The November Jobs Report Was Actually Quite Solid: One Should Not Expect More Going Forward

A.  Introduction

The Bureau of Labor Statistics (BLS) released its regular monthly “Employment Situation” report, for November 2021, on Friday, December 3.  The report is always eagerly awaited.  It provides estimates for the net number of new jobs created in the most recent month, as well as figures on the unemployment rate, certain wage measures, and much else.

The initial reaction to the report by the media was negative.  Net job growth, estimated at 210,000 in the month, was viewed as disappointing.  This was down from 546,000 net new jobs in October, and was well below Wall Street expectations (based on a survey of Wall Street firms by Dow Jones) that the figure for November would come to 573,000.  While it was noted that the unemployment rate also fell – to just 4.2% – the negative reaction contributed to a significant decline in the stock market that day, with the S&P 500 index, for example, down by over 2% at one point.

But the November jobs report was actually pretty solid.  In this post, we will look at what was reported and some factors to take into account when examining such figures.

B.  Monthly Job Gains in 2021

The chart at the top of this post shows the current BLS estimates of monthly net job growth this year, starting in February to cover the period of Biden’s presidency  The estimates are based on a survey of establishments by the BLS, that asks (along with much else) the number of employees on their payroll as of the middle week of each month.  Hence the January numbers would have been for before Biden’s January 20 inauguration.  The news reports following the release by the BLS of the November jobs report were often accompanied by charts such as this one, with the November figure showing a substantial reduction in the number of net new jobs compared to what was seen in earlier months.  The question of interest is whether this was significant.

A number of factors should be taken into account.  One is simply that there is substantial month to month variation, as seen in the chart.  This may be in part due to fluctuations in the economy, but may also be due to idiosyncratic factors (such as how the weather was in the week of the survey) and to statistical noise.  The figures are based on surveys, and surveys are never perfect.  Examined in context, the change in the November figure from the prior month is similar to the changes seen in other months this year.  Indeed, it was less than in several.

There will, however, always be limitations with any single estimate, and in part for this reason the BLS provides in its published document a few different estimates for employment growth. The measure shown in the chart at the top is rightly considered the best one.  It is based on a monthly survey (called the Current Employment Statistics, or CES, survey) of business and other establishments (including government entities as well as non-profits such as universities and hospitals) – whoever employs workers.  The sample size is huge:  144,000 different businesses and government entities, at almost 700,000 different worksites.   The BLS indicates this “sample” covers approximately one-third of all such jobs in the US.

The numbers are specifically for nonfarm payroll jobs, and hence exclude those employed on farms (which is now small in the US – about 1.4% of workers based on figures from other surveys) and more importantly the self-employed (about 6% of the labor force).  Given the large sample size, and also recognizing that those in the sample include not only small firms but also large entities employing thousands of workers, statistical noise is limited.  However, even with such a large sample size, the BLS states that the 90% confidence interval on the month to month changes in employment is +/- 110,000.  At the more commonly accepted 95% confidence interval it would be wider.

Finally, the figures for the prior two months in each report are preliminary and subject to change as more complete data comes in.  The November report, for example, indicated the estimate of net new jobs in October had been revised up by 15,000, and for September by 67,000.  And the October report last month indicated that its earlier estimate for September had been revised up by 118,000.  That is, the initial estimate for September had been 194,000 net new jobs, but this was revised up a month later to 312,000 net new jobs, and then revised again in the estimates published this month to 379,000.  Such revisions are routine, and one should expect that the initial estimate for November of 210,000 net new jobs will likely be revised in the coming months as more complete data becomes available.  While the revisions can in principle be positive or negative, in an expanding job market (as now) they are likely to be positive.  

The figures in the chart are also seasonally adjusted.  This is done via standard algorithms that estimate the normal annual pattern of employment changes in any given month based on historical data.  Employment growth is normally higher in certain months of the year (such as June, following the end of the school year) and normally lower in other months (such as January).  Analysts will therefore usually focus on the seasonally adjusted figures to see whether certain trends are developing outside of the normal seasonal fluctuations.

This is indeed appropriate.  However, it is also worth recognizing that due to Covid, with the resulting lockdowns, opening-ups, quite prudent changes in consumer behavior due to the health risks from Covid-19 even with all the protective measures taken that can be taken, and the truly historic fiscal relief measures provided through the government budget to support households in the light of all these disruptions, seasonal patterns this year (and last) are likely to be not at all similar to what they have been historically.  It is therefore of interest also to look at the underlying employment estimates, before the seasonal adjustment algorithms are run, to see what those numbers might be saying.

The next section will look at this, along with other measures of the change in employment.

C.  Alternative Measures, and Long-Term Limits on What Employment Growth Could Be

As noted, the BLS makes available in its monthly Employment Situation report several measures of how employment is estimated to have changed in the month, in addition to the one discussed above.  These additional measures should not be seen as better measures (at least in normal circumstances) than the seasonally adjusted measure based on the findings from the huge CES survey of establishments.  Rather, it is best to see them as supplementary measures, or alternative measures, that together help us understand what may be going on in terms of employment. There is always uncertainty in any individual measure, as they are all estimates.  It is better to look at several, to see what the overall story might be.

The estimated change in employment in November (or, more precisely, the change in nonfarm payroll), based on figures from the CES survey of establishments, was 210,000 after seasonal adjustment.  But three alternative estimates for employment growth in November were far higher, as depicted in this chart:

In the CES estimate before the normal seasonal adjustment, the growth in net new jobs in November was 778,000.  This difference between the seasonally adjusted and non-seasonally adjusted figures is substantially greater than what one has normally seen for November.  Seasonal adjustment is complicated, but a simple average of the difference between the seasonally adjusted figures for November and the non-seasonally adjusted figures over the 20 years from 2000 to 2019, is 205,000.  But in November 2021 it was 568,000, suggesting something unusual.  If the November 2021 increase in the number of jobs was adjusted by 205,000 rather than the 568,000 estimated by the algorithms, then the “seasonally adjusted” change in the number of jobs would have been 573,000 (= 778,000 – 205,000).  This is exactly what the pre-release expectation was on Wall Street (as noted at the start of this post).  That it was exactly the same as the Wall Street forecast is just a coincidence, but the fact it was close at all might be significant.  It may be suggesting that the standard seasonal adjustment calculations, built from patterns historically seen for the month, might not have captured well the circumstances in this highly unusual year.

Quite separately, the BLS also has an employment measure from the monthly survey of households conducted by the US Census Bureau (with BLS input on what is asked), called the Current Population Survey (CPS).  This survey of a sample of 60,000 households is used by the BLS to determine how many are in the labor force (i.e. are working or are looking for work), whether they are employed (including self-employed and on farms), and thus the number unemployed (those in the labor force but not employed).  The BLS uses this to determine the unemployment rate, but to get to that they have to first estimate, based on this survey, how many are employed.

The November estimates based on the CPS of net new employment were 1,136,000 for the seasonally adjusted figure and 831,000 for the figure before seasonal adjustment.  Why the seasonal adjustment led to a reduction in the job growth estimate from the CPS while it led to an increase in the job growth estimate from the CES is not clear (seasonal adjustment is complicated), but in any case, both figures are relatively close to the 778,000 estimate from the CES estimate before seasonal adjustment.  And all three are all well above the 210,000 seasonally adjusted estimate from the CES that we normally focus on.  Together they suggest that the 210,000 estimate, while usually the most reliable one, might in this case be on the low side.

I have also included in the chart four figures for what I have termed the “long-term limits” on what monthly job growth might be for an economy at full employment.  I included them on the same chart so that one can easily recognize the relative scale.

For an economy at full employment (with unemployment at frictional levels), employment growth cannot exceed the growth of the adult population.  And indeed it will be less, as not all adults (defined by the BLS as all those in the population at age 18 and above) will be in the labor force – some will be retired, some will be students in college, some will have voluntarily left the labor force to raise children or provide care for others, and for other reasons.  Examining what these limits are for the US will provide a sense of what monthly employment growth might be, on average, in the coming years.

First, on population:  Population growth is relatively steady and predictable.  For the ten-year period from November 2011 to November 2021, it averaged 180,000 per month in the US.  It will be similar to this in the coming years, and it sets a (very) crude upper limit on what job growth could be in a steady state.  But one can see even from this figure that it will not be possible to sustain forever monthly net new job growth of even 200,000.  There will not be that many new adults available each month.

But 100% of the adult population are not in the labor force.  As noted, some will be retired, some students, and so on.  The labor force participation rate (LFPR) is the ratio of those who choose to be in the labor force (employed or looking for employment) to the adult population.  In the November CPS figures, that LFPR was 61.8%.  If one assumes that it will remain at that rate, then the monthly growth in the labor force will not be 180,000 (the growth in the adult population) but 61.8% of this, or 111,000.  And if one assumes that unemployment will be something steady, at say 4% at full employment, then potential employment growth would be even less, at 107,000.

The implication is that if the labor force participation rate remains where it is now, one should not be surprised to see monthly figures on job growth of no more than roughly 100,000.  This follows by simple arithmetic.  It could be higher for some period (but not forever) if the labor force participation rate rises from the current 61.8%.  This is possible, and perhaps even likely in the very near term, but probably not for long.  The LFPR in fact rose in the November BLS report to 61.8% from 61.6% in the prior month.  It normally changes only slowly over time.  The disruption that followed from Covid-19 led to relatively wide swings at first, with the LFPR falling from 63.4% in January 2020 to 60.2% in April 2020 with the lockdowns.  But by June 2020 it was back to 61.4% and since has fluctuated in a relatively narrow range before rising the 61.8% of November 2021.

What no one knows is what will happen to the LFPR now.  It might rise a bit more, but the long term trend has been downward.  It peaked in the year 2000, with a steady increase up until then following from a rising participation rate of women in the labor force.  But since 2000 the participation rate for women has moved down, paralleling (but about 20% below) the slow downward trend seen for men since the mid-1950s.  (The factors behind this are discussed in some detail in this earlier blog post.)  It is due to this downward trend over the period of 2011 to 2021 that actual labor force growth over this period was just 67,000 per month (as depicted in the chart above) even though adult population growth was 180,000 per month over this same period.

The current 61.8% LFPR is in fact close to what a simple extrapolation of the trend since 2000 suggests it would be in November 2021.  While the LFPR has behaved unusually since 2016 (when it flattened out for several years and indeed then rose a bit until the start of 2020, before collapsing and then partially recovering in the spring of 2020 due to the Covid-19 crisis), it is now back roughly to what one would find by a simple extrapolation of the trend since the year 2000.

There may well be surprises in what now happens to labor force participation.  After the disruptions of the Covid-19 crisis, it may never revert to where it was just before the crisis.  Those who retired early may mostly choose to stay retired.  And many of those in low-paying jobs, particularly in cases of one spouse in a couple with young children, may have discovered during the Covid-19 crisis that one spouse dropping out of the labor force is not all that costly, and in a two-earner household they may be able to manage financially.

There is therefore a substantial degree of uncertainty on what will now happen to the LFPR.  If it goes up, with a substantial number of adults re-entering the labor force, there will be a transition period when the labor force (and hence the number employed) could rise by significantly more than the 107,000 per month that one would see at a constant LFPR.  Monthly changes in employment during this transition period could be substantial.  For example (and again, this is simple arithmetic), if the LFPR were to increase from the current 61.8% by one percentage point to 62.8% (which would put it back to where it was in much of 2016 through 2018), then the number in the labor force would increase by 2.5 million over what would follow from regular population growth.  Possible employment growth would be about the same 2.5 million if unemployment stays where it is now.  Thus there could be a transition period of five months during which employment could potentially grow by 600,000 per month (a fifth of the extra 2.5 million in the labor force under this scenario, on top of about 100,000 per month from natural population growth).  Or the transition period could be shorter or longer depending on the number of new jobs each month.

But the point is that even if the LFPR should rise, the impact would be a transitory one, after which one should expect employment growth each month of no more than 100,000 or so.  And as noted before, the trend over the last 20 years has been that the LFPR has been moving downward, not upward.

D.  Conclusion

The November jobs report was interpreted by many as disappointing, as the estimated number of net new jobs (based on the estimate normally used – and rightly so) was 210,000.  This was seen as low, and the stock market fell.  However, the report was in fact a pretty strong one, and analysts may have recognized this once they started to look at it more closely.  While one never knows with any certainty why the stock market moves as it does (and there will always be other factors as well), the S&P 500, after falling by over 2% at one point on December 3, started to recover partially by the end of the day.  And it then rose strongly on the next two trading days.

There are reasons to believe the estimate of 210,000 net new jobs in November may have been low.  Seasonal adjustment factors mattered more than normally, and other measures of job growth were significantly higher.  But even at 210,000, analysts need to recognize that as the economy returns to more normal conditions, monthly job growth will likely be a good deal less than that.  While monthly job growth during Biden’s presidency from February to November has so far averaged over a half-million per month (588,000 per month to be more precise), this was only possible because the unemployment rate could come down.  But unemployment is now low – it reached 4.2% in November – and cannot go much lower.  If the labor force participation rate stays where it is now, possible employment growth will only be around 107,000 per month.  If the LFPR rises, then this could go up for some transition period, but that transition period is limited in time and when it is over employment growth will then have to revert to something close to 100,000.

What is more likely is that the LFPR will now return to the longer-term trend seen since it reached its peak in the year 2000, and will fall slowly over time.  Monthly employment growth would then be less, at something less than 100,000 per month (where how much less depends on the pace at which the LFPR falls).

Expectations have to be reset.  Other than during a transition period should the labor force participation rate rise above where it is now, monthly net new jobs growth of 100,000 per month or so is likely to be the limit of what one will see.  But that would be a good performance in an economy that remains at full employment.  Only if unemployment shoots up due to some future downturn could one then see – during a recovery from that downturn – something more.

Polling Results Should Indeed be Worrying for the Democrats

The Washington Post and ABC News jointly sponsor a regular poll of American voters on a range of political issues, including whom they would vote for in upcoming Congressional elections.  Their most recent poll, released on November 14, showed that registered voters nationally would favor a Republican over a Democrat in their local congressional district race, by a margin of 10 percentage points.  The pollsters noted that this margin is the largest margin Republicans have enjoyed in all the polls they have conducted asking this question, in a series that goes back decades.

Numerous pundits soon followed with commentary on this, with articles such as “Why the generic ballot is so ominous for Democrats”, and “Democrats face a 2022 superstorm”, and “Democrats Shouldn’t Panic.  They Should Go Into Shock”.  Viewed in a longer-term context, and not simply in terms of the current polling margin in favor of the Republicans, should there indeed be such a concern if you are a Democrat?  As we will discuss in this post, the answer is yes.

The chart at the top of this post shows the responses that have been given to this question in the Washington Post / ABC News polls over the last two decades.  The figures for the past polling results are provided online here.  The polls asking this specific question have been undertaken in the periods leading up to each midterm election, normally starting about one year before the midterm election date and then repeated every few months until the November election.  There is then a gap – normally of three years – before they start the cycle for the next midterm election.

The specific question asked is:  “If the election for the U.S. House of Representatives were being held today, would you vote for (the Democratic candidate) or (the Republican candidate) in your congressional district? Would you lean toward the (Democratic candidate) or toward the (Republican candidate)?”  They also allow for possible responses of some other candidate, or neither, or would not vote, or no opinion, which I have combined into a single “Other / No Opinion” category in the chart.

Some points to note:

a)  The swing in preferences seen in the current midterm cycle is not unusual.  One saw a similar swing in favor of the Republicans in the period leading up to the 2010 election, and in favor of the Democrats in the periods leading up to the 2006 and 2018 elections.

b)  The initial polling results (one year before the midterm elections) were generally a pretty good predictor of the final outcome.  The polling results generally fluctuated within a relatively more narrow range in the multiple polls in the year leading up to the midterm itself.  The year 2014 was an exception, where the early indication was that Republican support had declined, but with it then recovering prior to the election.

c)  The “Other / No Opinion” category generally fluctuated within a range of roughly 5 percentage points, and in an understandable pattern.  Uncertainty on whom (if any) to vote for generally rose in the three years since the prior midterm, and then fell as one got closer to the election date.

d)  The substantial swings (from three years earlier) found in the initial polls in several of the cases (for the 2006, 2010, and 2018 midterms), coupled with the relatively smaller fluctuations then found in the year leading up to the election, were thus highly predictive.  The electoral results were net gains for the Democrats of 31 seats in 2006 and 41 seats in 2018, and a net gain for the Republicans of 64 seats in 2010.

Based on this pattern, Democrats can expect to lose a substantial number of seats in the 2022 midterms.  The closest parallel is to 2010, for several reasons.  Both 2010 and 2022 were (or will be) the first midterms of a newly elected Democratic president (Obama and Biden).  In the period leading up to 2010, Democrats saw their polling result fall (in the initial poll one year before the 2010 midterm, from three years before) by 9 percentage points, while that of the Republicans rose by 7 percentage points.   In the new poll one year before the 2022 midterm, the Democrats have seen the same fall of 9 percentage points in their polling result, while the preference for the Republicans rose again by 7 percentage points.

The Democrats had a net loss of 64 seats in the House of Representatives in 2010.  These early polling results suggest they could expect a similar result in 2022.

There are, of course, a number of provisoes:

a)  As they always say when investing in the stock market:  “Past performance is no guarantee of future returns.”  While there may have been this pattern in several of the recent midterm election cycles, there is no guarantee that pattern will continue.

b)  Furthermore, that pattern is based on a very small number of cases.  There have only been five midterm elections in the last 20 years, and substantial swings in voter preferences in only three of them.  And only one case (2010) with a Democratic president in his first term in office.  it is dangerous to generalize from figures for such a small number of election cycles.  But it would not be helpful to go back further in time as the political environment has changed, with more of a left-right polarization now than one had before.

c)  There will also be an impact from the substantial gerrymandering of congressional district lines now being redrawn to reflect the new census numbers.  The Supreme Court in 2019 ruled that it would not intervene in this practice when it considered two cases brought before it (of North Carolina and Maryland).  Each had egregiously gerrymandered district lines, and there were open and public statements in each from the politicians who had drawn those district lines that they had done so for the greatest possible partisan advantage that they could manage.

The Supreme Court nevertheless ruled that gerrymandering was not reviewable by any federal court, on a 5-4 vote where the five in the majority were the five Republican appointees to the court.  As a result, a number of states are now redrawing district lines to maximize partisan advantage.  And the ruling heavily favored Republicans, given their control of a larger number of states where politicians are allowed to draw the district lines that they will then be running in.  Republicans at the state level have full control of redrawing the lines for 184 congressional districts this year, while Democrats have full control in states where lines for just 75 districts will be redrawn.  In part this is because several of the larger Democratically controlled states (including California, Washington, and Colorado) now use independent, nonpartisan, commissions to draw the district lines.

The consequences of this gerrymandering will be on top of what one should expect from shifts in voter preferences.  And the margin in seats in the House that gives Democrats control of the chamber is only five following the 2020 election.  Already by this point, with only a small number of states having completed the redistricting process, a mid-November analysis at the New York Times concluded that Republicans will pick up a net of five congressional seats, and thus gain control of the House, even if the voting numbers in each locale were the same as what they were in 2020.  It would simply be a consequence of the newly drawn lines.

Coupling the gerrymandering with the shift in the preferences for Democrats vs. Republicans found in the polling results, there is every reason to expect Democrats will lose control of the House.  This in itself is not surprising.  Since the presidency of John Quincy Adams in 1826, the party of the incumbent president has lost seats in the House in the midterm after their first term election in all cases other than the sole exceptions of Franklin Roosevelt in 1934 and George W. Bush in 2002.  With the Democrats holding a majority in the House of just five seats, most have expected that Republicans will gain control after the 2022 midterms.

But the polling results, on top of the gerrymandering as well as the historical norm, suggest the Democratic losses are likely to be large.  Furthermore, the losses are most likely to be in the more competitive districts, which are more likely to be currently represented by the more moderate Democrats.  Thus the remaining Democrats in Congress following the 2022 election are likely to be the ones further to the left.  That is, the center of the Democrats is likely then to be shifted to the left, just as the center of the Republicans has shifted in recent years to the right.

Polarization, already large, would grow.

There Have Been Real Consequences From Not Taking Covid Seriously

A.  Introduction

Earlier posts on this blog have documented that vaccination rates against Covid-19 have been systematically lower in accordance with the share of a state’s vote for Trump in the 2020 election, and that mask-wearing to protect the individuals and those around them have also been systematically lower.  The higher the share voting for Trump in a state, the lower the share vaccinated and the lower the share wearing masks.

Those choices have had consequences.  As shown in the charts above, it should not then be surprising that states with a higher share of their vote for Trump have seen, on average, a higher number of cases of Covid-19 (per 10,000 of population) as well as a higher number of deaths.  The relationship is statistically a very strong one.  While many factors affect the likelihood of being infected with Covid-19 and of dying from it (including factors such as urban density, extent of travel, health status of the population, adequacy of the health care system, and more), political identification by itself appears to be a strong and independent factor.

In what is literally a life and death issue, one would have thought that rational self-interest would have dominated.  It has not.  Following a review of the data, this post will discuss some possible reasons why.

B.  The Relationship Between the Incidence of Covid-19 Cases and Deaths and the Share Voting for Trump

The figures at the top of this post plot the relationship between the number of cases of Covid-19 in a state (per 10,000 of population), or the number of deaths (also per 10,000), and the share in the state who voted for Trump in 2020.  The Covid data come from the CDC.  It was downloaded October 26, but since case and death counts from the states may not be fully reported to the CDC for up to a week to ten days, I used October 15 as the end date for the analysis here.  “Cases” are confirmed cases, and “deaths” are deaths as a consequence of Covid-19, both as defined in the CDC guidance for how these should be recorded.

For the start date I used July 1, 2020.  This came at the end of the first wave of Covid-19 cases and deaths.  Cases and deaths in this first wave were excluded for two reasons.  One is that the first wave arrived suddenly in mid-March and with an intensity that surprised many.  The nation was unprepared, with little done to prepare for the disease that was spreading around the world as Trump was claiming it was all under control, that it was “going to disappear”, and that it would soon “go away”.  Also, the CDC had bungled the initial testing (where testing was more readily accessible in parts of Africa than in the US in the key initial months), so the full extent of the developing problem was not clear until it hit.  The response, and the then only possible response, was to quickly institute lockdowns, and this was soon done in all 50 states.  The lockdowns were effective, albeit costly, and by late April the approach had succeeded in starting to bring down the daily number of new cases.  Case numbers continued to fall in May and into June.

But starting in early May, disparate decisions were taken across the different states on how fast to lift the lockdown measures.  Some opened up early and with little guidance on or advocacy for the wearing of masks, while others opened up more cautiously.  But with the opening up, and the refusal by a significant share of the population to wear masks and to follow social distancing recommendations, the daily number of new cases stopped falling and by around mid-June began to rise again.  The daily number of deaths followed a similar pattern but with a lag of about two weeks, and so began to rise around the end of June. Thus July 1 can be taken as a turning point – the end of the first wave and the start of the second.  While differences across the states had already started to develop from early May (when decisions were taken on how rapidly to open up), the consequences of the varying approaches only became clear as the second wave started to build.  On average across the nation, this was around July 1.

The second reason to exclude this first wave is that the quality of the data for that initial period was poor.  The Trump administration was slow in launching and then ramping up testing, with testing limited even well into April to those who showed obvious symptoms or who had been in close contact with someone with a confirmed case.  Thus many cases were missed.  While testing has been far from perfect throughout this pandemic, it was much worse in the earlier months than it was later.  For this reason as well, excluding the estimates from the earlier months will provide a better measure of how successful or not the different states were as they responded to the pandemic in their different ways after the initial lockdowns.

Excluding the first wave leads to the exclusion of 6% of confirmed cases and 18% of deaths from the overall totals as of October 15, 2021.  Most thus remained.  Note also the disparity in these figures.  That the official figures recorded that just 6% of the confirmed cases in the US (as of October 15) were in this initial, first wave, period, while this same period recorded 18% of deaths, strongly suggests that cases were significantly undercounted in that first wave.

The charts then show the incidence of total confirmed cases of Covid-19, or deaths from it, per 10,000 of population, over the period from July 1, 2020, to October 15, 2021, with this plotted against the share of the vote that Trump received in that state in 2020.  The relationship is a strong one:  The higher the share of the state vote for Trump, the higher the incidence of Covid-19 cases and of deaths.  Taking averages, the average number of confirmed cases over this period per 10,000 in the states won by Trump was 1,461 (i.e. 14.6% of their population) vs. 1,113 in the states won by Biden.  That is, there were on average 31% more cases in the states won by Trump.  The number of deaths from Covid-19 came to 21.2 per 10,000 in the states won by Trump vs. 15.3 in the states won by Biden, or 38% more in the states won by Trump.

But averaging across all the states won by Trump or by Biden is not terribly meaningful as there will be a mix of voters in every state.  Furthermore, there were a number of states where the vote was close to 50/50.

It is thus more meaningful to examine the trend across the different states, as a function of the share voting for Trump.  This trend is provided in the regression line shown in each chart, where simple, linear, ordinary least squares regression was used.  The statistical relationship found was very strong, and especially so for the regression for the number of cases of Covid-19.  The R-squared (a measure of how much of the variation in the values is accounted for by the regression line alone) was extremely high for such a cross-state sample as here – at 0.63 for the number of Covid-19 cases and a still high 0.36 for the number of Covid-19 deaths.  (R-squared values can vary between 1.0, in which case the regression line explains 100% of the variation across states, and 0.0, in which case the regression line explains none of the variation.)

The higher correlation (the higher R-squared) observed in the relationship for the number of cases than in the relationship for the number of deaths is what one would expect.  To die from the disease, one must first have caught it.  Hence this will depend on the number of cases in the state.  But deaths from it will then depend on additional factors such as the age structure of the population, general health conditions (obesity rates, for example), as well as the availability and quality of health care services (hospitals, for example).  These factors will vary by state, and hence add additional variation to that found for the number of confirmed cases.

The slope of the regression line is an estimate of how many additional cases of (or deaths from) Covid-19 to expect (per 10,000) for each 1% point higher share of the vote for Trump.  For each additional 1% point in the share of the vote for Trump in a state, there were on average 23.8 more cases (per 10,000 of population) of Covid-19 during the period examined, and on average 0.36 more deaths (per 10,000).  The t-statistics for these slope coefficients were both extremely high, at 9.1 for the number of cases and 5.2 for the number of deaths.  A t-statistic of 2.0 or higher is generally taken to be an indicator that the relationship found is statistically significant (as it implies that in 95% of the cases, the slope is something different from zero – a slope of zero would imply no relationship).  A t-statistic of 3.5 would raise that significance to 99.9%.  The t-statistics here of 9.1 and 5.2 are both far above even that mark.

One can also use the regression lines to address the question of what the impact would have been on Covid-19 cases and deaths if everyone behaved as Biden voters did (or as Trump voters did).  The regression lines look at how the incidence of cases or deaths change based on each additional percentage point in the vote for Trump.  If one extrapolates this to the extreme case of zero votes for Trump (and hence a “pure” Biden vote), one can estimate what cases and deaths would have been if all behaved as Biden voters did.

This is a straight line, i.e. linear, extrapolation of the effects, and the limitations from this assumption will be addressed in a moment.  But using linearity, the effects are easily calculated by simply inserting zeroes for the Trump share of the vote into the regression equations, so that one is left with the constants of +96.94 for the number of cases (per 10,000 of population) and -0.69 for the number of deaths.  That is, there would have been a predicted 97 (per 10,000) cases of Covid-19 over this period in the US rather than the actual figure of 1,261 (per 10,000).  This is 92% lower.  And the number of deaths would have been essentially zero (and indeed would have reached zero with still some share voting for Trump – based on the regression equation coefficients it would have been at the 2% point share for the Trump vote).

Are these results plausible?  Would cases and deaths have fallen by so much if all of the population had behaved (in terms of wearing masks, social distancing, getting vaccinated once vaccines became available, and other such behaviors) as the Biden voters did?  The answer is yes.  Indeed, the linear extrapolation is conservative, as infectious diseases such as Covid-19 spread exponentially.

If in some state each infected person infects, on average, two further people, the number infected will double in each time period for the disease.  This is exponential growth, with a reproduction rate of two in this example – a doubling in each period.  For Covid-19, the time period from when a person is infected to when that person may, on average, spread it to another, is a week and a half.  A person becomes infectious (can spread it to others) about one week after they became infected with the disease, and then can infect others for about a week (with the average then at the half-way point of that week).  Thus 100 cases of active infections in some region would double to 200 in that time period of a week and a half, then to 400 in the next time period, and so on.  If, in contrast, responsible behavior (such as vaccinations and mask-wearing) reduces the reproduction rate to one-half rather than two, then 100 cases will lead to 50 in the next time period, to 25 in the next, and so on down to zero.

In any given state there is a mix of Biden voters and Trump voters.  While there are many factors that matter, if these two identities reflect, on average, differing shares of people that do or do not choose to be vaccinated, wear masks, and so on, then the average reproduction rate will vary depending on the relative shares of such voters.  That average reproduction rate will be lower in states with a higher share of Biden voters, and for a sufficiently high share of Biden voters (a sufficiently low share of Trump voters), there will be an exponential decline in new infections from Covid-19.  The linear extrapolation based on the regression equations would thus be a conservative estimate of the number of cases to expect when most of the population behaves as the Biden voters have.

There are, of course, many factors that enter into whether a person is infected by someone with Covid-19, and whether they then die from the infection they got from someone.  But the charts and the regression results suggest that the share of the population in a state voting for Biden or for Trump is, by itself, strongly correlated with how likely that was.  Why?

C.  Personal Behavior and Political Identity

The fact of, and then the consequences from, this political divide for infection by Covid should not be a surprise to anyone.  As noted before, Trump voters are far less likely to be vaccinated or to wear masks to protect themselves and others from this highly infectious, and deadly, disease.  This then translates into higher infection rates, and the higher infection rates then to higher deaths.

One sees this unwillingness to be vaccinated also in surveys.  The most recent of the regular surveys by the Kaiser Family Foundation (published on October 28) found that 90% of Democrats had received at least the first dose of the Covid vaccine, while only 61% of Republicans had.  Furthermore, 31% of Republicans declared they would “definitely not” be vaccinated, while just 2% of Democrats held that view.  Gallup surveys have found similar results, with a survey from mid-September finding that 92% of Democrats had received at least the first dose of the Covid vaccine, but that only 56% of Republicans had.  And 40% of Republicans in that survey said they are not planning on being vaccinated ever, while only 3% of Democrats said that.

Not surprisingly, one then sees this reflected in state politics.  Republican governors (such as Abbott of Texas and DeSantis of Florida) have gone so far as to issue executive orders to block private companies from protecting their staff and their customers from this disease, and even to prohibit local school boards from taking measures to protect schoolchildren.

The direct result is that the virus that causes Covid-19 has continued to spread.  An infectious disease such as Covid-19 will only persist as long as it is being spread on to others.  It cannot survive on its own.  The issue, then, is not just that someone refusing to wear a mask or to be vaccinated is highly likely to catch the disease, but that that person is likely to spread it to others.  While Republican governors such as Abbott and DeSantis have said this is a matter of “personal freedom”, it is not that at all.  No one is free to do harm to others.  It is the same reason why there are laws against drunk driving.  Drunk drivers are more likely to cause crashes (not all of the time, but often), and those crashes will harm others, up to and including killing others.  Spreading Covid-19 is similar, up to and including that those who become infected may die from it.

For whatever causal reason, the facts themselves are clear.  But why has a significant share of the population chosen to behave this way?  This is now more speculative, and goes into an area that I openly acknowledge is not my area of expertise.  With that proviso, some speculation.

It is clear that political identity has played a central role, where Trump from the start treated the then developing pandemic as an issue where you were either with him – and his assertion that he had it all under control – or against him.  This started with Trump’s assertion in an interview on January 22, 2020 (from Davos, Switzerland) that he had no worries, that “we have it totally under control”, and that “It’s going to be just fine”.  This claim continued through February (as cases were growing in the US), where on February 27 he said “It’s going to disappear.  One day it’s like a miracle.  It will disappear.”  And in campaign rallies in February, he claimed to his cheering supporters that he had been doing a superb job in stopping the virus and that any charge to the contrary was simply a “hoax” coming from the Democrats.

Thus, from the start, Trump made the issue a political one.  If you were a true supporter of Trump you could not treat the disease as something of concern – Trump had taken care of it.  Any assertion that the developing pandemic was in fact serious, and needed to be addressed, was a “hoax” perpetrated by the Democrats.

Trump then continued to assert all would soon be well, saying on March 10 that “it will go away”, on April 29 that “This is going away.  It’s gonna go.  It’s gonna leave.  It’s gonna be gone.”, on May 11 that “we have prevailed”, on June 17 that “It’s fading away.”, and on July 19 that “It’s going to disappear”.  But more than 600,000 Americans have died since July 19, 2020, not far short of the 651,000 Americans who have died in battle in all of America’s wars since 1775.  From the start of the pandemic, more than 750,000 Americans have now died.

Trump’s politicization of Covid-19 was then amplified when, at the April 3 press conference in which he announced the CDC recommendation that everyone should wear face masks when going out, he immediately then added that he would not himself wear a face mask.  Face masks are highly effective in hindering the spread from person to person of the virus that causes Covid-19, and until vaccines became available, were the best way to hinder that spread.  But wearing a face mask is also highly visible.  For those who saw themselves as supporters of Trump, and believed what he said (that the virus was going away, that he had it under control, and that any concerns over this were merely a hoax promoted by the Democrats), then it was not surprising that many would follow Trump’s highly public example and not wear a mask either.  Some even went so far as to shoot, and kill, store personnel when told they should wear a face mask inside some store.

It is not surprising that such views would then carry over to vaccination.  Having rationalized not wearing a mask, it is easy to rationalize a refusal to be vaccinated.  And rationalizations could easily be found just by watching Fox News.  In the six months from April through September this year, for example, Fox News chose to air a claim undermining vaccination on all but two of those more than 180 broadcast days.  Many were also exposed to claims that can only be described as truly bizarre, such as that the vaccination will be secretly inserting a microchip into your body for the government to track you, with Bill Gates behind it all; or that it will make you magnetic with this managed through 5G telecom towers; or that it will re-write your body’s DNA; and more.

One can therefore easily come up with rationalizations not to be vaccinated, of varying degrees of plausibility, if you are predisposed against it.  But many of those providing such rationalizations must have realized that their rationalizations often did not make much sense.  Rather, their decisions appear to have been driven more by a visceral or emotional reaction (vaccinations just “feel” wrong) than as an outcome of a rational process.  That is, the decision not to be vaccinated was made first, based on emotions or feelings, with the rationalizations then arrived at later to justify a decision that had already been made.  (Such a process is in accord with the “social intuitionist” model of Jonathan Haidt, where decisions are made first, in a visceral reaction based on emotion, while rationalizations then come later to justify that decision.)

In the case of Covid-19, those decisions on vaccination (and earlier on wearing masks) were made in accordance with political identity – a perceived loyalty to Trump – rather than in recognition of the very real risks that would follow if one contracted Covid-19.  Wearing a mask or accepting a vaccination would simply be “wrong” and disloyal.

I have found it astonishing how strong this emotional reaction has apparently become.  Covid-19 is new (it did not even exist just two years ago), it is deadly (where on average about 1.5% of those infected have died – with a much higher fatality rate than this average for those who are older or who have other health issues), and may have serious long-term ill effects even for those who do not die from it.  Yet this visceral reaction appears to have been so powerful that many supporters of Trump still refuse to be vaccinated, despite the risk of genuine life and death consequences.

I should hasten to add that not all voters for Trump have refused to be vaccinated.  Indeed, according to the surveys, about 60% (a majority) have as of October.  There are also highly vocal partisans on the left who have refused to be vaccinated.  Their reasons are likely very different from that of the typical Trump voter, but the underlying cause appears still to be intuitive – the feeling that such vaccinations are simply “wrong”.  But the issue is that the relative shares of the two groups have been very different:  A far higher share of those who voted for Trump have refused vaccination than is the case for those who voted for Biden.  The consequences are as shown in the charts at the top of this post.

As noted before, the cause for this relationship cannot be known with certainty, and what I have presented here should be viewed as speculative on my part.  There may well be other explanations.  For example, a related but somewhat different explanation would be that a common third factor explains both the tendency of some to vote for Trump and also to be resistant to vaccinations.  Those in this group may put faith in conspiracy theories (including, but not limited to, terrible consequences from being vaccinated), distrust authority, proudly but stubbornly insist on doing the opposite of whatever is recommended, and for such reasons not only refuse to be vaccinated but also vote for Trump.

Whatever the explanation, the results have been tragic.  This has also been a lesson in how strongly some will keep to a held position, even as they have seen prominent figures, and sometimes friends or even family members, come down with this disease.  When an issue becomes one of identity, it appears that even with such tragic consequences there will be many who steadfastly refuse to change.