COVID-19 Megathread 5: The Trumps catch COVID-19

[Fmr. Gov. NickG]

I think there’s a bunch of explanations for why the death rate continues to decrease that don’t require something as deus ex machina as a mutation of the virus. 

The biggest one is probably simple increased testing, which reduced the death rate in two ways.  First, it identifies a much higher proportion of mild and asymptomatic cases, thus increasing the denominator of the death rate without increasing the numerator.  The second is that people discover they are infected much earlier, allowing them to seek medical attention sooner.

The second biggest is improved treatment.  Over the last few months, we have developed half a dozen new avenues for treating the virus.  Each of the independently may reduce deaths 10-20%, but across all treatment, especially when used in conjunction with each other, it is very plausible to me that more than half of patients who would have died in April are now recovering.

Third, there does seem to be evidence that the age of the average infected person has dropped, even independent of the increase in testing.  Perhaps it spread more among older professionals at first who brought it through airplanes and conferences before spreading now more to college students?  I think we’ve all also become more aware of the disproportionate impact on older people, and this may cause older people to socially distance more relative to younger people, and more relative to a few months ago.

[Fmr. Gov. NickG]

I read a post/article on Iran's COVID-19 response that seems relevant.  Iran experienced a second wave similar to what we are experiencing now (rising cases, lower average age of of cases, but not yet seeing a corresponding rise in deaths).  Deaths didn't spike initially but did 3-4 weeks later after the younger people passed the virus onto other subgroups (elderly, etc.) more likely to die from the virus.

This is not a second wave. This is a revival of the first wave, instigated by bad policy and worse community practices.

I wasn't claiming that the US is experiencing a second wave.  I agree we are still in the first wave.  I was referencing Iran's second wave.

"Iran experienced a second wave similar to what we are experiencing now." The "what" is ambiguous, and lends itself to be interpreted as a pronominal equivalent of "second wave."

Thanks for the clarification.

I think it's all semantics anyway. Cases are never going to go to 0, so it's subjective when the first wave has ended.

For a wave to end, the numbers have to plateau at a low non-zero number for a while, and then start a new curve. The U.S. never managed to do that to begin with.

It is somewhat semantic, but you do observe clear bimidal curves in a handful of places, Iran being one of them.  The only state I’ve seen where you could argue there’s a legitimate second wave right now is Louisiana.

[Fmr. Gov. NickG]

https://twitter.com/sahilkapur/status/1277008820066226176

Apparently the Nevada number is the result of a data delay error, and includes about 600 cases that should have been reported on Tuesday or Wednesday.  It’s still close to a record without them.

But as with many states, cases have been rising there for about 4 weeks, and their death rate has remained steady at around 3 deaths per day.

[Fmr. Gov. NickG]

You have got to watch this to believe. Looks like things are likely to keep getting worse, at least in Texas...

https://twitter.com/PeterHotez/status/1276886772287442944

Isn’t this just evidence that a lot of people are getting tested?

[Fmr. Gov. NickG]

You have got to watch this to believe. Looks like things are likely to keep getting worse, at least in Texas...

https://twitter.com/PeterHotez/status/1276886772287442944

Isn’t this just evidence that a lot of people are getting tested?

Watch the video. There are a lot of people there waiting in line to be tested... Which is good... But what is not good is that they are all standing in a closely packed line unnecessarily close right next to each other... Despite the fact that they are in line specifically at a COVID testing site... Where you would expect that a disproportionate number of people relative to the general population are in fact infected... And especially in a state/city in which there is a high positivity rate...

The question you should be asking yourself, given this, is why are they not standing 6 feet apart? TBH, why are they not standing further apart than that, like 12 feet (why is there this obsession with 6 feet anyway, as though it is perfectly fine to stand exactly 6 feet away if you could easily be standing a bit further away at no real additional cost???)

OK, but we've been seeing images and videos of people not social distancing for the last two months.  At least here they are outside and they are almost all wearing masks. 

It would be nice if they were standing farther apart.  But that's not why the video has been circulating so widely.  It's because people mistakenly believe these are sick people waiting for care.

[Fmr. Gov. NickG]

You have got to watch this to believe. Looks like things are likely to keep getting worse, at least in Texas...

https://twitter.com/PeterHotez/status/1276886772287442944

Isn’t this just evidence that a lot of people are getting tested?

This site is now listed as COVID-19 testing by appointment only. Some other sites by the same chain (of urgent care facilities) are indicating no appointment is necessary.

This is pretty close to Washington Avenue (bars/bar-hopping), and until recently the Heights (where the clinic is located was dry). Bars statewide were closed at Friday noon. It is possible that there were outbreaks at the bars, and on Friday might have urged would-be patrons to get tested and listed nearby testing locations.

Testing statewide had been pretty steady statewide around 30K per day, while the positivity rate stayed low around 5% or declined some. This could mean that testing had become fairly easy to get done, but there was less urgency when 95% were coming back negative. If a urgent care clinic wanted to drum up business, they might put on their sign out front "COVID-19 NO APPOINTMENT NEEDED".

Then all of a sudden there was a rush (statewide tests are now over 40,000 per day and positivity is up to 13%) which means people are actually experiencing symptoms before being tested just to confirm what they already suspected) and a line developed. If you have a fever or a cough, or a friend who was infected, you aren't going to think, I will wait until Monday to call my primary care physician (if you are in your 20s, particularly if male, you don't have a primary-care physician, and when you call your Mom, she will remind you that Dr. Jones was your pediatrician, and he retired 7 or 8 years ago). You are going to want it done now, and will get in the end of the line, figuring it was like the run on toilet paper.

So I drove by the urgent care facility to see if there was a line today. Absolutely nobody outside. Hah! I thought, they have switched to appointments! I pulled in the parking lot to look a little closer. On all the doors there was a white sheet of paper: "Out of COVID-19 Tests".

If they are rationing testing, that is definitely a very bad thing, and would go a long way to explaining a rise in positivity rate.  Obviously the more difficult you make it to get a test, the fewer asymptomatic people are going to seek one out.  

Texas should certainly be doing way more than 40k tests per day right now.  That puts them well below the per capita average of the US.

[Fmr. Gov. NickG]

https://twitter.com/StevenTDennis/status/1277295367827136514

Record cases on a day when usually we see lower case numbers than usual? That can't be good.

Case numbers don't experience nearly as much drop-off on Sundays as death numbers.
That said, these don't bode well for this state either.  Arizona reported nine deaths today, compared to just one death last Sunday.

[Fmr. Gov. NickG]

So, today appears to be the first day in a long time where there are reported more deaths than the same day a week before.

Indeed, week-over-week deaths were up from 271 to 285 according to worldometers.  They were also up slightly on Tuesday, though there was an old data dump that made that unclear.   Also, it seems the state most responsible for the spike today was New Jersey, which went from 14 to 27 week-over-week deaths.

It’s pretty sad that New Jersey, whose case average peaked on April 7, still has triple the per daily capita death rate of California, which has seen cases steadily rising nonstop for almost three months.

[Fmr. Gov. NickG]

And on worldometer, not only are the US deaths reported on Sunday higher than the previous Sunday, but also the week-over-week seven day average # of deaths is higher.

If you are just saying the 7-day average rose from yesterday, this is tautological (it will always rise when the week-over-week daily number rises).  If you are comparing the 7-day average to the previous Sunday, it has fallen, though not as quickly as some previous weeks (596 on June 28 vs. 636 on June 21 and 782 on June 14).

[Fmr. Gov. NickG]

https://twitter.com/ScottGottliebMD/status/1277656773982392320

What is “total predicted infections of 3.8 million” supposed to mean?  Haven’t 20-30 million Americans already been infected?

[Fmr. Gov. NickG]

I was interested in modeling how much of a lag there is between case numbers and death numbers in the US, so I plugged all of the case and death data into Stata and ran a regression, along with a linear time trend and fixed effects for days of the week.  The DV here is daily deaths, and there are five IVs related to cases: the seven-day case average on that day, and on the days 7, 14, 21, and 28 days earlier.  I've pasted the results below for days ranging from April 1 to June 28.

This will be inscrutable if you're not used to State data.  But the upshot is that the lagged effect of cases on deaths is pretty clear: 1-2 weeks (specifically the average of cases from 7-13 days earlier). The "avgminus7" variable has a large and significant effect, predicting an additional death for every 17 additional cases in the average one week earlier.  The average of the current week, along with the 2, 3, and 4, weeks earlier, all have a small and insignificant effect.  On top of this, deaths tend to drop about 15 per day, so an increase in 255 cases per day will actually result in a constant number of deaths, not an increase.

reg deaths days avgcases avgminus7 avgminus14 avgminus21 avgminus28 i.day if days>31

      Source |       SS           df       MS      Number of obs   =        88
-------------+----------------------------------   F(12, 75)       =     76.62
       Model |  32413040.9        12  2701086.74   Prob > F        =    0.0000
    Residual |  2643942.96        75  35252.5728   R-squared       =    0.9246
-------------+----------------------------------   Adj R-squared   =    0.9125
       Total |  35056983.8        87  402953.837   Root MSE        =    187.76

------------------------------------------------------------------------------
      deaths |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
        days |  -15.10064   1.274004   -11.85   0.000    -17.63858   -12.56269
    avgcases |  -.0038266   .0089625    -0.43   0.671    -.0216808    .0140275
   avgminus7 |    .058964   .0141994     4.15   0.000     .0306774    .0872506
  avgminus14 |   .0049107   .0153738     0.32   0.750    -.0257156    .0355369
  avgminus21 |    .002581   .0134527     0.19   0.848    -.0242181      .02938
  avgminus28 |  -.0053081   .0078224    -0.68   0.499    -.0208912    .0102749
             |
         day |
          2  |   152.0686   76.66254     1.98   0.051    -.6509597    304.7882
          3  |   826.6271   76.69768    10.78   0.000     673.8375    979.4167
          4  |    818.107   75.29342    10.87   0.000     668.1148    968.0992
          5  |   675.9876   75.26829     8.98   0.000     526.0454    825.9297
          6  |   595.3323   75.31052     7.91   0.000     445.3061    745.3586
          7  |   404.5534    75.4019     5.37   0.000     254.3451    554.7617
             |
       _cons |   549.2656    197.128     2.79   0.007     156.5665    941.9647
------------------------------------------------------------------------------

The graph below shows the predicted deaths generated from the model against actual deaths (starting from April 1). I think it fits pretty well.  The exception is that because I am constraining a constant effect onto each weekday, the weekday effects are too small when deaths are exceptionally high and a bit too large when deaths are much lower.  

[Fmr. Gov. NickG]

For what it's worth, I've noticed that any time the number of known cases in a state or county gets to about 1% of the state or county's total population, the new cases start cratering.

I believe Arizona passed this threshold yesterday, so you’ve got a good test of your theory to look forward to this week.

[Fmr. Gov. NickG]

These states that are shutting down again had better have learned their lesson about "saving the economy" and minimizing the "authoritarianism". The more times they open and shut, the worse off the economy will be. Close early and stay closed until the right time without rushing it.

How do you know when the time is right?

[Fmr. Gov. NickG]

But the upshot is that the lagged effect of cases on deaths is pretty clear: 1-2 weeks (specifically the average of cases from 7-13 days earlier).

...

The graph below shows the predicted deaths generated from the model against actual deaths (starting from April 1). I think it fits pretty well.  The exception is that because I am constraining a constant effect onto each weekday, the weekday effects are too small when deaths are exceptionally high and a bit too large when deaths are much lower.  

https://i.imgur.com/LPUY62j.png

Nice to see some effort put into some statistical analysis - well done!

There are, however, some important omitted variables. That would certainly include some that you yourself have emphasized in earlier posts - 1) improvements in medical care that lower the fatality rate and 2) increased testing detecting more cases and 3) reduced median age of the newer cases as opposed to the ones from March/April. All of these would tend to lower the ratio of deaths to confirmed cases in more recent observations as compared to earlier observations.

Because all 3 of those omitted variables would lower the ratio of deaths to confirmed cases in more recent time periods, that would seem to imply that if those were taken into account, the guesstimated lag time you would find would be longer than the one you came up with from this regression that omitted them. So that would mean that your results don't really suggest that the true lag period is 1-2 weeks, they suggest that the lag period is longer than 1-2 weeks by some unknown amount.

It would be hard/impossible to get reliable data to quantify how much the "true" fatality rate has declined as a result of better treatment. It would also probably be hard (though theoretically doable if you had median age data for each day and could make some assumptions about the age distribution of cases) to take into account the impact of the median age on fatality rate. However, you could fairly easily include the positivity rate as a variable indicating how much testing has improved.

So if you feel like adding it in, I wonder, what happens to the regression if you include the positivity rate? I haven't thought about it carefully, but it also seems at first glance like that would actually be linear in its effects on the lag (alternatively to the positivity rate, you could include tests as a variable, which seems like it ought to have the same effect as including the positivity rate, due to the formula for positivity rate being calculated from tests and cases - and if you were to include both, you should get multicollinearity).

So does including the positivity rate (or alternatively tests) as a variable raise or lower the lag that you would estimate (I would expect it would raise the lag time)?

If you had data to take into account changes in median age and changes in quality of care/availability of new treatments, those should have effects in the same direction.

The linear time term (the “days” variable) is likely very crudely picking up effects of all of these omitted variables right now.  Including the positivity rate is a good idea, and I will need to think about how to incorporate it; it would probably needed to be interacted with cases (and lagged cases) in order to be useful.   I’m not sure how I would disentangle the treatment/patient age effects since I don’t know that there is daily national data on that available.

[Fmr. Gov. NickG]

For those counting cases, it looks like there are some optimistic signs today that the surge in cases over the last several days in the south may have peaked, especially considering that Tuesday is usually the heaviest day.  All of these states are reporting casee numbers today below yesterday’s 7-day average:

Florida: 6093 cases today vs. 6589 average over last 7 days
Arizona: 2228 today vs. 3200 over last 7 days
Georgia: 1874 today vs. 1927 over last 7 days
North Carolina: 1229 today vs. 1428 over last 7 days
Alabama: 870 today vs. 961 over last 7 days

All of these states are also reporting a drop in week-over-week deaths.

The only state I can see at first glance with a new record is South Carolina, which reported 1755 cases versus a 7-day average around 1300. (I didn’t try to look up really small states though)

It’s too soon to know about Texas and California today.

[Fmr. Gov. NickG]

6/29 (Yesterday):

• Cases: 2,681,802 (+44,725 | Δ Change: ↑10.32% | Σ Increase: ↑1.70%)
• Deaths: 128,779 (+342 | Δ Change: ↑20.00% | Σ Increase: ↑0.27%)

6/30 (Today):

• Cases: 2,727,853 (+46,051 | Δ Change: ↑2.96% | Σ Increase: ↑1.72%)
• Deaths: 130,122 (+1,343 | Δ Change: ↑292.69% | Σ Increase: ↑1.04%)

Deaths jumped by a thousand today, compared to yesterday. What explains this? Perhaps it is the inevitable outcome of the recent surge in cases which we've seen.

Worldometers included this note today:
“On June 30, the count of New Yorkers who have died of COVID-19 increased by 692. Most of that increase is due to new information we received from the NYS Department of Health about city residents who died outside the city. The vast majority of these deaths occurred more than three weeks ago."

These were added to the total number of US deaths, but not the daily total for yesterday, which was reported as 764.  I’m a little confused about why these were added, as it seems like it would lead to double-counting in many cases.  Are the states consistent about counting deaths among residents of their state versus deaths that actually occured in that state?

[Fmr. Gov. NickG]

Well I definitely spoke to soon yesterday about the numbers looking a bit better in Arizona.   Both the case and death figure there are absolutely terrible.  Florida still seems like it might have stabilized, with still no noticeable increase in week-over-week deaths.

[Fmr. Gov. NickG]

https://twitter.com/VaughnHillyard/status/1278349068817833984

I think this might be the highest number of per capita new cases that any state has ever reported.

Edit: For reference, this would be the equivalent of 220,000 new cases if scaled up to the US population.

[Fmr. Gov. NickG]

I updated my Stata model to include test positivity rate and its interaction with case.  These are the results, estimating daily deaths from April 1 to June 30.

Here, "avgcases" is the average number of cases over the previous 7 days, while "averagecase7" is the average for the 7 days before that (etc. for 14, 21, and 28 days before that). "{osperc" is test positivity rate on the given day, while "posminus7" is the lagged positivity rate 7 days earlier (etc for 14, 21, and 28 days earlier).  "Casespos" (and its lagged variables) are the interaction between total cases and positivity rate.  "Deathminus7" is the number of deaths 7 days earlier.  I also include fixed effects for each day of the week.

The overall impact of lagged cases is still very similar.  Average cases lagged one week are the only variable with any substantive or statistical significance on their own.  

For the most part, adding positivity rate has no significant effect, with one exception.  Positivity is significant in interaction with 14-day lagged cases.  So while cases lagged 7-days always have an effect on deaths, cases lagged 14-days only have a major effect when test positivity (also lagged 14-days) was high.  For example, an additional 1000 cases 14 days earlier would result in 8 additional deaths when test positivity was 3%, but 21 additional deaths when positivity was 10%.  Current cases, as well as cases 3 weeks ago or more, never have a significant effect on deaths.

. reg deaths deathminus7 days avgcases avgcases7 avgcases14 avgcases21 avgcases28 posperc posminus7 p
> osminus14 posminus21 posminus28 casespos casespos7 casespos14 casespos21 casespos28 i.weekday if da
> ys>31&days<123

      Source |       SS           df       MS      Number of obs   =        91
-------------+----------------------------------   F(24, 66)       =     60.11
       Model |  35728644.6        24  1488693.53   Prob > F        =    0.0000
    Residual |  1634438.67        66  24764.2223   R-squared       =    0.9563
-------------+----------------------------------   Adj R-squared   =    0.9403
       Total |  37363083.3        90   415145.37   Root MSE        =    157.37

------------------------------------------------------------------------------
      deaths |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
 deathminus7 |   .4674046    .104832     4.46   0.000     .2581007    .6767085
        days |  -11.89638   5.265929    -2.26   0.027    -22.41015   -1.382614
    avgcases |  -.0095963   .0120319    -0.80   0.428    -.0336188    .0144262
   avgcases7 |   .0395809   .0173289     2.28   0.026     .0049827    .0741792
  avgcases14 |  -.0491985   .0171413    -2.87   0.006    -.0834222   -.0149747
  avgcases21 |   .0159292   .0149175     1.07   0.289    -.0138546    .0457129
  avgcases28 |  -.0051158   .0109589    -0.47   0.642     -.026996    .0167645
     posperc |  -1690.344   2372.583    -0.71   0.479    -6427.358     3046.67
   posminus7 |   715.4029   2176.551     0.33   0.743    -3630.221    5061.027
  posminus14 |  -5678.463   2039.865    -2.78   0.007    -9751.185   -1605.741
  posminus21 |   676.9361   1301.756     0.52   0.605    -1922.104    3275.976
  posminus28 |  -99.96956   868.0994    -0.12   0.909    -1833.186    1633.247
    casespos |   .0746608   .0778795     0.96   0.341    -.0808307    .2301522
   casespos7 |  -.0269732   .0704437    -0.38   0.703    -.1676185    .1136721
  casespos14 |   .2120465   .0679098     3.12   0.003     .0764602    .3476329
  casespos21 |  -.0446484   .0502626    -0.89   0.378    -.1450009    .0557042
  casespos28 |   .0143544   .0445213     0.32   0.748    -.0745353     .103244
             |
     weekday |
          2  |   82.38582   73.89428     1.11   0.269    -65.14887    229.9205
          3  |   420.2182   111.0866     3.78   0.000     198.4268    642.0097
          4  |   343.6183   116.7902     2.94   0.004     110.4391    576.7975
          5  |   227.4863   113.5001     2.00   0.049     .8759877    454.0966
          6  |    136.587   126.1011     1.08   0.283     -115.182     388.356
          7  |   101.4743   93.32886     1.09   0.281    -84.86273    287.8114
          8  |  -281.1558   156.3947    -1.80   0.077    -593.4079    31.09634
             |
       _cons |    1633.96   693.1888     2.36   0.021     249.9637    3017.956
------------------------------------------------------------------------------

[Fmr. Gov. NickG]

Here's a graph of the new model fit:
(The x-axis "Days" is days since March 1.  Because of all the lags, I entered data starting March 1, but only started estimating deaths starting April 1.)

[Fmr. Gov. NickG]

Here’s an article published in the NY Times today on the potential for human challenge trials...I was beginning to lose hope about this, but the article was more optimistic than I expected.

https://www.nytimes.com/2020/07/01/health/coronavirus-vaccine-trials.html

Sign up at 1DaySooner.org!

[Fmr. Gov. NickG]

I updated my Stata model to include test positivity rate and its interaction with case.  These are the results, estimating daily deaths from April 1 to June 30.

Here, "avgcases" is the average number of cases over the previous 7 days, while "averagecase7" is the average for the 7 days before that (etc. for 14, 21, and 28 days before that). "{osperc" is test positivity rate on the given day, while "posminus7" is the lagged positivity rate 7 days earlier (etc for 14, 21, and 28 days earlier).  "Casespos" (and its lagged variables) are the interaction between total cases and positivity rate.  "Deathminus7" is the number of deaths 7 days earlier.  I also include fixed effects for each day of the week.

The overall impact of lagged cases is still very similar.  Average cases lagged one week are the only variable with any substantive or statistical significance on their own.  

For the most part, adding positivity rate has no significant effect, with one exception.  Positivity is significant in interaction with 14-day lagged cases.  So while cases lagged 7-days always have an effect on deaths, cases lagged 14-days only have a major effect when test positivity (also lagged 14-days) was high.  For example, an additional 1000 cases 14 days earlier would result in 8 additional deaths when test positivity was 3%, but 21 additional deaths when positivity was 10%.  Current cases, as well as cases 3 weeks ago or more, never have a significant effect on deaths.

. reg deaths deathminus7 days avgcases avgcases7 avgcases14 avgcases21 avgcases28 posperc posminus7 p
> osminus14 posminus21 posminus28 casespos casespos7 casespos14 casespos21 casespos28 i.weekday if da
> ys>31&days<123

      Source |       SS           df       MS      Number of obs   =        91
-------------+----------------------------------   F(24, 66)       =     60.11
       Model |  35728644.6        24  1488693.53   Prob > F        =    0.0000
    Residual |  1634438.67        66  24764.2223   R-squared       =    0.9563
-------------+----------------------------------   Adj R-squared   =    0.9403
       Total |  37363083.3        90   415145.37   Root MSE        =    157.37

------------------------------------------------------------------------------
      deaths |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
 deathminus7 |   .4674046    .104832     4.46   0.000     .2581007    .6767085
        days |  -11.89638   5.265929    -2.26   0.027    -22.41015   -1.382614
    avgcases |  -.0095963   .0120319    -0.80   0.428    -.0336188    .0144262
   avgcases7 |   .0395809   .0173289     2.28   0.026     .0049827    .0741792
  avgcases14 |  -.0491985   .0171413    -2.87   0.006    -.0834222   -.0149747
  avgcases21 |   .0159292   .0149175     1.07   0.289    -.0138546    .0457129
  avgcases28 |  -.0051158   .0109589    -0.47   0.642     -.026996    .0167645
     posperc |  -1690.344   2372.583    -0.71   0.479    -6427.358     3046.67
   posminus7 |   715.4029   2176.551     0.33   0.743    -3630.221    5061.027
  posminus14 |  -5678.463   2039.865    -2.78   0.007    -9751.185   -1605.741
  posminus21 |   676.9361   1301.756     0.52   0.605    -1922.104    3275.976
  posminus28 |  -99.96956   868.0994    -0.12   0.909    -1833.186    1633.247
    casespos |   .0746608   .0778795     0.96   0.341    -.0808307    .2301522
   casespos7 |  -.0269732   .0704437    -0.38   0.703    -.1676185    .1136721
  casespos14 |   .2120465   .0679098     3.12   0.003     .0764602    .3476329
  casespos21 |  -.0446484   .0502626    -0.89   0.378    -.1450009    .0557042
  casespos28 |   .0143544   .0445213     0.32   0.748    -.0745353     .103244
             |
     weekday |
          2  |   82.38582   73.89428     1.11   0.269    -65.14887    229.9205
          3  |   420.2182   111.0866     3.78   0.000     198.4268    642.0097
          4  |   343.6183   116.7902     2.94   0.004     110.4391    576.7975
          5  |   227.4863   113.5001     2.00   0.049     .8759877    454.0966
          6  |    136.587   126.1011     1.08   0.283     -115.182     388.356
          7  |   101.4743   93.32886     1.09   0.281    -84.86273    287.8114
          8  |  -281.1558   156.3947    -1.80   0.077    -593.4079    31.09634
             |
       _cons |    1633.96   693.1888     2.36   0.021     249.9637    3017.956
------------------------------------------------------------------------------

If you click on "Source Mode" in the lower right, and then paste from a spreadsheet, it will be formatted so the columns line up better. Note clicking on "Source Mode" actually toggles you into "Display Mode". You can click on "Display Mode" to return you to Source Mode and edit any text.

I'd probably also trim some of your digits.

Trying to make sense of the numbers. Is it the fact that coef has the largest absolute value that indicates that deaths are lagging detection by about 14 days?? That the coef for 7 days is the next largest suggests that if might be less than 14, but to to refine the number may be impossible due to the weekly periodicity of reporting??

I repasted only the significant effects here:

------------------------------------------------------------
      deaths |      Coef.   Std. Err.      t    P>|t|     
-------------+---------------------------------------------

 deathminus7 |   .4674046    .104832     4.46   0.000 
 
        days |  -11.89638   5.265929    -2.26   0.027   

  avgcases7 |   .0395809   .0173289     2.28   0.026   

  avgcases14 |  -.0491985   .0171413    -2.87   0.006
  posminus14 |  -5678.463   2039.865    -2.78   0.007   
  casespos14 |   .2120465   .0679098     3.12   0.003     

   _cons |    1633.96   693.1888     2.36   0.021   

The positive coefficient for 7-day lagged cases (avgcases7), with no significant effects on 7-day test positivity, is straightforward.  For every additional case in the 7-day lagged average, you get an additional .04 deaths (25 cases=1 death).

The effect of the 14-day lagged period is more complicated and can only be understood by looking at the three variables in conjunction with each other.  "Casepos14" is the product of 14-day lagged cases and 14-day lagged positivity rate (an interaction term).  When you have a positive coefficient for the interaction term, but a negative coefficient for both uninteracted terms, that indicates that you should only see a positive effect when both terms are high.   So 14-day lagged cases increase deaths much more when the 14-day lagged test positivity rate was also high.   

It's also worth noting that when we talk about positivity rate being "high", we're not so much referring to the variation over the last few weeks, when national positivity rate has been in the 4-7% range, but rather in March and April, when it was 15-20%.

Finally, I should also mention that cases and positivity rate end up explaining only a very small portion of variation in deaths.  You'll see the R^2 value of the model above is .956.  That's really high, and means the model explains about 96% of the variation in the data.  However, you can get to an R^2 value of .916 in a model that includes only data on lagged deaths and day of the week, with no information about cases or tests whatsoever.  I've posted the predictions of this model, which I called "crudepredict" below (green line), contrasted with the actual data (blue line) and the full model (red line).  You'll see the green line is a slightly less accurate predictor of the blue line data than the red line, but they both do pretty well. 

Just knowing the previous death rate and the day of the week gets us 92% of the way there.  Adding in those 15 extra variables on cases and tests only gets us an additional 4%.

[Fmr. Gov. NickG]

It is really strange to me how much the coverage of the pandemic has shifted away from reporting death numbers and toward just reporting case numbers, and away from a national perspective toward a focus on specific states.

It seems like this should be the reverse; we should report more on cases early on when we know little about the spread of the virus and its effect, and more on deaths later, especially as we learn more about what actually causes and prevents deaths.   

As an example, all during April and May, the Washington Post has a very prominent graphic on its front page tracking daily deaths from the virus.  But at some point in June, they removed this graphic and replaced it with a graphic of "places with most reported cases per capita", with no mention of deaths at all.

It really seems like our reporting is stuck in the weeds of a few states with surging cases in the moment, with little perspective on the overall course and impact of the virus going forward.

[Fmr. Gov. NickG]

Massachusetts is reporting 51 deaths today (14 more than Arizona and just 13 less than Florida) just two days after they reported zero deaths.  That seems like some real reporting voodoo.

[Fmr. Gov. NickG]

Does anyone have an explanation for what is going in on Sweden right now?

They never really saw any reduction in their cases, and in fact cases have been rising steadily for about five weeks now (from about 600/day on average to 1000/day).

Yet deaths have been declining consistently for over three months, from a 7-day average of almost 100/day, to an average yesterday of six.

They almost look like an exagerrated version of the US, with cases continuing to explode but deaths almost entirely disappearing.