Comment by danso | Hacker News Reader

Comment by danso | Hacker News Reader

Very nice and well-written writeup. Here's one graf that randomly provoked some thoughts:

> But looks can be deceptive. The fact that it looks okay is hiding a dark secret that, if you’re not aware of the fact, won’t even get noticed. The map is using totals (absolute values). There are very very few golden rules in cartography but this is one of them: you cannot map totals using a choropleth thematic mapping technique. The reason is simple. Each of our areas on the map is a different size, and has a different number of people in it. These two innate characteristics of all thematic maps means you simply cannot compare like for like across the map.

> The label tells us that Hubei region has over 65,000 cases of coronavirus. It sounds a lot. But does Hubei have 100,000 people, or possibly 100,000,000 people living there?

I definitely agree with the author: that there are very few "golden rules" in visualization, and that not depicting absolute numbers in a choropleth map is one of them. However, the author does an excellent job (with a bar chart and revised map) showing how this anti-pattern severely obfuscates how much the Hubei region is an extreme outlier.

IMO, during the first half of an epidemic, when a small portion of the population is infected and infections are growing exponentially, it makes sense to use the absolute number of infections. Then later when infection is widespread and the curve looks logistic, it makes sense to give the proportion of infected. I think we are clearly in the first half when it comes to coronavirus.

With time it makes sense to report the number of current cases as the cumulative number includes more and more people who are now cured.

The complaint wasn't about absolute numbers per se, but absolute numbers in a choropleth presentation. Which I think is correct.

> during the first half of an epidemic, when a small portion of the population is infected and infections are growing exponentially, it makes sense to use the absolute number of infections.

That's even a bit of an understatement, because three infected in a city of millions can easily be an early stage pandemic whereas three infected in the middle of nowhere would just be a three very unlucky persons.

I can see why people might be mislead, but absolute value is a very understandable metric.

If you start moving to things like per-capita, i actually think that has the potential to be more confusing for more numbers of people.

so yes, absolute values will be highly correlated with population, but again it just depends on what you really want to highlight and communicate

Maybe you really do want the absolute value.

I guess I'm having a hard time thinking of reasons why absolute value is more important than rate, especially in this scenario, when we're measuring the impact of an infectious disease that spreads person-to-person. I suppose in the hypothetical situation, where there are 10,000 cases in Wyoming and 10,000 cases in New Jersey. A choropleth map by rate would shade Wyoming 15-18 times darker than New Jersey. And this would obfuscate the likely reality that 10,000 cases in New Jersey is imminently a far bigger story – because N.J. is not only 15-18 times more populous, but ~25 times more dense. (The fact that Wyoming is ~15x bigger by land mass would make the issue worse, in terms of visual distraction)

But I'm not sure how shading this by absolute totals – in which case, Wyoming and N.J. would be the same shade – would provide significantly more value? Sure, N.J.'s situation wouldn't be effectively invisible in the totals map, compared to the rates map. But now we have to imagine a scenario in which a person-to-person virus managed to sicken so many people (proportionally speaking) in such a large rural state compared to an extremely urban state. It's very hard to imagine a scenario in which we don't want to focus attention on Wyoming. For 10,000 Wyoming people to be infected – and only 10,000 affected in New Jersey – would almost certainly mean that the infection's original epicenter is Wyoming, and that someone from Wyoming had direct contact via travel with a New Jersey resident, (especially if Wyoming and N.J. are outliers in terms of absolute totals by state).

I'm not saying absolute is necessarily better than rate. I'm saying that it's a choice that highlights different information.

there is no objective right answer, and it's about intent of the communicator. in this case, rate may very well be what the intent needs to be.

> Maybe you really do want the absolute value.

What for though? What can you tell from that value, other than the value itself?

You cannot tell whether it's common or rare, you cannot tell the risk of anyone in a certain area to be affected, you will have a hard time showing trends because people will react to the phenomena and avoid a certain high-risk area which will then result in fewer cases in that area.

When the presence of the contagion is the risk, absolute numbers communicate a lot. Relative counts are less meaningful right now.

The majority of meaningful information received from such a chart right now is the presence or absence of the virus. Secondary is the number of cases to indicate the stage of spread (e.g. 1 suggests maybe an outlier, 2-10 suggests early stages of contact spreading, etc).

Communicating information with an inherently exponential growth rate is just entirely different beast.

Why is "has province reported any cases?" the most meaningful information? Ignoring the current reality of every province having reported cases since January, a simple boolean shading would obfuscate nearly every vital insight realistically conceivable. If it were the case that 3 months after the Hubei outbreak, Hubei had 100,000 reported cases, and all bordering provinces had 1-100, that is an extremely important distinction to make when assessing the effectiveness of containment policies (and/or the trustworthiness of official government numbers).

I agree with what you're saying about reporting cases in China. My point was in the context of reporting elsewhere in the world where most areas have zero cases and so having cases or not is the most important information, followed by the number of cases. I should have been clearer.

> Secondary is the number of cases to indicate the stage of spread (e.g. 1 suggests maybe an outlier, 2-10 suggests early stages of contact spreading, etc).

if you want to indicate that info, why not indicate that info, e.g. by marking current cases of unknown, known local transmission, imports only etc? the absolute count becomes irrelevant.

why use a proxy to speculate about what you can say directly?

But how do you do that in a way that is superior to simply just showing the numbers? The point of the OP was to argue that absolute numbers aren't meaningful in topological maps. But how do you communicate these more specific facts in topologically useful ways aside from just showing the numbers?

"How likely is the area to infect other areas"

"How many hosts does the virus have in which to mutate"

"How much will the global economy be affected by the cases in this area"

Is a virus going to spread equally fast in a sparsely populated area vs a densely populated area?

Isn't there implicit knowledge going into two of those? 100 cases in the Yukon territory feel less likely to infect other areas than 100 cases in NYC. For the global economy, you'll also want additional data, e.g. how relevant to the global economy is that area, how many people are working there, is is the working population that's affected or is it primarily old people etc.

You are arguing a digression.

The argument isn't that you never want to present absolute values.

The argument is that you never want to present absolute values on a choropleth map, explicitly because it always obfuscates the data in a way that is misleading.

The data in question should probably be in a table, if the idea is to be able to compare Hubei Province to the others. The bar chart makes the difference very clear, but it also makes it difficult to interpret the data for the other provinces.

> this anti-pattern severely obfuscates how much the Hubei region is an extreme outlier.

For this reason I think the 3D projection graph is actually not as bad as made out. Sure, it's hard to tell anything about any of the other provinces compared to Hubei, but it really highlights the difference between there and other provinces.

Reply on news.ycombinator.com