When your main attraction is noise

Peter K. asked me about this 538 chart, which is a stacked column chart in which the percentages appear to not add up to 100%. Link to the article here.

Here's my reply:

They made the columns so tall that the "rounding errors" (noise) disclosed in the footnotes became the main attraction.

***

The gap between the highest and lowest peaks looks large but mostly due to the aspect ratio. The  gap is only ~2% at the widest (101% versus 99%) so it is the rounding error disclosed below the chart.

The lesson here is to make sure you suppress the noise and accentuate your data!

 

 

Comments

[Dan Vargo]

What do you think about re-ordering the series by moving the top (green) to the very bottom (orange) so that they are next to each other? This would show the vast majority of the difference comes from the interplay of those two groups while the rest is just noise.

[Kaiser]

DV: Like the way you're thinking about this. Figure out the key messages and then order your data accordingly. I am thinking a Bumps chart is better; will test it out when I get a chance.

[Ken]

DV: All the white groups are showing a strong decline, just greatest for the evangelical protestant.

I would just go for a line chart for the non summed percentages. Alternatively a clustered bar chart with the clustering by religious group. Anything but the stacked bar chart.

I would also like to see how much missing data there was. In Australia religion is a voluntary question, is that the case in America?