5
$\begingroup$

I have been given a set of data points $(x_i,y_i)$. I have to plot a scatter plot and determine if there are any outliers. But I haven't been taught a method to measure which data point is an outlier and which is not. So how can I do it for example in Sage or R? I found by Google that there is at least two tests to do that, Dixon's and Grubbs's test, so which one should I learn in this problem?

x = c(1,34,6,47,10,49,23,32,12,16,29,49,28,8,57,9,31,10,21,26,31,52,21,8,18,5,18, 
     26,27,26,32,2,59,58,19,14,16,9,23,28,34,70,69,54,39,9,21,54,26) 
y = c(47,76,33,78,62,78,33,64,83,67,61,85,46,53,55,71,59,41,82,56,39,89,31,43,29,55, 
     81,82,82,85,59,74,80,88,29,58,71,60,86,91,72,89,80,84,54,71,75,84,79)
$\endgroup$
  • $\begingroup$ post some data and you'll get interesting answers. $\endgroup$ – user603 Mar 25 '13 at 11:27
  • $\begingroup$ possible duplicate of Is there a simple way of detecting outliers? - if the responses there, or here or here (since you describe the data as $x,y$ pairs), don't answer your question, please be more specific. $\endgroup$ – Macro Mar 25 '13 at 13:45
  • $\begingroup$ This question appears to be a special case of the situation addressed in the question on identifying multivariate outliers, where a large number of possible solutions are proposed. (However, the two questions are not quite duplicates in my mind, because bivariate data may lend themselves to special techniques that do not generalize to higher dimensions.) $\endgroup$ – whuber Mar 25 '13 at 13:46
  • $\begingroup$ @whuber/@gung: as I tried to show, when there are just two variables, one can do things that wouldn't work in higher dimensions. $\endgroup$ – user603 Mar 25 '13 at 16:17
  • $\begingroup$ Outliers are really only outliers with respect to some model (even if the model is implicit); points that would be highly unusual under one model are just typical points under another; if you specify some bivariate model it will help you identify the points that don't correspond to it (are highly unlikely, given the model). $\endgroup$ – Glen_b -Reinstate Monica Mar 25 '13 at 22:59
4
$\begingroup$

Here is how I would approach this. Your problem is bivariate. I would use a bagplot (1), which is a bivariate generalization of the boxplot (and so more of a visual exploratory tool).

In R the code to do this is:

library(aplpack)
bagplot(cbind(x,y),pch=16,cex=2)

yielding the plot below:

bagplot

You can read this plot as you would read a boxplot: the orange central region is the bivariate median, the dark blue region 'the bag' is the bivariate IQR (it contains the 50% most central points) and the light region 'the fence' contains the points that are further away (but not enough that they would be considered outliers.)

There are no data points outside the fence so no clear outliers as far as the bagplot is concerned.

(1) P. J. Rousseeuw, I. Ruts, J. W. Tukey (1999): The bagplot: a bivariate boxplot, The American Statistician, vol. 53, no. 4, 382–387

$\endgroup$
  • $\begingroup$ Looks good but I don't know how to install that 'aplpack' to my computer (sage 5.8, r-base 2.15.1-5ubuntu1). $\endgroup$ – guest Mar 25 '13 at 13:37
  • $\begingroup$ try method 2 here $\endgroup$ – user603 Mar 25 '13 at 13:53
  • $\begingroup$ One question, how can you label cases in the bagplot? $\endgroup$ – Pedro Pires Nov 7 at 12:45

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.