogical convenience: for populations that are roughly symmetric and light-tailed — a uniform, a flat die roll, a mild mixture — the average is already close to bell-shaped by about thirty draws, so textbooks adopted the number and it stuck. It was never a theorem, and treating it as one is one of the most consequential everyday errors in applied statistics. ## What actually controls the speed The accuracy of the normal approximation to the distribution of an average is governed by how asymmetric the parent distribution is. The leading correction term to the normal approximation is proportional to `skewness / sqrt(n)`. Setting that error below a fixed tolerance and solving for n shows the required sample size grows with the **square** of the skewness. A widely quoted operational guideline, attributed to Cochran, is: `n > 25 * skewness^2` Run some numbers through it: - A symmetric distribution has skewness 0 — no skew penalty; the rule reduces to 'a modest n is fine'. - An exponential distribution has skewness exactly 2, giving `n > 100`. - A lognormal with log-scale parameter 1 has skewness above 6, giving `n > 900`. - Revenue-per-user distributions with a handful of whales routinely show sample skewness of 10 or more, pushing the requirement into the thousands. Kurtosis (heavy tails without asymmetry) matters too, entering at the next order, but skewness dominates in practice because business metrics are bounded below at zero and unbounded above. ## The revenue-per-user case Consider per-user revenue over a month. A large share of users spend nothing at all, giving a spike at zero. Most paying users spend a small amount. A tiny fraction spend hundreds or thousands of times the median. The distribution is a spike plus a long right tail. Average thirty such users and one of two things happens. Usually no whale is in the sample, and the average lands well below the true mean. Occasionally a whale is included, and the average jumps far above it. The resulting distribution of averages is not a bell: it is itself right-skewed, with a dense body below the true mean and a stretched upper tail. That is the CLT working correctly and slowly — it *will* become normal, just not at thirty. The practical damage is directional. Because the distribution of the average is right-skewed at small n, normal-curve reasoning misprices both tails: it understates how often the average lands moderately low and understates how far the average can travel high. Any tail probability read off a normal curve is then systematically wrong, and the error does not average out across repeated analyses — it biases the same way every time. ## How to check rather than assume The honest answer in an interview is that you diagnose rather than defer to a number. 1. **Look at the parent distribution.** Plot it. A visible spike-plus-tail shape immediately kills the n = 30 assumption. 2. **Compute the sample skewness** and put it through the `25 * skewness^2` guideline for an order-of-magnitude sample-size requirement. 3. **Check the tail's concentration.** If the top 1% of observations contribute a large share of the total, a small sample's average is effectively a lottery on whether that 1% appears. 4. **Resample.** Draw many samples of your actual n from the data you have and plot the resulting averages. If that histogram is visibly skewed, the approximation is not there yet at that n. ## Remedies When n cannot be increased far enough, the options are to change the quantity or change the method. Analysing a transformed scale (a log, for instance) tames the skew, but note that the mean of the logs is not the log of the mean, so the business question must genuinely be about the transformed quantity. Trimming or winsorising the extreme tail reduces skew at the cost of changing what you are estimating and needs an explicit, pre-registered rule. Reporting a different summary — a median or a trimmed mean — sidesteps the issue when the business question tolerates it. Alternatively, aggregate at a coarser unit, since averages of already-averaged groups are far better behaved. ## What the interviewer is listening for They want to hear that n = 30 is a heuristic and not part of the theorem, that skewness is the driver, that the required n scales roughly with skewness squared, and that you would inspect the data rather than assert a threshold. A candidate who says 'we had 30 samples so the CLT applies' has revealed that they learned a slogan rather than a theorem.