Statistics › Section B · Descriptive statistics › Topic B.7
Range, variance, and standard deviation
A centre alone tells half the story. This topic measures how far data travel from their centre - the range, the variance, the standard deviation - and adds the z-score, the ruler that compares values across different data sets.
Centre is not enough
Two cities can share a mean daily temperature of 20 degrees while one barely moves and the other swings between freezing and scorching. A centre alone hides that; every centre needs a companion number describing spread. This topic builds the spread toolkit: the range, quick but fragile; the standard deviation and its square, the variance, the workhorses for symmetric data; and the z-score, which turns spread into a ruler for comparing values across different data sets.
The range
The range is the largest value minus the smallest: one subtraction, and the full extent of the data in a single number. Its weakness is that it uses exactly two values, ignoring everything between: change one extreme - a single typo, or one genuinely unusual day - and the range leaps. The range suits quick summaries and small data sets. For anything robust, quartiles do better: the interquartile range from the percentile topic measures the middle half and is barely moved by the extremes, which is precisely what the range cannot manage.
The standard deviation
The standard deviation \(s\) measures the typical distance of the data values from their mean. It is never negative, and it is zero only when every value is identical. Small \(s\): values cluster tightly around the mean. Large \(s\): values scatter widely. Two bakeries can both average a 5-minute wait, but with \(s \approx 1\) the first serves most customers within a minute or so of 5, while with \(s \approx 3\) the second's waits of 2 or 8 minutes are routine. The standard deviation is what turns "same average" into a real comparison.
Computing s, and why n − 1
The recipe: 1. Find the mean. 2. Take each deviation \(x_i - \bar{x}\); the deviations always total zero, so squaring is used to remove the signs. 3. Square each deviation - which also makes distant values count heavily. 4. Add the squares and divide by \(n - 1\): that is the sample variance \(s^2\). 5. Take the square root to return to the original units: that is \(s\). Dividing by \(n-1\) rather than \(n\) corrects a sample's tendency to underestimate the population's spread, since deviations measured from the sample's own mean come out slightly small. The population versions, \(\sigma^2\) and \(\sigma\), measure deviations from \(\mu\) and divide by the population size \(N\).
Worked example: computing s from a list
Delivery minutes for five orders: 3, 5, 7, 7, 8. 1. Mean: \(\bar{x} = \frac{30}{5} = 6\). 2. Deviations: \(-3, -1, 1, 1, 2\) - they total zero, as the balance point promises. 3. Squares: \(9, 1, 1, 1, 4\), totalling 16. 4. Variance: \(s^2 = \frac{16}{5-1} = 4\). 5. Standard deviation: \(s = \sqrt{4} = 2\) minutes. 6. Check: the values run from 3 to 8, so a typical deviation of about 2 minutes from the mean is plausible - and \(s\) sits sensibly below the range of 5.
Worked example: s from a frequency table
The pet-ownership survey from the centre topic: 6 households own no pets, 9 own one, 4 own two, and 1 owns three; the mean is 1 pet. 1. Deviations from the mean: \(-1, 0, 1, 2\); squares: \(1, 0, 1, 4\). 2. Weight each square by its frequency: \(6(1) + 9(0) + 4(1) + 1(4) = 14\). 3. Divide by \(n - 1 = 19\): \(s^2 = \frac{14}{19} \approx 0.737\). 4. Take the root: \(s = \sqrt{0.737} \approx 0.86\) pets. 5. Check: households own between 0 and 3 pets, so a typical distance of under one pet from the mean fits the table.
Position in SD units
The standard deviation becomes a ruler with one habit: any value can be written as the mean plus some number of standard deviations. With mean 6 and \(s = 2\), the value 10 sits at \(6 + 2(2)\): two SDs above the mean. The value 4 sits at \(6 + (-1)(2)\): one SD below. The multiplier need not be a whole number: 8 is exactly one SD up, and 5 is half an SD below the mean. This "how many SDs from the mean" reading is exactly what the z-score formalises next.
Worked example: z-scores across two classes
Maya scored 72 on a test whose class mean was 65 with \(s = 4\); Liam scored 88 on a different test with mean 80 and \(s = 10\). Who performed better relative to their class? 1. Maya: \(z = \frac{72-65}{4} = 1.75\). 2. Liam: \(z = \frac{88-80}{10} = 0.8\). 3. Maya sits 1.75 SDs above her class mean and Liam only 0.8 above his, so Maya's performance is stronger - even though Liam's raw score is higher. 4. Check: the same scale, SDs from the mean, is what makes unlike tests comparable. A z-score is nothing but the multiplier in value = mean + k·s.
How much data lies within k SDs
Two rules turn the standard deviation into guaranteed or typical shares of the data. Chebyshev's rule holds for any data set, whatever its shape: at least 75 percent of values lie within two SDs of the mean, and at least 8/9, about 89 percent, within three. The empirical rule applies only to bell-shaped, symmetric data: about 68 percent lie within one SD, about 95 percent within two and about 99.7 percent within three. For delivery times with mean 6 and \(s = 2\), Chebyshev guarantees at least 75 percent between 2 and 10 minutes; if the times were bell-shaped, about 68 percent would fall between 4 and 8. So in bell-shaped data, a value with \(|z|\) above 2 is fairly unusual: only about 5 percent lie that far out.
Choosing centre and spread together
The spread measure should match the centre. Symmetric quantitative data: mean and standard deviation - both use every value, and together they summarise cleanly. Skewed data or data with outliers: median and IQR - both resist extremes, which the mean and \(s\) do not, since a single tail value inflates both. Categorical data has no centre-and-spread pair: report the mode and a frequency table. The pairing habit matters because each pair describes the data from the same philosophy - count everything, or ignore the extremes - and mixing philosophies misleads.
Common mistakes to avoid
First, forgetting the square root: the variance lives in squared units (minutes squared, pets squared) and only \(s\) is comparable to the data. Second, dividing by \(n\) instead of \(n-1\) for sample data - the classic small-sample error. Third, reporting \(s\) for heavily skewed data, where one tail value inflates it. Fourth, comparing raw scores across data sets with different means and spreads instead of using z-scores. Fifth, quoting a centre without its spread partner. Sixth, trying to measure spread with the deviations themselves: they always total zero, which is why they are squared first.
Key takeaways
The range is maximum minus minimum - quick, but hostage to two values. The standard deviation \(s\) is the typical distance from the mean, computed as the square root of the sample variance \(s^2 = \frac{\sum (x_i - \bar{x})^2}{n-1}\); the \(n-1\) corrects a sample's tendency to understate spread. Variance is in squared units; taking the root returns to the data's units. Every value can be written as the mean plus k standard deviations, and the z-score \(\frac{x - \bar{x}}{s}\) makes that multiplier a standard ruler for comparing across data sets. Chebyshev's rule bounds the share within k SDs for any shape; the empirical rule gives about 68, 95 and 99.7 percent for bell-shaped data. Pair mean with \(s\) for symmetric data, median with IQR for skewed data.
Flashcards
Flashcard
Maximum minus minimum
Test your understanding
Question 1 of 5
The data 4, 9, 2, 12, 6 have what range?
Question 2 of 5
For the sample 2, 4, 6, 8, 10 (mean 6), the squared deviations total 40. The sample variance is:
Question 3 of 5
Why divide by n − 1 rather than n when computing a sample variance?
Question 4 of 5
Maya scored 72 on a test with mean 65 and s = 4. Her z-score is:
Question 5 of 5
Household incomes are right-skewed with a few huge outliers. Which centre-spread pair should a summary report?
