8 topics. Each one opens with its key idea, then the detail you need to use it.
1. Describing a distribution: shape, center, spread, outliers
Key idea: Describe shape, center, spread and outliers, always with the variable's name and units.
A distribution tells you what values a variable takes and how often. On the exam, "describe" means four things in context: shape, center, spread and unusual features. Shape usually comes first. A histogram is symmetric if the left and right sides roughly mirror each other. It is skewed right if the long tail points toward high values, and skewed left if the tail points toward low values. Count the peaks too, since a distribution with one peak is unimodal and one with two peaks is bimodal. A bimodal shape with a gap usually means two different groups are mixed together. For example, a neighborhood of condos near $350,000 and large houses near $700,000 can produce two peaks with an empty stretch between them. To estimate the median from a histogram, find the value that splits the total area of the bars in half. Do not use the tallest bar, because that marks the mode. A cumulative relative frequency graph shows the same information another way. Where the curve climbs steeply, many observations fall in that interval, and where it flattens out, very few observations do. The median sits where the curve crosses 0.5.
2. Mean versus median, and why skew separates them
Key idea: The mean chases the tail; the median holds its place, so the mean sits on the side of the skew.
The mean is the balance point of the data, so every value pulls on it, while the median is the middle value of the ordered list and only cares about order. That difference is why skewness pulls them apart. In a right-skewed distribution, a few large values pull the mean to the right, while the median stays near the bulk of the data and the mode sits at the peak on the left. The order is typically mode, then median, then mean. In a left-skewed distribution the order typically flips to mean, then median, then mode. The firm rule is that the mean is pulled toward the tail, so the mean is greater than the median for right skew and less than the median for left skew. In a symmetric, single-peaked distribution all three are about equal. You can run this logic backward too. If a table reports a mean of 72 and a median of 75, the mean is below the median, which suggests a left skew. When data are strongly skewed or have outliers, report the median as the center and the IQR as the spread. When data are roughly symmetric with no outliers, the mean and standard deviation are the better pair, because they use all the information.
3. Spread, outliers and box plots
Key idea: The IQR and median resist outliers; fences sit 1.5 IQRs beyond the quartiles, and whiskers stop at the last value inside them.
The standard deviation is roughly the typical distance of a value from the mean. Because it is built on the mean, it is not resistant to extreme values. The IQR is Q3 − Q1, the spread of the middle 50% of the data. To find the quartiles, split the ordered data at the median and take the median of each half. For 3, 5, 7, 9, 11, 13, 15, 17, 19, the median is 11. The lower half 3, 5, 7, 9 gives Q1 = 6, the upper half gives Q3 = 16, and so the IQR is 10. The resistance gap shows up when you drop an outlier. Remove the 100 from 5, 8, 12, 15, 20, 24, 30, 35, 40, 100 and the mean falls from 28.9 to 21, while the median only moves from 22 to 20. To flag outliers, compute the fences Q1 − 1.5 × IQR and Q3 + 1.5 × IQR. If the IQR is 18, the fences sit 27 units beyond each quartile. A box plot draws the five-number summary. The box runs from Q1 to Q3 with a line at the median, and each whisker runs to the most extreme value still inside its fence. Outliers appear as separate points, and the mean and standard deviation are never used. Read skew from the long side, so a long upper whisker with high outliers, like phone call lengths, means right skew. A box plot cannot show two peaks or a gap, so a histogram can reveal what it hides.
4. Comparing distributions
Key idea: Every sentence in a comparison needs a comparing word, both groups, and the context.
A comparison question wants a real comparison, not two separate descriptions. Use words like "greater than", "less than" or "about the same as" in every sentence. Compare all four features, which are shape, center, spread and outliers, and then put numbers on the comparison. Suppose Team A scores a median of 102 points per game with an IQR of 14, and Team B has a median of 98 with an IQR of 22. A good answer says Team A's median score is 4 points higher than Team B's. It adds that Team A's scores are more consistent, because its IQR of 14 is smaller than Team B's 22. "Consistent" is a word about spread, not center, which trips up many students. Pick measures that suit both groups. If one distribution is skewed, compare medians and IQRs for both, so you are comparing like with like. On side-by-side box plots, you can compare medians directly and see which box is wider. On back-to-back stemplots or stacked dotplots, also look for clusters and gaps.
5. Z-scores, linear transformations and Normal plots
Key idea: Adding a constant moves center but not spread; multiplying scales both.
Distance from the mean means nothing until you know the spread, and a z-score fixes that: z = (value − mean) / standard deviation. A score of 92 on an exam with mean 72 and standard deviation 10 has z = (92 − 72)/10 = 2.0, so it sits two standard deviations above the mean. In a roughly Normal distribution, that is about the 97.7th percentile, which is clearly unusual. Being 17 points above the median tells you much less, because it ignores spread. Changing units is a linear transformation, new value = a + b × old value. Adding a constant a shifts every measure of center and location by a. It leaves the standard deviation, IQR and range alone, because every value moves together. Multiplying by b multiplies the center and the spread by b. So converting a mean of 20°C to Fahrenheit gives 32 + 1.8 × 20 = 68°F. A spread of 5°C, however, becomes 1.8 × 5 = 9°F, with no 32 added. A Normal probability plot checks whether data are close to Normal. If the points fall close to a straight line, a Normal model is reasonable. Curving away from the line signals skew, and points far off at the ends mark outliers.
6. Sampling methods and survey bias
Key idea: Strata are similar inside, clusters are mixed inside; nonresponse is who stays silent, response bias is who answers wrong.
The population is the whole group you want to learn about, and the sampling frame is the list you actually choose from. When the frame leaves people out, you get undercoverage. A frame listing only schools that subscribe to a state database misses students at other schools and homeschooled students. A simple random sample gives every group of n individuals the same chance of selection. A stratified sample splits the population into similar groups, such as grade levels, and takes an SRS from each, which helps when the groups differ on the response. A cluster sample randomly picks whole groups that each look like a mini-population, such as homerooms, and surveys everyone in them, which saves time and money. A systematic sample takes every kth individual after a random start. Convenience and voluntary response samples show sampling bias, because people who are easy to reach or who choose to answer often differ from everyone else. Undercoverage is another form of sampling bias, since the method itself leaves some people out. Good survey design also needs honest answers from everyone chosen. Nonresponse bias happens when selected people refuse or cannot be reached. If 150 of 1,000 mailed renovation surveys come back, the repliers are probably the keenest renovators, so interest is overstated. Response bias happens when people answer but answer inaccurately. Loaded wording, such as calling a uniform rule "costly and failed", pushes people toward no, and sensitive questions about cheating lead to underreporting. On the exam, name the bias, explain its mechanism in context and state its direction.
7. Observational studies, confounding and causation
Key idea: Only a randomized experiment supports a cause-and-effect conclusion.
In an observational study, researchers measure variables but do not impose any treatment, while in an experiment they deliberately assign treatments and watch the response. This difference decides whether you can claim causation. Suppose a study finds that students who eat breakfast earn higher grades. The students chose whether to eat breakfast, so other differences between families come along with that choice. Family income, sleep habits or how organized a family is might affect both breakfast and grades. A lurking variable is one the study did not measure that affects the relationship. Two variables are confounded when their effects on the response cannot be separated. Here, you cannot tell whether breakfast or a stable home raised the grades. An observational study can show an association, but it cannot prove cause and effect, no matter how large the sample is. When the exam asks for a confounding variable, you must name one and explain both links. Explain how it relates to the explanatory variable, and how it affects the response. Naming a variable without both links usually loses the point.
8. Designing experiments: randomization, control, blocking and matched pairs
Key idea: Randomize to balance unknown variables; block to remove a known one.
A well-designed experiment has four parts. It compares two or more treatments, which can include a placebo or an existing method as a control. It uses random assignment, so lurking variables tend to balance out across groups. It controls other variables by keeping conditions the same for every group. It also replicates, using enough experimental units that chance differences average out. The response variable is what you measure, and the treatments are the levels of the explanatory variable. Blocking goes a step further. If you know a variable like age or gender affects the response, group the units into blocks by that variable. Then randomly assign treatments separately within each block. Blocking removes that variable's effect from the comparison, which makes a real treatment effect easier to see. A matched pairs design is a special block design. Either each subject receives both treatments in a random order, or similar subjects are paired and a coin flip decides who gets which. The analysis then uses the differences within each pair. Blinding adds one more protection against bias, and in a double-blind study neither the subjects nor the people measuring the response know who got which treatment. Finally, check generalizability. Volunteers in a randomized experiment support a causal conclusion, but only for people like those volunteers, unless they were also randomly sampled from a larger population.