StatCrunch Residual Plots- Step-by-Step Guide
What Residual Plots Actually Are
A residual plot shows the difference between what your regression model predicts and what the actual data points are. Each dot on the plot represents one observation. The horizontal axis shows your predicted values. The vertical axis shows the residuals (errors).
That's it. Nothing fancy. Just a scatter plot of prediction errors.
Why You Can't Skip These Plots
Regression output gives you R-squared values and p-values. Those numbers lie sometimes. Residual plots show you whether your model actually works.
Here's what residual plots catch:
- Non-linear patterns your model missed
- Unequal variance (heteroscedasticity)
- Outliers messing up your coefficients
- Data entry errors
Your p-value could be 0.001 and your model could still be garbage. Residual plots tell you the truth.
Creating Residual Plots in StatCrunch
Step 1: Run Your Regression
Go to Stat > Regression > Simple Linear (or whichever regression type fits your data).
Select your Y variable (response) and X variable (predictor). Click Calculate.
Step 2: Access the Residual Output
After your regression runs, look at the results table. Click the blue link that says "Save" next to the residuals column.
A menu opens. Choose To worksheet. This adds the residuals as a new column in your data table.
Step 3: Plot the Residuals
Go to Graph > Scatter Plot.
For the X variable, select your predicted values (the fitted values from the regression). For the Y variable, select the residuals you just saved.
Click Calculate. That's your residual plot.
Step 4: Add a Reference Line (Optional But Helpful)
On your scatter plot, go to Options > Edit.
Under Horizontal Lines, add a line at 0. This makes it easier to see if residuals are evenly distributed above and below the center line.
Reading Your Residual Plot
Here's what to look for:
Good Pattern: Random Scatter
Points should look like a cloud with no obvious shape. They're spread evenly above and below the zero line. This means your model assumptions hold.
Bad Pattern: U-Shaped or Curved
Residuals curve upward or downward. Your model missed something nonlinear. You need a quadratic term or a different model type.
Bad Pattern: Funnel Shape
Points fan out as you move right. The spread gets bigger. This is heteroscedasticity—your predictions get less accurate at higher values. This affects your p-values.
Bad Pattern: Outliers
Points far from the cluster need investigation. One bad data point can tank your entire regression. Check if it's a typo or if it represents something real.
Bad Pattern: Points All on One Side
If residuals are mostly positive or mostly negative in one area, something's wrong with your model specification.
What Your Residual Pattern Means
| Pattern | What It Means | What To Do |
|---|---|---|
| Random scatter around 0 | Model fits well | Nothing. You're done. |
| U-shaped curve | Missing nonlinearity | Add polynomial terms or transform variables |
| Inverted U (hill) | Missing nonlinearity | Same fix—add curvature to your model |
| Wide funnel (right side spreads out) | Unequal variance | Use weighted regression or transform Y |
| Narrow funnel (right side squeezes) | Inverse variance pattern | Consider log transformation of Y |
| Isolated points far away | Outliers present | Verify data, consider robust regression |
| Most points above zero in one region | Model misspecified | Check if you need different predictors |
Quick Diagnostic Checklist
Every time you run a regression, check your residual plot against these five questions:
- Are points roughly symmetric above and below zero? ✅ or ❌
- Does spread stay roughly constant across all X values? ✅ or ❌
- Is there any obvious curve or pattern? ✅ or ❌
- Are there any points that look way off? ✅ or ❌
- Do about half the points fall above and half below zero? ✅ or ❌
All five should be yes. If any are no, your model needs work.
The Bottom Line
StatCrunch makes residual plots easy to generate. The steps take about two minutes once you know where to click. The interpretation takes practice, but the rules above cover 90% of what you'll encounter.
Run your regression. Save the residuals. Plot them against fitted values. Look for patterns. If you see randomness, your model is probably fine. If you see structure, dig deeper.
That's the whole game.