English:Correlation and Lines of Best Fit
Correlation and Lines of Best Fit
Introduction
Welcome to Correlation and Lines of Best Fit. In this aiMOOC for Grades 9–10, you will learn how to read scatter plots, describe correlation, draw and interpret a line of best fit, make predictions, calculate simple residuals, and judge when a linear model is useful. You will also learn an essential rule of data literacy: a relationship between two variables does not by itself prove that one variable causes the other.
A scatter plot represents paired numerical data. Each point has an x-coordinate and a y-coordinate, so one point represents one observation. The overall pattern can show a positive association, a negative association, little or no linear association, clusters, gaps, or unusual points.
The image above compares scatter plots with different Pearson correlation coefficients. Use it to focus on two ideas: direction and strength. Direction tells you whether y tends to increase or decrease as x increases. Strength tells you how tightly the points follow a straight-line pattern.
Learning Goals
By the end of this course, you should be able to:
- Scatter plot: Describe the direction, form, and strength of a relationship shown by paired numerical data.
- Correlation: Distinguish positive, negative, and weak or absent linear correlation.
- Line of best fit: Draw or interpret a reasonable trend line and explain what it represents.
- Linear equation: Use an equation of the form to estimate values.
- Residual: Calculate the difference between an observed value and a predicted value.
- Interpolation: Make cautious predictions inside the observed data range and distinguish them from extrapolation.
- Causality: Explain why correlation alone is not enough to establish cause and effect.
Reading Scatter Plots
Direction, Form, and Strength
When you examine a scatter plot, ask three questions.
Direction: If points tend to rise from left to right, the variables have a positive association. If they tend to fall from left to right, they have a negative association. If there is no clear upward or downward pattern, the linear association may be weak or absent.
Form: A line of best fit is most useful when the pattern is roughly linear. A curved pattern can represent a real relationship even when a straight line does not describe it well.
Strength: A strong linear association has points that lie relatively close to an imagined straight line. A weak linear association has points spread more widely around that line.
For Pearson's correlation coefficient, written as , values range from −1 to +1. Values near +1 indicate a strong positive linear relationship, values near −1 indicate a strong negative linear relationship, and values near 0 indicate little linear relationship. A value near 0 does not prove that the variables have no relationship at all; the relationship might be curved or otherwise non-linear.
Correlation Is Not Causation
Suppose students who spend more time studying tend to earn higher test scores. A scatter plot might show a positive association, but the graph alone cannot prove that study time caused every score difference. Other factors, such as prior knowledge, sleep, teaching quality, motivation, or access to resources, may also matter.
When you see a correlation, ask whether a hidden or confounding variable could influence both variables. Also ask whether the direction of cause could be reversed or whether the pattern could have arisen by chance.
Good statistical language says, "The variables are associated" or "Higher x-values tend to occur with higher y-values." Use causal words such as "causes" only when the study design and evidence justify them.
Lines of Best Fit
What a Best-Fit Line Does
A line of best fit summarizes the central linear trend in a scatter plot. It usually does not pass through every point, and it does not have to pass through the origin. When you draw a trend line by eye, aim for a line that follows the direction of the data with points reasonably balanced above and below it.
A more formal best-fit line is the least-squares regression line. For ordinary least-squares regression of y on x, the chosen line minimizes the sum of the squared vertical residuals. This gives a consistent rule for deciding which line fits the observed points best.
Equation of a Best-Fit Line
A linear model is often written as:
Here, is the predicted value of y, is the slope, and is the y-intercept.
The slope tells you the predicted change in y for each increase of 1 unit in x. The intercept is the predicted y-value when x equals 0. An intercept may be mathematically correct but meaningless in context if x = 0 is outside the sensible or observed range.
Consider this sample data about study time and a test score:
| Study time in hours | Test score |
|---|---|
| 1 | 58 |
| 2 | 63 |
| 3 | 69 |
| 4 | 75 |
| 5 | 80 |
| 6 | 88 |
A least-squares model for these data is approximately . The slope suggests that each additional hour of study is associated with an increase of about 5.91 score points in the model. For 4.5 hours, the predicted score is about .
Notice the wording: the model describes an association and makes a prediction. It does not prove that adding exactly one hour of study will cause every student's score to increase by 5.91 points.
Interpolation and Extrapolation
Interpolation means predicting within the range of x-values represented by the data. In the study-time example, predicting a score for 4.5 hours is interpolation because 4.5 lies between 1 and 6 hours.
Extrapolation means predicting outside the observed x-range. Predicting a score for 12 hours would be extrapolation. It is usually less reliable because the pattern may change beyond the data you actually observed.
A line can continue forever on a graph, but the real-world relationship often cannot. Always check whether the prediction is reasonable in context.
Residuals and Model Fit
What Is a Residual?
A residual measures the vertical difference between an observed y-value and the value predicted by the model:
A positive residual means the observed point is above the line. A negative residual means it is below the line. A residual near zero means the prediction is close to the observed value.
Using the study-time model, the predicted score for 5 hours is approximately . The observed score is 80, so the residual is approximately . The point lies about one score point below the model's prediction.
Why Least Squares Uses Squares
If positive and negative residuals were simply added, they could cancel one another. Least squares instead squares every residual and then adds those squared values. Squaring makes every contribution non-negative and gives larger errors more influence.
For Grades 9–10, the key idea is more important than the full algebraic derivation: a least-squares line is selected by an objective rule that tries to keep the vertical prediction errors collectively small.
Outliers and Unusual Patterns
Outliers Can Matter
An outlier is a point that lies far from the overall pattern. An outlier can change the slope, intercept, correlation coefficient, and predictions of a fitted line, especially when the dataset is small.
Never delete an outlier simply because it is inconvenient. First investigate whether it is a data-entry error, a measurement problem, or a genuine but unusual observation. If it is valid, it may contain important information.
Always Look at the Graph
Summary statistics can hide important structure. Anscombe's quartet is a famous example: four datasets have the same basic summary statistics, including the same correlation and regression line, yet their scatter plots look very different.
This is why responsible data analysis combines calculations with visual inspection. A single number cannot tell you whether a relationship is curved, dominated by an outlier, clustered, or otherwise unsuitable for a simple straight-line model.
A Practical Analysis Routine
When you are given paired numerical data, use this sequence:
- Context: Identify what x and y represent, including units.
- Scatter plot: Plot every pair and inspect the overall pattern.
- Association: Describe direction, form, and strength without claiming causation.
- Outlier: Identify unusual points and investigate them rather than automatically removing them.
- Line of best fit: Fit or draw a line only when a linear model is reasonable.
- Prediction: Use the line equation for estimates and state whether the prediction is interpolation or extrapolation.
- Residual: Compare observed and predicted values to judge individual prediction errors.
- Conclusion: Interpret the result in context and state important limitations.
Interactive Tasks
Quiz: Test Your Knowledge
What does a positive correlation mean on a scatter plot? (As x tends to increase y also tends to increase) (!As x tends to increase y always decreases) (!Every point lies exactly on one line) (!One variable definitely causes the other)
Which description best matches a strong negative linear correlation? (Points lie fairly close to a downward sloping line) (!Points lie fairly close to an upward sloping line) (!Points form a perfect circle) (!The variables must have equal values)
What is the main purpose of a line of best fit? (To summarize the central linear trend and support predictions) (!To connect every point in order) (!To prove that one variable causes another) (!To make all residuals equal to zero)
What does the slope in a linear prediction model describe? (The predicted change in y for a one unit increase in x) (!The total number of plotted points) (!The largest observed y value) (!The distance from every point to the origin)
Which prediction is an interpolation? (A prediction made within the observed range of x values) (!A prediction made far beyond the observed range) (!A prediction made without using data) (!A prediction that must be exactly correct)
How is a residual calculated? (Observed value minus predicted value) (!Predicted value minus the x value) (!Observed value plus predicted value) (!Slope divided by intercept)
What does a negative residual mean? (The observed point lies below the predicted line) (!The observed point lies above the predicted line) (!The scatter plot has no correlation) (!The slope of the line must be negative)
Why can extrapolation be risky? (The relationship may change outside the observed data range) (!All lines stop at the final data point) (!Correlation coefficients cannot be used with numbers) (!Every extrapolated value has a zero residual)
What should you do when you notice an outlier? (Investigate the point and its context before deciding how to handle it) (!Delete it immediately) (!Move it onto the best fit line) (!Assume it proves causation)
Why is it important to inspect a scatter plot even when summary statistics are available? (Different data patterns can have similar summary statistics) (!Graphs always prove causal relationships) (!Summary statistics cannot contain numbers) (!Scatter plots make every prediction exact)
Memory Game
| Scatterplot | Graph that displays paired numerical observations as points |
| Positivecorrelation | Pattern in which larger x values tend to occur with larger y values |
| Negativetrend | Pattern in which larger x values tend to occur with smaller y values |
| Bestfitline | Straight model that summarizes a central linear pattern |
| Residual | Observed outcome minus its model prediction |
| Outlier | Observation that lies unusually far from the overall pattern |
Drag and Drop
| Match the correct terms. | Topic |
|---|---|
| Upward pattern | Positive association |
| Tightly grouped around a straight trend | Strong linear relationship |
| Prediction inside the observed range | Interpolation |
| Prediction beyond the observed range | Extrapolation |
| Association without proof of cause | Correlation warning |
...
Crossword Puzzle
| Scatterplot | What graph shows paired numerical observations as points? |
| Correlation | What word describes the direction and strength of a linear association? |
| Residual | What is the observed value minus the predicted value called? |
| Outlier | What is an unusual point far from the overall pattern called? |
| Slope | What part of a linear equation gives the predicted change in y per unit of x? |
| Extrapolation | What is prediction outside the observed data range called? |
LearningApps
Cloze Text
Open-Ended Tasks
Easy
- Scatter Plot Description: Find or create a small scatter plot with at least eight points and write three sentences describing its direction, form, and strength.
- Human Best-Fit Line: Draw a scatter plot from classroom data, add a reasonable line of best fit by eye, and explain why you placed the line where you did.
- Correlation Photo Hunt: Photograph or sketch two real-world situations that could produce paired numerical data and predict whether each relationship would be positive, negative, or weak.
- Prediction Check: Use a given best-fit equation to make two interpolated predictions and explain the meaning of each result in context.
Standard
- Class Data Investigation: Collect one safe, non-sensitive numerical pair from volunteers, such as hand span and height, create a scatter plot, describe the association, and discuss limitations.
- Residual Table: For a dataset and its best-fit equation, calculate predicted values and residuals for at least six observations, then identify the largest absolute residual.
- Outlier Experiment: Create a small dataset, fit or estimate a trend line, add one extreme point, and compare how the apparent relationship and predictions change.
- Correlation Interview: Interview a teacher, coach, scientist, technician, or business professional about a situation in which two measured variables are related, then explain whether the evidence is correlational or causal.
Advanced
- Digital Regression Project: Use a spreadsheet or graphing tool to create a scatter plot, calculate a linear regression equation and correlation coefficient, and interpret both in context.
- Model Comparison: Find data with a visibly curved pattern and compare a straight-line model with a non-linear description, explaining why correlation alone can miss structure.
- Media Claim Audit: Find a news or social-media claim involving two numerical variables, investigate the original data if possible, and produce a short written or video critique of correlation, causation, sampling, and possible confounders.
- Anscombe Investigation: Recreate or import the four datasets in Anscombe's quartet, calculate or verify their common summaries, compare the graphs, and present what the example teaches about responsible data analysis.
Learning Assessment
- Interpret a New Scatter Plot: Given an unfamiliar scatter plot, describe direction, form, strength, and any unusual points, then justify whether a linear model is appropriate.
- Build and Use a Model: From a small table of paired data, draw a scatter plot, fit a reasonable line, write its equation, and use it for one interpolation.
- Evaluate Extrapolation: Decide whether a prediction beyond the observed range is defensible, and explain what assumptions would have to remain true.
- Analyze Residuals: Calculate residuals for several observations and use their signs and sizes to identify where the model overpredicts or underpredicts.
- Challenge a Causal Claim: Given a strong correlation, propose at least two alternative explanations and describe what additional evidence would be needed to support causation.
- Assess an Outlier: Compare conclusions with and without an unusual point and argue whether the point should be retained, corrected, or excluded based on evidence rather than convenience.
Evidence of Learning
Strong evidence of learning includes:
- Knowledge: Accurate explanations of scatter plots, positive and negative correlation, linear association, slope, intercept, interpolation, extrapolation, residuals, and outliers.
- Graphical Skill: Correct construction and interpretation of scatter plots and sensible placement or use of best-fit lines.
- Quantitative Skill: Correct substitution into a linear model, calculation of predictions, and calculation and interpretation of residuals.
- Statistical Reasoning: Appropriate distinction between association and causation, including recognition of confounding variables and the limitations of extrapolation.
- Data Product: A clear graph, spreadsheet, report, poster, presentation, or video that communicates an investigation using paired numerical data.
- Transfer: Ability to evaluate a new real-world claim by checking the graph, model, prediction range, unusual points, and strength of the available evidence.
OERs on the Topic
Explore the English Wikipedia overview of correlation for additional examples, definitions, and links to related statistical ideas.
You can also connect this course to Linear regression, Pearson correlation coefficient, Data visualization, Statistics, and Scientific method.
Linked Learning Areas