A social media analytics company builds a model to predict daily user engagement from post frequency. After collecting data from 5,000 users, they run the regression and receive good news: the relationship is statistically significant (). But then they look at one more number: . The model explains only 4% of the variance in engagement.
Should the company base its strategy on this model?
Statistical significance tells you the relationship is real — not just noise. But it says nothing about whether the model is useful. A massive dataset can make even a trivially weak relationship significant. Before trusting any regression model for prediction, you need to ask: How well does it actually explain the data? Is my prediction inside the range where the model was fit? And are the model’s assumptions even satisfied?
This lesson gives you the tools to answer those questions rigorously.
By the end of this lesson, you will be able to:
Interpret the slope and intercept of a regression equation precisely, using the three required phrases.
Distinguish interpolation from extrapolation and explain why extrapolation is risky.
Read a residual plot to diagnose non-linearity, heteroscedasticity, outliers, and influential points.
Perform a five-step significance test for using the -distribution with .
Distinguish statistical significance from practical significance using .
Section 2: Prerequisites
▾
What you need coming in — and why it matters today:
Regression equation (REG-2): You know how to compute and . Today you go deeper — interpreting what those numbers mean in context and knowing when predictions from that equation can be trusted.
Residuals (REG-2): You computed individual residuals in REG-2. Today you will read entire residual plots — patterns in the residuals reveal whether the model’s assumptions hold.
Conditions for regression (REG-2): Linearity, independence, equal variance, and near-normality of residuals. Today’s residual plot diagnostics are the practical tool for checking linearity and equal variance.
Five-step hypothesis test framework (inf-5):, , test statistic, -value, decision and conclusion. Today’s test for follows exactly this structure — the only new element is the -distribution and a different test statistic formula.
Decision rule (inf-5): Reject if ; fail to reject if . “Fail to reject accept.” This rule applies identically to the correlation test.
Quick check — can you recall these?
Which of the following is the correct interpretation of the slope in ?
Success Factor:
What changes in this lesson: In REG-2, you built the regression equation and computed residuals. Here you ask: Can this equation be trusted? That means reading residual patterns visually, classifying predictions as safe or risky, and formally testing whether the linear relationship is real in the population. The five-step framework from inf-5 carries over exactly — only the test statistic formula and distribution change.
Retrieval Warm-up — from earlier lessons
An environmental scientist fits a regression line to data on river flow rate (, m³/s) and suspended sediment concentration (, mg/L) for 18 measurement stations. She gets , with and . She wants to verify her arithmetic before proceeding. Which check should she perform?
A researcher states: “I ran a hypothesis test and got with . I conclude that the null hypothesis is true.” Which error in reasoning is present?
Section 3: Core Concepts
▾
How this section is organized: Ten concepts build the complete toolkit for evaluating and using a regression model.
C1–C2: Interpreting the slope and intercept precisely (what the numbers mean)
C3–C4: Safe vs. risky prediction — interpolation and extrapolation
C5–C6: Residual plot diagnostics — checking model assumptions visually
C7–C8: Outliers and influential points — when one observation changes everything
C9–C10: The significance test for and the statistical-vs.-practical distinction
C1 — Slope Interpretation (Precision)
The slope in is more than a number — it is a statement about how two variables are related, on average, in the population the data represent.
A complete slope interpretation requires three specific phrases. Each is non-optional.
Slope Interpretation — Required Form
”For each 1-unit increase in , the predicted changes by units, on average.”
If : replace “changes by” with “increases by.” If : “decreases by units.”
Always name the units of both and in your sentence.
Mini-example:, where = study hours and = exam score (out of 100).
Correct: “For each additional hour of study, the predicted exam score increases by 3.70 points, on average.”
Three phrases present: ✓ “additional hour of study” (1-unit increase in ) ✓ “predicted exam score” (predicted ) ✓ “on average”
Three traps in slope interpretation:
(1) Missing “on average”: The line predicts the mean response for all students with a given number of hours — not what any specific student will score.
(2) Causation language: “Studying 1 more hour causes the score to increase by 3.70 points” — regression shows association, not causation. Use “predicted” or “is associated with,” not “causes.”
(3) x and y reversed: “For each 3.70-point increase in score, hours increase by 1” — always describe responding to , never the reverse.
C2 — Intercept Interpretation
The intercept is the predicted value of when . Whether that is meaningful depends on whether makes sense in context.
Intercept Interpretation
The intercept gives when . It is contextually meaningful only if falls within the observed data range — that is, only if predicting at is interpolation, not extrapolation.
If is outside the observed range, the intercept is a mathematical anchor that keeps the line positioned correctly — it is not a reliable real-world prediction.
Mini-example:, where = fertilizer (g) and = tomato yield (kg), observed range g.
is within the range → meaningful: “The predicted tomato yield with no fertilizer is 1.60 kg.”
Contrast:, age (years) → reaction time (ms), observed range . represents a newborn — far outside the data. The intercept 177.4 ms is not a meaningful prediction for a newborn.
Do not interpret the intercept as a real-world prediction just because it has a plausible numerical value. The check is whether is inside (or very near) the range of the data that was used to fit the model. If not, the intercept is just a placement parameter.
C3 — Interpolation
Interpolation means predicting for an value inside the observed data range .
Interpolation
A prediction is an interpolation if .
Interpolation is generally reliable: the model was fit to data in this region, so the linear pattern has been empirically verified there.
Mini-example: Model fit on study hours . Predicting for hours: → interpolation. The model can be trusted here.
C4 — Extrapolation
Extrapolation means predicting for an value outside the observed range.
Extrapolation
A prediction is an extrapolation if or .
Extrapolation is risky: the linear relationship observed within may not extend beyond it. Predictions can be implausible or physically impossible.
Mini-example: Same model ( hours). Predicting for hours: → extrapolation. The model assumes linearity continues indefinitely, but real exam scores are capped at 100 — the linear trend cannot hold.
Predicting for any is mathematically possible — the arithmetic always works. The danger is in interpreting the result as a reliable estimate. Always check whether is inside or outside before using a prediction. This is the single most important habit in applied regression.
C5 — Residual Plots: Linearity Check
A residual plot graphs the residuals against the fitted values (or against ). It makes systematic patterns visible.
Residual Plot — Linearity Check
Plot residuals on the vertical axis vs. fitted values on the horizontal axis.
Good sign: Residuals bounce randomly above and below with no systematic curve → linearity assumption holds.
Bad sign: A curved band (e.g., U-shape, arch) → the relationship is non-linear; a straight-line model is inappropriate.
Key diagnostic: A U-shaped curve in the residual plot means positive residuals at small and large , with negative residuals in the middle. The model consistently underestimates at the extremes. The fix: transform the data or fit a non-linear model.
C6 — Residual Plots: Homoscedasticity Check
Even if the residuals are random (no curve), their spread should stay constant across all fitted values. Non-constant spread is called heteroscedasticity.
Residual Plot — Homoscedasticity Check
Homoscedasticity (good): The vertical spread of residuals looks similar for small, medium, and large values.
Heteroscedasticity (bad): A fan shape — residuals tightly clustered for small , widely spread for large (or vice versa). This means the model’s precision varies across the range of predictions, which invalidates standard errors.
A fan-shaped residual plot is not just “a few outliers.” It indicates that the entire variance structure of the model is wrong. Standard errors, confidence intervals, and p-values from a heteroscedastic model are not reliable.
Residual Plot Explorer: A residual plot is the primary diagnostic tool for regression. The left panel shows the original scatter plot with the regression line; the right panel shows the corresponding residual plot, revealing what the model is missing.
Scatter Plot
Residual Plot
Residuals bounce randomly around e = 0 with consistent spread. Both the linearity and
homoscedasticity assumptions appear satisfied.
C7 — Outliers in Regression
A regression outlier is a point with a large residual — it falls far from the regression line in the -direction.
Regression Outlier
An observation is a regression outlier if is unusually large compared to the typical residual size.
Outliers inflate and can pull the regression line toward them, distorting the slope and intercept.
Mini-example: If every residual is between and but one observation has , that point is a clear regression outlier. It single-handedly increases by .
C8 — Influential Points and Leverage
An influential point is one that, if removed, would substantially change the slope or intercept. Points with extreme -values have high leverage — they can be influential even without a large residual.
Influential Points and Leverage
An influential point is one whose removal would substantially change or .
A point has high leverage when its -value is far from . High leverage points control the slope — the line is “anchored” to them.
Key distinction: A high-leverage point that happens to fall exactly on the regression line has a residual of 0 — it has the potential to control the slope (it anchors that end of the line), but because it sits on the trend, removing it changes the fit very little. Leverage outlier, and leverage influence.
Do not confuse outliers and influential points. An outlier has a large residual (far from the line vertically). An influential point changes the slope if removed. A point can be: (a) an outlier only; (b) influential only (high leverage, small residual); or (c) both — the most dangerous case.
C9 — Significance Test for (Five-Step)
We can test whether the population correlation (rho) is zero — i.e., whether there is a real linear relationship in the population, or whether the observed could be due to chance alone.
Significance Test for the Population Correlation (ρ = 0)
Step 1 — Hypotheses:
(no linear relationship in the population)
(two-tailed), or / (one-tailed)
Step 2 — Check conditions: are approximately bivariate normal; observations are independent.
Step 3 — Test statistic:
Step 4 — p-value: Use the -distribution with . For two-tailed: .
Step 5 — Decision and conclusion: Reject if . State the conclusion in context.
Why the -distribution? In inf-5, we used because we assumed was known. Here we are estimating the population correlation from the sample — there is additional uncertainty. The -distribution with accounts for this. This is the same principle as inf-6’s -test for a mean with unknown .
Mini table (, two-tailed):
df
6
8
10
13
18
23
28
2.447
2.306
2.228
2.160
2.101
2.069
2.048
C10 — Statistical vs. Practical Significance
A statistically significant result () means we have evidence that in the population. It does not mean the model is useful for prediction.
Statistical vs. Practical Significance
Statistical significance (): The sample provides sufficient evidence that the linear relationship is non-zero in the population.
Practical significance: The model explains enough variance to be useful for prediction. This is measured by — the proportion of variability in explained by .
A large can make even a very weak relationship statistically significant. Always report alongside the -value.
The key question: After finding , ask: “What is ?” If , the model explains only 4% of the variance in . The remaining 96% of the variation in is left unexplained — it comes from factors the model does not capture. That model is not useful for prediction, even though the relationship is statistically real.
Two traps with practical significance:
(1) “Significant p-value means good model” — With , even produces a significant -value, yet explains essentially nothing.
(2) “r² = 0.85 means predictions are 85% accurate” — measures the proportion of variance explained, not prediction accuracy for individual observations.
Section 4: Worked Examples
▾
Example 1 — Fully Worked: Interpreting Slope and Intercept
Context: A researcher fits a regression of exam score () on study hours () using data from 30 students. The observed range is hours. The regression equation is .
Interpret the slope and intercept.
Full solution with reasoning:
Slope: I notice and = study hours, = exam score.
I need three phrases: “1-unit increase in ” → “additional hour of study”; “predicted ” → “predicted exam score”; and “on average.”
Interpretation: “For each additional hour of study, the predicted exam score increases by 3.70 points, on average.”
Intercept:; means zero hours of study. The observed range starts at , so lies below the lower boundary.
I check: is inside ? No — zero hours is below , so predicting at is extrapolation.
Interpretation: “The intercept 56.90 is not a reliable real-world prediction: (a student who did not study) falls outside the observed range of 1–8 hours, so the model was never fit to data there. Here 56.90 functions as a mathematical anchor that positions the line — not as a trustworthy baseline score.”
Example 2 — Partial Scaffold: Testing
Context: A researcher collects data on pairs and finds . Test against at .
Critical value:.
Your turn: Before looking at the solution, try substituting , into the formula .
Predict first: Do you expect this result to be statistically significant? looks strong — but does the sample size matter?
Show Solution
Step 1: vs. , .
Step 2: Conditions assumed met (data are approximately bivariate normal, independent).
Step 3:
Step 4:, so .
Step 5: Reject . There is statistically significant evidence of a linear relationship between the two variables in the population.
Note on C10: — the model explains about 56% of the variance in . This is both statistically significant and moderately practically significant.
Example 3 — Prediction Checkpoint: Interpolation vs. Extrapolation
Context: A researcher uses the model (fertilizer in grams → tomato yield in kg), fit to data with g. Two predictions are requested: g and g.
Predict the risk level before computing: Which prediction do you expect to be reliable? Which do you expect to be risky? Why?
Show Solution
g: → Interpolation (reliable).
g: → Extrapolation (risky).
The prediction of 17.35 kg may be unreliable. The linear relationship observed up to 20 g of fertilizer may not continue to 35 g — at high fertilizer levels, yield often plateaus or decreases due to nutrient toxicity. The model has no data to support linearity in this region.
Example 4 — Find the Error
A researcher uses (age in years → reaction time in ms), fit to data from adults aged 20–65 years (, , ).
The researcher reports:
Researcher’s analysis:
“The regression proves that aging causes slower reactions, confirming the biological mechanism."
"Since the p-value is 0.001, the model is statistically significant, so we can trust all predictions from it."
"A person aged 80 years will have a predicted reaction time of ms. This is a reliable clinical prediction.”
Identify all errors in the researcher’s analysis.
Show Solution
Error 1 — Causation language: Regression shows association, not causation. Saying the regression “proves aging causes slower reactions” is incorrect. The observed association could be due to confounders (e.g., health conditions correlated with age). Use “is associated with” or “predicts.”
Error 2 — Extrapolation misuse: years is outside the observed range years. This is extrapolation. The linear trend observed in adults 20–65 may not hold at age 80 — neurological and physical changes at extreme ages may create non-linearities. Reporting 312.6 ms as a “reliable clinical prediction” is incorrect.
Error 3 — Conflating statistical significance with prediction reliability: means the correlation is real in the population (not zero). It does not mean the model can be trusted for all predictions, especially extrapolated ones. Statistical significance applies to the data range used for fitting.
Note: — the model does explain 66% of variability within the observed range. But none of that applies to predictions at .
Section 5: Guided Practice
▾
Problem 1 — Slope and Intercept Interpretation
Problem 2 — Interpolation and Extrapolation
Problem 3 — Residual Plot Diagnosis
Three residual plots are described below. For each description, select the correct diagnosis.
(a) “The residuals bounce randomly above and below zero with no discernible pattern. The spread looks roughly the same for all fitted values.”
(b) “For small fitted values the residuals are positive; for middle fitted values they cluster near zero; for large fitted values they become positive again, forming a U-shape.”
(c) “For low fitted values the residuals are tightly clustered within ±2; for high fitted values the residuals range from −15 to +15.”
Problem 4 — Significance Test for
Section 6: Independent Practice
▾
Problem 1 — Full Interpretation Chain
Problem 2 — Significance Test and
Problem 3 — Find the Error
Problem 4 — Prediction Risk
Problem 5 — Multi-Step Synthesis
Mixed Review — Retrieval from Earlier Lessons
Section 7: Mastery Check
▾
Question 1 — Feynman Test
In your own words, explain why a statistically significant regression relationship is not automatically useful for prediction. Include what contributes to that judgment.
Compare with a model response
Statistical significance is evidence that the population relationship is unlikely to be exactly zero. It does not say the relationship is large enough to make accurate or useful predictions. The value gives the proportion of variation in the response explained by the model, so a small can reveal that a statistically significant model still has weak predictive value.
Question 2 — Apply
A regression model predicts course mark from weekly study hours. Its slope is 2.4 points per hour. Which is the most precise interpretation of that slope?
For n = 25 paired observations, a study finds r = −0.62. Its correlation test statistic is t ≈ −3.79, and the two-tailed critical value is 2.069. Which conclusion is correct at α = 0.05?
A regression model was fitted using x-values from 10 to 40, and its residual plot shows random scatter with roughly constant spread. A learner wants a prediction at x = 28. What is the best recommendation?
Question 3 — Error Analysis
Section 8: Boss Fight
▾
You’re the analyst
A decision-maker has sent you a data brief and wants a recommendation. The brief gives you a regression equation, the data range, a correlation result, and a residual description—but it does not tell you which method to use.
Start by working out a complete recommendation yourself. If you want support, you can ask for the steps after committing to your own analysis. “Try a new case” gives you a fresh brief, not the same numbers again.
Section 9: Challenge Problems
▾
These are cold-transfer problems. Commit to each answer before reading its feedback; the scenarios are new, so choose the method from the evidence rather than from a labelled recipe.
Problem 1 — The Effect of an Influential Point
Problem 2 — How Sample Size Affects Significance
Problem 3 — What Does Actually Measure?
Section 10: Solutions Reference
▾
Worked solutions for Section 4 and durable reference notes for the generated practice in Sections 5–9 are on the solutions page.