7SSGN110 Chap.8 Linear Regression and Model Diagnostics
Linear Regression and Model Diagnostics
Prediction begins with a fitted relationship
Simple linear regression represents a response with an intercept, a slope times one predictor and a residual component. Least squares chooses coefficients that minimise squared vertical residuals. Before fitting, name both variables and units, confirm the independent observational unit and record the observed predictor range.
The line describes the sampled design; it does not by itself justify a population or causal claim.
Interpret coefficients with scale
The slope is the expected response change for one predictor-unit increase, while the intercept is the predicted response when the predictor equals zero. An intercept may lack environmental meaning if zero sits far outside the data.
Rescaling metres to kilometres changes the numerical slope but not the relationship. State the subtraction and unit conversion explicitly so signs and magnitudes remain consistent in equations, figures and prose.
Link predictions with residuals
A fitted value lies on the model line at a specified predictor value.
A residual is observed response minus fitted response, so a negative residual places an observation below the line. Keep that convention fixed. RMSE expresses typical residual size in the response unit, and R squared describes the share of sample response variation accounted for by the model.
Neither statistic shows that the mean relation is linear or that uncertainty is well calibrated.
Read diagnostics as different questions
Residuals versus fitted values reveal curvature and changing spread. A QQ plot checks whether residual shape is compatible with the normal model used for inference.
Scale-location views emphasise variance patterns, while leverage and Cook's distance identify observations capable of changing the fit. One repair does not answer every diagnostic.
Refit and recheck after a transformation, variance adjustment or data correction.
Protect the range and the causal boundary
Prediction outside the observed predictor range assumes that the same line continues where no data were collected. Flag extrapolation or decline to use the model.
A prediction interval for a new observation is wider than an interval for the mean response because it includes individual residual variation. Influential sites should be verified and examined through sensitivity analysis, not deleted automatically.
Report the slope, uncertainty, fit, diagnostic conclusion, sampled range and remaining confounding in one coherent results paragraph.
Make the model reproducible
Archive the prepared data, model call, coefficient table, diagnostic plots and prediction inputs together. A reported prediction should be recoverable from the exact coefficient version used in the text.
If sensitivity analysis changes the slope or conclusion, retain both fits and explain the final choice. Keep factor references, transformations and missing-record rules explicit. This audit prevents a polished figure from being generated by a different model than the one whose statistics appear in the results.
For revision, move between equation, fitted point, residual and diagnostic plot without software.
Change one predictor value and predict how the fitted response moves; then introduce curvature or a funnel and name the challenged assumption. Finish by checking whether every proposed prediction lies inside the observed range.
What this chapter covers
- 01
Intercept
- 02
Slope
- 03
Least squares
- 04
Residual
- 05
R squared
- 06
RMSE
- 07
Diagnostic plots
- 08
Extrapolation
Worked example · free
Predict and calculate a residual
- 3Substitute 400 into the fitted equation.
- 2Calculate observed minus fitted.
- 3Interpret slope and residual with units and scope.
Key terms
- Regression Slope
- A Regression Slope is the expected response change for one predictor-unit increase under the fitted linear model.
- Regression Residual
- A Regression Residual is the observed response minus its fitted value.
- Cook's Distance
- Cook's Distance measures how strongly deleting an observation would alter the fitted model.
Linear Regression and Model Diagnostics FAQ
What does the regression slope mean?
The slope is the fitted change in response for a one-unit increase in the predictor, with all units stated and within the model's scope. It describes association unless the design and assumptions support a causal interpretation.
What should residual plots look like?
Residuals versus fitted values should form a roughly patternless band around zero with stable spread. Curvature suggests missing nonlinearity, a funnel suggests heteroscedasticity, and isolated high-leverage points require verification and sensitivity analysis rather than automatic deletion.
Why is extrapolation risky?
A fitted line is supported over the observed predictor range. Outside that range, the relationship may bend, reach a physical boundary or reflect a different process. Flag extrapolation explicitly and avoid point predictions that imply unsupported certainty.
Assessment move
For each model, write the equation with units, hand-calculate one prediction and residual, then inspect every diagnostic plot before reading R squared. Practise matching curvature, funnels, QQ departures and leverage to a specific assumption and possible repair.
Working through Linear Regression and Model Diagnostics in 7SSGN110? Sia is AskSia’s AI Statistics tutor — ask any 7SSGN110 Linear Regression and Model Diagnostics question and get a clear, step-by-step explanation grounded in how 7SSGN110 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.