What does the intercept term in a fitted linear regression model represent?

## The literal definition A fitted linear model is `y_hat = b0 + b1*x1 + b2*x2 + ... + bk*xk`. Set every predictor to zero and everything but `b0` vanishes, so the intercept is by definition the **predicted value of the outcome when all predictors equal zero**. That is the whole of the definition; everything else is about whether that statement is worth saying. ## When it is meaningless, and why that is fine Regress house price on square footage. The intercept is the predicted price of a zero-square-foot house. There are no such houses, zero is far outside the observed range of the predictor, and the fitted line has no obligation to behave sensibly out there - it is a straight line extended into territory the data never visited. The intercept can even be negative, giving a "negative price", which alarms people who think it is a claim about the world. It is not a claim about the world. It is the height at which the line must sit so that the slope fits the data in the region where the data live. Predictions at realistic square footages are completely unaffected. The right response to an odd-looking intercept is to say what it is - an anchoring constant, not an estimate of anything observable - rather than to change the model. The intercept becomes directly interpretable only when zero is a real, populated value of every predictor: hours of advertising spend, number of prior purchases, minutes of downtime. Then "the predicted outcome with no advertising at all" is a sentence someone can act on. ## Centring: making the intercept mean something Subtract each predictor's mean from it before fitting. The predictor `x` becomes `x - mean(x)`, so a value of zero on the new scale is the average value on the old one. The slopes do not change at all - shifting a predictor moves only the intercept - but the intercept now reads as **the predicted outcome for a case at the average of every predictor**. In OLS with an intercept, that value is exactly the sample mean of the outcome, because the fitted surface passes through the point of means. This is a free improvement in readability. You keep the same fit, the same residuals, the same R-squared and the same t-statistics, and you gain an intercept that can be quoted to a stakeholder: "a typical house in this market is predicted at 420,000 dollars, and each extra square foot adds about 210." A related trick is to centre at a meaningful reference rather than the mean - a policy threshold, a baseline year, a target headcount - so the intercept reads as the prediction at that reference point. ## Do not drop the intercept A common error is to remove the intercept because it "has no meaning" or because it is not statistically significant. Fitting without an intercept forces the regression line through the origin, which imposes the substantive claim that the outcome is exactly zero when the predictors are zero. Unless that claim is genuinely true and important, the constraint pulls the line and biases the slopes, sometimes dramatically. It also breaks the usual residual bookkeeping: the residuals no longer sum to zero, and R-squared as reported for a no-intercept model is computed differently and is not comparable with the ordinary one. The intercept's own p-value is usually uninteresting for the same reason its value often is: testing whether the predicted outcome at an unobservable all-zero point differs from zero rarely answers a question anyone asked. ## Reading the intercept with several predictors With many predictors, the intercept requires **all** of them to be zero simultaneously. That combination is far less likely to be observed than any single zero, so with a handful of predictors the intercept is almost always extrapolation. Centring helps again here: centre all of them and the intercept describes the average case on every dimension at once, which is a point that at least sits in the middle of the data cloud. ## What to say in an interview Give the definition first, then immediately qualify it. "The intercept is the predicted outcome when every predictor is zero. In this model that is a zero-square-foot house, so I would not quote it as a finding; it is the constant that positions the line. If I needed an interpretable constant I would centre the predictors, which leaves the slopes untouched and makes the intercept the prediction for an average case. I would not drop it - a through-the-origin fit imposes an assumption I have no reason to make." That answer shows you can distinguish an estimate of something real from a piece of machinery, which is the entire point of the question.