Linear Regression

Linear regression models a numerical outcome as a weighted sum of input features plus an intercept. Training chooses coefficients that minimize a measure of prediction error, commonly squared error. The coefficient for each feature describes how the prediction changes when that feature changes while the others are held constant, though this should not be interpreted as causation without a valid research design. Linear regression is fast, interpretable, and valuable as a baseline, but it assumes the selected representation can express the relevant relationship. Outliers, correlated features, nonconstant error variance, missing interactions, and distribution shifts can weaken both predictions and statistical conclusions.