I have measuments of vegetation coverage on Y plotted against surface height (and hence flooding frequency) on X. The vegetation often has two herb layers, which are estimated seperately. If only one layer is present, the coverage of the upper layer is 0 (and if no vegetation is present, both are 0). Therefore, the upper layer has many zero estimates in the graph.

There are two treatments: grazed and non-grazed (exclosure).

I've plotted a loess curve with 95%CI to show the general trend and differences between the treatments. However, I know Loess is a non-parametric method, but was wondering if it's usage in this case is correct, especially for the high vegetation layer.

Can I use a loess curve with confidence interval regardless of the number of zeros? If not, how can I check if it is appropriate in this case?

enter image description here

  • $\begingroup$ Given that your data are bounded between 0 and 1 I would recommend using either a GAM (mgcv package) with binomial error structure or the locfit package (locfit or locfit.raw functions) with family=binomial (this fits a local polynomial using max likelihood on a logit link scale). In the latter case you would still need to use cross validation to determine the optimal span. GAM models have inbuilt adaptive smoothing, so are more convenient in that sense, and they would also allow extra factors to be added... In both cases you would need to use the raw counts that your proportions are based on. $\endgroup$ – Tom Wenseleers Aug 22 '18 at 17:00
  • $\begingroup$ @Tom I think you have misunderstood my data structure. There are no counts; the dots in my figures are the raw data and represent percentage coverage of the vegetation (80% means that there is 20% bare soil, uncovered by vegetation). $\endgroup$ – RHA Aug 22 '18 at 18:14
  • $\begingroup$ Yes I gathered that - but is this percentage cover still not calculated say from GIS imagery, and ultimately not based on counted pixels in the GIS image or something? If so those raw counts should be used... Alternatively you can also work with your percentages but then you have to do your smoothing on a transformed scale (e.g. a logit transformed Y scale, but taking care to replace 0 with e.g. 1E-5 and 1 by 0.99999), after which you can backtransform your fit to the original scale. This might require specifying appropriate weights to take into account non-homogeneous variances. $\endgroup$ – Tom Wenseleers Aug 22 '18 at 19:56
  • 1
    $\begingroup$ Ha well no problem - but in that case it would be best to do your smooth on a logit transformed Y scale (qlogis() in R) and then backtransform using the inverse logit (plogis() in R) - then you'll never get predictions that go outside the allowable [0%,100%] domain... $\endgroup$ – Tom Wenseleers Aug 22 '18 at 20:24
  • 1
    $\begingroup$ If you do your smooth using a GAM you can use the "by=" syntax to produce different curves for your different combinations of your factors, see stat.ethz.ch/R-manual/R-devel/library/mgcv/html/gam.models.html for some examples $\endgroup$ – Tom Wenseleers Aug 22 '18 at 23:40

A Loess confidence interval doesn't mean much unless the Loess parameters have been cross-validated (which usually is not the case). When you use Loess for exploration, as it was originally intended, understanding how to control it will help you guide your exploration and interpret its results better.

Consider this small study of a synthetic dataset which has only $0$ or $1$ as responses: it is an extreme example of your situation. The data, plotted as black points, are outcomes of Bernoulli$(p)$ variables ("coin flips") where $p$ varies in a damped sinusoidal manner with the horizontal coordinate $x$, as shown by the white reference curve in each panel. The panels vary only by the "span" of the Loess smooth, which determines how local each Loess estimate is: smaller spans produce estimates that are more localized; that is, they reflect the responses for the closest neighbors of each $x$ value much more than for distant neighbors. The smooth is shown in blue and its surrounding confidence band in dark gray.


The lefthand panel uses the default span of $0.75$. This causes the Loess estimate at each point to depend on most of the points in the plot: it is a heavy smooth for these data. In many cases the white plot lies outside the shaded confidence band, showing this confidence band may be misleading.

It is clear that only with the final span of $0.25$ does the smooth come at all close to the true values: here, the white graph is contained within the shaded gray area. Unfortunately, in practice we do not have access to any true underlying curve: that's precisely what we're trying to estimate.

All three of these smooths are perfectly valid, insofar as they are efforts to sketch out the overall trend in the response ("y") relative to the regressor ("x"). The heavy smooth at the left suggests the response rate is approximately stable (which, on average, it is). The lighter smooth at the right captures higher-frequency variation. In practice, it might not be apparent whether what it shows is "real" or is "noise."

In practice, we never accept just one default level of smoothing: we vary the amount of smoothing, exactly as illustrated here, in order to learn about the data at varying levels of local resolution. We might also vary the smoothing in order to create different kinds of visual descriptions of the data, guiding the viewer's eye to global trends (as at the left) or local behaviors (as at the right), as we see appropriate.

The best tool for "checking appropriateness" is to study the residuals of the smooth in the context of a particular analytical or visualization objective. Good books on Exploratory Data Analysis, such as John Tukey's EDA, provide a wealth of techniques for computing and analyzing smooths and their residuals.

If you would like to experiment, here is the R code that created these illustrations.

# Generate data.
n <- 2e2
x <- 1:n
p <- (sin(x/100 * 2*pi)^2 - 1/2)*exp(-x/n) + 1/2
y <- rbinom(n, 1, p)
df <- data.frame(x=x, y=y, p=p)
# Set up for drawing.
spans <- c(0.75, 0.5, 0.25)
k <- length(spans)
viewports <- lapply(1:k, function(i) 
  grid:::viewport(width=1/k, height=1, x=(i-1/2)/k, y=1/2))
names(viewports) <- spans
# Create the plots.
g <- ggplot(df, aes(x, y)) + geom_point(aes(x,y), df, alpha=0.25) + 
for (i in 1:k) {
  print(g + geom_smooth(method="loess", span=spans[i])  +
    geom_line(aes(x,p), df, color="White", lwd=1) + 
    labs(title=paste("Span =", spans[i])),


John W. Tukey, EDA. Addison-Wesley, 1977.

| cite | improve this answer | |
  • 1
    $\begingroup$ Thank you. I was already fiddling with this span parameter. Btw: nice way of plotting with viewports, if facetting is impossible. I've learned two things today :-) $\endgroup$ – RHA Feb 12 '17 at 21:35
  • $\begingroup$ My first thought was to use faceting, but it seemed like it might be too cumbersome for this purpose. If you know of a relatively simple way to create these graphics using faceting, I would like to hear of it. $\endgroup$ – whuber Feb 12 '17 at 21:40

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.