A stakeholder points at a number on your slide and asks the only question that matters: is that real, or is it noise? Sign-ups went from 400 to 430 last week. Is that growth you can bank on, or the kind of wobble that would have happened anyway? Answering that, honestly and out loud, is what statistics buys you. Not a wall of formulas, just a small set of tools that tell you what a number means and how much you can trust it.
You do not need a degree in statistics to do good analyst work, and you will use maybe a dozen ideas again and again. This part walks through the ones that earn their keep: how to describe the middle of your data, how to describe its spread, how the normal curve makes both easier to reason about, and how a sample of a few hundred rows can speak, with a stated margin of error, for a population of millions. Everything here builds on the exploratory habits from Part 10, where you first met the mean, the median and the interquartile range.
Key takeaways
Statistics splits into two jobs: describing the data you have, and using a sample to say something about the population you cannot see.
To describe a column, name three things every time: its centre, its spread, and its shape. One number for the middle is never enough on its own.
Every estimate from a sample carries uncertainty. A grown-up answer is a range with a confidence level, not a single point pretending to be exact.
Two jobs statistics does for you
Statistics comes in two flavours, and knowing which one you are doing keeps you out of trouble. Descriptive statistics summarise the data in front of you. If you have every order your shop took last month, and you compute the average order value, that average is a fact about those orders, complete and certain. You are describing, not guessing. The mean, the median, the standard deviation and the percentiles from Part 10 are all descriptive; they compress a table into a few honest numbers.
Inferential statistics do the harder thing. You rarely have all the data. You have a sample, a few thousand survey replies out of a million customers, and you want to say something about the whole population, the full group you care about but cannot measure. Inference is the set of tools that lets a sample speak for the population while being honest about how wrong it might be. The moment you say most of our customers prefer the new layout based on a survey of two thousand, you have crossed from describing into inferring, and you owe your reader a margin of error. Most analyst mistakes come from treating an inferential claim as if it were a descriptive fact.
Which average should you report?
People say average as if there is only one, but you have three ways to describe the middle of a column, and picking the wrong one quietly misleads. The mean is the arithmetic average: add everything up and divide by the count. The median is the middle value when you line the numbers up in order, so half sit below and half above. The mode is simply the value that shows up most often, and it is the only one of the three that works on words as well as numbers, which is why you use it for things like the most common signup city.
The choice matters most when the data is skewed, which in business data it almost always is. The mean is dragged toward extreme values; a single whale customer or one data-entry error can pull it far from where most of the data sits. The median shrugs those off. If the mean and the median are close, your column is roughly symmetric and either is fine. If they diverge, that gap is itself a finding, and the median is usually the more honest headline. The table below shows the three measures side by side and when each earns its place.
| Measure | What it answers | Reacts to outliers? | Reach for it when |
|---|---|---|---|
| Mean | The arithmetic centre of gravity | Strongly, pulled by extremes | Data is symmetric, no wild outliers |
| Median | The typical, middle value | Barely, it is resistant | Money, counts, anything skewed |
| Mode | The most common value | Not applicable | Categories, or a clear peak |
Three measures of the middle. On a right-skewed spend column where mean is 1,080 and median is 980, the median is the number you put on the slide.
How spread out is the data?
A middle value on its own can fool you. Two teams can both average 50,000 rupees in monthly sales, yet one team is steady around that figure while the other swings from 10,000 to 90,000. Same centre, completely different story, and the difference is spread. The simplest measure of spread is the range, the largest value minus the smallest, but one freak value blows it up, so it is a blunt instrument. The interquartile range from Part 10, the middle-half width, is far steadier.
The measure you will quote most is the standard deviation. It answers a plain question: on average, how far do the values sit from the mean? A small standard deviation means the data huddles close to its centre; a large one means it is scattered wide. Variance is its close cousin, the standard deviation squared, and you will see it in formulas, but for talking to humans the standard deviation wins because it is in the same units as the data, rupees not rupees-squared. The two curves below share an identical mean and differ only in standard deviation, which is exactly the thing a single average would hide from you.
Gotcha
Do not confuse the standard deviation with the standard error; they look alike and mean opposite things. The standard deviation measures how spread out the individual values are, and it does not shrink as you collect more data. The standard error measures how precise your estimate of the mean is, and it does shrink as the sample grows. Report the standard deviation to describe your data. Report the standard error, or the margin of error built from it, when you are estimating something about a population you did not fully measure.
The normal curve and the 68, 95, 99.7 rule
Many natural and business measurements pile up in a familiar bell shape called the normal distribution: most values cluster near the mean, and they thin out symmetrically on both sides. It matters to you not because your data is always normal, it often is not, but because the normal curve comes with a rule of thumb so handy you will use it weekly. Once you know the mean and the standard deviation of a roughly normal column, you know where almost everything lives.
The rule is called the 68, 95, 99.7 rule, or the empirical rule. About 68 percent of the values fall within one standard deviation of the mean, about 95 percent within two, and about 99.7 percent within three. So if delivery times average 30 minutes with a standard deviation of 5, then roughly 95 percent of orders arrive between 20 and 40 minutes, and an order taking 47 minutes is beyond three standard deviations, a genuine rarity worth investigating. The chart below draws those bands.
Percentiles and the z-score
Two tools let you place a single value in context, and both are worth having at your fingertips. A percentile tells you the share of values that fall below a point. If a customer sits at the 90th percentile of spend, they spend more than 90 percent of customers, a phrasing anyone in the room understands instantly. Percentiles are the honest way to talk about position in skewed data, which is why you see the 50th percentile, the median, and the 25th and 75th, the quartiles, everywhere in analyst work.
The z-score answers a different question: how many standard deviations is this value from the mean? A z-score of 0 sits exactly on the mean, a z-score of plus 2 is two standard deviations above it, and a z-score of minus 1.5 is one and a half below. Because it strips away the original units, the z-score lets you compare across different columns; a customer who is 2 standard deviations above average on spend and 3 above on visits is more unusual on visits, even if the raw numbers are not comparable. Standardising this way is the same move that powers the outlier rules and, later, many machine learning steps you will meet down the road.
How a sample speaks for a population
Here is the idea that makes surveys and experiments possible. You cannot ask all ten million customers, so you ask a random sample of a few thousand and use their answer to estimate the whole. The catch is that a different random sample would give a slightly different answer. The standard error measures exactly that wobble: how much your sample estimate would jump around if you drew the sample again and again. A smaller standard error means a more precise estimate, and the single biggest lever on it is sample size.
The crucial and slightly unfair truth is that precision improves with the square root of the sample size, not in step with it. To halve your margin of error you need roughly four times the data, not twice. That is why a national poll of a couple of thousand people can carry a margin of a few points, and why chasing a tiny bit more precision gets expensive fast. It is also why one clean random sample beats a huge but biased one; size cannot fix a sample that systematically leaves people out. The flow below is the decision I walk through before quoting any sample figure.
Confidence intervals and the margin of error
A confidence interval is how you report an estimate honestly. Instead of claiming 43 percent of users prefer the new design, you say 43 percent, with a 95 percent confidence interval of 40 to 46 percent. The half-width of that range, the 3 points, is the margin of error, and the 95 percent is the confidence level, the long-run share of such intervals that would contain the true value. The standard reading is plain: your best estimate is 43 percent, and the true figure is very probably between 40 and 46.
People trip on one point, so say it clearly: the 95 percent describes the method, not this one interval. It means that if you repeated the whole sampling process many times, about 95 in 100 of the intervals you built would capture the true value. In day to day work you can read it as high confidence the truth sits in this range, but never round it up to certainty. The table below shows how the margin of error tightens as the sample grows, for a survey where the true split is near 50 percent, the hardest case to pin down.
| Sample size | Approx margin of error (95%) | What it means |
|---|---|---|
| 100 | about 10 points | Only good for a rough read |
| 400 | about 5 points | Usable for many decisions |
| 1,000 | about 3 points | The typical national poll |
| 4,000 | about 1.5 points | Four times the data for half the margin |
Margin of error at 95 percent confidence for a proportion near 50 percent. Notice you need 4,000 responses to reach half the margin of 1,000, the square-root law in action.
What statistically significant really means
Come back to the sign-ups that rose from 400 to 430. Statistical significance is the tool that answers whether that jump is more than the data would wobble by chance. The logic runs backwards, which trips everyone at first. You start by assuming nothing really changed, the null hypothesis, then ask how surprising your result would be if that were true. That surprise is measured by the p-value: the probability of seeing a jump at least this big when nothing actually changed. A small p-value, by convention below 0.05, means the result is unlikely to be a fluke, so you call it statistically significant.
Two cautions save you from the classic misreads. First, statistically significant does not mean important. With a huge sample, a trivial difference of 0.1 percent can be significant yet not worth a meeting. Always look at the size of the effect, not just the p-value. Second, a result that is not significant is not proof that nothing happened; it may just mean your sample was too small to tell. Significance is one input to a decision, never the whole verdict. You will meet all of this properly when we design experiments in Part 18; for now, knowing what the phrase does and does not promise is enough to keep you honest in a meeting.
Report a range, name the uncertainty
If you carry one habit out of this part, make it this: never hand over a single number from a sample without saying how sure you are of it. Describe your data with a centre, a spread and a shape; the median and the interquartile range will serve you honestly nine times out of ten. When you cross into estimating something about a population, attach a confidence interval and say the confidence level out loud. And when someone asks whether a change is real, look at both the p-value and the size of the effect before you answer. Those few moves are the difference between analysis that holds up and a number that falls apart the moment someone pushes on it.
You now have the working statistics an analyst leans on daily: two averages and a mode, the standard deviation for spread, the normal curve and its 68, 95, 99.7 rule, percentiles and z-scores for context, and the sampling ideas behind margins of error and significance. Part 12 takes the natural next step and separates correlation from causation, the trap that catches even careful analysts when two columns move together. Keep the Data Analyst guide open as your map, and revisit Part 10 whenever a distribution surprises you.
This week, take one number you have already reported and redo it as a range. Find its confidence interval, or if it came from a sample, estimate its margin of error from the table above, and rewrite the sentence with the uncertainty stated. It will feel less tidy and far more true, and that is the shift from quoting numbers to doing statistics.
References
- Margin of Error: Formula and Interpreting, Statistics By Jim
- Uncertainty and How We Measure It, Office for National Statistics
- The 68, 95, 99.7 rule, Wikipedia
About The Author
DrJha