No significant long-term change in smartphone-tracking indicators can be detected since the Los Angeles Mayor’s order of December 2 or the California public health officer’s regional orders of December 3 and 6.
Figure 1 below shows the average time spent at home by smartphones located within the City of Los Angeles, within the remainder of Los Angeles County, and within Orange County before and after the issuance of city- and statewide stay-at-home orders at the beginning of December.
Mayor Eric Garcetti’s December 2 order applied only to the City of Los Angeles, which is contained within Los Angeles County. The state public health officer’s order of December 3, followed by a supplemental order on December 6, applied to the entire Southern California region, including all of Orange and Los Angeles counties. (Municipalities within Los Angeles County but outside the City of Los Angeles include the cities of Beverly Hills, West Hollywood, Pasadena, Long Beach and Santa Monica, to name a few.)
Figure 2 below shows the corresponding trends in the percentage of devices staying completely at home.
Let’s look first at the trends outside the City of Los Angeles. Comparing Sunday November 29 with the following Sunday December 6 in Figure 1, we see that the December 2 and 3 orders may have been associated with about a half-hour increase in average stay-at-home time. But this short-term effect appears to have been dissipated during the subsequent two weeks. What’s more, Figure 2 suggests that if there was indeed a decrease in social mobility in Orange County or the rest of Los Angeles County, it started before the orders were issued.
Examining the blue trend lines for the City of Los Angeles in both Figures 1 and 2, we find it even harder to discern an impact. Again, if there was indeed a small short-term effect after the orders were issued, it appears to be gone two weeks later.
These two graphs do not show impressive drops in social mobility after the city- and statewide orders. There may have been some minor short-term reductions in out-of-home movement, but they have not been persistent.
Why So Little Impact?
The December 2-6 orders did more than require people to stay at home. Mayor Garcetti’s order, for example, prohibited “all public and private gatherings of any number of people from more than one household,” except for outdoor faith-based services and outdoor protests while wearing a mask. Acting State Public Health Officer Eric Pan’s order, for example, provided that “all retailers may operate indoors at no more than 20% capacity.” Perhaps these accompanying provisions have had a favorable effect in reducing coronavirus transmission.
Still, Garcetti’s order was entitled Targeted Safer at Home Order. Right off the bat, the order reads, “Subject only to the exceptions outlined in this Order, all persons living within the City of Los Angeles are hereby ordered to remain in their homes.” Pan’s order was entitled Regional Stay At Home Order. Paragraph #2 reads, “All individuals living in the Region shall stay home or at their place of residence except as necessary to conduct activities associated with the operation, maintenance, or usage of critical infrastructure.” Yet the evidence is that residents of Los Angeles and Orange counties did not stay at home any more or less than did before these stay-at-home orders.
The reflex interpretation of these findings is that they are simply one more manifestation of COVID-19 burnout and pandemic fatigue. Our own view is that there is now a serious problem of signal versus noise.
There have been so many orders and revised orders and supplemental orders that it has become nearly impossible to ascertain what restrictions on mobility are actually in effect. If we could have performed a focused survey of public awareness or an analysis of social media content, we’d want to know how many Southern Californians were even aware that the new stay-at-home orders were in effect.
Whatever the interpretation, these findings reinforce a critical conclusion.
Our calculations are derived from the SafeGraph Social Distancing database, which follows the GPS pings of an anonymous panel of smartphones equipped with location-tracking software. Each mobile device is assigned a home or origin based on the census block group (CBG) where it commonly spends the night. All CBG codes in Orange County begin with the 6-character string 06059, while all CBG codes in Los Angeles County begin with the 6-character code 06037. For example, the CBG codes for the three census block groups within census tract 1011.10 in the City of Los Angeles would be: 060371011101, 060371011102, and 060371011103. We can thus use census tracts to distinguish CBGs belonging to the City of Los Angeles from other municipalities within Los Angeles County.
For each calendar day from November 1 through December 22 and each CBG in Los Angeles and Orange counties, we extracted a record showing the total number of devices (device_count), the number of devices staying completely at home (completely_home_device_count), and the mean in-home dwell time (mean_home_dwell_time). For each calendar date and each of the three major geographic subdivisions (City of Los Angeles, rest of Los Angeles County, Orange County), we computed the overall mean time staying at home (Figure 1) and the overall percentage of devices staying completely at home (Figure 2).
Los Angeles County is fast becoming the epicenter of the pandemic in the United States. ⦿ Multi-generational households remain the principal pathway for coronavirus transmission. ⦿ Community health centers will be critical to getting through the hard winter to come.
Go Where the Virus Is.
The map below captures the main message of this article: Go where the virus is.
The rose-shaded areas are the Los Angeles County communities where the per-capita incidence of newly confirmed COVID-19 cases grew the fastest during the three weeks ending December 17. While these communities together contain half of the population of the county, their residents now make up nearly two-thirds of all newly diagnosed infections.
The black circles mark the locations of 353 member clinics of the Community Clinic Association of Los Angeles County (CCALAC). These community health centers are uniquely situated to confront the epidemic surge that is already overwhelming the county’s acute-care hospital capacity.
Actually, these health centers don’t have to move an inch to go where the virus is. They’re already there.
The Winter Surge Has Arrived.
Figure 2 below updates the weekly incidence of newly diagnosed COVID-19 cases in Los Angeles County, running from the week starting March 1 through the week starting December 6, 2020. When we last studied the Los Angeles County epidemic, we could see the emergence of a fourth phase during September. Now, the Phase IV surge is obvious. The incidence of newly confirmed COVID-19 cases per 100,000 during the week of December 6 is nearly four-fold the peak incidence seen during the week of July 12.
Los Angeles Mayor Eric Garcetti’s recent stay-at-home order may indeed retard the growth of new cases in the weeks to come. But it will not change the fundamentals of SARS-CoV-2 transmission. Herd immunity from mass vaccination is likely to be several months off.
There is, to put it graphically, no easy way to extinguish the virus-flames that will rage during the winter months to come.
Multi-Generational Household Transmission
Figure 3 compares two maps of Los Angeles County. Each map is broken down into countywide statistical areas (CSAs), a hybrid geographic classification of independent municipalities such as the City of Beverly Hills, neighborhoods of Los Angeles such as Hollywood, and unincorporated places such as Hacienda Heights.
On the left, the CSAs are color-coded according to the number of newly confirmed COVID-19 cases per 100,000 recorded during November 27 – December 17, 2000. On the right, the same CSAs are coded according to the proportion of households that we’ve identified as at risk for multi-generational transmission. As explained in this detailed report (which, for those who care, remains under peer review), we classified a household as at risk for multi-generational transmission if it had at least four persons, at least one person 18–34 years of age and another person was at least 50 years of age.
The two maps in Figure 3 show striking a concordance. Those communities with the highest prevalence of at-risk households, as depicted on the right, had the highest incidence of recent COVID-19 infection, as shown on the left. While we’ve previously noted the concordance with cumulative incidence through September 19 and through November 26, the comparison here is with the number of newly incident cases during the last three weeks of Phase IV.
In short, transmission within multi-generational households continues to dominate the Los Angeles County pandemic.
For those who don’t see the straight visual comparison of two maps as convincing, the technical section below offers a more rigorous demonstration.
The Data Fit With What We See on the Battlefield.
There isn’t the remotest doubt at this juncture that the numbers of confirmed cases reported through voluntary testing substantially understate the actual numbers of incident SARS-CoV-2 infections. Asymptomatic persons, we now know, are responsible for at least 40-45 percent of cases and play a dominant role in disease transmission. Still, the official public statistics fit with what we’re now seeing, as they say, en el campo de batalla.
The clinical histories, one by one, are remarkably the same. One person – usually an asymptomatic or pre-symptomatic young adult – has unknowingly imported the infection into the household. By the time the first member of the household displays any symptoms, all of them – children, adolescents, parents, aunts, uncles, grandparents – have already been infected.
In one household, a woman in her 40’s, having just come down with body aches, fever, and loss of smell, isolated her daughter and her mother in one bedroom and her son and father in another bedroom. The 23-year-old son, who turned out to be patient zero in this particular family, infected his 83-year-old grandfather roommate, who naturally became the focus of considerable clinical attention. (I have intentionally altered the ages of family members.)
The supreme irony of this and so many other cases is that nearly everyone in the household was taking special care to avoid contagion. The parents would make limited trips to the market, always wearing masks. Many grandparents would never go out at all.
Are the young adults who unintentionally import their infections into their multi-generational families simply being careless or ignorant? Not exactly. A young adult who turned out to be patient zero in his own family always wore his mask at work. When he went jogging with his buddies after work, they kept their distance. After one run, they all went for some cold refreshments.
One of the friends ordered hot tea. He just had allergies, he said.
It’s all in the math. Multi-generational households at risk for viral transmission make up 13.8 percent of all households in Los Angeles County. With an estimated 3.3 million households in the county, we’re talking about 455,000 multi-generation households where an asymptomatic or mild SARS-CoV-2 infection in a younger household member would put older household members at significant risk.
Let’s round off the size of the average multi-generation household to just 5 members. During the last three weeks, well over 1 percent of these households were infected. And let’s say the typical super-spreader event generates 100 cases, at least during the first round of transmission. The rest is arithmetic.
During the past three weeks, propagation of the virus through multi-generational households in Los Angeles County has been the equivalent of 200 so-called super-spreader events.
Community Health Centers Are the Key.
The Google Map below is a navigable version of the screenshot shown at the beginning of this article. It is worth repeating that the hot spots with the highest incidence of new infection are precisely the places where multi-generational families are most prevalent.
What’s more, the members of those multi-generational families at highest risk for transmission are, by and large, already patients of nearby community health centers.
Here Are Some of the Things Health Centers Could Do.
Here is an abbreviated list of the things that these community health centers could do – if only they were immediately given adequate resources.
Community health centers need to be equipped to perform high-volume SARS-CoV-2 testing of their patients, with a focus on bringing in all household members at the same time.
It is widely acknowledged that testing needs to be more proactive. We can’t continue to sit back and wait for symptomatic people to come through the door. Workplaces and educational institutions have engaged in routine testing of asymptomatic people. Community health centers need to be the next locus of routine testing. Health centers know who are their high-risk patients. They are ideally positioned to bring in the entire families of these high-risk patients for routine testing.
Community health centers need to be equipped with telemedicine technology, so that providers can rapidly respond to patients’ inquiries and concerns.
Surveys give us only an incomplete picture of the public’s knowledge of such basic issues as how COVID-19 is spread. The plain fact is that patients have an abundance of unanswered questions that apply to their specific circumstances.
Intelligent, useful answers to these questions require detailed knowledge of the incubation period of the virus, the time profile of viral shedding before and after symptoms appear, the likelihood of false-positive or false-negative tests, and the emergence of serious complications in the second critical week . Meaningful answers require clinical knowledge of whether COVID-19 can cause diarrhea, earache, body rashes, back pain. and anxiety due to a low oxygen level. Healthcare providers at community health centers are ideally positioned to answer these questions.
Healthcare providers at community health centers are uniquely positioned to respond rapidly to patients who are at risk of deterioration.
If any reader has the remotest doubt where people in the United States turn when they feel really sick from COVID-19, here is the answer. They call 911.
And this is by no means an abuse of the public safety system. In a December 16 news release announcing that Los Angeles County now had more than 5,000 COVID-19 hospitalizations, the Department of Public Health stated in plain English: “If you are having difficulty breathing, go to an emergency room or call 911.” A December 22 news release announcing another 1,000 COVID-19-related deaths in the past two weeks stated in plain Spanish: “Si tiene dificultad para respirar, vaya a la sala de emergencias o llame al 911.”
Healthcare providers in community health centers are ideally positioned to address patients’ concerns. Who else is going to tell a COVID-19-stricken patient who can hardly eat anything whether she should still take her blood pressure or her diabetes pills? Who else is going to intelligently advise a patient when chest tightness becomes so severe that it’s time to go to the ER? The Department of Public Health has urged providers: “Do not send patients to emergency departments unless absolutely medically necessary.” Physicians, nurse practitioners, physicians’ assistants, and nurses know their patients. They can respond to the challenge if given the resources to do so.
Community health centers could distribute pulse oximeters to their patients and instruct them on their use.
The pulse oximeter is now widely acknowledged as a critical tool in the outpatient evaluation of a patient acutely ill with COVID-19. Community health centers are ideally situated to distribute these devices to patients most in need. New York City is distributing oximeters to healthcare providers, and Los Angeles County health centers should get them, too. In abundance.
Community health centers are well situated to provide timely outpatient treatment to high-risk COVID-19 patients.
There is growing evidence that monoclonal antibody infusions need to be given as early as possible to high-risk patients during the initial viremic phase of their illness. Recent controlled trials suggest that these medications may be ineffective if we wait until the patient is so ill as to need hospitalization. The same may be true for the antiviral drug remdesivir.
We need to identify high-risk patients at the onset of symptoms and establish outpatient facilities where trained personnel can administer these therapies. When it comes to Los Angeles County, community health centers could serve as the ideal locus for this new model of care. It is not difficult to imagine a health center-based nurse administering an infusion to a 70-year-old diabetic with a history heart and kidney problems within a couple of days of symptoms.
What About Social Distancing Measures?
How does this fit in with the now-enormous literature on the effectiveness of the social distancing measures? No one disputes that a great many of the emergency measures enacted by governments worldwide have substantially reduced the propagation of the virus. But the time has come to realize that these now-conventional measures – stay-at-home orders, closures of bars, restaurants, gyms, hair salons, movie theaters, restrictions on large gatherings – are no longer enough. And what’s worse, their enforcement may represent a misallocation of limited resources.
Los Angeles County authorities need to start thinking beyond their standard public health toolbox. They need to adapt their policies to the unique facts on the ground.
If public authorities in Los Angeles County diverted substantial additional resources to community health centers, we’d still have a real chance at putting out the virus-flames.
Technical Details on Multi-Generational Household Transmission
Figure 4 offers another way visualize the critical role of intra-household transmission in the Los Angeles epidemic. We’ve graphed the number of newly recorded cases per 100,000 during November 17 – December 17 against the corresponding number of new COVID-19 cases recorded during October 17 – November 26, 2000. Each data point represents a CSA. The larger the circle, the higher proportion of at-risk multi-generational households in the community.
The linear relationship in the figure is striking. It tells us that those communities recording the most new cases at the start of the recent surge continue to serve as foci of new COVID-19 cases as the surge continues. And the fact that the data points at the upper right are the fattest, while those at the lower left are the skinniest, confirms that CSAs with lots of multi-generational households continue to drive the epidemic.
Figure 5 below displays an updated version of the graphs that we showed in our October 17 and November 28 posts. The graph relates the recent incidence of COVID-19 infection on the vertical axis to the prevalence of at-risk households across some 300 CSAs in Los Angeles County, as measured on the horizontal axis. The graph tells us that a 10-percentage point increase in the prevalence of at-risk households is associated with a 51-percent increase in COVID-19 diagnoses. This relationship has not let up since the days of Phase II, when the epidemic began to concentrate in specific communities.
The TPR is uninformative in an epidemic dominated by asymptomatic transmission. And it doesn’t give us a clue what to do in Philadelphia during the hard winter ahead.
During the ongoing COVID-19 epidemic, public health practitioners and policymakers have increasingly relied on the test positivity rate (TPR) to decide whether to impose or relax constraints on social mobility and how much to expand testing capacity. Yet there is a genuine question whether the TPR – which basically gauges the number of positive cases as a percent of all persons tested – may be steering us off course.
During an epidemic where testing for infection is non-random and voluntary, the TPR may indeed tell us how well our testing program is identifying more severe, symptomatic cases. But in the current COVID-19 epidemic where asymptomatic persons are responsible for at least 40-45 percent of cases and play a dominant role in disease transmission, the TPR may not tell us how well we’re identifying infectious individuals.
When Testing Is Mandatory and Random
When testing is mandatory and random – for example, when the University of Wisconsin-Madison required all on-campus students to be tested regularly during an outbreak – the TPR may indeed accurately gauge the infection rate. If, say, one percent of all persons randomly tested during the past week come out positive, then we can reasonably estimate the current incidence rate per week to be about one percent. To be sure, there can still be false positives and false negatives, and there can be retesting of the same individuals. And we certainly don’t want to mix tests for active infection with antibody tests. While these concerns are valid, they are not fundamental.
When Testing is Voluntary and Non-Random
When testing is voluntary and non-random, however, it is widely acknowledged that the sickest, most symptomatic individuals will queue up for testing first. This behavioral observation has led to the conclusion that as testing capacity is expanded, the TPR will decline. If only the individuals in the first row in Figure 1 are tested, the TPR will be 40 percent (that is, 4 out of 10). If we expanded out capacity to test both rows, the TPR will be down to 30 percent (that is, 6 out of 20). This straightforward syllogism seems to imply that a low TPR is a favorable sign that our testing program “is casting a wide enough net.” Most authorities, including the World Health Organization (WHO), put the cutoff for a sufficiently low TPR at 5 percent, but some analysts peg the cutoff at 3 percent. Whatever cutoff is designated, the TPR is to be interpreted as “a measure of whether we’re doing enough testing.”
The Catch: A Large (and Growing) Pool of Asymptomatic Infected Persons
But there’s a catch. In the current COVID-19 epidemic, a lower TPR means our testing program is adequately identifying the sick people who are voluntarily queuing up to be tested. A further expansion of voluntary testing capacity may indeed lower the TPR, but it will not necessarily identify asymptomatic infectious individuals because they’re not even getting in line.
In Figure 2, we’ve included another group of 20 such asymptomatic people. Eight of them are positive and they don’t know it. If we really could test everyone, the TPR would be back up to 35 percent (that is, 14 out of 40).
In more technical terms, the critical assumption of a declining marginal yield to expanded testing breaks down when there is a significant pool of asymptomatic infected individuals with an extremely low propensity to get tested voluntarily. And as we’ll see shortly, the continued growth of this pool of asymptomatic infected individuals is the principal challenge underlying the hard winter months ahead.
Some Real Data from Philadelphia
Some real data from the city of Philadelphia in the United States, recently downloaded from its online dashboard, will clarify the point. At issue here is not whether Philadelphia’s current caseload is representative of other cities, but whether the problems in interpreting Philadelphia’s TPR are typical.
Figure 3 shows the weekly incidence of test-positive COVID-19 cases per 100,000 population in the city. To facilitate the presentation, we have partitioned the graph into four phases. (We’ve similarly partitioned the COVID-19 epidemic in Los Angeles County into four phases.)
In Phase I, from the week starting March 9 through the week starting March 30, 2020, highlighted in burgundy red, cases were rising rapidly from 3 to 178 per 100,000, equivalent to an average doubling time of 4.7 days. In Phase II, from the week starting April 6 through the week starting May 25, highlighted in green, the case incidence rate fell as emergency social distancing measures enacted in mid-March took effect.
During Phase III, from the week starting June 1 through the week starting September 14, highlighted in orange, the case incidence rate remained stable at about 50 per 100,000. Since then, with the arrival of Phase IV, highlighted in navy blue, the case incidence rate has topped 400 per 100, rising with an average doubling time of about 18 days. While reporting delays temporarily render the data for the last week in November unreliable, the case incidence has surely continued to rise.
The One-Case-per-Thousand-per-Week Threshold
Figure 3 includes an additional, horizontal dashed line at the level of 100 per 100,000 cases per week, equivalent to 1 case per 1,000 per week. This line crosses the epidemic curve during the week of March 23 (Phase I), the weeks of May 11 and May 18 (Phase II), and the week of October 12 (Phase IV). At each crossing point, the case incidence rate was the same, but the underlying dynamics of the city’s epidemic were different. Crossing the threshold from above (Phase II) is not the same as crossing from below, and crossing the threshold during the early surge of cases (Phase I) is not the same as a re-crossing when previously relaxed social distancing measures have proved insufficient (Phase IV).
Dynamic Pueyo Plot
Figure 4 plots the case incidence rate against the TPR, once again calculated on a weekly basis. The data points have been color-coded to correspond to the four phases identified in Figure 3. With the exception of a few points tightly clustered together in Phase III, the corresponding week is noted beside each point. (For visual clarity, the data point for the week of March 9 has been omitted.) The area of each point (not the diameter) is proportional to the total number of tests performed during each week, ranging from 2,147 tests during the week of March 16 to 60,980 tests during the week of November 16.
We’re calling Figure 4 a dynamic Pueyo plot because Tomas Pueyo appears to be the first analyst to plot case incidence against TPR. Pueyo’s version of the plot, however, was a static, cross-sectional comparison of different countries at a particular point in time, and not a dynamic rendering of the course of the epidemic in a single locality.
The Critical Question
The critical question is: What, precisely, does the additional information on the TPR in Figure 4 add to our knowledge of the dynamics of the disease propagation, the adequacy of testing, or the most appropriate epidemic-control measures in Philadelphia during the hard winter ahead?
Figure 4 tells us that the first time the epidemic curve crossed the 100-cases-per-100,000 threshold during the week of March 23, the TPR was 28 percent, far above the WHO’s acceptable level of 5 percent. The second time, during the weeks of May 11 and May 18, the TPR was 12 percent, still beyond the acceptable range. The implication is that the observed decline in case incidence during Phase II in the weeks after the mid-March declaration of emergency may have been an artifact of inadequate testing. Such a conclusion would defy the now-enormous literature on the effectiveness of the social distancing measures put into effect during Phase II.
The fact that the orange data points in Figure 4 are situated at or near the 5-percent TPR threshold would seem to imply that Philadelphia had adequate testing during Phase III. But if that conclusion has any substantive meaning at all, then it must apply as well to the first four weeks of Phase IV, when the number of cases per week more than doubled. Does this mean that no additional aggressive testing measures were required?
It is now evident that during the week of September 21, Philadelphia was on the verge of a new and more deadly resurgence of infections, yet its TPR of 2.5 percent was telling us everything was just fine.
The raw number of positives indeed obscures the number of negatives or the number tested. And there isn’t the remotest doubt that current counts of positive test results substantially understate the actual incidence of infection. Still, when it comes to assessing the dynamic state of the epidemic in Philadelphia or anywhere else, the number tested may be a more misleading denominator than the standard population count.
One might argue that the Philadelphia data are entirely consistent with the TPR paradigm. The expansion in testing seen in the last two months – as indicated by the enlarging blue data points in Figure 4 – was accompanied by an increase in the TPR. This observation, some analysts may contend, precisely illustrates the value of the TPR in identifying a dangerous outbreak. The problem with this logic is that both the number of positive tests and the number of total tests are endogenous variables. Without a natural experiment, we can’t tell whether new positive cases were pushing total testing, or whether expanded testing was pulling the number of case counts. Another analyst might observe, “The percent positive will be high if the number of positive tests is too high, or if the number of total tests is too low.” But that’s just having it both ways.
The Hard Winter Ahead
The first equation of the classic SIR model of epidemics teaches us that the incidence of new cases is a product of two basic factors: (1) the number of contagious individuals and (2) the rate at which these contagious individuals transmit their infections to other susceptible individuals.
During Phase I of the epidemic in Philadelphia, Los Angeles, New York City, San Antonio,Broward County, Florida, and countless other metropolitan areas throughout the world, the surge in cases was driven by factor #2. At the very start, after all, we had no social distancing measures, so that a relatively small number of contagious persons could propagate their infections widely.
During this Phase IV of the epidemic – which appears to be continuing into the cold, hard winter to come – the new surge in cases is being driven by factor #1. The growing number of asymptomatic individuals has now overwhelmed what social distancing measures we have put in place. Despite marginal improvements in testing capacity of the order of magnitude seen in Philadelphia, we haven’t come anywhere near the level of testing we need.
Without routine, repeated testing of asymptomatic carriers through widespread rapid testing on a far more massive scale than we’ve seen so far, we’ll continue to be “operating in the dark,” no matter what the test positivity rate is.
Comments and Responses
In response to the thoughtful comment below by Prof. Raphael Thomadsen. As an indicator of the adequacy of testing, the TPR appears to have little or no informative value. When it comes to gauging the severity of the epidemic, the TPR might serve as a proxy for the case incidence rate. But that just begs the question: What does the graph of TPR in Figure 4 tell us that we haven’t already learned from the graph of the case incidence rate in Figure 3? As to whether there were different test regimes, take a look at the graph below. There was clearly a different regime during Phase I, when the CDC was still struggling to come up with a reliable PCR test for coronavirus infection. After that, the total number of tests has climbed steadily at about 5 percent per week.
Even during the current surge, communities with the highest proportion of multi-generational households have seen the largest increase in new COVID-19 cases. To break the chain, we need to test entire households, and not just individuals.
Figure 1 below shows the weekly incidence of newly diagnosed COVID-19 cases in Los Angeles County, running from the week starting March 1 through the week starting November 8, 2020. When we last studied the Los Angeles County epidemic, we saw three distinct phases. Now, we can clearly see four.
During Phase I, which ran through the week of April 4–10, the epidemic spread radially from initial foci of infection located in relatively affluent communities such as the Brentwood and Beverly Crest neighborhoods of Los Angeles and the City of West Hollywood. During Phase II, which ran through the week of July 5-11, COVID-19 incidence rose at slower rate, as coronavirus infections became increasingly concentrated in areas at higher risk of intra-household transmission. During Phase III, which continued until the week of August 30-September 5, COVID-19 incidence gradually declined, while cases continued to accumulate in the same high-risk areas.
Now, we’re in the midst of Phase IV, and cases are again surging. We’ve cut off our graph at the week of November 8-14, as reporting delays prevent us from accurately gauging the full extent of the recent surge. But the incidence of new cases will undoubtedly continue to rise.
Figure 2 compares two maps of Los Angeles County. Each map is broken down into countywide statistical areas (CSAs), a hybrid geographic classification of independent municipalities such as the City of Beverly Hills, neighborhoods of Los Angeles such as Hollywood, and unincorporated places such as Hacienda Heights.
On the left, the CSAs are color-coded according to the age-adjusted cumulative incidence of COVID-19 per 100,000 population as of November 26, 2020, which we derived from the surveillance dashboard of the Los Angeles County Department of Public Health. On the right, the same CSAs are coded according to the proportion of households that we’ve identified as at risk for multi-generational transmission, which we derived from the 2018 public use microsample of the U.S. Census Bureau’s American Community Survey. As explained in this detailed report, we classified a household as at risk for multi-generational transmission if it had at least four persons, at least one person 18–34 years of age and another person was at least 50 years of age.
The two maps in Figure 2 show just as striking a concordance as when we last performed the comparison. Those communities with the highest prevalence of at-risk households had the highest cumulative incidence of COVID-19 infection.
Within-Household Transmission Has Sustained the Los Angeles County Epidemic
Figure 3 duplicates a pair of graphs that we displayed in our last look at the Los Angeles County epidemic. Both relate the cumulative incidence of COVID-19 infection on the vertical axis to the prevalence of at-risk households across some 300 CSAs in Los Angeles County, as measured on the horizontal axis.
The graph on the left covers cases of COVID-19 diagnosed during Phases I and II, from March 1 through July 11, 2020, when weekly incidence rates were continuing to rise. The graph on the right, by contrast, covers cases diagnosed during Phase III, from July 12 through October 16, when weekly incidence rates turned around and gradually began to fall. Both graphs show a significant positive relationship between the prevalence of at-risk, multi-generational households and the incidence of newly diagnosed coronarvirus infections.
Figure 4 repeats the same analysis, covering the cumulative incidence of new cases during Phase IV, from October 17 though November 26. Once again, there is a significant positive relationship across communities between the prevalence of multi-generational households and the incidence of new infections.
The only difference between the plots in Figure 3, which cover Phases I through III of the epidemic, and Figure 4, which covers Phase IV to date, is that the slope of the Phase IV plot is flatter. It remains to be seen whether the slope will stay flatter as backlogged case reports finally make it to the LA County dashboard. If so, it could be an indicator that, while multi-generational household transmission is still a critical vehicle for sustaining the epidemic, the virus has begun to spread outside its established areas of concentration. We need to follow this relationship closely in the coming weeks.
The Negative Feedback Loop
A number of factors, operating in combination, may have triggered the Phase IV surge in Los Angeles County. While it’s nowhere near as cold in Southern California as other parts of the country, nevertheless people are beginning to spend more time indoors. Rising infection rates elsewhere may have led to more importations from outside the county. And then, as we have repeatedly seen in Los Angeles and elsewhere, there is the obvious negative feedback loop. When the epidemic becomes more severe, public officials impose restrictions on social mobility and people take more protective measures. And when the epidemic starts to dissipate, public officials relax restrictions and people drop their guard. That’s why we see the cycles underlying the epidemic curve in Figure 1.
Whatever the underlying trigger for Phase IV, the high prevalence of multi-generational households in many communities in the Los Angeles area operates as a multiplier. As we noted in our last look, when a younger person, having contracted COVID-19 outside the household, brings his or her infection back home, the impact is magnified by the presence of cohabitants of multiple generations. During our own clinical work, we have recently noted a significant resurgence in the number of infected patients seeking care and advice. Every one of these patients, without exception, has been part of a multi-generational household.
Test Households, Not Just Individuals
A symptomatic patient calls seeking medical advice. She started to feel ill yesterday, but today she has body aches, spiking fevers, and can’t smell her food. She wants to know whether she should get tested. Except as a formality, testing this symptomatic patient is completely uninformative. If the test comes out positive, well, we already knew from the patient’s clinical presentation that she was infected. (A sudden loss of smell in the absence of a completely blocked nose is, by itself, a highly specific test for COVID-19.) And if the test comes out negative, then it was quite likely a false negative and needs repeating.
So, what do we advise the patient? Absolutely everyone in the household has to be tested as soon as possible. Everyone from the toddlers up to the grandparents. The patient may be the first to seek care, but with the high rate of asymptomatic transmission, we need to proactively find out who introduced the infection into the household, who remains susceptible, and who is at risk for the development of severe symptoms in the coming days.
To better understand how the coronavirus can spread rapidly, we need to think more broadly about clusters or networks of places, and not just one place at a time. The idea is that infected individuals can move easily and rapidly across multiple places within the cluster or network. The component places are in turn linked together by close geographic proximity, or by an efficienttransportationnetwork.
Last spring, to take a salient example, South Korean authorities reported an outbreak of 34 cases after a 29-year-old patient visited five clubs and bars in the Itaewon district of Soeul from the night of May 1 to the early hours of the following morning. Case tracking eventually identified a total of 246 primary and secondary infections.
This super-spreader model underlies our analysis of the possible role of a cluster of off-campus bars in a recent outbreak at the University of Wisconsin-Madison, where nearly three thousand students tested positive for SARS-CoV-2 during September 2020. For a detailed exposition of our data sources, analytical methods and findings, see our National Bureau of Economic Research Working Paper 28132.
Green, Red, Purple and Yellow
Figure 1 shows a screenshot of a section of the university campus map, focusing on the eastern end of the campus. The superimposed solid lines mark the external boundaries between the census tracts, while the dashed lines mark the internal boundaries between the four census block groups within census tract 16.06.
The pair of green buildings within census block group 16.06-4 represent the two on-campus residence halls, Sellery and Witte, which were subject to a lockdown when approximately 20 percent of their residents became infected.
The pair of red buildings further to the south within census block group 16.06-3 represent two other residence halls, Ogg and Smith, that were not overrun with infections and not subject to quarantine.
The solid purple circles mark the locations of a cluster of 20 nearby off-campus bars, located mostly in census tracts 16.03 and 16.04.
The yellow circles show a comparison group of 68 coffee houses, inexpensive and medium-priced restaurants located in the same area. These venues are closer to Sellery and Witte than to Ogg and Smith. The most remote bar was only a 13-minute walk from Sellery.
Case Control Study
These four graphic elements – the green pair of residence halls , the red pair of residence halls , the purple cluster of bars , and the yellow group of restaurants – form the basic components of a case control study.
In the typical case control study, researchers want to find out whether exposure to a particular toxin is associated with the development of a particular disease. To that end, the researchers identify two categories of individuals. Those who have come down with a disease are the cases , and those who did not get the disease are the controls. The researchers then determine the odds that someone from the cases was exposed to the toxin. Similarly, they determine the odds that someone from the controls was exposed to the toxin. The ratio of these two quantities is called the odds ratio, which is routinely computed in a research report of the study.
For example, in the classic case control study of smoking and lung cancer reported by Richard Doll and Bradford Hill in 1950, the cases were hospitalized patients with lung cancer and the controls were patients without lung cancer. The odds that a patient with lung cancer (one of the cases) was a smoker were 2.97 times the odds that a patient without lung cancer (one of the controls) was a smoker.
In our study, the cases were residents of Sellery and Witte , while the controls were residents of Ogg and Smith . While not everyone in Sellery and White tested positive for COVID-19, it’s enough to know that the rate was much higher Sellery and White than in Ogg and Smith, so much higher that Sellery and White had to be put under quarantine.
In the classic Doll-Hill study, hospitalized patients were interviewed about their smoking habits. In our study, we used anonymized smartphone tracking data from the SafeGraph Patterns database to determine how many devices originating in Sellery-White (census block group 16.06-4) and Ogg-Smith (census block group 16.06-3) visited bars (the purple circles ) or restaurants (the yellow circles ). The odds that a Sellery-White resident (one of the cases) visited a bar were 2.95 times the odds that an Ogg-Smith resident did so. By contrast, the corresponding odds ratio for visiting the restaurants (the yellow circles) was only 1.55.
The SafeGraph data tell us only how many devices originating in a particular census block group visited a particular point of interest (in this study, a bar or a restaurant). The data aren’t broken down by individual smartphone, so we don’t know how many devices visited specific combinations of bars or restaurants. Still, we do know that the odds ratio for visiting a bar was almost twice the odds ratio for visiting a restaurant.
Observational Studies versus Experiments
Our case control study is an observational study. It is not an experiment in which subjects are randomly assigned to visit bars or restaurants to see who comes down with COVID-19. Accordingly, it is at least arguable that people who go to bars are less likely to wear masks, maintain social distancing, and take protective measures generally. The same criticism could be applied to a study of people who attended a political rally, a motorcycle rally, or a large wedding reception. Still, our comparison of residence halls – rather than individuals – tends to blunt this criticism.
There are numerous factors that go into a student’s decision to live in one residence hall versus another – whether the rooms are singles or doubles, whether there is more than one bathroom on a floor, whether the floors are mixed coed, whether the rooms have carpeting or air conditioning, and whether the student can cohabit with his or her friends, not to mention the price. It would be a stretch to argue that these factors readily correlate with a lack of protective behavior.
One could, of course, speculate about other possible explanatory factors. Sellery and Witte were regarded by some observers as party dorms, at least during prior semesters. This raises the possibility that some smartphone visits to the 20-bar cluster were to purchase alcoholic beverages to bring back to residence hall parties. But that wouldn’t negate the causal role of the bars in facilitating the parties.
Another possible explanation is that residents of Sellery and Witte partied at off-campus fraternities and sororities. But the smartphone tracking data from the SafeGraph Social Distancing database do not show a large number of visits from census block group 16.06-4 (where Sellery and White was located) to census block groups 16.04-1 and 16.04-4 (where the fraternities and sororities were concentrated).
A better explanation is that, with the preventive measures taken by the university, a substantial portion of the alleged partying was shifted off-campus. While some of the partying could have taken place in off-campus private residences, the smartphone tracking data tell us that a significant proportion was shifted to off-campus bars .
Sellery and Witte had more incoming freshman than other residences. To the extent that Wisconsin’s legal age limit of 21 was strictly observed, it would tend to reduce the visitation rate of these residence halls to local bars. Incoming students are required to enroll in an on-campus meal plan, which might help to explain why off-campus restaurant visitation appeared to be a less important predictor of the higher rate of infections at Sellery and Witte. On the other hand, Smith Residence Hall had its own Starbucks, which would tend to reverse that trend.
If there is any factor that clearly distinguishes Sellery and Witte from Ogg and Smith, it is that one pair of residence halls is nearer and the other pair is farther away. But that begs the question: Nearer to or farther from what? Our findings fit with the conclusion that the residents of Sellery and White suffered significantly more infections than the residents of Ogg and Smith because they were nearer to the 20-bar cluster, and not because they were also nearer to coffee houses, inexpensive and moderately priced restaurants, or nearer to some classrooms.
The Timing is Right
Figure 2 contains two concurrent plots covering each day from August 16 through October 11. The orange line, measured on the left vertical axis, shows the total number of daily visits by all devices, without restriction on origin, to any one of the eleven bars in the cluster within a 10-minute walk of Sellery. Since the SafeGraph Patterns database tracks the GPS pings emitted from a limited panel of devices, only the relative changes in visit counts have significance. Thus, during the week of August 16, there was an average of 176 daily visits among device-holders in the SafeGraph panel to the eleven bars. During August 29-30, the number peaked at 471, a relative increase of 2.7-fold.
The blue line, measured on the right vertical axis in Figure 2, shows the daily number positive COVID-19 cases reported by the Wisconsin Department of Health Services (WDHS) for census tract 16.06. (WDHS reported COVID-19 cases by census tract, but not by census block group.)
Comparison of the orange and blue data series shows a double-peaked surge in the volume of bar attendance, with the first peak occurring on August 29-30 and the second on September 5. Soon thereafter, the double-peaked surge in the volume of bar attendance was followed by a double-peaked surge in positive SARS-CoV-2 cases, with the first peak on September 10 and the second on September 14.
The timing is entirely consistent with a causal relationship between the two trends. The delay between the two curves in Figure 2 represents the latency period between initial infection and the subsequent onset of detectable disease.
In Figure 3, we have zoomed all the way out on our campus map. Now we can see the entire campus, along with all the surrounding census tracts. This time, we’ve annotated the campus map with blue bubbles indicating the number of positive COVID-19 cases in each surrounding census tract reported by WDHS during the two-month interval August 16 – October 16. The cumulative number of cases is proportional to the area (not the diameter) of each bubble. The four largest bubbles are: tract 16.04 (870 cases); tract 16.06 (726 cases); tract 16.03 (488 cases); and tract 11.01 (266 cases). These four census tracts comprised 76.7 percent of the 3,065 cases in campus-area census tracts compiled by WDHS during this 2-month interval.
The data in Figure 3 make clear that the cluster of 20 bars – concentrated in tracts 16.03 and 16.04 – was located directly within the geographic epicenter of the outbreak.
Census Tract Study
Figure 3 suggests another way of analyzing the data. Rather than comparing individual residence halls, we could treat individual census tracts as the units of analysis.
That’s exactly what we’ve done in Figure 4 below. The each lilac-colored point on the graph corresponds to a census tract. The vertical axis measures the number of positive coronavirus tests per 1,000 population in each census tract, while the horizontal axis measures the number of bar visits per 1,000 population by the census tract of origin of the bar visitor. Both axes are on a logarithmic scale so that, for example, the distance between 2 and 10 (that is, a 5-fold increase) is the same as the distance from 10 to 50.
With the exception of one outlier (tract 17.04), the points are aligned along the green fitted line. The slope of the fitted line was 0.87. Economists interpret this slope as an elasticity. For every 10-percent increase in bar visits per capita, the incidence of COVID-19 cases per capita rises an estimated 8.7 percent.
When we constructed the same graph for the per-capita number of visits to the group of 68 restaurants, the data did not fit anywhere near so tightly as in Figure 4. When we performed a multivariate regression, we again found a significant relation to bar visits (with an elasticity of 0.90) and no significant relation to restaurant visits. This finding was entirely consistent with the results of our case control study.
What Have We Learned
In contrast to studies of outbreaks that focus sharply on a single location, our study concentrated on a cluster of places, in this case a cluster of bars right in the geographic epicenter of the outbreak.
While we posited that a cluster or network of places can serve as a super-spreader, we did not have sufficient data in this study to map out the individual connections between places. We had some data on the movements of smartphone holders from one census block group to another, but we would need an even finer geographic grid to ascertain how the connections between bars within a census block group unfold. Still, it would be inappropriate to assume that, in absence of hard data, each establishment in the 20-bar cluster had no more than an independent effect on the risk of coronavirus propagation.
Our findings should not be interpreted broadly to mean that restaurants are entirely free of risk while bars are the sole source of contagion. The narrower interpretation is that a specific, centrally located cluster of bars appeared to be a significantly greater vehicle for propagation of the virus than restaurants in a particular university-based outbreak. Neither do our findings point the finger at all bars. Among the 51 bars throughout the campus area, we focused sharply on a cluster of 20 bars at the geographic epicenter of the outbreak.
Future studies of college and university outbreaks need to concentrate harder on the dynamics of viral transmission, and not simply on how many cases ended up in dormitories, athletic teams, fraternities and sororities. Retrospective case-tracking needs to expand its scope to ask an infected individual not just whether he went to a bar, but also whether he went bar-hopping, whether his roommates also went to bars, and if so, to what bars on what nights.
We need to think more about the ways that multiple places within a cluster may act synergistically to enhance viral propagation. The outbreak in the Itaewon district of Soeul, cited above, points to a model where an index case moves from one place to another within the cluster. More generally, when there is high mobility between places, they may function effectively as one place. If an average of 20 patrons attended each of five interconnected bars on a single night, we would end up with an “event” with 100 attendees. When a student says to his friends, “Let’s go to bar A and if we can’t get in, we’ll just go to bar B,” we have a classic network externality where the mere availability of bar B enhances the demand for, and thus the transmission potential of, bar A.
Even more broadly, we need to think about systemic factors that influence viral propagation, and not simply the characteristics of individuals or the places they go to. The epidemic in Los Angeles County has been sustained in great part by intra-household transmission among multigenerational families. But the larger question is what public policies have enhanced or mitigated these housing conditions. The spread of coronavirus in New York City and other metropolises may have been enhanced by individuals of high mobility. But the larger question is what transportationnetworks carried them from one place to another.
We went back and looked at COVID-19 incidence and bar attendance in Dane and Milwaukee counties.
A lot has happened in the state of Wisconsin since we last reviewed its struggles with the continuing COVID-19 epidemic. (Our last review was subsequently published here.)
On October 6, Wisconsin Department of Health Services Secretary-designee Andrea Palm issued Emergency Order #3, restricting indoor gatherings in stores, restaurants, bars and other public venues to 25-percent capacity. “The State of Wisconsin is in the midst of a deadly, uncontrolled, and exponentially growing spike in cases of COVID-19,” noted the order’s preamble. “Some Wisconsin hospitals are already struggling to keep up with care demands – both because of bed space and staffing shortages – and we have to do what we can to slow down the spread of this disease so our health care workers can keep up,” noted the accompanying FAQ. Five days later, on October 11, the weekly state report of the White House Task Force declared, “Wisconsin has seen a sustained peak of epidemic activity in the last week with an ongoing health emergency.”
The following day, St. Croix County Circuit Judge R. Michael Waterman upheld Gov. Tony Evers’ August 1 emergency order mandating the use of masks in enclosed public spaces. At least for the moment, it looked like the state government was finally recovering from the aftermath of the Wisconsin Supreme Court’s decision back on May 13 to nullify Secretary Palm’s original statewide safer-at-home order of March 24.
But then on October 14, Sawyer County Circuit Judge John Yackel temporarily blocked Secretary Palm’s October 6 order limiting the indoor capacity of stores, bars and restaurants. That same day, the state opened a field hospital in State Fair Park as the census of hospitalized COVID-19 patients crossed the 1,000 threshold. On Monday, October 19, Secretary Palm was to appear in court to defend her order.
COVID-19 Incidence in Dane and Milwaukee Counties
Figure 1 below shows the daily incidence of newly confirmed COVID-19 cases in Wisconsin’s two most populous counties from March 15 through October 16. When we last took a snapshot on July 24, the incidence of new cases was starting to come back down in the wake of a July 7 local order requiring the use of face masks while taking public transportation in the city of Madison and the rest of Dane County, as well the July 13 adoption by the Milwaukee City Council of the Milwaukee Cares Mask Ordinance.
From the first week of August onward, however, new COVID-19 cases resumed their upward climb. We’re still investigating whether the apparent bump in Dane County cases (purple data points) in early September can be wholly attributed to an outbreak at the University of Wisconsin in Madison. Still, the point is clear, the local mask-related orders of July had at best a temporary effect.
The database records the movements of holders of smartphones with location-tracking software. For every day from February 17 through September 30, we computed the number of entries into each of 240 Milwaukee County bars and 230 Dane County bars. To make the two series compatible, we normalized the numbers of entries so that the mean for the period February 17 – March 13 was equal to 100. The figure shows the normalized series from March 1 onward.
When we last looked at the bar-attendance data, the gap in attendance between the two counties from mid-March through the end of May had already disappeared, and the number of visits in both counties was hovering around 60 percent of its pre-epidemic baseline. Since then, bar visitation has risen to around 70 percent of baseline, with attendance on some weekends exceeding 90 percent.
The temptation here is simply to assign all the blame to the bars. A fairer interpretation is that the bar-attendance data are no more than an indicator of a broader pattern of increasing social activity in the face of repeated governmental efforts to promote mask wearing and reduce crowding in public venues. At least in the two most populous counties in Wisconsin, the data suggest that these governmental efforts have had only limited, temporary effectiveness, with their impact repeatedly wearing off in a matter of weeks.
At the peak of the COVID-19 epidemic in Los Angeles County during the first half of July, newly confirmed COVID-19 diagnosis were running at about 210 per 100,000 population per week. Since then, the incidence rate has dropped to about 70 per 100,000 per week. But for the high prevalence of multi-generational families at risk for intra-household transmission, we’ve maintained that the COVID-19 incidence rate in Los Angeles would now be much lower. In Wisconsin, by contrast, the most recent data are running in the range of 225 confirmed cases per 100,000 per week in Dane County and 300 per 100,000 per week in Milwaukee County.
One explanation is that there is so little political consensus in Wisconsin that state and local governments are effectively paralyzed, or at least severely constrained into taking half-measures. Another is that governmental orders to wear masks and stay out of bars are effective only if accompanied by compelling messages. Admonitions to protect yourself or to protect others, we’ve maintained, need to be replaced by messages to protect your family. In any case, our research needs to stop asking whether public policies work and start inquiring when and where they work.
Public health policy needs to be reoriented from a focus on protecting the individual to a focus on protecting the household.
In our clinical work at a community health center here in downtown Los Angeles, the classic COVID-19 presentation is not that of a single patient, but of an entire household that has come down with the virus within the space of a few days.
With so much variability in the duration of the incubation period from infection to symptoms, it’s not terribly informative which household member happened to get sick first. But if you take a careful medical history, you’ll invariably identify a younger, socially mobile family member with few or no symptoms.
We’ve lost count how many times a fifty-something patient, struggling with body aches, loss of smell and chest tightness, expresses relief that her millennial son seems absolutely fine, even though he has been in constant, direct contact with everyone else from grandpa on down. Ironically, it is painfully clear who imported virus into the household.
Two Maps, Same County
Figure 1 below compares two maps of Los Angeles County. Each map is broken down into countywide statistical areas (CSAs), a hybrid geographic classification of independent municipalities such as the City of Beverly Hills, neighborhoods of Los Angeles such as Hollywood, and unincorporated places such as Hacienda Heights.
On the left, the CSAs are color-coded according to the age-adjusted cumulative incidence of COVID-19 per 100,000 population as of September 19, 2020, which we derived from the surveillance dashboard of the Los Angeles County Department of Public Health. On the right, the same CSAs are coded according to the proportion of households that we’ve identified as at risk for multi-generational transmission, which we derived from the 2018 public use microsample of the U.S. Census Bureau’s American Community Survey. As explained in this detailed report, we classified a household as at risk for multi-generational transmission if it had at least four persons, at least one person 18–34 years of age and another person was at least 50 years of age.
The two maps in Figure 1 show a striking concordance. Those communities with the highest prevalence of at-risk households had the highest cumulative incidence of COVID-19 infection.
Figure 2 below shows the weekly incidence of newly diagnosed COVID-19 cases in Los Angeles County as a whole, running from the week starting March 1 through the week starting October 4, 2020. We have divided the epidemic into three phases. During Phase I, which ran approximately through the week of April 4–10, the epidemic spread radially from initial foci of infection located in relatively affluent communities such as the Brentwood and Beverly Crest neighborhoods of Los Angeles and the City of West Hollywood.
During Phase II, which ran through about the week of July 5-11, COVID-19 incidence rose at slower rate, as COVID-19 infections became increasingly concentrated in areas at higher risk of intrahousehold transmission. Since the week starting July 12, COVID-19 incidence has been gradually declining, while cases continue to accumulate in the same areas where the prevalence of at-risk households remains higher. We can see the evolution of the epidemic from March 28 through September 19 in the color-coded animation in Figure 3 below.
Within-Household Transmission Sustained the Epidemic Even As COVID-19 Diagnoses Were Declining.
Figure 4 below shows two graphs. Both relate the cumulative incidence of COVID-19 infection on the vertical axis to the prevalence of at-risk households across some 300 CSAs in Los Angeles County, as measured on the horizontal axis.
The graph on the left covers cases of COVID-19 diagnosed during Phases I and II, from March 1 through July 11, 2020, when weekly incidence rates were continuing to rise. The graph on the right, by contrast, covers cases diagnosed during Phase III, from July 12 through October 16, when weekly incidence rates have turned around and gradually begun to fall.
During Phases I and II of the epidemic through mid-summer, when weekly case counts were still rising, a 10-percentage-point increase in the prevalence of at-risk households was associated with a 46-percent increase in COVID-19 diagnoses. During Phase III when weekly case counts were declining, the same 10-percentage point increase in the prevalence of at-risk households was associated with a 53-percent increase in COVID-19 diagnoses.
Even as new COVID-19 cases were coming back down, there remained a strong relationship between COVID-19 incidence and the prevalence of at-risk households. In fact, the relationship got stronger. Even as more stringent social distancing measures took effect and the governor launched a wear-a-mask campaign in early July, the epidemic was sustained in Los Angeles County by continued within-household transmission.
Figure 5 below maps relative gym attendance during the month of April 2020. As described in our detailed report, we used the SafeGraph Patterns Data on the movements of devices with location-tracking software to count monthly visits to any one of two thousand gyms in relation to the geographic home base of each device. Each CSA is color-coded according to its gym attendance in April 2020 – the month with the largest overall decline in gym visits – as a percentage of the CSA’s baseline gym attendance rate in February 2020.
The gym-attendance map in Figure 5 certainly doesn’t look like the cumulative incidence maps in Figures 1 and 3. While social distancing may still be an important determinant of the overall trend in COVID-19 incidence seen in Figure 2, we can’t explain differences between CSAs solely on the basis of gym attendance. It turns out, however, that the strong relationship between at-risk household prevalence and COVID-19 incidence, seen in Figure 4, is even stronger among those CSAs with higher gym attendance. That is, there appears to be a synergy between the rate of gym attendance and the prevalence of at-risk households in determining COVID-19 case counts.
The story underlying this synergy is straightforward. Gym attendance is an indicator of social mobility of younger persons. Higher social mobility means a higher risk of contracting COVID-19. When a younger person, having contracted COVID-19 outside the household, brings his or her infection back home, the impact is magnified by the presence of cohabitants of multiple generations.
A Skeptical Note on Selective Social Distancing Policies
Things would be a lot simpler if older persons were all sequestered in retirement communities or assisted living facilities. But the data here demonstrate that this is not the reality of Los Angeles County. The overall, countywide prevalence of at-risk households is 13.8 percent. With an estimated 3.3 million households in the county, we’re talking about 455,000 multi-generation households where an asymptomatic or mild SARS-CoV-2 infection in a younger household member would put older household members at significant risk.
A New Focus for Public Health Policy
Our findings require us to view the household rather than the individual as the foremost target of healthcare policy. The message “protect yourself” (protégete in Spanish) needs to be reconfigured as “protect your family” (protege a tu familia). Protecting your family is a far more immediate and personal concept than “protecting others” (proteger a los demás).
When a healthcare provider encounters a new patient with suspected or established COVID-19, the clinical interview needs to turn quickly to questions about other household members. “Who do you live with?” “Is anyone else sick?” “How old are they?” “Do they have other medical problems?” “Do they have their own doctor? Or a health plan?” The widely recognized model of the patient-centered medical home needs to be replaced by the family- and household-centered medical home.
Corrected for reporting delays, the daily incidence of newly confirmed cases appears to have doubled.
As noted in COVID-19 Reporting Delays: Whither New York City?, we’ve been following the daily counts of newly confirmed cases of COVID-19 as they are regularly reported by the New York City health department. As a result of delays in reporting, we’ve observed, the most recent counts routinely fall below the actual number of cases to date. In fact, the health department cautions on its COVID-19 data dashboard, “Due to delays in reporting, which can take as long as a week, recent data are incomplete.”
Using a statistical method first applied to reporting delays of AIDS cases in the 1990s and recently updated in a technical report, we have filled in the missing data and projected the actual number of cases diagnosed to date. Our statistical approach cannot predict any single individual’s pending test result, but it can give us a reasonably accurate estimate of recent, new COVID-19 cases at the population level.
In Figure 1 above, the gray data points show the numbers of cases so far reported as diagnosed on each day from June 21 through October 2. The periodic dips in the data arise from reduced testing over the weekends. As a result of reporting delays, the most recent gray data points give the false impression that the epidemic has petered out. The pink data points show that, once all the case reports come in, the number of cases is expected to be at least double – if not triple – the approximately 300 cases per day seen during the past 2–3 months.
Reduced Reporting Delays, But Still Not Enough
Figure 2 shows the updated cumulative distribution of reporting delays, based on data over the most recent two months. These new data show a significant reduction in reporting times. When we last checked on August 15, only 81.3 percent of test results had been reported by 10 days, compared to 96.6 percent during the last two months. The mean reporting time is now 3.44 days, compared to 5.43 days as of August 15.
The problem, however, is that by just two days after testing, only 47.2 percent of the results – less than half – are reported. When it comes to a rapid public health response to an outbreak, two days can be an eternity. As of our cutoff date of October 2, the health department reported that 353 newly confirmed cases had been diagnosed two days earlier on September 30. That would mean 353 ÷ 47.2% = 748 cases will eventually be reported for September 30. This projection has been marked by the arrow in the upper right corner of Figure 1.
Why Our Projections Might Be Wrong
The main reason why our projections could be wrong is that the health department has abruptly reduced its reporting delays, but this improvement is not captured in our analysis of its reporting patterns during the past two months. The most likely possibility would be a substantial, recent increase in the demand for rapid testing, perhaps related to the recent reopening of schools. Thus far, we cannot find any data to support this speculation.
Nonetheless, the increase projected in Figure 1 is so substantial that we think it’s appropriate now to post our findings. When we add up the projected counts over the past 10 days, we’re talking about 2,500 excess cases above baseline.
What Might Be Happening
Alarms have been raised about newly emerging foci of infection in Brooklyn and Queens. The new data, however, suggest that something else may be happening on a larger scale.
Figure 3 is an update of a graphic we’ve already displayed, but now with two more months of data. This figure does not incorporate any of our projections. It’s simply a rendering of data buried in the health department’s archive in multiple daily files named by-age.csv. The incidence of COVID-19 in the younger adult age group, ages 18–44, has now clearly overtaken the incidence in the older group. In the period through June 20, as we previously noted, younger adults had an incidence that was on average 40 percent lower than that of their older counterparts. At our last report, the incidence among younger New Yorkers after June 20 was about 20 percent greater. Now it’s more than 30 percent greater.
This shift in age distribution is not likely to be the result of the emergence of a recent hot spot, or the reopening of schools in the last two weeks. While we have no data on the age distribution of those who fled the city when the epidemic was raging, we doubt that we’re now witnessing a novel twist on the Return of the Native.