8 topics. Each one opens with its key idea, then the detail you need to use it.
1. Reading a scatterplot
Key idea: Describe direction, form, strength and unusual features, in context, before you compute anything.
Put the explanatory variable on the horizontal axis and the response variable on the vertical axis. The explanatory variable is the one you use to predict, and the response is the one you predict. Then describe four things, always in context.
Direction is positive if y tends to rise as x rises, and negative if y tends to fall. Form is linear if the points cluster around a straight line, or nonlinear if they bend, level off or curve back. Strength is how tightly the points hug that form. Unusual features include outliers, gaps and separate clusters.
Form matters most, because everything else in the unit assumes a linear form. Picture car speed against fuel economy. Mileage rises up to about 50 miles per hour and then falls. That relationship is strong, but it is not linear, so a straight line through it would miss badly at both ends.
A complete description names the variables in every part. For engine size and highway mileage, you would report a strong, negative, linear association, plus one car that gets much better mileage than its engine size suggests. Naming the variables turns a generic sentence into an answer that earns credit.
2. Correlation r and its properties
Key idea: r is a unitless measure of linear strength and direction. Shifts, rescaling by a positive constant and swapping the axes do not change it, and multiplying one variable by a negative constant flips its sign.
The correlation r measures the direction and strength of a linear association. It runs from -1 to 1. Values near 1 or -1 mean the points lie close to a line, and values near 0 mean little or no linear association. So r = 0.85 between study hours and test scores describes a strong, positive linear association, while r = -0.12 means there is almost no linear pattern at all.
The exam tests these properties directly. First, r has no units, so changing feet to meters or Fahrenheit to Celsius leaves it alone. Second, adding a constant to every value, or multiplying by a positive constant, does not change r. Multiplying by a negative number flips its sign. Third, r is the same if you swap x and y. Fourth, r only measures linear association, so a perfect U-shape can give r = 0. Fifth, r is not resistant, which means a single outlier can pull it up or down a lot.
The number r is not a slope and not a percentage. If r = 0.85, you cannot say y goes up 0.85 per unit of x, and you cannot say 85% of anything is explained. Both of those claims need other quantities, which the next topics cover.
3. Building the least-squares regression line
Key idea: b = r(sy/sx), and the line passes through (x-bar, y-bar); find the slope first, then the intercept.
The LSRL is written y-hat = a + bx, where y-hat is the predicted value of y. It is the one line that makes the sum of the squared residuals as small as possible. That is why it is called "least squares."
You can build it from five summary numbers. The slope is b = r(sy/sx), with the spread of y on top. The line always passes through the point (x-bar, y-bar), so the intercept is a = y-bar - b(x-bar).
Try it with ages and heights of children. Suppose x-bar = 8 years, sx = 2 years, y-bar = 130 cm, sy = 12 cm and r = 0.9. The slope is 0.9(12/2) = 5.4 cm per year. The intercept is 130 - 5.4(8) = 86.8 cm. So the line is predicted height = 86.8 + 5.4(age).
The common slips are predictable. Using sx/sy gives the slope for predicting x from y, which is a different line. Using r alone as the slope ignores the units of both variables. Forgetting the factor r gives a slope of sy/sx. That slope is at least as steep as the correct one, and it is always positive, so a negative association gets a line that slopes the wrong way.
The formula also explains a hard property question. If you add 10 to every y-value, sy does not change, so the slope stays the same. Only y-bar moves, and that shifts the intercept up by 10.
4. Interpreting slope and intercept in context
Key idea: Slope = predicted change in y per one-unit increase in x; intercept = predicted y at x = 0, which may have no real meaning.
The slope is the predicted change in the response for each one-unit increase in the explanatory variable. A full answer names one more unit of x, says the predicted y rises or falls, and gives the amount b in the units of y, on average. Suppose a line predicts monthly heating cost from outdoor temperature as cost = 150 - 2.5(temperature). The slope means that for each 1-degree rise in temperature, the predicted heating cost drops by $2.50.
Two errors cost points here. One is reversing the variables, for example by claiming that each extra dollar of cost lowers the temperature by 2.5 degrees. The slope always describes y per unit of x, not the other way round. The other is leaving out "predicted" or "on average," which turns a model statement into a promise about every month.
The intercept is the predicted y when x = 0. In the heating line, the predicted cost at 0 degrees is $150. The intercept is not the mean of y. The line passes through y-bar at x-bar, not at x = 0.
Always ask whether x = 0 makes sense. A line predicting weight from height has an intercept for a person 0 inches tall. That is far outside the data, so the intercept is just a number that places the line, not a real prediction.
5. Residuals
Key idea: Residual = observed - predicted; positive means the point sits above the line, so the model underestimated it.
A residual is the error the line makes for one data point. It equals the observed y minus the predicted y, or residual = y - y-hat. To find one, plug the point's x into the line, then subtract that prediction from the actual y.
Here is an example. A line predicts quiz score from minutes of study as score-hat = 52 + 0.6(minutes). A student studied 40 minutes and scored 70. The prediction is 52 + 0.6(40) = 76, so the residual is 70 - 76 = -6 points. The negative sign means the point lies below the line, and the line overestimated her score by 6 points.
The sign carries the meaning. A positive residual means the actual value was higher than predicted, so the point is above the line. A negative residual means the actual value was lower than predicted, so the point is below it. Read the context carefully. For a race time, a positive residual means the runner was slower than predicted, because a bigger time is worse.
Try one where lower is better. A line predicts marathon time from training as time-hat = 300 - 1.2(hours), in minutes. A runner trained 60 hours, so her predicted time is 300 - 72 = 228 minutes. She finished in 236 minutes, so her residual is 236 - 228 = +8 minutes. She ran 8 minutes slower than the model predicted, even though her residual is positive.
For any LSRL, the residuals add up to 0. Points with large residuals are outliers in the y-direction, because the line predicts them poorly. If you get the order backwards and compute predicted minus observed, every sign flips and your interpretation becomes wrong.
6. Residual plots and model fit
Key idea: Random scatter means a line fits; a curve means a line is not appropriate; a fan means predictions vary in precision.
A residual plot puts x (or the predicted values) on the horizontal axis and the residuals on the vertical axis. A horizontal line at 0 marks where a perfect prediction would sit. The plot removes the linear trend so you can see what is left over.
If a line is a good model, the residuals form random scatter with no pattern around 0. A curve, such as a U-shape, means the relationship is nonlinear. The line overpredicts in some regions and underpredicts in others in a systematic way, so a linear model is not appropriate. Note that this is a pattern across all the points, not one strange point. A nonlinear model would describe the data better than a straight line.
A fan shape, where the residuals spread out as x grows, is called heteroscedasticity, or non-constant variance. The line may still follow the trend, but predictions are less precise where the residuals spread out. A single isolated residual far from 0 is an outlier, which is a different problem from a curve.
Put together, a linear model is appropriate when the scatterplot looks roughly linear and the residual plot shows random scatter with roughly constant spread.
Do not confuse a residual plot with a normal probability plot, which only shows whether the residuals look roughly normal and says nothing about whether a line is the right shape.
7. R squared, the coefficient of determination
Key idea: R squared = share of variation in y explained by the line; it carries no sign and says nothing about form.
R squared is the fraction of the variation in the response that the linear model with x explains. If r squared = 0.81 for house price on square footage, then 81% of the variation in house prices is accounted for by the linear relationship with square footage. The other 19% comes from things the model leaves out, such as location or age of the house.
The word "variation" is not optional. R squared does not mean 81% of houses are predicted correctly, and it does not mean prices are 81% accurate. It compares how spread out the residuals are with how spread out y is on its own.
To move between r and r squared, square or take the square root. If r = 0.6, then r squared = 0.36. If r squared = 0.81, then r is either 0.9 or -0.9. You pick the sign from the slope or the scatterplot, because squaring wipes out the direction. A frequent trap reports r = 0.85 and offers "85% of the variation is explained" as an answer. The correct figure is 0.85 squared, which is about 72%.
A high r squared does not prove that a line is the right shape. Curved data can still give r squared = 0.9, so check the residual plot as well. Computer output often also lists s, the typical size of a residual in the units of y.
8. Prediction, extrapolation and causation
Key idea: Predict only inside the data range, and never claim cause from observational data, however strong r is.
To predict, plug an x-value into the line. Use the prediction with care, because the line is only a summary of the data you have.
Extrapolation means predicting for an x-value outside the range of the data. It is unreliable because nothing guarantees the linear pattern continues. A line fitted to children aged 2 to 12 might predict a 40-year-old to be about 3 meters tall. Growth levels off, so the straight line fails. Predictions inside the data range are called interpolation, and they are much safer.
The second limit is causation. A strong correlation in observational data does not show that x causes y. Shoe size and reading score are strongly correlated among elementary students. Bigger feet do not improve reading. Age is a lurking variable that drives both, because older children have bigger feet and read better. When a correlation comes from a lurking variable or from pure coincidence, it is called a spurious correlation.
Only a randomized experiment, where the researcher assigns x, can support a causal claim. A correct conclusion for observational data uses the word "associated," not "causes" or "leads to."
The model's other limitations come from what it leaves out. A line predicting house price from square footage ignores location, so two same-size houses in different towns get the same prediction. More houses make the slope more precise, but they cannot add a variable the model never measured. For the same reason, a bigger observational sample cannot remove a lurking variable. And the line gives the average price for a given size, not the price of one house.