The Normal Distribution: Meet the Bell Curve
The normal distribution — the famous “bell curve” — describes how many real-world measurements naturally spread around an average. In this lesson you’ll learn to read the curve, use the 68–95–99.7 rule, and turn any measurement into a z-score you can interpret in seconds.
Learning objectives
By the end of this lesson you will be able to:
- Identify the key features of a normal distribution (symmetry, one peak, mean = median = mode).
- Explain what the mean (μ) and standard deviation (σ) control in a bell curve.
- Apply the 68–95–99.7 rule to estimate what proportion of data falls in a given range.
- Calculate a z-score for any measurement and interpret what it means.
- Select situations where the normal model is (and is not) a sensible assumption.
Prerequisites
You should be comfortable with the idea of an average (mean) and basic arithmetic (subtraction and division). That’s it. If you have met the standard deviation before, great — but we recap it below, so no prior statistics knowledge is required.
The idea in plain English
Measure the height of 1,000 adults and draw a bar chart of the results. You won’t get a random mess — you’ll get a very particular shape: most people are close to the average, and the further you move from the average, the fewer people you find. The result is a smooth, symmetric hill: the bell curve.
What it is: the normal (or Gaussian) distribution is a mathematical model for data that clusters symmetrically around an average. It is completely described by just two numbers: the mean (where the centre is) and the standard deviation (how spread out the data is).
Why it is used: it shows up everywhere. When a result is the sum of many small, independent influences — genetics plus diet plus environment for height, or dozens of tiny process variations for a machined part — the combined outcome tends to be bell-shaped. This is why heights, exam marks, measurement errors, IQ scores, and daily temperatures all look roughly normal.
What question it answers: “How common or unusual is a particular value?” If you know the mean and standard deviation, the curve tells you what fraction of results you should expect in any range.
Where it is used in practice: quality control limits, grading and performance comparisons, tolerance and specification analysis, forecasting error bands, and as the foundation of most classic statistical tests.
What it does not tell us: the normal curve is a model, not a law of nature. Plenty of real data is not normal — incomes, house prices, call-waiting times, and defect counts are usually lopsided (skewed). Assuming normality without checking is one of the most common mistakes in data analysis, and we’ll show you how to spot it.
Key terminology
| Term | Plain-language meaning | Technical meaning | Practical interpretation |
|---|---|---|---|
| Mean (μ) | The average. | The balance point of the distribution; for a normal curve it sits exactly at the peak. | “A typical value.” Slide μ and the whole curve moves left or right. |
| Standard deviation (σ) | A typical distance from the average. | The square root of the variance; the distance from the mean to the curve’s inflection point. | Small σ = tall, narrow curve (consistent data). Large σ = short, wide curve (variable data). |
| Normal / Gaussian distribution | The bell curve. | A continuous, symmetric distribution fully defined by μ and σ. | A model for data where values cluster evenly around an average. |
| z-score | How many standard deviations a value is from the mean. | z = (x − μ) / σ; a value from the standard normal distribution (μ = 0, σ = 1). | z = +2 means “unusually high”; z = 0 means “perfectly average.” |
| Empirical rule | The 68–95–99.7 shortcut. | For normal data: about 68% of values lie within ±1σ, 95% within ±2σ, 99.7% within ±3σ. | Lets you estimate proportions instantly, with no tables or software. |
| Skewness | Lopsidedness. | Asymmetry of a distribution; a long tail on one side. | A warning sign that the normal model (and the empirical rule) may not apply. |
| Percentile | The share of values below a given point. | The cumulative probability up to a value. | “90th percentile” = higher than about 90% of results. |
The formula (gently)
You will rarely compute this by hand — software does it for you — but seeing it once removes the mystery. The height of the bell curve at any value x is:
f(x) = 1σ√2π · e−(x − μ)² / 2σ²
- f(x) the height (density) of the curve at value x — taller means “more likely around here”
- μ the mean of the data (same units as x, e.g. cm, marks, ml)
- σ the standard deviation of the data (same units as x; must be greater than 0)
- π the constant pi ≈ 3.14159
- e Euler’s number ≈ 2.71828
Plain-language interpretation: the exponent measures how far x is from the mean, in units of σ. Values close to μ get a small exponent and a tall curve; values far from μ get a large negative exponent and the curve drops toward zero. The fraction at the front simply scales the curve so the total area underneath equals 1 (i.e. 100% of the data).
Assumptions — use this model when: the data is continuous (or nearly so), roughly symmetric with one peak, and free of hard boundaries near the bulk of the data.
Do not use it when: the data is clearly skewed (incomes, waiting times), is a count of rare events (defects per unit), has two peaks (mixed populations), or is capped hard at a boundary (e.g. percentages piling up at 0% or 100%).
The one formula you will use: the z-score
z = x − μσ
- z the z-score: distance from the mean, measured in standard deviations (no units)
- x the value you are examining (in original units)
- μ the mean; σ the standard deviation (σ > 0)
Plain-language interpretation: subtract the average, then divide by the “typical distance.” The answer tells you how many typical distances your value sits above (+) or below (−) the mean. Standardising like this lets you compare apples with oranges: a z-score of +1.5 on an exam means the same “degree of unusualness” as a z-score of +1.5 in height.
The 68–95–99.7 rule
Here is the superpower of the normal distribution. For any normal data, no matter the mean or standard deviation:
- About 68% of values fall within 1 standard deviation of the mean (34.1% on each side).
- About 95% fall within 2 standard deviations (adding 13.6% on each side).
- About 99.7% fall within 3 standard deviations (adding 2.1% on each side, leaving roughly 0.1% in each extreme tail).
Instant example: adult heights with mean 170 cm and standard deviation 8 cm. Then about 68% of people are between 162 and 178 cm, about 95% are between 154 and 186 cm, and only about 0.15% are taller than 194 cm (that’s beyond +3σ).
Interactive: shape the curve
The mean and standard deviation are the curve’s only two “control knobs.” Drag the sliders and watch what each one does. The shaded band always covers μ ± 1σ — the middle 68%.
Method: the curve plots f(x) = (1 / (σ√2π)) · e−(x−μ)²/2σ² over a fixed axis from 40 to 160. Note how a larger σ makes the curve wider and flatter — the total area under the curve always stays at 100%.
Step-by-step: answering “how unusual is this value?”
- Check the shape first. Plot a histogram (or dot plot) of your data. You are looking for one peak and rough symmetry. If it’s clearly skewed or has two peaks, stop — the normal model doesn’t apply.
- Find the mean (μ) and standard deviation (σ). From your data, from software, or from a known specification.
- Choose your tool. If the value of interest sits neatly at ±1, 2, or 3σ, the empirical rule is enough. Otherwise, compute a z-score.
- Calculate. z = (x − μ) / σ. Keep the sign: negative means below the mean.
- Convert to a proportion. Use the empirical rule, a z-table, or software
(Excel’s
NORM.S.DIST) to turn z into “% of values below/above this point.” - Interpret in context. Translate back to plain language: “a fill volume of 372 ml is 2.2 standard deviations above target — fewer than 2 bottles in 100 should be this full.”
- Document. Record the mean, standard deviation, sample size, and the date/source of the data, so the analysis can be checked and repeated.
Worked example: is this exam score exceptional?
The problem: a certification exam has scores that are approximately normally distributed. Priya scored 88. Her team lead asks: “How good is that, really?”
The data: historical results show a mean of μ = 72 marks and a standard deviation of σ = 8 marks. The histogram of past scores is single-peaked and symmetric, so the normal model is reasonable.
The method: standardise the score with a z-score, then use the empirical rule.
The formula:
z = x − μσ
Substitute the values:
z = 88 − 728
Intermediate calculation: the numerator is 88 − 72 = 16. Then 16 ÷ 8 = 2.
Final answer: z = +2.0 — Priya’s score is exactly 2 standard deviations above the mean.
Practical interpretation: by the empirical rule, 95% of scores fall within ±2σ of the mean, so 5% fall outside — and by symmetry, half of that 5% (about 2.5%) is above +2σ. Priya beat roughly 97.5% of past candidates. That is a genuinely exceptional result, and you worked it out with nothing but subtraction, division, and one memorised rule.
Limitations: this assumes past scores really are close to normal and that the historical mean and standard deviation still describe current exams. If the exam changed recently, the comparison is weaker.
Interactive: z-score calculator
Now you try. Enter any mean, standard deviation, and value — the calculator standardises it, places it on the curve, and tells you how unusual it is.
Method: z = (x − μ) / σ. The percentile is the cumulative standard normal probability Φ(z), computed with the Zelen–Severo polynomial approximation (accurate to about 7 decimal places). Assumes your data is approximately normal.
Interpretation guide
| When you see… | It means… |
|---|---|
| z = 0 | The value is exactly average. |
| z between −1 and +1 | Completely ordinary — about 68% of all values live here. |
| z beyond ±2 | Unusual — only about 5% of values are this far from the mean. Worth a closer look. |
| z beyond ±3 | Rare (about 0.3%). In quality control this typically triggers an out-of-control signal. |
| A histogram with one peak and mirror-image sides | The normal model is plausible; the empirical rule should work well. |
| Mean noticeably larger than the median | Right skew — a long tail of high values. Be cautious with the normal model. |
| Data piled up against a boundary (e.g. 0) | Likely skewed; the bell curve would predict impossible negative values. |
| Two separate peaks | Probably two mixed groups (two machines, two shifts). Split the data before modelling. |
Common mistakes and exam traps
1. Assuming everything is normal Assumption error
Why it’s wrong: many datasets (incomes, waiting times, defect counts) are skewed, so normal-based percentages will be badly off. How to recognise it: a lopsided histogram, or a mean far from the median. How to avoid it: always plot the data before applying the empirical rule or z-scores.
2. Mixing up standard deviation and variance Calculation error
Why it’s wrong: variance is σ² — the standard deviation squared. Dividing by variance instead of σ gives a meaningless z-score. How to recognise it: z-scores that look absurdly small (e.g. 0.03 for an obviously extreme value). How to avoid it: check units — σ is in the same units as the data; variance is in squared units.
3. Reading “below” when the question asks “above” Interpretation error
Why it’s wrong: tables and software usually give the area below a value. If the question asks for the proportion above, you must subtract from 1 (or 100%). How to recognise it: an answer like “97.7% of bottles are overfilled” when overfilling should be rare. How to avoid it: sketch the curve and shade the region you actually want before calculating.
4. Dropping the minus sign Calculation error
Why it’s wrong: z = −1.5 and z = +1.5 are equally “unusual” but on opposite sides of the mean — one is below average, one above. How to recognise it: conclusions that don’t match common sense (“the shortest person is above average height”). How to avoid it: always compute x − μ carefully and keep the sign through the whole calculation.
5. Treating 68–95–99.7 as exact for any dataset Exam trap
Why it’s wrong: the rule holds only for (approximately) normal data, and the numbers themselves are rounded (the true values are 68.27%, 95.45%, 99.73%). How to recognise it: exam questions that sneak in a skewed dataset and expect you to notice. How to avoid it: state the normality assumption whenever you use the rule — examiners often award marks for it.
6. Using the density instead of the cumulative probability in Excel Software-output error
Why it’s wrong: NORM.DIST(x, mean, sd, FALSE) returns the curve’s
height, not a probability. Only the cumulative form (TRUE) gives “% of values below x.”
How to recognise it: “probabilities” that don’t land between sensible bounds, or results that change wildly with units.
How to avoid it: for proportions, always set the last argument of NORM.DIST to TRUE.
Software guidance: Excel and Minitab
Software should confirm and speed up what you now understand — not replace the thinking.
Excel
| Task | Function | Inputs | Output that matters |
|---|---|---|---|
| z-score for a value | =STANDARDIZE(x, mean, sd) |
Your value, the mean, the standard deviation | The z-score (compare against ±1, ±2, ±3) |
| % of values below x | =NORM.DIST(x, mean, sd, TRUE) |
Value, mean, sd, and TRUE for cumulative |
A proportion between 0 and 1 (multiply by 100 for %) |
| % below a z-score | =NORM.S.DIST(z, TRUE) |
The z-score, and TRUE |
The percentile as a proportion |
| Value at a given percentile | =NORM.INV(p, mean, sd) |
The percentile as a proportion (e.g. 0.95) | The cutoff value in original units |
Interpreting the result: =NORM.DIST(182, 170, 8, TRUE) returns about 0.9332,
meaning roughly 93.3% of values fall below 182 — matching a z-score of +1.5.
Common Excel mistakes: passing FALSE when you want a probability (see mistake 6 above),
and entering the variance where the function expects the standard deviation.
Minitab
Where to find it: Graph › Probability Distribution Plot › View Probability lets you shade regions of a normal curve visually. For numbers, use Calc › Probability Distributions › Normal, choose Cumulative probability, and enter the mean, standard deviation, and your value as the input constant.
Output that matters: the P(X ≤ x) value — the proportion below your value. To check whether your data is normal in the first place, run Stat › Basic Statistics › Normality Test and look at the probability plot: points close to a straight line suggest normality. Common Minitab mistake: reading the probability density output as if it were a probability.
Practice activity: the bottling line
Scenario: a bottling machine fills juice bottles with a mean of μ = 350 ml and a standard deviation of σ = 10 ml. Historical fill volumes are approximately normal.
Your task: using only the empirical rule (no calculator needed), answer:
- What percentage of bottles contain between 340 ml and 360 ml?
- What percentage of bottles contain more than 370 ml?
- The legal minimum on the label is 320 ml. Roughly how often should a bottle fall below it?
Work each one out (a quick sketch of the curve helps), then check yourself.
1. About 68%. 340 and 360 are exactly μ − 1σ and μ + 1σ (350 − 10 and 350 + 10). The empirical rule says about 68% of values fall within ±1σ.
2. About 2.5%. 370 = 350 + 2×10 = μ + 2σ. About 95% of bottles fall within ±2σ, leaving 5% outside. The curve is symmetric, so half of that 5% — about 2.5% — is above +2σ.
3. Almost never — about 0.1–0.15% of bottles. 320 = 350 − 3×10 = μ − 3σ. About 99.7% of bottles fall within ±3σ, leaving 0.3% outside, split between two tails. Below −3σ is roughly 0.15% — about 1 or 2 bottles in every 1,000.
The reasoning pattern: convert each boundary into “how many σ from μ?”, then read the proportion straight off the 68–95–99.7 rule, remembering to halve any leftover percentage when the question asks about one tail only.
Mini quiz
Five questions. Pick an answer for each, then press Check my answers.
Final summary
Key takeaways
- The normal distribution is a symmetric, single-peaked model where values cluster around the mean.
- Two numbers describe it completely: the mean μ (location) and the standard deviation σ (spread).
- The empirical rule: about 68% of values fall within ±1σ, 95% within ±2σ, 99.7% within ±3σ.
- A z-score, z = (x − μ) / σ, converts any value into “standard deviations from the mean” so different measurements can be compared fairly.
- |z| beyond 2 is unusual; |z| beyond 3 is rare and usually worth investigating.
- Software gives the area below a value by default — subtract from 100% when you need the area above.
- Always plot your data first: skewed, bounded, or two-peaked data breaks the normal model.
Decision rule
If the histogram shows one peak and rough symmetry, use the normal model: boundaries at whole multiples of σ → empirical rule; anything else → z-score plus a table or software. If the histogram is skewed, bounded, or two-peaked, do not use the normal model.
Suggested next lesson: Descriptive Statistics: Mean, Median & Standard Deviation (if you want to solidify the building blocks) or Confidence Intervals for Beginners (to see the normal curve put to work in estimation).
References
- NIST/SEMATECH e-Handbook of Statistical Methods — Normal Distribution: itl.nist.gov/div898/handbook
- OpenStax, Introductory Statistics (open textbook), chapter on the Normal Distribution: openstax.org