35

I'm trying to plot a smooth line that runs directly through all my data points and has a gradient based on another variable. Theoretically polynomial interpolation would get the job done but I'm not sure how I would do that with ggplot. This is what I've come up with so far:

DATA:

 dayofweek hour impressions conversions      cvr
         1    0     3997982       352.0 8.80e-05
         1    1     3182678       321.2 1.01e-04
         1    2     2921004       248.6 8.51e-05
         1    3     1708627       115.6 6.77e-05
         1    4     1225059        98.4 8.03e-05
         1    5     1211708        62.0 5.12e-05
         1    6     1653280       150.0 9.07e-05
         1    7     2511577       309.4 1.23e-04
         1    8     3801969       397.8 1.05e-04
         1    9     5144399       573.0 1.11e-04
         1   10     5770269       675.6 1.17e-04
         1   11     6936943       869.8 1.25e-04
         1   12     7953053       996.4 1.25e-04
         1   13     8711737      1117.8 1.28e-04
         1   14     9114872      1217.4 1.34e-04
         1   15     9257161      1155.2 1.25e-04
         1   16     8437068      1082.0 1.28e-04
         1   17     8688057      1047.2 1.21e-04
         1   18     9200450      1114.0 1.21e-04
         1   19     8494295      1086.8 1.28e-04
         1   20     9409142      1092.6 1.16e-04
         1   21    10500000      1266.8 1.21e-04
         1   22     9783073      1196.4 1.22e-04
         1   23     8225267       812.0 9.87e-05

R CODE:

ggplot(d) + 
  geom_line(aes(y=impressions, x=hour, color=cvr)) +
  stat_smooth(aes(y=impressions, x=hour), method = lm, formula = y ~ poly(x, 10), se = FALSE)

So I can get the gradient I want using geom_line but its not smooth. With stat_smooth I get a smooth line but it doesn't run through all the data points and doesn't have the gradient I want. Any ideas of how to accomplish this?

enter image description here

zx8754
  • 52,746
  • 12
  • 114
  • 209
moku
  • 4,099
  • 5
  • 30
  • 52

1 Answers1

64

A polynomial interpolation in the sense that you are using it is probably not the best idea, if you want it to go through all of your points. You have 24 points, which would need a polynomial of order 23, if it should go through all the points. I can't seem to use poly with degree 23, but using a lesser degree is already enough to show you, why this won't work:

ggplot(d) + 
  geom_point(aes(x = hour, y = impressions, colour = cvr), size = 3) +
  stat_smooth(aes(x = hour, y = impressions), method = "lm",
              formula = y ~ poly(x, 21), se = FALSE) +
  coord_cartesian(ylim = c(0, 1.5e7))

enter image description here

This does more or less go through all the points (and it would indeed, if I managed to use an even higher order polynomial), but otherwise it's probably not the kind of smooth curve you want. A better option is to use interpolation with splines. This is also an interpolation that uses polynomials, but instead of using just one (as you tried), it uses many. They are enforced to go through all the data points in such a way that your curve is continuous.

As far as I know, this can't be done directly with ggplot, but it can be done using ggalt::geom_xspline.

Here I show a base solution, where the spline interpolation is produced in a separate step:

spline_int <- as.data.frame(spline(d$hour, d$impressions))

You need as.data.frame because spline returns a list. Now You can use that new data in the plot with geom_line():

ggplot(d) + 
  geom_point(aes(x = hour, y = impressions, colour = cvr), size = 3) +
  geom_line(data = spline_int, aes(x = x, y = y))

enter image description here

Axeman
  • 32,068
  • 8
  • 81
  • 94
Stibu
  • 15,166
  • 6
  • 57
  • 71
  • Ah very cool I was looking at splines but I didn't know how to integrate them into my plot. – moku Feb 04 '16 at 18:13
  • 9
    I think you can get a similar result with `stat_smooth` by using `method = "gam"` and changing the number of knots (or dimension of the smoothed term, I guess): `stat_smooth(aes(y=impressions, x=hour), formula = y ~ s(x, k = 24), method = "gam", se = FALSE)` – aosmith Feb 04 '16 at 18:13
  • 1
    @aosmith Now it would be nice to have a way to refer to the number of values in the x axis automatically instead of having to hard-code the `k = 24` number. Any idea? – slhck Apr 30 '19 at 08:33
  • @slhck Maybe something like `k = nrow(data)` is what you need? – aosmith Apr 30 '19 at 14:13
  • 2
    With my data, I couldn't get the `gam` approach to work, and the spline approach in the answer above tended to overshoot. This answer (https://stackoverflow.com/questions/58352475/scatter-smooth-like-in-excel-ggplot2-plotly) with `spline_shape = -0.5` worked a treat. – Mark Neal Apr 25 '20 at 21:19