# How to interpret Residuals vs. Fitted Plot

I am investigating the effects of weather on restaurant demand. Currently, I am testing the model assumptions for my multiple linear regression model.

My model specification (simplified) is as follows: lm(Visitor ~ Temperature + Temperature_Squared + Pressure + Clouds + Sun + Rain + Day_Fri + Day_Sat + Day_Sun + Day_Mon + Day_Tue + Day_Wed + Hour_00 + Hour_01 + Hour_02 + Hour_13 + Hour_14 + Hour_15 + Hour_16 + Hour_17 + Hour_18 + Hour_19 + Hour_20 + Hour_21 + Hour_22 + Hour_23 + Holiday, data=dat)

After running the model, I obtained the following two graphs:

1. The residuals vs. fitted plot appears to be relatively flat and homoskedastic. However, it has this odd cutoff in the bottom left, that makes me question the homoskedasticity. What does this plot signal and, more importantly, what does it mean for my interpretation? Is multiple linear regression the correct model?

2. How do I interpret the "bump" in the top-right part of the QQ plot?

NB: The data is complete and does not have unreasonable outliers. Initial results indicate only 1 (out of 6) IVs to be significant, while all control variables are significant. Also, no issues with multicollinearity were detected.

• Visitors is a count variable with support $\{0, 1, \ldots\}$ but the OLS assumes it is normal with support $(-\infty,\infty)$. For low predicted (fitted) visitor counts, the prediction error (residual) can only get so low, hence the cutoff in the plot. A more apt specification might be a Poisson regression or another regression model based on a count outcome. Nov 6, 2019 at 16:57
• Probability plots often 'wobble' towards the extremes because data is relatively sparse there. Nov 6, 2019 at 17:23
• You probably should mention that you're using the R language, for those few folks at this site who don't recognize it. :-) Nov 7, 2019 at 14:35

Both the cutoff in the residual plot and the bump in the QQ plot are consequences of model misspecification.

You are modeling the conditional mean of the visitor count; let’s call it $$Y_{it}$$. When you estimate the conditional mean with OLS, it fits $$E(Y_{it}\mid X_{it})=\alpha+\beta X_{it}$$. Notice that this specification assumes that if $$\beta>0$$, you can find a low enough $$X_{it}$$ that pushes the conditional mean of the visitor count into the negative region. This however cannot be the case in our everyday experience.

Visitor count is a count variable and therefore a count regression would be more appropriate. For example, a Poisson regression fits $$E(Y_{it}\mid X_{it})=e^{\alpha+\beta X_{it}}$$. Under this specification, you can take $$X_{it}$$ arbitrarily far towards negative infinity, but the conditional mean of the visitor count will still be positive.

All of this implies that your residuals can't by their nature be normally distributed. You seem to not have enough statistical power to reject the null that they are normal. But that null is guaranteed to be false by knowing what your data are.

The cutoff in the residual plot is a consequence of this. You observe the cutoff because for low predicted (fitted) visitor counts the prediction error (residual) can only get so low.

The bump at the end of your QQ plot also follows from this. OLS underpredicts in the right tail because it assumes that the relationship between $$X_{it}$$ and the outcome is linear. Poisson would assume it is multiplicative. In turn, the right tail of the residuals in the misspecified model is fatter than that of the normal distribution.

I think @BruceET is making a good point that a “wobble” is natural for any estimator, and the question is whether the wobble is outside of a valid confidence bound. But in this case it also signals model misspecification.

Here are a dozen normal probability plots in R, each for a sample of size 100 from a known standard normal population. Each plot is roughly linear, but most have a 'wobble' or two, especially toward the extremes.

set.seed(116)
par(mfrow=c(3,4))
for(i in 1:12) {
z = rnorm(100); qqnorm(z, pch=20) }
par(mfrow=c(1,1))


Repeat the code (without the set.seed statement) for more examples. Examples of normal probability plots in textbooks seem, on average, to be better behaved than the plots one typically sees in practice -- even when normality assumptions are very nearly true.

Addendum: Six additional plots with reference lines as suggested in Comment by @Henry.

set.seed(117)
par(mfrow=c(2,3))
for(i in 1:6) {
z = rnorm(100)
qqnorm(z); qqline(z, col=2) }
par(mfrow=c(1,1))


• It might be worth adding qqline(z) into your loop Nov 7, 2019 at 7:12
• @Henry, Tried that, but in the small plots I thought it looks too cluttered. Will experiment. // Works fine with fewer, larger plots. Nov 7, 2019 at 8:31

Let's assume "visitors" is the total number of visitors and thus whole positive numbers. Let us assume, the model predicts zero visitors and there are zero visitors,then the residual is zero. If there are more then zero visitors, the the residuals must be positive. If the modell predicts a negative number of visitors, then the residual must be at least of absolute value as the prediction.

In general: as the visitors are bound to a positive or zero value, there is a lower limit to the residuals.

The bump in the QQ plot is minimal and probably not worth worrying about with regards to regression assumptions.

• The bump is because the right tail is fatter than it is for the normal distribution. This would be the case if the visitor count was conditionally Poisson distributed. The bump and the cutoff are signs that OLS is not the right model. Nov 6, 2019 at 17:44
• @Student is right except that OLS is not a model; it's an estimation method. Nov 6, 2019 at 17:46
• Linear regression model works for me. You just need to spell out the assumptions, or as I often prefer to say, the ideal conditions. Nov 6, 2019 at 21:08
• @NickCox I learned about your "ideal conditions" terminology only recently and I like it. Nov 6, 2019 at 21:22
• So that’s you and me on board. All we need now is the rest of the world to join in. Nov 6, 2019 at 21:30