Visualizing data spread along both y- and x-axis in 2D scatterplot

Question

I'm currently in the process of doing a literature review of a brain atrophy measurement, called BPF, commonly used in research regarding neurodegenerative disorders.

I have searched for publications having investigated BPF and for each study population I have come across I have noted population size, mean BPF, mean age and the standard deviations of these two means. I do not have access to source data.

I would like to visualize all these populations in a 2D bubble-plot (scatterplot with variable dot-sizes), with BPF on one axis and age on the other, each data point representing one population. I would like the size of the points (bubbles) to be representing population size. I would like to include error bars for each point representing +-SD of age (bar running along the age axis) and +-SD of BPF (bar running along the BPF-axis).

I am currently using SPSS21 and have myself come so far as to be able to create the bubble-plot (using size(PopulationSize) for the dots) with error bar for one of the variables (using SPSS high-low-close graph), but I can not seem to get a second error bar along the other axis in there.

Any suggestions on how to accomplish making this graph, using SPSS or other software, would be greatly appreciated.

Thank you for your help!

If you want help specifically with SPSS code, your question would be off-topic here. There is an interesting underlying statistical question here, but that means the question should be thrown open to people regardless of what software they use. As it stands, your question seems more the first, rather than the second, so I suggest that you either broaden it or withdraw it and seek help in an SPSS forum. (Stack Overflow is not really for this kind of question, unless you show code and give a reproducible example.) — Nick Cox, Dec 12 '13 at 11:37
Thank you for your feedback, Nick! I have edited the question so it is not limited to SPSS. I went with SPSS since that is the software I'm accustomed to, but any tips on accomplishing this using other software is of course also greatly appreciated! My database can of course be exported to another software. /Mattias — Mattias, Dec 12 '13 at 12:55

score 3 · Accepted Answer · edited Jul 14 '15 at 12:30

3

Here is a solution using R:

enter image description here

R Code:

#Make up data
age<-runif(10, 30,70)
agesd<-runif(10, 0.1,5)
bpf<-runif(10, 0,1)
bpfsd<-runif(10, 0.01,.2)    
pop.size<-runif(10,5,50)

#The plot
plot(age,bpf, pch=16, cex=log(pop.size), col=rainbow(length(pop.size)), 
ylim=c(0,1),xlim=c(20,90))
segments(age+agesd,bpf,age-agesd,bpf, lwd=2)
segments(age,bpf+bpfsd,age,bpf-bpfsd, lwd=2)    
legend("topright", legend=paste("Study",1:10), 
col=rainbow(length(pop.size)), pt.cex=1.5, pch=16)

edited Jul 14 '15 at 12:30

bud.dugong

103
2

answered Dec 12 '13 at 14:12

Flask

1,711
1
14
24

Thank you! I have previously only very briefly worked with R, but now I seem to have reason to revisit it! – Mattias Dec 12 '13 at 15:02

score 3 · Answer 2 · edited Apr 13 '17 at 12:44

If your errors are really orthogonal in both directions I would probably suggest using boxes instead of error bars (if they aren't ellipses (Friendly et al. 2013) are probably a better choice). Below the image is SPSS code illustrating how to plot boxes in GPL.

enter image description here

data list free / x y pop.
begin data
1 4 5
2 3 10
3 2 15
4 1 20
end data.
compute xlow = x - (x*.4)/sqrt(pop).
compute xhigh = x + (x*.4)/sqrt(pop).
compute ylow = y - (y*.4)/sqrt(pop).
compute yhigh = y + (y*.4)/sqrt(pop).
formats ALL (F1.0).

*Boxes.
GGRAPH
  /GRAPHDATASET NAME="graphdataset" VARIABLES=x y xlow xhigh ylow yhigh pop
  /GRAPHSPEC SOURCE=INLINE.
BEGIN GPL
 SOURCE: s=userSource(id("graphdataset"))
 DATA: x=col(source(s), name("x"))
 DATA: y=col(source(s), name("y"))
 DATA: pop=col(source(s), name("pop"))
 DATA: xlow=col(source(s), name("xlow"))
 DATA: xhigh=col(source(s), name("xhigh"))
 DATA: ylow=col(source(s), name("ylow"))
 DATA: yhigh=col(source(s), name("yhigh"))
 TRANS: casenum = index() 
 GUIDE: axis(dim(1), label("x"))
 GUIDE: axis(dim(2), label("y"))
 SCALE: pow(aesthetic(aesthetic.size)) 
 ELEMENT: polygon(position(link.hull((xlow + xhigh)*(ylow + yhigh))), color.interior(color.grey), split(casenum), transparency.interior(transparency."0.5"))
 ELEMENT: point(position(x*y), color.interior(color.black), size(pop))
END GPL.

I can show you how you would plot error bars in SPSS, but I think this is much clearer a picture. For scatterplots of more and overlapping cases things to keep in mind are to make the box elements semi-transparent so overlap is more clear. Also in the first graph plotting the point at the center of the rectangle may seem superflous, but this becomes more necessary as the area of the box shrinks, places that have higher certainty in the estimates shrink to nothing in the plot otherwise.

Below is a similar example extended to 246 observations.

enter image description here

set seed 10.
input program.
loop #i = 5 to 250.
  compute x = RV.UNIFORM(5,50).
  compute y = -.5*x + RV.NORMAL(0,10).
  compute pop = #i.
  end case.
end loop.
end file.
end input program.
compute xlow = x - (x*.4)/sqrt(pop).
compute xhigh = x + (x*.4)/sqrt(pop).
compute ylow = y - (y*.4)/sqrt(pop).
compute yhigh = y + (y*.4)/sqrt(pop).
formats ALL (F1.0).

*Boxes.
GGRAPH
  /GRAPHDATASET NAME="graphdataset" VARIABLES=x y xlow xhigh ylow yhigh pop
  /GRAPHSPEC SOURCE=INLINE.
BEGIN GPL
 SOURCE: s=userSource(id("graphdataset"))
 DATA: x=col(source(s), name("x"))
 DATA: y=col(source(s), name("y"))
 DATA: pop=col(source(s), name("pop"))
 DATA: xlow=col(source(s), name("xlow"))
 DATA: xhigh=col(source(s), name("xhigh"))
 DATA: ylow=col(source(s), name("ylow"))
 DATA: yhigh=col(source(s), name("yhigh"))
 TRANS: casenum = index() 
 GUIDE: axis(dim(1), label("x"))
 GUIDE: axis(dim(2), label("y"))
 SCALE: pow(aesthetic(aesthetic.size)) 
 ELEMENT: polygon(position(link.hull((xlow + xhigh)*(ylow + yhigh))), color.interior(color.grey), split(casenum), transparency.interior(transparency."0.5"))
 ELEMENT: point(position(x*y), color.interior(color.black), size(pop), transparency.interior(transparency."0.1"), transparency.exterior(transparency."1"))
END GPL.

Below is some code showing how to plot error bars instead of boxes in SPSS. Ellipses is more difficult, as you need some estimate of the correlation between the errors. Here I've posted some code that provides a function to draw the points for an ellipse in SPSS given the centroid, major and minor axis, and the degree of the slope. See this post on how to calculate an ellipse given a covariance matrix.

*Error Bars.
GGRAPH
  /GRAPHDATASET NAME="graphdataset" VARIABLES=x y xlow xhigh ylow yhigh
  /GRAPHSPEC SOURCE=INLINE.
BEGIN GPL
 SOURCE: s=userSource(id("graphdataset"))
 DATA: x=col(source(s), name("x"))
 DATA: y=col(source(s), name("y"))
 DATA: xlow=col(source(s), name("xlow"))
 DATA: xhigh=col(source(s), name("xhigh"))
 DATA: ylow=col(source(s), name("ylow"))
 DATA: yhigh=col(source(s), name("yhigh"))
 TRANS: casenum = index() 
 GUIDE: axis(dim(1), label("x"))
 GUIDE: axis(dim(2), label("y"))
 ELEMENT: edge(position((xlow + xhigh)*y)))
 ELEMENT: edge(position(x*(ylow + yhigh)))
 ELEMENT: point(position(x*y), color.interior(color.black))
END GPL.

Thank you very much! This is an excellent suggestion, I had not thought of using boxes for this purpose. I think that the errors are not necessarily truly orthogonal, since the probability of a single case in a population deviating 1SD or more regarding both age and BPF should be lower than the opposite, or is that incorrect? Ellipses might be even better. The reference you linked to, by Friendly et. al., seems excellent, even though parts are a bit beyond my current statistical understanding. — Mattias, Dec 12 '13 at 15:13
I have tried the graph with boxes, as per your supplied syntax, with great results! The only drawback is that the data points are a bit clustered, so that even with the transparency a lot of occlusion is going on. If you have the time to show me the syntax for adding error bars instead of boxes, and perhaps even to try ellipses instead, I would greatly appreciate it. It would be great to have a look at what those two graphing options would look like with the data I have. They might make the clustered data points easier to distinguish. — Mattias, Dec 12 '13 at 15:14

Visualizing data spread along both y- and x-axis in 2D scatterplot

2 Answers2