More graphs of mortality trends

Corinne Riddell writes:

In late March you released a series of plots visualizing mortality rates over time by race and gender. For almost a year now, we’ve been working on a similar project and have compiled all of our findings into an R shiny web app here, with a preprint of our first manuscript here. I’m also attaching our preprint’s web appendix as it fails to load on bioRxiv at the moment. The web app takes about 30 seconds to load, but if you can stand the wait, you’ll be able to interact with the app to investigate changes to life expectancy over time by state, race, and cause of death.
In the spirit of eliciting peer review, I would be happy to have you and your readers have a look at the app and/or the preprint.

In the spirit of providing feedback, I started clicking through and first came across this graph:

Uh oh! Something went terribly wrong here. I reloaded the page and got this:

I like certain aspects of the visual layout but to my eye there’s just too much going on with the colors, the dotted lines, the multiple columns, all happening at once. Also I don’t like how the top and bottom graphs blur together: visually it doesn’t work at all for me. I think they could have a much cleaner display here.

The next one I also think is just way too tricky:

I find myself doing lots of mental arithmetic trying to add and subtract and compare these different numbers.

In any case, I think Riddell and her colleagues are doing it exactly right: they’re putting their graphs out there and getting comments so they can do better. I certainly don’t think my own graphs are perfect and I’m glad to see various people taking their cracks at making these data accessible in different ways.

19 thoughts on “More graphs of mortality trends

  1. I like the time series plots just fine. Actually they might be better without the legend, which is redundant and just adds some clutter. Females on the left, Males on the right, labeled lines for black and whites, clearly labeled line showing the difference…it’s hard to see how they could make this much simpler. Andrew doesn’t like the dotted lines…I guess they could be solid color as long as the colors aren’t a problem for colorblind people. I certainly wouldn’t object if they were solid.

    The bar plots take a bit more interpretation, but actually I’m OK with that too. Once I got the hang of them I had no problem. I do wonder why the gap runs the other direction for 85+ in all of the states that I looked at, but that’s not a graphical problem really.

    I would never say a plot couldn’t be improved, but these work for me!

    • Hi Phil,

      The mortality cross-over that you’ve noticed (i.e., the contribution to the gap flipping for 85+) has received a lot of speculation over the years. Folks had thought that it was due to mis-measurement of age on death certificates, especially among black folks without birth certificates. Thus, some people think that the mortality cross-over reflects measurement bias. More recently, researchers have concluded that it is a true effect that cannot be explained by measurement error… Definitely something I’m going to investigate further!

  2. I also find the graphs work fairly well. The bar charts are a bit too cluttered, and as Phil says, improvement is always possible. I am more concerned about the substance of what is being displayed – and, I haven’t read the paper so it is possible that this is addressed there. Two issues stand out to me. First, the causes of death are not clearly defined. Given the attention on addiction to opiods, it isn’t obvious to me whether those deaths are considered “non-communicable” or “all other.” I thought there was more granular data on causes than that. My second concern is related: the graphs refer to “causes of death” while the actual data addresses “contribution to the life expectancy gap.” Maybe it’s just too early in the morning for me (or maybe I’m just getting too old), but it is far from obvious to me that these two mean the same thing.

    • Right, I am interested in suicide data, and I have no idea where it falls in the “cause” spectrum. I honestly think that should be pulled out on its own. I mean, suicide from injury (jumping, gun, slitting wrists?) vs suicide from poisoning (acetaminophen overdose, opiod overdose, etc) might fall in two different categories? I’m guessing that “non-communicable” means “disease” so things like ALS or Cancer or Anaphalaxis or something.

      Also, I don’t like that it focuses only on black/white comparison, and there’s no way to directly compare male-female except flipping back and forth which doesn’t work in this presentation (it takes something like 10 seconds on my computer, which admittedly is running a Stan model)

      • Hi Daniel,

        Suicide is categorized as injury as well. We agree that disentangling the injury category is a priority for us and something we’re looking into in the next iteration of this work. As I mentioned in my response to Dale, the data that we used becomes suppressed when the death count is between 1 and 9 so pulling out specific causes became a problem but something we’d like to do.

        Just to clarify, the JAGS models have already been run. The app is admittedly slow at the moment, and something I’m working on improving.

        • Oh, when I said running a Stan model, I meant my computer is bogged down because a model is running, not that your web app had to run the model ;-)

          If you’re planning to divide out the injury category, I’d suggest

          AutomobileAccident
          WorkRelatedAccident
          AccidentalNonAuto
          Criminal
          IntentionalSelfHarm/suicide

          Of course you probably were already thinking that. Just saying those separate categories are things I’d be interested in.

    • Hi Dale,

      Drug overdoses are part of the injuries category. A table containing the categorization of causes of death are available in the pre-print, but I think I’ll add this information to the app since the cause categorizations are so general. You’re right that there is more granular data on cause available. Since we were already stratifying by several factors, we sacrificed granularity for better estimation since the data becomes suppressed when the count of deaths is between 1 and 9.

      In regards to your second question, we’re estimating the *cause of death* “contribution to the life expectancy gap”. This is performed using Arriaga’s method, (a demographic technique). Briefly, the technique partitions the total gap (in years) into pieces that are attributable to different causes of death that occur because of differences in cause-specific mortality rates between blacks and whites.

  3. I don’t like it either. I don’t get why there’s a pale/dark red and pale/dark blue for the genders. Is pale supposed to mean C.I.? If so it’s unnecessary. Also, there should at least be a white space or a black line between the top graphs and the bottom graphs because they don’t represent the same thing. The legend also is not relevant much to the above graph and then the dotted line is for black and black line is for white. Why not put a grey for white and black for black. Or pale color for white and dark color for black. The plot is really nice looking but very confusing.

  4. After reading your paper with Phil on why all maps of parameter estimates are misleading, I was curious when I saw they had a map portrayal under Trends in Cause Contribution. I quite like the way they did that; it seems to be true to the mapping goal and to avoid the concerns you two once had.

    On the cross-sectional cause contribution, I struggle to see the solid black vertical line when it’s at the very end of the horizontal bar. If it were just a bit wider, that might help–or perhaps there’s something funny about the way it’s rendered at the far right.

    When exploring a state, I was temporarily confused because the Population Growth graphs use a solid line for blacks, while the Trends in Life Expectancy graphs used a dashed line for blacks. Consistency could help.

    I don’t know if it’s color-blind safe.

    Other than that, I’m more interested in what it’s saying than how it’s saying it, which, I think, is the (or a major) goal of good graphics.

    • Oh, and they could probably round a few of the estimates that pop up on hover (at one point, the life expectancy gap in Washington State was 4.809874 years).

      I do like the Download plot as a png option that captures both of a pair of graphs together. I’m not as sure of the result when I zoom in: it seems to put one graph partially on top of the other.

      I also like the availability of the spike lines.

  5. On the first, I don’t think they need a separate graph for difference – the trend is perfectly clear from looking at the gap between the plots at the top. The separate graph adds a lot of visual noise, but little new information.

    • Wasn’t one of Cleveland’s conclusions that we can discern horizontal lines better than sloped ones, and so it should be easier to sense if a difference is steady (horizontal) or not by plotting that difference rather than by plotting the more likely sloped individual lines?

    • Hello gdanning,

      One of the benefits of the separate graph for the difference is that we can plot the CI for the difference. If you select a small state (say Rhode Island) the CI is quite wide and directly seeing the CI for the difference is better than trying to discern its CI for the preceding graph.

  6. When it comes to life expectency, the place to go is Brad Plummer’s chart number 7 at

    https://www.washingtonpost.com/news/wonk/wp/2013/05/24/these-31-charts-will-destroy-your-faith-in-humanity/?utm_term=.6a1561c9c976

    There you will find that as of 1998 in the U.S. as a whole, life expectancy was already 77. Or, as Plummer put it, “More 77-year-olds are dying than ever before.” Therefore, the Alabama curves indicate that its famous motto, “Thank God for Mississippi” is well deserved.

Leave a Reply to gdanning Cancel reply

Your email address will not be published. Required fields are marked *