# Points Outside Linear Regression Confidence Band

I have done a linear regression of predicted measurements (of my model) vs. observed measurements and plotted the confidence band. Can I draw any conclusions about the points that lie outside the band?

If I am interpreting confidence bands correctly, if a point does not lie within the confidence band it means that there is 95% chance that its not within the range of the mean predicted value for that specific $x$ value (observed measurement) and nothing else (I cannot say anymore).

• A confidence interval isn't an interval for a point. Keep in mind the distinction between a confidence interval for the mean and a prediction interval for an observation. e.g. see here. In addition, even if every assumption was exactly correct, you'd still expect some points outside a prediction interval. [Of course if you're looking at points you used in the fitting an appropriate interval would be different from a PI, and its meaning is somewhat trickier.] – Glen_b May 18 '17 at 7:55

No, you essentially cannot infer anything from a data point lying outside the confidence band.

I think your interpretation of the confidence and the prediction bands may be off.

• The 95% confidence band is a band that contains the true unknown mean response for a particular predictor value 95% of the time if you were to repeat your experiment many, many times.
• The 95% prediction band is a band that contains 95% of future observable realizations if you were to repeat your experiment many, many times.

Note the difference: the confidence band applies to unobservable parameter estimates, the prediction band to observables. The confidence band only includes uncertainty in estimating the mean; the prediction band includes both this uncertainty and residual variation around this mean. You may want to look at the tag wikis for and .

Here is an illustration. Note how the confidence band gets smaller as we increase the number $n$ of observations, because we can estimate the mean more and more precisely. Conversely, the prediction band does get smaller, but not so much, because while the uncertainty around estimating the mean gets smaller, the residual variation stays the same.

opar <- par(mfrow=c(2,2))
for ( nn in c(20,80) ) {
set.seed(1)

xx <- sort(runif(nn,-1,1))
yy <- 0.5*xx+0.2*rnorm(nn)
model <- lm(yy~xx)
conf <- predict(model,interval="confidence")
pred <- predict(model,interval="prediction")

plot(xx,yy,type="n",ylim=c(-1,1),main=paste(nn,"data points, with confidence band"))
polygon(c(xx,rev(xx)),c(conf[,"lwr"],rev(conf[,"upr"])),col="lightgrey",border=NA)
abline(model)
points(xx,yy,pch=19)

plot(xx,yy,type="n",ylim=c(-1,1),main=paste(nn,"data points, with prediction band"))
polygon(c(xx,rev(xx)),c(pred[,"lwr"],rev(pred[,"upr"])),col="darkgrey",border=NA)
abline(model)
points(xx,yy,pch=19)
}
par(opar)


(Note that you'd typically not use a prediction interval for those observations you trained your model on, and R rightly complains about this. Conversely, looking at the confidence interval in-sample makes perfect sense.)

• Maybe my misunderstanding, but your description of a prediction interval doesn't seem correct to me. To simplify what I mean, consider a single estimation at a particular x, as opposed to the full regression band. My understanding of a prediction interval is that if you do your experiment, form your interval, then make a single subsequent observation, 95 % of these will be within the interval - when the whole process is repeated. Which is not quite the same as saying that 95 % of all subsequent observations will be in the interval - the latter is the definition of a tolerance interval. – Mooks May 18 '17 at 11:47