# Predicition interval for LOESS smoothed data

Is the concept of a prediction interval for non-parameteric fits such as LOESS meaningful? I can't see why it wouldn't be - and the fact that it is implemented in R also implies it is.

https://mclust-org.github.io/msir/reference/loess.sd.html

But I can't find any reference on how to derive it (including in the book referenced in the link above) however, which makes me think I am missing something - unlike for parameteric modelling where the prediction interval is heavily documented.

I would like to derive prediction intervals for a various fitting techniques, including LOESS, which is why I'd like to understand it better- guidance welcome

• You're correct. The concept of a prediction interval is independent of any model you might be working with.
– whuber
Jan 19, 2022 at 22:44
• Perfect thanks - do you know of any references that explain how to go about the implementation of the prediction interval with LOESS? My googling wasnt very productive.. Jan 20, 2022 at 9:38
• Loess is an exploratory tool. Any prediction interval offered by software ought to be viewed as a kind of ad hoc sketch that might, on occasion, be useful as an indicator of the magnitude of local variation. If you need a true nonparametric prediction interval related to Loess, then bootstrapping would be an attractive method.
– whuber
Jan 20, 2022 at 15:15

For reference, the specific function you've linked basically calculates the non-parametric variance function of the Loess fit.

The intervals constructed are referred to as variance smooths by Cook & Weisberg (1999, pp. 52 - 53) for the Loess, and are calculated as:

$$\hat{E}(y | x) \pm \sqrt{\widehat{Var}(y | x)}$$

Where $$\hat{E}(y | x)$$ is the estimated mean function and $$\sqrt{\widehat{Var}(y | x)}$$ is the estimated standard deviation function, calculated by a procedure as follows (outlined in Section 2.5.2 of the same text, pp. 36 - 37):

1. Calculate $$E(y | x)$$ using Loess with a given smoothing parameter, $$\alpha$$

2. Create a new variable representing the squared deviation of each observation from the mean function

$$\hat{e}^2 = (y - \hat{E}(y | x))^2$$

1. Calculate $$\hat{E}(\hat{e}^2 | x)$$ using Loess with the same $$\alpha$$ - this is our estimate of $$\widehat{Var}(y | x)$$ and conceptually is like smoothing a scatterplot of the squared deviations on $$x$$

2. Take the square root of the variance function to calculate the standard deviation function

If you dig into the loess.sd function in msir, you'll find this is basically what it does (also allowing you to scale the standard deviation function using the nsigma argument).

But I'll work through an example manually using R to demonstrate. I'm using the cholesterol data from the bootstrap package which evaluated the effects of cholostyramine compliance on blood cholesterol levels.

# Load the Data
cholesterol <- bootstrap::cholost
colnames(cholesterol) <- c("compliance", "improvement")
with(cholesterol, plot(compliance, improvement))

# 1. Fit loess to find mean function E(y | x)
alpha <- 0.6
m1 <- loess(improvement ~ compliance,
data = cholesterol,
span = alpha)

# Generate and plot predictions from mean function
new_dat <- data.frame(compliance = seq(0, 100))
preds <- predict(m1, new_dat)
lines(seq(0, 100), preds)

# 2. Find squared deviations
cholesterol\$e <- residuals(m1)^2

# 3. Fit loess to find variance function E(e^2 | x)
m2 <- loess(e ~ compliance,
data = cholesterol,
span = alpha)

# 4. Estimate Std. Deviation function across values of x
sd <- sqrt(predict(m2, new_dat))

# Find upper and lower limits and plot
ll <- preds - sd
ul <- preds + sd
lines(new_dat$$compliance, ll, lty = 2) lines(new_dat$$compliance, ul, lty = 2)


Which graphically looks like this:

#### References

Cook, R. D., & Weisberg, S. (1999). Applied regression including computing and graphics. John Wiley & Sons.

• Thats fantastic thanks. I'm actually having to implement it in .NET, not R which is why the explanation and references are particularly useful. Cheers! Jan 20, 2022 at 11:09