Assumptions for Multiple Linear Regression
What you'll learn: You'll review the core assumptions that must hold for multiple linear regression to produce valid, trustworthy results.
The Same Foundation, Extended
Remember the five core assumptions you learned for simple linear regression? They all still apply when you add more predictors. Multiple linear regression isn't a different framework—it's the same model extended to handle multiple X variables simultaneously.
Let's review each assumption in the multiple regression context:
1. Linearity
The relationship between each predictor and the response must be linear. With multiple predictors, the combined effect should also be linear. You're still fitting a flat surface (or hyperplane), not a curved one.
2. Independence
Observations must be independent of each other. Adding more predictors doesn't change this—each data point should still provide unique information.
3. Homoscedasticity
Residual variance must remain constant across all fitted values. Even with multiple predictors creating a more complex fitted surface, the spread of errors should stay consistent.
4. Normality
Residuals should follow a normal distribution. This assumption concerns the errors, not the predictors themselves, and remains unchanged with multiple X variables.
5. No Perfect Multicollinearity
Here's where multiple regression gets trickier: your predictors shouldn't be perfectly correlated with each other. If one predictor is an exact linear combination of others, the model breaks down mathematically—you can't uniquely estimate coefficients. (You've already explored detecting this with VIF and correlation matrices.)
Why These Assumptions Matter
Violating these assumptions leads to biased estimates, inflated standard errors, incorrect p-values, and unreliable predictions. The diagnostics you learned (residual plots, Q-Q plots, VIF) help you verify these conditions hold.
Key Takeaway: Multiple linear regression relies on the same five core assumptions as simple regression: linearity, independence, homoscedasticity, normality of residuals, and—critically—no perfect multicollinearity among predictors.