Georgia 21 Day COVID-19 Death Forecasts

Nothing much going on in the tropics (couple of suspect areas off the west coast of Mexico), no recent earthquakes, so let’s take a closer look at Georgia COVID-19 fatality numbers, and how to to use them to get a look into the future.  Yes, there will be math! And an equation!!

I’m really tired of hearing people say “the spike is because of testing” or  “the fatality rates are dropping!” implying that things are not headed the wrong way. Well, the rates are dropping are a little, but that’s not helping as much as you might think.  Lets look at the ratio of people dying to new positives. That’s OK in concept, but most people doing those graphs are failing to account for how the disease progresses.  On average, it is three weeks between someone testing positive and the date they expire (and exposure to expiration time lag is probably more like 4 weeks).  If you don’t take in to account that 21 day time lag, you can reach a wrong and dangerous conclusion.  Here is what the wrong calculation (just taking the deaths and dividing by positives) ratio, and a much better calculation (taking in to account the three week delay) looks like if we plot it.  Orange is the 21 day lag ratio, and blue is the “wrong” way that is duping people …

So, yes, better treatment and more younger people catching this is probably improving survival rates.  But it’s not as dramatic as the proponents of that view suggest.  The proof, of course, is in the prediction.  Let’s go back in time to the days of yore, May 30th.  At that time the lag21 Positive Fatality Ratio was 6.111%.  It had been slowly dropping, so we compute the trend and end up with an equation that looks like this:

Forecast = (0.0611-days_since_may_30th*(delPFR – K*days_since_may_30th)*L21P
— where:
delPFR:  the change in Positive Fatality Ratio as of May 30th (2.8e-4)
K: constant to adjust for improving survival rates (calculated over 1 to 30 May, 9.2e-7)
L21P:  the number of positive tests 21 days before the date you are trying to forecast

so with this equation, we can forecast up to 21 days in the future.  It’s a very simple model (we have more sophisticated ones), but is easy to explain and only uses reported information.  How did it work, and what does it show for the next three weeks?  Here’s that plot …

Yesterday the reported deaths were 2849.  On June 11th, three weeks ago, the prediction for yesterday was 2913, only a 2.25% error.  Properly using simple tools, we can do a pretty good forecast out three weeks into the future.  It has worked reliably over the last month, so we can expect it will continue to work well unless something dramatic changes, which is pretty amazing, isn’t it?  I mean, except for the fact it is showing that because a bunch of morons aren’t taking all this seriously, and doing very simple things like wearing a mask in congested places, etc., it is very possible that almost as many people will die over the next three weeks than died over the last three months.  And we can’t do much about it because it depends on what we did in the past, and even if everybody changes their behavior today, it will take weeks for it to show up in the statistics. But aside from that, isn’t math great?  Sigh.

6 thoughts on “Georgia 21 Day COVID-19 Death Forecasts

  1. Good analysis, but I have to disagree with this line. “Yesterday the reported deaths were 2849. On June 11th, three weeks ago, the prediction for yesterday was 2913, only a 2.25% error.” This is not an accurate way to measure error. You are making a prediction for the next 3 weeks: what matters is that you predicted 538 deaths over 3 weeks vs 474 actual, which is a 13.5% error. This is much more than 2.25%.

    • I totally disagree – you are comparing something different from what that model was trying to predict. The model is trying to predict cumulative deaths on a given day, not cumulative deaths over the time period. Each day is an independent forecast so it is incorrect to do cumulative error. If you want to be pedantic and really assess model error, strictly speaking you should be comparing each day over the period then assessing the distribution of errors since the target variable is deaths on a given day. Done that way, the average daily error was -0.27%, the peak overestimate 3.05%, the peak underestimate -2.96%. (or, if you prefer, the average is an undercount of 6 people, the peak overcount was 87, the peak undercount 78).

  2. But cumulative deaths is just current deaths + future deaths, and current deaths are not a variable, they are a constant. There is no accuracy involved in predicting current deaths. If you asked me to predict US GDP 3 weeks from now, it’s current GDP +/- a very small change, and my prediction will undoubtedly be close to the answer in absolute terms, but it’s also worthless because anybody can calculate that. The real value of a model is how well it can predict unknowns. Your comment about daily error is strictly correct, but since the variable itself is noisy I agree that 3-week predictions are more useful.

    • Well, now what you’re really talking about skill, which is a whole different discussion we can have separately. I’d also point out this isn’t a small delta – over the model period, deaths have gone up by 38%.

      Which raises a point: the model isn’t using current deaths. Look at the equation: it is predicting deaths three weeks in the future from current POSITIVE TEST RESULTS (yes, based on the past relationship between positive test results and deaths with a 21 day lag, but not using current deaths dynamically). If that relationship were changing, the results would diverge (which if you use a static ratio it does rather dramatically.

      I’ll gently suggest that you’re missing the point here: the relationship between test results and deaths. This wasn’t intended to be a scientific paper, or some brilliant/innovative model, but a discussion accessible to the somewhat literate general public to try to get across several points, the biggest being that there is a three week lag between increased positive death results and those deaths showing up in the daily death totals. And if that holds, the next few weeks will be depressing.

      The argument some are making is that more testing is causing more positives, or that relationship is changing radically because more younger people are showing up positive. Neither seems to be true – if it were, this relationship would break down. So it’s not “worthless” because people are denying this fundamental relationship, and this hopefully shows that.

  3. Pingback: Georgia COVID deaths: the critical week ahead | Enki Research

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.