Behavioral Correlates of the COVID-19 Burden in the States
The prestige media has paid near-exclusive attention to case counts and death rates. The former is measured with a great deal of noise. And even if there weren't a substantial measurement error — how many go uncounted? — cases are a poor measure of both health burdens for the population and stress on the hospital system. The measurement errors themselves propagate to case fatality rates, which are thus too dicey to hang much on.
I have myself used death rates (not case fatality rates) as when I showed that the international cross-section of COVID-19 death rates is upside-down: COVID-19 death rates are positively associated with per capita income, mean adult height, life expectancy, state capacity, and practically any measure of living standards and institutional competence. I have been criticized for ignoring non-fatality burdens. I have also been criticized for not adjusting for age differentials — since age is such a robust and significant risk factor both for severe covid and fatality, this is an important critique.
Anyway, back to the prestige media. There was an atrocious New York Times article whose basic storyline was that the very bad, no good, stupid red [racist] states are paying for their stupidity. The entire case relied on cherry picking the categories of tight, intermediate, and lax controls (see the first graph below). The obvious alternate hypothesis, that stringency was a function of burdens earlier in the year (as is clear from the graph if you pay careful attention) and that rising case counts in lax states reflects the geographic dispersion of the pandemic away from the Metropolitan outward-facing states, was not even considered.
The alternate hypothesis is, in fact, the right one. This can be seen by "blindly" (or rather, honestly) stratifying the states into three quantiles by stringency of repressive measures. See the next graph. Here we use a direct measure of hospital stress, bed occupancy (staffed bed occupancy and ICU occupancy read similarly). This is the appropriate measure if one is interested in the question of whether states and cities need to be shut down, and more generally, whether more repressive measures are called for. For the only reason to repress social intercourse is hospital stress — if our hospitals are close to getting overwhelmed, we should repress social intercourse; otherwise we should not. For the protection of society against the pandemic must be balanced by the mandate to reproduce an open society.
Stringency does not predict burdens because people's behavior mediates between government measures and the rapidity of the propagation of the pandemic. In what follows, we'll pay exclusive attention to these mediating variables and their impact on burdens. We have good reason to pay attention to the mediating behavioral variables. People have adapted their lives in quite dramatic ways in response to the pandemic. They are wearing masks in public, avoiding being out and especially, crowds and events; maintaining social distance; and working from home (something only non-manual laborers can do). Do these behavioral adaptations help? Is the New York Times right? Are lax states paying for their laxness with higher covid burdens? I have finally found data that can help us test this set of hypotheses.
We have panel data on adaptive behavior from CovidStates by state and season (spring, summer, fall). This allows us to test which, if any, behaviors predict covid burdens. For our response we choose Covid-19 hospitalization rates — this is a more inclusive measure than deaths (h/t Rebecca). We obtain state-level hospitalization count data from the Covid Tracking Project, and state-level population estimates together with percent of population above seventy from Social Explorer. We obtain rates of hospitalization per hundred thousand. We shall also adjust for age using the percentage of the state's population that is over 70 (h/t Jamie).
I'll share graphs for all features later in this post. And I have put the code and the data on my Github. But it is worth our while to examine the main features. Take mask wearing. People have been wearing masks more and more over the course of the year. All the states have moved up on this score.
However, mask-wearing does not predict lower covid burdens. To the contrary, mask-wearing rates are positively associated with covid burdens. It is obvious that the causal arrow goes from burdens to mask-wearing rates rather than the other way around — as the New York Times egregiously alleged. That is, mask-wearing is more prevalent in harder-hit states relative to the others. This is true within season and overall, as the next graph shows.
This upside-down relationship holds for many of the mediating behavioral variables. The social distancing index, an aggregate measure of protective measures, has fallen across the board since the spring shock.
But there is no relationship between the social distancing index and covid burdens — both within seasons and overall. If anything the association is positive. We estimate four sets of regressions. First, we use OLS with and without controlling for age and seasonal fixed effects. Second, we estimate quantile regressions with and without adjusting for age and seasonal fixed effects. The social distancing index is positively associated with hospitalizations per hundred thousand, once we control for age differentials and seasonal effects.
More generally, Table 1 and the next two graphs reports our main results. We report z-scores for all four of our specifications together with the diachronic pattern (for which we provide box plots later). Interstate differences in people reporting that they attended events, visited friends and returned to the work place are negatively associated with covid burdens. This is upside-down: if the causal arrow ran from behavior to burdens, then the associations should be positive. But that is not the case. The correlations suggest that the chain of causation runs the other way: states where burdens are lower report higher rates of attendance, visitations and return to the workplace. Once we entertain the idea that covid burdens are largely random, if geographically correlated, this makes sense: people are being more careful where the pandemic is more virulent. This interpretation strengthens when we examine the positive associations. People are avoiding contact, avoiding crowds, disinfecting surfaces, wearing a mask and washing their hands at higher rates in states where the burden is higher.
The next two graphs report the slope coefficients and standard errors for our estimates. The first is OLS, of course. We report both simple linear estimates and gradient estimates controlling for seasonal and age adjustments.
And the following one reports the slope coefficients and standard errors for our quantile regression estimates. These two graphs contain the same information as Table 1. They are provided as a visual aid to the reader.
The bottom line is that we have here a clear case of reverse causation: 'y->X'. The New York Times and other professional class scolds have it the wrong way around. It is not lax behavior that causes higher burdens. Rather, higher burdens translate into people being more careful in all sorts of ways.
We collect the diachronic pattern of behavioral patterns by state in the following gallery.
Next, we report the scatter plots in another gallery. You can find all these graphs on my Github separately.
The moral of the story is that the professional class scold, particularly one writing for the paper of record, needs to hold her tongue. There is no evidence to back up the notion that stupider states are paying for their stupidity in the third wave.
Still, the third wave appears to be much larger than the first one in the spring. Of course, right? Yes and no. It is true that the toll taken by the third one is more serious in terms of deaths and hospitalizations in the aggregate. However, as an argument for lockdown, and more generally, the repression of social intercourse, it is mistaken. The reason is very simple. The first wave was geographically very concentrated. This meant that hospitals in some places, like New York City, were indeed overwhelmed. The failure to tame the first wave, whatever the reasons for it, and we should be skeptical of evidence-free claims, did cost unnecessary lives at the margin. This is not true in the third wave so far because the wave is much more dispersed, meaning that the likelihood of hospitals getting overwhelmed in lots of places is proportionately lower. Of course, things may come to a head in some cities — and they should indeed consider coercive measures. (Carl Schmitt got some things right about the state of emergency, h/t Lee). Our hearts go out to the overworked hospital staff. But a glance at hospital stress (whether measured by bed occupancy, staffed bed occupancy or ICU occupancy, h/t Stefan) makes it clear that the probability of American hospitals getting overwhelmed in this cycle is lower than the first time around. The prestige media needs to check its hyperventilation.
Postscript. What I tried to do in the post above is to debunk the egregious story being peddled by the New York Times that stupid red states are paying for their stupidity in the third wave. I showed how, in the cross-section of US states, the associations between protective behaviors and hospitalization rates is inconsistent with the moralized story one finds in the prestige media. Rather, the pattern of correlations is consistent with the idea that people are more careful where the burdens are higher.
A commenter on the original post, writing under the pseudonym NearPerfectTaco, flagged a paper published in the Journal of Econometrics two months ago. Titled Causal impact of masks, policies, behavior on early covid-19 pandemic in the U.S., the paper tries to identify the causal diagram between policy, behavior and outcomes. It basically builds on the insight documented in my original post — any causal effect of protective behavior such as mask wearing on outcomes like hospitalization or death rates is badly confounded by the fact that, as I established, people respond to higher burdens by becoming more careful. In order to disentangle these confounding effects, the authors propose a structural model as follows.
Nature moves first and generates information — case counts and death rates. Policymakers observe the information and choose policies — mask mandates, business and school closures, etc — with a lag. Then, people respond to both information and policy by modifying their behavior, also with a lag. Finally, outcomes are a function of information, policy and behavior, all lagged. This is a clean way to isolate the effect of policy on outcomes. Unfortunately, the authors do not consider hospitalization rates as an outcome variable, choosing instead to look at case counts and death rates. Still, at least the latter is a fair proxy of Covid burdens. Since they are interested in case counts as an outcome variable, they also control for testing rates. Finally, they have mobility data from Google which allows them to identify the effect of policy on mobility behavior.
They find that both policies and behaviors have large total effects on case counts and deaths. They also find that policy strongly affects behavior. This may be due to the direct effect of shutdowns, as well as the indirect effect of draconian measures reinforcing the perception that things are bad and that it is better to be careful. In their baseline model, reported in Table 4, the direct effect of policy is b = -0.346, P < 0.05. That is, policy measures have directly reduced death growth by 34.6 percent. The direct effect of behavior is estimated at b = -0.837, P < 0.01. So, in terms of direct effect, the measures taken independently by people were more than twice as important in tempering the growth of fatalities. Finally, they estimate that the total effect of policy measures, reported in Table 5, both direct and indirect by changing people's behavior, to be b = -0.885, P < 0.01. Based on this estimate, they reckon that policy measures dramatically tempered COVID-19 fatalities in the United States. Their counterfactual estimates must be taken with a pinch of salt since they rely on the stability of the identified structural equation model. The main issue is the assumption that behavioral change predicted by policy innovations would not have obtained in the absence of said innovations, even if burdens were considerably higher. That's not very plausible. So the exact counterfactual estimates are likely to be overestimates.
Thanks to NearPerfectTaco for flagging the paper. It neatly explains the upside-down associations we documented.
Having digested these results, is it fair to say that we should impose even more draconian measures since they are effective in controlling the spread of Covid and tempering fatalities? Not quite. The social loss function is not reducible to case counts and death rates. All the policies and behavioral changes have real costs — social isolation and attendant mental health burdens, children falling back in their education (especially those from poor families), small businesses closing down (nearly half are expected to close for good by the time this is over), poor economic performance, fiscal burdens and other economic dislocations, indolence and the enormous opportunity costs of sitting idle. All of these, and other costs that I haven't mentioned, must be traded off against the protective gains of more repressive policies and more fearful behaviors. The decision is not up to the economists, health scientists, public health experts and other technocrats. It are up to our elected officials advised by these technocrats. And, ultimately, it are up to us as a civic society.