Assumptions of Simple Linear Regression
Checking LINE assumptions in simple regression
Simple linear regression uses a straight line for the conditional mean, while its usual small-sample inference also depends on assumptions about the errors. The mnemonic LINE prompts four distinct checks: linearity, independence, normality, and equal conditional variance. Each check asks a different question, so one reassuring display cannot certify the whole model.
What each condition means
Linearity means that E(Y∣X=x)=β0+β1x describes the mean response at each value of x. A residual is an observed response minus its fitted response; systematic movement of residual centers across x can reveal a missed mean pattern. Independence concerns relationships among errors across observations and depends heavily on sampling and measurement design.
Normality concerns the error distribution conditional on x, particularly for exact small-sample t procedures under the usual regression model. Equal conditional variance means that error spread is approximately the same across predictor values. These are modeling conditions, not requirements that the raw response have a bell-shaped histogram or that the correlation be large.
Read the right diagnostic
Plot residuals against x or a fitted value to look separately at their center and spread. A changing center suggests a mean-model problem; a widening vertical band suggests changing variance. A normal quantile plot examines error-shape plausibility through the fitted residuals, while a plot against collection order can expose temporal patterns.
No plot alone establishes independence; the sampling design also needs scrutiny. When spread varies, ordinary constant-variance standard errors can be misleading; variance-aware inference or a better variance model may be appropriate.
Worked example
Imagine an instrument calibration with independent specimens measured once at each of several input levels.
If a fitted response is 12 units and the observed response is 14 units, its residual is 14−12=2 units; an observed response of 13 at a fitted value of 15 gives 13−15=−2 units.
Plot all residuals against input level and inspect whether their centers stay near zero and their vertical spreads remain comparable, rather than deciding from these two points.
A residual-versus-fitted plot is centered near zero at all fitted values, but its vertical spread grows steadily to the right. What is the best assessment?
Researchers measure the same patient each week for 12 weeks, then fit ordinary simple linear regression of a lab measure on week. The residual histogram is roughly bell-shaped. Which concern remains especially important?