I’ve Got the Seeing the Reds and Greens as One Blues

Today I want to highlight a print article from the New York Times I received about two weeks ago. It’s been sitting in a pile of print pieces I want to sit down, photograph, and then write up. But as we begin to return to normal, I need my second dining room chair back because at some point I’ll have guests over.

The article in question examined the rates of Covid-19 vaccination across the United States. And on the front page, above the fold no less, we can compare the vaccination rates for Covid-19 to those of the 2019–2020 flu and if you unfold it to its full-length glory we can add in the 2009–2010 H1N1 swine flu outbreak.

Front page graphics

First thing I want to address is the obvious. Look at those colours. Who loves a green-to-red scale on a choropleth? Not this guy. They are a pretty bad choice because of green-to-red colour blindness. (There’s two different types as well as other types of colour blindness, but I’m simplifying here.) But here’s what happen when I pull the photo into Photoshop and test for it. (This is a screenshot, because I’m not aware of a means of exporting a proof image.)

Reds and greens become yellows and greys.

You can still see the difference between the reds and greens. That’s good. And it’s because colour is complicated. In red-green colour blindness, the issue is sensitivity to picking up reds and greens. (Again, oversimplifying for the sake of a blog post.) Between those two colours in the spectrum we have yellow. To the other side of green we have blue.

So if a designer needs to use a red-green colour scheme—and any designer who has worked in data visualisation will have undoubtedly have had a client asking for the map/chart/whatever to be in red and green—there’s a trick to making it work.

I don’t know if this is true, but growing up, I learned that green was the one colour the human eye evolved to distinguish the most. Now for a print piece like this, you are working in what we call CMYK space (cyan, magenta, yellow, and black). Red is a mixture of magenta and yellow. Green a mixture of cyan and yellow. If you remember your school days, it’s similar to—but not the same as—mixing your primary colours. So if you need to make red and green work, what can you do? First, you can subtract a bit of yellow from your green, because that exists between red and green. But then, and this is why CMYK is different from your primary school primary colours, we can adjust the amount of magenta. Magenta is not a “pure” red, instead it’s kind of purplish and that means has some blue in it. Adding a little bit of magenta, while it does add “red” into the green, it’s also adding more blue to the blue present in the cyan. Now you can spend quite a bit of time tweaking these colours, but very quickly I can get these two options.

Reds and greens.

Great, you can still see them as both red and green. Your client is probably happy and probably accepts this greenish-blue as green, because we have that ability to distinguish so many types of green. But what about those with red-green colour blindness? Again, I can’t quite do a straight export, so the best is a screenshot, but we can compare those two options like so.

I can see the differences significantly more clearly here.

You can probably still tweak the green, but by going for that simple tweak, you can make the client happy—even though it’s still just better to avoid the red and green altogether—and still make the graphic work.

There’s a bit more to say about the rest of the article, which has some additional graphics inside. But that’ll have to wait for another day. As will clearing down the pile of print pieces to share, because that keeps on growing.

Credit for the piece goes to Lazaro Gamio and Amy Schoenfield Walker.

Covid-19: A Global Update

I’ve been trying to limit the amount of Covid-19 visualisations I’ve been covering. But on Sunday this image landed at my front door, above the fold on page 1 of the New York Times. And it dovetails nicely with our story about the pandemic’s impact on Pennsylvania, New Jersey, Delaware, Virginia, and Illinois.

Some not so great looking numbers across the globe.

For most of 2020, the United States was one of the worst hit countries as the pandemic raged out of control. Since January 2021, however, the United States has slowly been coming to grips with the virus and the pandemic. Its rate is now solidly middle of the pack—no longer is America first.

And if you compare the chart at the bottom to those that I’ve been producing, you can clearly see how our five states have really gotten this most recent wave under control to the point of declining rates of new cases.

However, you’ve probably heard the horror stories from India and Brazil where things are not so great. It’s countries like those that account for the continual increase in new cases at a global level.

Credit for the piece goes to Lazaro Gamio, Bill Marsh, and Alexandria Symonds.

Politicising Vaccinations

Yesterday I wrote my usual weekly piece about the progress of the Covid-19 pandemic in the five states I cover. At the end I discussed the progress of vaccinations and how Pennsylvania, Virginia, and Illinois all sit around 25% fully vaccinated. Of course, I leave my write-up at that. But not everyone does.

This past weekend, the New York Times published an article looking at the correlation between Biden–Trump support and rates of vaccination. Perhaps I should not be surprised this kind of piece exists, let alone the premise.

From a design standpoint, the piece makes use of a number of different formats: bars, lines, choropleth maps, and scatter plots. I want to talk about the latter in this piece. The article begins with two side by side scatter plots, this being the first.

Hesitancy rates compared to the election results

The header ends in an ellipsis, but that makes sense because the next graphic, which I’ll get to shortly, continues the sentence. But let’s look at the rest of the plot.

Starting with the x-axis, we have a fairly simple plot here: votes for the candidates. But note that there is no scale. The header provides the necessary definition of being a share of the vote, but the lack of minimum and maximum makes an accurate assessment a bit tricky. We can’t even be certain that the scales are consistent. If you recall our choropleth maps from the other day, the scale of the orange was inconsistent with the scale of the blue-greys. Though, given this is produced by the Times, I would give them the benefit of the doubt.

Furthermore, we have five different colours. I presume that the darkest blues and reds represent the greatest share. But without a scale let alone a legend, it’s difficult to say for certain. The grey is presumably in the mixed/nearly even bin, again similar to what I described in the first post about choropleths from my recent string.

Finally, if we look at the y-axis, we see a few interesting decisions. The first? The placement of the axis labels. Typically we would see the labelling on the outside of the plot, but here, it’s all aligned on the inside of the plot. Intriguingly, the designers took care for the placement—or have their paragraph/character styles well set—as the text interrupts the axis and grid lines, i.e. the text does not interfere with the grey lines.

The second? Wyoming. I don’t always think that every single chart needs to have all the outliers within the bounds of the plot. I’ve definitely taken the same approach and so I won’t criticise it, but I wonder what the chart would have looked like if the maximum had been 35% and the grid lines were set at intervals of 5%. The tradeoff is likely increased difficulty in labelling the dots. And that too is a decision I’ve made.

Third, the lack of a zero. I feel fairly comfortable assuming the bottom of the y-axis is zero. But I would have gone ahead and labelled it all the same, especially because of how the minimum value for the axis is handled in the next graphic.

Speaking of, moving on to the second graphic we can see the ellipsis completes the sentence.

Vaccination rates compared to the election results

We otherwise run into similar issues. Again, there is a lack of labelling on the x-axis. This makes it difficult to assess whether we are looking at the same scale. I am fairly certain we are, because when I overlap the graphics I can see that the two extremes, Wyoming and Vermont, look to exist on the same places on the axis.

We also still see the same issues for the y-axis. This time the axis represents vaccination rates. I wish this graphic made a little clearer the distinction between partial and full vaccination rates. Partial is good, but full vaccination is what really matters. And while this chart shows Pennsylvania, for example, at over 40% vaccinated, that’s misleading. Full vaccination is 15 points lower, at about 25%. And that’s the number that needs to be up in the 75% range for herd immunity.

But back to the labelling, here the minimum value, 20%, is labelled. I can’t really understand the rationale for labelling the one chart but not the other. It’s clearly not a spacing issue.

I have some concerns about the numbers chosen for the minimum and maximum values of the y-axis. However, towards the middle of the article, this basic construct is used to build a small multiples matrix looking at all 50 states and their rates of vaccination. More on that in a moment.

My last point about this graphic is on the super picky side. Look at the letter g in “of residents given”. It gets clipped. You can still largely read it as a g, but I noticed it. Not sure why it’s happening, though.

So that small multiples graphic I mentioned, well, see below.

All 50 states compared

Note how these use an expanded version of the larger chart. The y-minimum appears to be 0%, but again, it would be very helpful if that were labelled.

Also for the x-axis in all the charts, I’m not sure every one needs the Biden–Trump label. After all, not every chart has the 0–60% range labelled, but the beginning of each row makes that clear.

In the super picky, I wish that final row were aligned with the four above it. I find it super distracting, but that’s probably just me.

Overall, this is a strong piece that makes good use of a number of the standard data visualisation forms. But I wish the graphics were a bit tighter to make the graphics just a little clearer.

Credit for the piece goes to Danielle Ivory, Lauren Leatherby and Robert Gebeloff.

Choropleths…Again

Admittedly, I was trying to find a data set for a piece, but couldn’t find one. So instead for today’s post I’ll turn to something that’s been sitting in my bookmarks for a little while now. It’s a choropleth map from the US Census Bureau looking at population change between the censuses.

Unequal growth

The reason I have it bookmarked is for the apportionment map, but I will save apportionment for another post because, well, it’s complicated. But map colours are a thing we’ve been discussing of late and we can extend that conversation here.

What I find interesting about this map is how they used a very dark blue-grey colour for their positive growth and an orange that is a fair bit brighter for negative growth, or population loss. And because of that difference in brightness, the orange really jumps out at you.

To be fair, that’s ideal if you’re trying to talk about where state populations are shrinking, because it focuses attention on declines. But, if you’re trying to present a more neutral position, like this seems to be, that colour choice might not be ideal.

Another issue is that if you look at the legend it simply says loss for that orange. But, look above and you’ll see four bins clearly delimited by ranges of percents for the positive growth. If we are trying to present a more neutral story, the use of the orange places it visually somewhere near the top of that blue-grey spectrum.

If you look at the percentages, however, Michigan’s population decline was 0.6% and Puerto Rico’s 2.2%. If this map used a legend that treated positive and negative growth equally, you would place that one state and one should-be state in a presumably light orange. The scale of their negative growth is equal to something like Ohio, which is in the lightest blue-grey available.

Consequently, this map is a little bit misleading when it comes to negative growth.

Credit for the piece goes to the Census Bureau graphics team.

Choropleths and Colours Part 2

Last Thursday I wrote about the use of colour in a choropleth map from the Philadelphia Inquirer. Then on Sunday morning, I opened the door to collect the paper and saw a choropleth above the fold for the New York Times. I’ll admit my post was a bit lengthy—I’ve never been one described as short of words—but the key point was how in the Inquirer piece the designer opted to use a blue-to-red palette for what appeared to be a data set whose numbers ran in one direction. The bins described the number of weeks a house remained on the market, in other words, it could only go up as there are no negative weeks.

Compare that to this graphic from the Times.

More choropleth colours…

Here we are not looking at the Philadelphia housing market, but rather the spread of the UK/Kent variant of SARS-CoV-2, the virus that causes COVID-19. (In the states we call it the UK variant, but obviously in the UK they don’t call it the UK variant, they call it the Kent variant from the county in the UK where it first emerged.)

Specifically, the map looks at the share (percent) of the variant, technically named B.1.1.7, in the tests reported for each country. The Inquirer map had six bins, this Times map has five. The Inquirer, as I noted above, went from less than one week to over five weeks. This map divides 100% into five 20-percent bins.

Unlike the Inquirer map, however, this one keeps to one “colour”. Last week I explained why you’ll see one colour mean yellow to red like we see here.

This map makes better use of colour. It intuitively depicts increasing…virus share, if that’s a phrase, by a deepening red. The equivalent from last week’s map would have, say, 0–40% in different shades of blue. That doesn’t make any sense by default. You could create some kind of benchmark—though off the top of my head none come to mind—where you might want to split the legend into two directions, but in this default setting, one colour headed in one direction makes significant sense.

Separately, the map makes a lot of sense here, because it shows a geographic spread of the variant, rippling outward from the UK. The first significant impacts registering in the countries across the Channel and the North Sea. But within four months, the variant can be found in significant percentages across the continent.

Credit for the piece goes to Josh Holder, Allison McCann, Benjamin Mueller, and Bill Marsh.

Choropleths and Colours

In many cities through the United States, real estate represents a hot commodity. It’s not difficult to understand why, as have covered before, Americans are saving a bit more. Coupled with stay-at-home orders in a pandemic, spending that cash on a home down payment makes a lot of sense for a lot of people. But with little new construction, it’s a seller’s market.

The Philadelphia Inquirer covers that angle for the Philadelphia region and in the article, it includes a map looking at time to sell a house. And it’s that interactive map I want to look at briefly this morning.

Red vs. blue

Primarily I want to discuss the colours, as you can gather from this post’s title. We have six bins here, each indicating an amount of time in one-week intervals. So far so good. Now to the colours, we have red for homes that sell in one week or less and blue for homes that sell in five weeks or more.

Blue to red is a pretty standard choice. You will often see it in maps where you have positive growth to negative growth or something similar, I’ve used it myself on Coffeespoons a number of times, like in this map of population growth at the county level here in Pennsylvania.

In those scenarios, however, note how you have positive values and negative values. The change in colour (hue) encodes the change in numerical value, i.e. positive vs. negative. We then encode the values within that positive or negative range with lighter/darker blues and reds. Most often the darker the blue or red, the greater the value toward the end of the spectrum. For example, in Pennsylvania, the dark blue meant population growth greater than 8% and red meant population declines in excess of 8%.

As an aside you’ll note that there are no dark blue counties in that map and that’s by design. By keeping the legend symmetrical in terms of its minimum and maximum values, we can show how no counties experienced rapid population growth whilst several declined rapidly. If dark blue had meant greater than 4% growth, that angle of the story would have been absent from the map.

Back to our choropleth discussion, however. How does that fit with this map of selling times for homes in the Philadelphia region?

Note first that five weeks is a positive value. But so is one week or less. The use of the red-blue split here is not immediately intuitive. If this map were about the change or growth in how long homes sell, certainly you could see positive and negative rates and those would make sense in red and blue.

The second part to understand about a traditional red-blue choropleth is that at some point you have to switch from red to blue, a mid-point if you will. If you are talking positive/negative like in my Pennsylvania map, zero makes a whole lot of sense. Anything above zero, blue, anything below zero red.

Sometimes, you will see a third colour, maybe a grey or a purple, between that red and blue. That encodes a fuzzier split between positive and negative. Say you want to give a margin of 1%, i.e. any geographic area that has growth between +1% and -1%. That intrinsically means the bin is both positive and negative at the same time, so a neutral colour like grey or a blend of the two colours, a purple in the case of red and blue, makes a whole lot of sense.

Here we have nothing like that. Instead we jump from a light yellow two-to-three weeks to a light blue three-to-four weeks.

What about that yellow? In a spectrum of dark blue to light blue, you will see lighter blues than darker blues. But in a red spectrum, that light red becomes pinkish or salmonish depending on that exact type of red you use. (Conversation for another day.) Personal preferences will often push clients to asking a designer to “use less pink” in their maps. I can’t tell you the number of times I’ve heard that.

If that comes up, designers will often keep their blue side of the legend from the dark to light—no complaints there, or at least I’ve never heard any. But for the red side, they’ll switch to using hue or type of colour instead of dark to light red.

Not all colours are as dark as others. Blue and red can be pretty dark. Yellow, however, is a fairly light colour. Imagine if you converted the colours to greyscale, you’ll have very dark greys for blue and red, but yellow will be consistently far lighter than the other two.

The designer can use the light yellow as the light red. But to link the yellow to red, they need to move through the hues or colours between the two. There’s a whole conversation here about colour theory and pigment and light absorption vs. pixels and light emission, but let’s go back to your colours you learned in primary school (pigment and light absorption). Take your colour wheel and what sits between red and yellow? Orange.

And so if a client objects to a light pink, you’ll see a pseudo dark-to-light red spectrum that uses a dark red, a medium orange, and a light yellow. Just like we see here in this Inquirer map.

Back to the two-to-three week and three-to-four week switch, though. What’s the deal? This is my sticking point with the graphic. I am looking for the explanation of why the sudden break in colour here, but I don’t see any obvious one.

Why would you use this colour scheme where blue and red diverge around a non-zero value? Let’s say the average home in the region sells in three weeks, any of the zip codes in red are selling faster than average, hot markets, and those taking longer than average are in blue, cold markets. Maybe it’s the current average, however. What if it were the average last year? Or the national average? These all serve as benchmarks for the presented data and provide valuable context to understand the market.

Unfortunately it’s not clear what, if any, benchmarks the divergence point in this map reflects. And if there is no reason to change colours mid-legend, with only six bins, a designer could find a single colour, a blue or purple for example, and then provide five additional lighter/darker shades of that to indicate increasing/decreasing levels of speed at which homes sell.

Overall, I left this piece a wee bit confused. The general trend of regional differences in how quickly homes are selling? I get that. But because there’s a non-logical break between red and blue here—or at least one I fail to see in the graphic—this map would work almost as well if each bin were a separate colour entirely, using ROYGBIV as a base for example.

Credit for the piece goes to John Duchneskie.

Biden’s Biggest Pyramids

Yesterday we looked at an article from the Inquirer about the 2020 election and how Biden won because of increased margins in the suburbs. Specifically we looked at an interactive scatter plot.

Today I want to talk a bit about another interactive graphic from the same article. This one is a map, but instead of the usual choropleth—a form the article uses in a few other graphics—here we’re looking at three-dimensional pyramids.

All the pyramids, built by aliens?

Yesterday we talked about the explorative vs. narrative concept. Here we can see something a bit more narrative in the annotations included in the graphic. These, however, are only a partial win, though. They call out the greatest shifts, which are indeed mentioned in the text. But then in another paragraph the author writes about Bensalem and its rightward swing. But there’s no callout of Bensalem on the map.

But the biggest things here, pun intended, are those pyramids. Unlike the choropleth maps used elsewhere in the article, the first thing this map fails to communicate is scale. We know the colour means a county’s net shift was either Democratic or Republican. But what about the magnitude? A big pyramid likely means a big shift, but is that big shift hundreds of votes? Thousands of votes? How many thousands? There’s no way to tell.

Secondly, when we are looking at rural parts of Bucks, Chester, and Montgomery Counties, the pyramids are fine. They remain small and contained within their municipality boundaries. Intuitively this makes sense. Broadly speaking, population decreases the further you move from the urban core. (Unless there’s a secondary city, e.g. Minneapolis has St. Paul.) But nearer the city, we have more population, and we have geographically smaller municipalities. Compare Colwyn, Delaware County to Springfield, Bucks County. Tiny vs. huge.

In choropleth maps we face this problem all the time. Look at a classic election map at the county level from 2016.

Wayb ack when…

You can see that there is a lot more red on that map. But Hillary Clinton won the popular vote by more then 3,000,000 votes. (No, I won’t rehash the Electoral College here and now.) More people are crowded into smaller counties than there are in those big, expansive red counties with far, far fewer people.

And that pattern holds true in the Philadelphia region. But instead of using the colour fill of an area as above, this map from the Inquirer uses pyramids. But we face the same problem, we see lots of pyramids in a small space. And the problem with the pyramids is that they overlap each other.

At a glance, you cannot see one pyramid beind another. At least in the choropleth, we see a tiny field of colour, but that colour is not hidden behind another.

Additionally, the way this is constructed, what happens if in a municipality there was a small net shift? The pyramid’s height will be minimal. But to determine the direction of the shift we need to see the colour, and if the area under the line creating the pyramid is small, we may be unable to see the colour. Again, compare that to a choropleth where there would at least be a difference between, say, a light blue and light red. (Though you could also bin the small differences into a single neutral bin collecting all small shifts be them one way or the other.)

I really think that a more straight forward choropleth would more clearly show the net shifts here. And even then, we would still need a legend.

The article overall, though, is quite strong and a great read on the electoral dynamics of the Philadelphia region a month ago.

Credit for the piece goes to John Duchneskie.

Choose Your Own FiveThirtyEight Adventure

In case you weren’t aware, the US election is in less than a week, five days. I had written a long list of issues on the ballot, but it kept getting longer and longer so I cut it. Suffice it to say, Americans are voting on a lot of issues this year. But a US presidential election is not like many other countries’ elections in that we use the Electoral College.

For my non-American readers, the Electoral College, very briefly, was created by the country’s founding fathers (Washington, Jefferson, Adams, Franklin, et al.) to do two things. One, restrict selection of the American president to a class of individuals who theoretically had a broader/deeper understanding of the issues—but who also had vested interests in the outcome. The founders did not intend for the American people to elect the president. The second feature of the Electoral College was to prevent the largest states from dominating smaller states in elections. Why else would Delaware and Rhode Island surrender their sovereignty to join the new United States if Virginia, Pennsylvania, and New York make all the decisions? (The founders went a step further and added the infamous 3/5 clause, but that’s another post.)

So Americans don’t elect the president directly and larger states like California, New York, and Texas, have slightly less impact than smaller states like Wyoming, Vermont, and Delaware. Each state is allotted a number of Electoral College votes and the key is to reach 270. (Maybe another time I’ll get into the details of what happens in a 269–269 tie.) Many Americans are probably familiar with sites like 270 To Win, where you can determine the outcome of the election by saying who won each state. But, even though the US election is really 50 different state elections, common threads and themes run through all those states and if one candidate or another wins one state, it makes winning or losing other states more or less likely. FiveThirtyEight released a piece that attempts to link those probabilities and help reveal how decisions voters in one state make may reflect on how other voters decide.

The interface is fairly straightforward—I’m looking at this on a desktop, though it does work on mobile—with a bunch of choices at the top and a choropleth map below. There we have a continually divergent gradient, meaning the states aren’t grouped into like bins but we have incredibly subtle differences between similar states. (I should also point out that Maine and Nebraska are the two exceptions to my above description of the Electoral College. They divide their votes by congressional district, whoever wins the district gets that Electoral College vote and then the state overall winner receives the remaining two votes.)

Below that we have a bar chart, showing each state, its more/less likely winner state and the 270 threshold. Below that, we have what I’ve read/heard described as a ball plot. It represents runs of the simulation. As of Thursday morning, the current FiveThirtyEight model says Trump has an 11 in 100 chance of winning, Biden, conversely, an 89-in-100 chance.

But what happens when we start determining the winners of states?

Well, for my non-American readers, this election will feature a large number of voters casting their ballots early. (I voted early by mail, and dropped my ballot off at the county election office.) That’s not normal. And I cannot emphasise this next point enough. We may not know who wins the election Tuesday night or by the time Americans wake up on Wednesday. (Assuming they’re not like me and up until Alaska and Hawaii close their polls. Pro-tip, there’s a potentially competitive Senate race in Alaska, though it’s definitely leaning Republican.)

But, some states vote early and/or by mail every year and have built the infrastructure to count those votes, or the vast majority of them, on or even before Election Day. Three battleground states are in that group: Arizona, Florida, and North Carolina. We could well know the result in those states by midnight on Election Day—though Florida is probably going to Florida.

So what happens with this FiveThirtyEight model if we determine the winners of those three states? All three voted for Trump in 2016, so let’s say he wins them again next week.

We see that the states we’ve decided are now outlined in black. The remainder of the states have seen their colours change as their odds reflect the set electoral choice of our three states. We also now have a rest button that appears only once we’ve modified the map. I’m also thinking that I like FiveyFox, the site’s new mascot? He provides a succinct, plain language summary of what the user is looking at. At the bottom we see what the model projects if Arizona, Florida, and North Caroline vote for Trump. And in that scenario, Trump wins in 58 out of 100 elections, Biden in only 41. Still, it’s a fairly competitive election.

So what happens if by midnight we have results from those three states that Biden has managed to flip them? And as of Thursday morning, he’s leading very narrowly in the opinion polls.

Well, the interface hasn’t really changed. Though I should add below this screenshot there is a button to copy the link to this outcome to your clipboard if, like me, you want to share it with the world or my readers.

As to the results, if Biden wins those three states, Trump has less than a 1-in-100 chance of winning and Biden a greater than 99-in-100.

This is a really strong piece from FiveThirtyEight and it does a great job to show how states are subtly linked in terms of their likelihood to vote one way or the other.

Credit for the piece goes to Ryan Best, Jay Boice, Aaron Bycoffe and Nate Silver.

Where Are the Votes?

I’m not working for a good chunk of the next few days. But, I did want to share with my readers an analysis of Pennsylvania’s missing votes. Broadly, Trump needs to win the Commonwealth of Pennsylvania next week—yes, the US election is now one week away. Though, Pennsylvania allows mail-in ballots postmarked on Election Day to arrive within a few days and still be counted. So we may not have final tallies for the state until the weekend or Monday after Election Day.

Pennsylvania, of course, narrowly voted for Donald Trump over Hillary Clinton in 2016 with 44,000+ votes making the difference. In 2020, polling has consistently placed Joe Biden above Donald Trump by 5+ points. But, can Trump again pull off an upset victory?

I argue that yes, he can. And fairly easily too. (If you want to see why I think Pennsylvania is really Trumpsylvania, I recommend checking out my longer, more in-depth analysis.) So where would the votes come from? I mapped the 2016 difference between votes cast and registered voters, i.e. people who could have voted, but did not for whatever reason. I then coloured the map by the county’s winner in 2016. Red counties voted for Trump by more than 10 points and blue for Clinton by more than 10 points. The purple counties are those that were competitive, plus or minus 10 points for either candidate.

In the purple counties, both candidates will want to drive out as many voters as possible. But in the blue counties, Biden has reliably Democratic votes and in red Trump has reliably Republican votes. So why on Monday did Trump visit Allentown, Lititz, and Martinsburg? Because that’s where those votes are.

Allentown, in Lehigh County, is competitive. In fact, neighbouring Northampton Co. will be a key swing county next week and one I will be following closely as the returns come in. But Lititz, Lancaster Co., and Martinsburg, Blair Co., are in reliably red counties. (Though in my Trumpsylvania piece I argue Lancaster Co. is undergoing a transition to a competitive, albeit lean Republican county.)

In Lancaster Co., which went to Trump by nearly 20 percentage points in 2016, there were still just short of 100,000 voters who didn’t vote in 2016. Not all of those voters would have voted for Trump, but for sake of argument, just say 50% would have. That makes just short of 50,000 potential Trump votes—more than Trump’s entire state margin.

Blair Co. is in the Pennsyltucky region of the state, relatively rural, but in Blair’s case, its county seat Altoona is the state’s 10th largest city. While the total number of votes—and the total number of non-voting voters—are smaller than in Lancaster Co., add up all the available votes and it’s a large number.

If you add up all those red counties’ missing votes, you get a total of just shy of 840,000 missing votes. Far more than enough to drastically swing the Commonwealth to Trump in 2020.

Of course, Biden’s counting on driving out turnout in Philadelphia and Pittsburgh and their suburbs, along with other cities in the state, like Allentown, Scranton, Harrisburg, and Erie. In those blue counties, there were 927,000 missing votes, so the potential for a Biden win is also there.

But, if Democratic voters don’t vote again in 2016, Trump has plenty of potential votes to pick up across the state.

Credit for the piece is mine.

Covid Migration

Yep, Covid-19 remains a thing. About a month or so ago, an article in City Lab (now owned by Bloomburg), looked at the data to see if there was any truth in the notion that people are fleeing urban areas. Spoiler: they’re not, except in a few places. The entire article is well worth a read, as it looks at what is actually happening in migration and why some cities like New York and San Francisco are outliers.

But I want to look at some of the graphics going on inside the article, because those are what struck me more than the content itself. Let’s start with this map titled “Change in Moves”, which examines “the percentage drop in moves between March 11 and June 30 compared to last year”.

Conventionally, what would we expect from this kind of choropleth map. We have a sequential stepped gradient headed in one direction, from dark to light. Presumably we are looking at one metric, change in movement, in one direction, the drop or negative.

But look at that legend. Note the presence of the positive 4—there is an entire positive range within this stepped gradient. Conventionally we would expect to see some kind of red equals drop, blue equals gain split at the zero point. Others might create a grey bin to cover a negative one to positive one slight-to-no change set of states. Here, though, we don’t have that. Nor do we even get a natural split, instead the dark bin goes to a slightly less dark bin at positive four, so everything less than four through -16 is in the darker bin.

Look at the language, too, because that’s where it becomes potentially more confusing. If the choropleth largely focuses on the “percentage drop” and has negative numbers, a negative of a negative would be…a positive. A -25% drop in Texas could easily be mistaken with its use of double negatives. Compare Texas to Nebraska, which had a 2% drop. Does that mean Nebraska actually declined by 2%, or does it mean it rose by 2%?

A clean up in the data definition to, say, “Percentage change in moves from…” could clear up a lot of this ambiguity. Changing the colour scheme from a single gradient to a divergent one, with a split around zero (perhaps with a bin for little-to-no change), would make it clearer which states were in the positive and which were in the negative.

The article continues with another peculiar choice in its bar charts when it explores the data on specific cities.

Here we see the destinations of people moving out of San Francisco, using, as a note explains, requests for quotes as a proxy for the numbers of actual moves. What interests me here is the minimalist take on the bar charts. Note the absence of an axis, which leaves the bars almost groundless for comparison, except that the designer attached data labels to the ends of the bars.

Normally data labels are redundant. The point of a visualisation is to visualise the comparison of data sets. If hyper precise differences to the decimal point are required, tables often are a better choice. But here, there are no axis labels to inform the user as to what the length of a bar means.

It’s a peculiar design decision. If we think of labelling as data ink, is this a more efficient use with data labels than just axis labels? I would venture to say no. You would probably have five axis labels (0–4) and then a line to connect them. That’s probably less ink/pixels than the data labels here. I prefer axis lines to help guide the user from labels up (in this case) through the bars. Maybe the axis lines make for more data ink than the labels? It’s hard to say.

Regardless, this is a peculiar decision. Though, I should note it’s eminently more defensible than the choropleth map, which needs a rethink in both design and language.

Credit for the piece goes to Marie Patino.