How can I simplify/beautify a busy stacked bar chart comparing two groups within yearly cohorts?

[samp945]

Hello all,

I need to produce a graph comparing time-to-failure between two groups within yearly cohorts over a 20-year period. That's necessarily a lot of data. I will be presenting these data in other ways, but to begin, I must present them as a whole.

The attached graph shows two groups for each yearly cohort but it is difficult to distinguish between the group A (left) and group B (right). I have attempted to do this by adding a gradient and transparency, but perhaps there is a better way. Any suggestions?

Additionally, at present there is no way to distinguish between Group A (left) and Group B (right) aside from the footnote. Is there a way for me to somehow label the left (solid) vs. right (hatched) groups other than doubling the list of values in the legend? For example, a supplementary legend in the top-right corner that states: Solid = Group A / Hatched = Group B?

Maybe I shouldn't be using VBAR for this type of graph? If so, what other graph option would work better?

Thanks for your time!

ods graphics / reset 
		height=6.5in imagename="FigureZ" imagefmt=PNG ATTRPRIORITY=NONE;
	Footnote justify=left height=1.2 font="Times New Roman" "Figure X.XX. Time-to-Failure Rates - Group A (Left) vs. Group B (Right)";
	Title;
 
	proc sgplot data=Failure noborder nocycleattrs;

	vbar Cohort/ 
		response=GroupA
		group=TTF
		grouporder=ascending
		groupdisplay=stack
		barwidth=0.4
		name="GroupA"
		legendlabel="GroupA"
		discreteoffset=-0.4;

	vbar Cohort/ 
		response=GroupB
		group=TTF
		grouporder=ascending
		groupdisplay=stack
		barwidth=0.4
		name="GroupB"
		legendlabel="GroupB"
		fillattrs=(transparency=0.9)
		filltype=gradient
		fillpattern
			fillpatternattrs=(pattern=x4);

	keylegend "GroupA"
		/ title="Time-to-Failure (Years)" Titleattrs=(size=12) Valueattrs=(size=12) location=outside position=bottom noborder;

	xaxis  Valueattrs=(size=12) discreteorder=data display=(nolabel noticks);
	yaxis Valueattrs=(size=12) Labelattrs=(size=12) grid values=(0 to 75 by 5) label="Failure Rate of Cohort (Percent)";
	run;

[yabwon]

Maybe something like:

data Failure2;
  set Failure;
  GroupB  = - GroupB;
run;


ods graphics / reset 
		height=6.5in imagename="FigureZ" imagefmt=PNG ATTRPRIORITY=NONE;
	Footnote justify=left height=1.2 font="Times New Roman" "Figure X.XX. Time-to-Failure Rates - Group A (Left) vs. Group B (Right)";
	Title;
 
	proc sgplot data=Failure2 noborder nocycleattrs;

	vbar Cohort/ 
		response=GroupA
		group=TTF
		grouporder=ascending
		groupdisplay=stack
		barwidth=0.7
		name="GroupA"
		legendlabel="GroupA"
		discreteoffset=-0.1
    ;

	vbar Cohort/ 
		response=GroupB
		group=TTF
		grouporder=ascending
		groupdisplay=stack
		barwidth=0.7
		name="GroupB"
		legendlabel="GroupB"
		fillattrs=(transparency=0.9)
		filltype=gradient
		fillpattern
			fillpatternattrs=(pattern=x4);

	keylegend "GroupA"
		/ title="Time-to-Failure (Years)" Titleattrs=(size=12) Valueattrs=(size=12) location=outside position=bottom noborder;

	xaxis  Valueattrs=(size=12) discreteorder=data display=(nolabel noticks);
	yaxis Valueattrs=(size=12) Labelattrs=(size=12) grid 
    values=(-60 to 60 by 5) 
    valuesdisplay=(
    %macro loop(s,e,b,f=ABS);
      %local i;
      %do i = &s. %to &e. %by &b.;
        "%sysfunc(&f.(&i.))"
      %end;
    %mend;
    %loop(-60,60,5)
    )  

    label="Failure Rate of Cohort (Percent)";
	run;

I had to generate data myself:

data Failure;
  call streaminit(123);
  retain GroupA GroupB 10; 
  do Cohort = 2010 to 2020;

    do TTF = 1,2,3,4,5,6;
      GroupA + (-rand("uniform",0,.25));
      GroupB + (-rand("uniform",0,.25));
      output;
    end;

  end;
run;

Result:

Bart

[DanH_sas]

One alternative would be to transpose your data so that your GroupA and GroupB responses are folded together with a GROUP column containing the values "GroupA" and "GroupB". Then, you could something like the following with PROC SGPANEL:

proc sgpanel data=sashelp.prdsale;
panelby year / layout=rowlattice novarname;
vbar month / response=actual group=product;
run;

[samp945]

Thanks @yabwon and @DanH_sas for your suggestions!

However, I am hoping to find a strategy that plots each group side-by-side so that each yearly cohort can be compared without referring to any values on the Y-axis.

For example, my existing vbar plots show that for cohorts 2000 - 2004, group A (left) had a higher Time-to-Failure than group B (right), but this trend reversed for cohorts 2005 onward (group B had a higher TTF) than group A).

[DanH_sas]

Given that goal, I would use a two overlaid VLINEs instead of overlayed VBARs, and I would drop the TTF group. That way, you can clearly see the transition you described. Just use the default STAT of SUM. I would also use the MARKERS option so you will see the markers at each year.

[Reeza]

Since the segments starting position don't always align, this graph doesn't really do that. 

I'd consider some variation on a slope graph with small multiples possibly.

Slope Graph

https://www.storytellingwithdata.com/blog/2020/7/27/what-is-a-slopegraph

Small Multiples

https://kieranhealy.org/blog/archives/2020/03/27/a-covid-small-multiple/

[Ksharp · Accepted solution]

I would like to use Medal Graph:

Sochi Medal Graphs - Graphically Speaking (sas.com)

/*
https://blogs.sas.com/content/graphicallyspeaking/2014/02/12/sochi-medal-graphs/
*/
data Failure;
  call streaminit(123);
  retain GroupA GroupB 10; 
  do Cohort = 2010 to 2020;

    do TTF = 1,2,3,4,5,6;
      GroupA + (-rand("uniform",0,.25));
      GroupB + (-rand("uniform",0,.25));
      output;
    end;

  end;
run;

proc sgpanel data=Failure noautolegend;
  panelby TTF / layout=panel rows=1 onepanel sort=data  spacing=5;
  hbar Cohort / response=GroupA  nostatlabel transparency=0.5 name='a';
  hbar Cohort / response=GroupB  nostatlabel transparency=0.5 barwidth=0.4 name='b';
  keylegend 'a' 'b';
  rowaxis discreteorder=data display=(nolabel) fitpolicy=none;
  colaxis integer display=(nolabel);
  run;

[Quentin]

I generally hate stacked bar charts.  It's really hard to see trends in the size of each section of a bar, unless the differences are huge.

I think it would help to think about "what is the purpose of this plot" or, "what is the story I want this graph to tell."  It's often easier for a graph to tell one story clearly, rather than try to make it tell multiple stories.

I agree with Dan, if your goal is to compare failure rates, I would use a line plot. 

When I think time-to-failure, I think Kaplan-Meier survival plot.  What if you made a paneled survival plot?  Each plot could show the survival curve for your two groups.  And there could be one plot per year.  I guess 20 plots in a panel might be a bit much...