In a residual-vs-fitted plot from an OLS fit, what does a clean U-shaped curve indicate?

## What you are looking at Each point is one observation: horizontal position is the model's prediction `yhat_i`, vertical position is the residual `e_i = y_i - yhat_i`. A U shape means the smoother through the cloud dips below zero in the middle of the fitted range and rises above zero at both ends. Because a residual is observed minus predicted, positive residuals mean the model predicted too low. So the U reads as: **too low at both ends, too high in the middle**. An inverted U (a frown) is the same defect mirrored — too high at the ends, too low in the middle. ## Why it is a specification problem, not a noise problem Least squares removes the *linear* part of the relationship between the residuals and everything in the model. What it cannot remove is a shape the model has no way to express. If the truth is `y = f(x) + error` with `f` curved, and you fit a straight line, the leftover curvature has nowhere to go and appears in the residuals as a systematic arc. The assumption that fails is the one that gives the coefficients their meaning: that the expected error is zero at every value of the predictors. Once the expected residual depends on where you are in the fitted range, the estimated slope is not an estimate of a stable relationship — it is the slope of the best straight-line approximation to a curve, and it depends on where your data happen to be dense. ## Why this is worse than most diagnostic findings It is worth grading violations by what they damage: - A wrong mean function damages the **estimates** themselves. Predictions are systematically wrong in whole regions, and extrapolating past the data range makes it worse, sometimes dramatically. - Most other diagnostic findings damage the **uncertainty** around the estimates while leaving them centred correctly. Curvature therefore usually goes to the top of the fix list. It also generates false confidence: because the fit still passes through the middle of the cloud, the overall R-squared can be high, and a summary table gives no hint that the model is wrong in a patterned way. ## Diagnosing which predictor is responsible The fitted value blends every predictor, so a U shape tells you the model is curved somewhere without saying where. The standard next step is to plot the residuals against each predictor in turn. Whichever one shows the same arc is the culprit. Plotting residuals against a variable you did *not* include is equally useful: a clear pattern there is evidence of an omitted variable. A subtlety: a U shape can appear even when every predictor genuinely enters linearly. A missing interaction, or an omitted variable correlated with the fitted value, can bend the residual cloud. So the finding is really 'the mean function is incomplete', of which 'this predictor needs a curve' is only the most common cause. ## How to respecify Options, roughly in order of how much structure they assume: 1. **Add a squared term** in the suspect predictor. Simple, interpretable, and it captures exactly one bend. Keep the linear term in as well. 2. **Transform the predictor** — a logarithm is natural when the effect of the predictor is proportional rather than additive, or when its distribution is strongly right-skewed. 3. **Transform the response** — appropriate when the model is really multiplicative. Note this changes what the coefficients mean and also changes the error structure, so re-check the diagnostics afterwards rather than assuming you fixed one thing without moving another. 4. **Add a spline** or another flexible term when the shape has more than one bend or you do not want to commit to a functional form. 5. **Add the missing interaction or variable**, when the curve is a symptom of the model being incomplete rather than of one predictor being curved. ## Traps when acting on the finding - **Do not delete the points at the ends.** They are not outliers; they are the evidence. Removing them hides the curve and leaves you with a model that is still wrong and now also fitted on a truncated range. - **Do not chase every wiggle.** With enough points a smoother will always show something. What matters is a systematic, monotone-then-reversing shape that is large relative to the vertical spread of the residuals. - **Beware polynomial degree.** High-degree polynomials behave violently at the edges of the data and are a common way to trade a visible bias for an invisible variance problem. - **Re-run the full check after refitting.** A quadratic term can flatten the residual plot while the spread or the tail behaviour still needs attention. ## The one-line answer A U in the residual-vs-fitted plot says the model's shape is wrong, the predictions are biased by region rather than merely noisy, and the fix belongs in the specification of the mean function.