Crowding Effects Dominate Demographic Attributes in COVID-19 Cases

Graphical abstract

Highlights

• •Cross-sectional studies of infection rates are critical to understand the drivers behind the COVID epidemic.
• •This study sought to identify demographic and socio-economic indicators that drive the incidence rate of COVID-19.
• •The paper assesses these factors at zip-code granularity.
• •The average household size is the single most important explanatory variable; the percentage of the population 65 or older, and that below the poverty line are also strongly positively associated.
• •Population density, per se, does not have a significantly positive impact on incidence rates.

Abstract

Objective

With an eye toward possible public policy implications, our objective is to identify the socio-economic and demographic factors that drive the large variation in COVID-19 incidence rates observed within relatively compact geographic regions, and to quantify the relative impact of each of these factors. We use international comparisons as a starting point.

Methods

New York City, consisting of some 175 zip codes, is an ideal arena to pursue the above study given the large variation in case incidence rates across zip codes. We conducted systematic regression studies employing data with zip code granularity. Our model specifications are based on a well-established epidemiologic model that explains the effects of household sizes on R0.

Results

Average household size emerges as the single most important driver behind the large variation in COVID-19 incidence rates. It independently explains 62% of the variation. The percentage of the population above the age of 65 and the percentage below the poverty line are also strongly positively associated with zip code incidence rates. As to ethnic/racial characteristics, the percentages of African Americans, Hispanics and Asians within the population are significantly associated, but the magnitude of the impact is smaller. (The proportion of Asians within a zip code has a negative association.) Contrary to common belief, population density, by itself, does not have a significantly positive impact (other than when a high population is driven by large household sizes).

Conclusion

Our findings support implemented and proposed policies to quarantine patients and separate infected individuals from families or dormitories; they also support newly revised nursing home admission policies.

Introduction

The world continues to search for a fundamental understanding of the dynamics of the current pandemic. For example, we try to understand why, in the United States (US), as of May 17, 2020, 192,000 COVID-19 cases have been reported in New York City (NYC) alone, while the country-wide total has been mercifully restricted to 1.5 million. Thus, a city representing 2.5% of the US population accounted for 12.8% of the reported cases. Identifying the main drivers of the disease spread has important implications for public policies to contain the current epidemic and mitigate the widely expected “second wave”.

Population density is widely believed to be a main driving force. This theory has intuitive appeal. After all, the number of infections in a given region depends on the basic reproductive number R0, the average number of cases directly generated by one individual, in a population where all individuals are susceptible. R0, in turn, depends, in part, on the number of individuals with whom a single case has physical contact during the time interval in which she is contagious, presumably positively correlated with the population density.

However, the theory is challenged, first of all, on an international basis. Many cities with population density as great as or greater than NYC’s 10,198 residents per square kilometre (sq. km) have reported much lower case rates. For example, Manila, Baghdad, Mumbai, Seoul, Mexico City and Singapore have, respectively, 46128, 32874, 32103, 16000, 9800 and 8358 residents/sq. km. Their case incidence rates per 100,000 residents vary, however, between 9.4 (Seoul) and 635.4 (Singapore), as compared to NYC’s rate of 2286 cases per 100,000 residents.

These cities’ lower rates may be explained by ex-ante differences in international traffic patterns in and out of the country, affecting the cluster of “imported” cases, or specific containment and testing policies adopted by the respective governments. However, among states within the US, California had one of the highest population densities (251.3 residents per square mile) but one of the lowest COVID-19 mortality rates (8 per 100,000 residents), while Louisiana, with a population density 2.5 times lower than that of California, has reported 49 COVID-19 deaths per 100,000 residents.

And stark differences are apparent within NYC itself. (Wadhera et al., 2020) reported recently that among the city’s five boroughs, Manhattan had by far the fewest hospitalizations per 100,000 residents, but the greatest population density, 2.5 times the citywide average (25,106 vs 10,198 residents/sq.km). (Because the percentage of confirmed cases that require hospitalization is remarkably robust throughout the country, hospitalization rates can be viewed as proxies for incidence rates.) In fact, at zip code granularity, the rates of reported cases and the population densities are negatively correlated, with a correlation coefficient of – 33% (Table 2).

Table 2.

Correlation coefficients among the variables.

	Population density= # residents per sq. mile	Average household size	Percent below poverty line	Percent above age 65	# confirmed cases/100,000 residents
Population density=# residents per sq mile	1	−0.324	0.285	−0.064	-0.325
Average household size	−0.324	1	0.189	−0.207	0.622
Percent below poverty line	0.285	0.189	1	−0.367	0.229
Percent above age 65	−0.064	−0.207	−0.367	1	0.092
# confirmed cases/100,000 residents	-0.325	0.622	0.229	0.092	1

Several authors [(Barr & Tassier, 2020) and (Bassett, 2020)] have cautioned against relying on population density as a prime explanatory variable.

First and foremost, populations consist of “households” and the inter-household and intra-household dynamics of an epidemic differ fundamentally. In particular, the contact rate between a pair of individuals living in the same household tends to be much greater than that between two individuals from different households. Starting with the seminal papers (Bartoszynski, 1972) and (Becker, 1977) this observation has been at the core of many epidemiological models. See Section II for a seminal model by (Becker & Dietz, 1995). (Rader et al., 2020) report that infection patterns across China can only be explained with “hierarchically-structured populations where individuals had high rates of contact within their households”.

The Novel Coronavirus Pneumonia Emergency Response Epidemiology Team (2020) reported that 80% of transmission clusters of the coronavirus occurred within households. Nearly three weeks into the initial outbreak in Wuhan, the government instituted social distancing and travel bans, reducing R0 from 3.88 to 1.26. However, it was not until Wuhan instituted a centralized quarantine where anyone with COVID-19 -like symptoms would be centrally quarantined at hotels or dormitories, that R0 was reduced to 0.32, a 75% improvement over social distancing alone! Indeed, this intervention was the final step in achieving full containment in China.

In NYC, average household size varies by zip code, between 1.57 and 3.97 (Table 1 ). It is also highly positively correlated with the number of confirmed cases per 100,000 residents (Table 2 ). And although Manhattan is the most densely populated of all boroughs, it is also the one with the smallest average household size. Thus, any statistical model intended to explain the variation in case intensities across the City should include the average household size as an explanatory variable.

Table 1.

Descriptive statistics by NYC zip code.

	Minimum	Maximum	Average
# confirmed cases per 100,000 residents	436.91	4226.51	2077.79
Population density= # residents per sq. mile	1389.08	126,067.69	38 480.75
Average household size	1.57	3.97	2.65
Percentage below poverty line	2.20	45.90	16.83
Percentage above the age of 65	0.46	28.98	14.26
Percentage of total Hispanic population	1.12	75.77	26.35
Percentage total Black non-Hispanic population	0.37	90.51	20.06
Percentage total Black population	0.37	93.81	23.69
Percentage total Asian population	0.07	72.62	14.78

More broadly, we need to identify demographic and socio-economic indicators that have significant predictive power. NYC is an ideal arena in which to pursue this investigation; as mentioned, the city has, by far, the highest confirmed incidence rate in the country, but it also exhibits remarkable diversity with respect to many potentially relevant demographic and socio-economic factors. Combining various data sources, described below, we have been able to assess these factors for 174 of the 177 NYC zip codes; see Table 1.

A basic household-based epidemiology model

(Becker and Dietz, 1995) consider a population with households of varying sizes. Let μ_H, σ_H and $c{v}_H={\sigma }_H/{\mu }_H$ denote the mean, standard deviation and coefficient of variation of the distribution of household sizes, respectively. Let b denote the mean number of infectious contacts that an individual makes with individuals outside her own household during her entire infectious period. An infectious contact is one that is close and intensive enough to result in disease transmission.

The following closed form expression for R0, the basic reproductive number of the epidemic,

was derived from [6, p.211]:

$$R_0=\frac{b}{2}\left[1+{\left(1+\frac{4\left({\mu }_H\left(1+c{v}_H^2\right)-1\right)}{b}\right)}^{\frac{1}{2}}\right]$$ (1)

In a population where all individuals live alone, ${\mu }_H=1$ and $c{v}_H=0$, so that $R0=b$, as in elementary models that do not account for household clusters. In contrast, where people live in households of size 4 (e.g., married couples with 2 children), ${\mu }_H=4$ and, if $b>=1$ (i.e., an average infected individual infects at least one individual outside her own household):

$$\begin{array}{l}{R}_0=0.5\left(b+\sqrt{b^2+12b}\right)=b+0.5\left(\sqrt{b^2+12b}-b\right)\\ =b+\frac{6b}{\left(\sqrt{b^2+12b}+b\right)}=b+\frac{6}{\sqrt{1+12b^{-1}}+1}\ge b+1.3\end{array}$$

adding at least a full 1.3 points to the R0 value! And when household sizes vary, as they typically do, the household effect is even greater, see (1). For example, if the household distribution approaches a geometric distribution,

$$c{v}_H=\sqrt{1-\frac{1}{{\mu }_H}}=\sqrt{0.75}=0.87,R_0=0.5\left(b+\sqrt{b^2+24b}\right)=b+\frac{12}{\sqrt{1+24b^{-1}}+1}\ge b+2$$

Population density, by itself, may affect R0 via the parameter b, depending on the interactions among individuals in different households. But if it does, (1) shows that it interacts strongly with average household size (and its coefficient of variation) as well.

Methods

The numbers of reported COVID-19 cases per zip code were obtained from the NYC Department of Health's GitHub repository, and population sizes and population densities from

http://zipatlas.com/us/ny/zip-code-comparison/population-density.htm.

The demographic and socio-economic characteristics below were extracted from https://github.com/BuzzFeedNews/2020-05-covid-city-zipcodes/blob/master/data/processed/census/zip_census_data.csv, a database created from 2018 5-year American Community Service (ACS) estimates, obtained from the US Census Bureau.

Average household size

Age

• •age_60_and_over
• •age_65_and_over
• •age_75_and_over
• •median_age

Economic

• •pct_more_than_one_occupant_per_room
• •pct_below_poverty_level
• •household_median_income
• •household_mean_income

Racial/Ethnic

• •total_non_hispanic,
• •total_white_nonhispanic,
• •total_black_nonhispanic,
• •total_asian_alone,
• •total_black,
• •total_hispanic

We applied standard linear regression to estimate best fit regression equations, assuming independent noise terms, all with an identical normal distribution. We have conducted our study in two stages: in the first stage, we focused on crowding, age and income related factors; in the second stage, we investigated to what extent any of the racial/ethnic characteristics have additional explanatory power.

Based on the empirical and theoretical observations above, we specified a model with (a) population density and (b) average household size as potential explanatory variables, along with (c) an interaction term between them, (d) the percentage of the population above the age of 65, and (e) the percentage below the poverty line. The relevance of the latter two demographics is both intuitive and is explained in our Discussion section. In the remainder of this paper, we employ the following variable names:

• • population_density: Population density=# residents per sq. mile
• • avg_household_size: Average household size
• • pct_below_poverty_line: Percent below poverty line
• • pct_age>65: Percent above the age of 65
• • cases_per_100k: # confirmed cases/100,000 residents

As mentioned, the correlation between population density and cases_per_100k is -33%. The correlation between the standard interaction term population density*avg_household_size and cases_per_100k is also negative –albeit weaker (-12%). We therefore specified the interaction term as (population density*avg_household_size)² which is positively correlated with cases_per_100k.

We have based our regression studies on the reported cases on May 12, 2020, and report the results for that day. However, to assess the robustness of our results, we have repeated the studies for a full 15 day window around this target date, starting with May 5, 2020 and ending with May 19, 2020.

Results

As the correlation figures in Table 2 suggest, average household size is, by far, the single most important explanatory variable in the model. Indeed, it independently explains 62% of the variation in incidence rates. In a model that includes only household size as an explanatory variable, the following equation provides the best fit.

# confirmed cases/100,000 = 450.5 + 986.9* (average household size -1)

The coefficient in front of the average household size variable is highly statistically significant. The equation indicates that in zip code A, where the average household has one more member than that in otherwise comparable zip code B, the case rate would be 986.9 cases/100,000 higher than in zip code B (Table 3 ).

Table 3.

Regression results with average household size as the single explanatory variable; avg_household size_adj = average household size-1.

	Coefficients	Standard Error	t Stat	P-value	Lower 95%	Upper 95%
Intercept	450.51	163.46	2.76	0.006	127.86	773.154542
avg_household_size_adj	986.90	94.79	10.41	5.41E-20	799.80	1174.01

However, a significantly superior fit can be obtained by adding the remaining explanatory variables (Table 4 ). This model accounts for 72% of the variation in case rates, while R² and the adjusted R² are 51% and 50%, respectively, resulting in the following equation:

Table 4.

Full Regression Results without Interaction Effects.

	Coefficients	Standard Error	t Stat	P-value	Lower 95%	Upper 95%
Intercept	−284.40	353.05	−0.805	0.42	−981.40	412.59
Population_density	−0.006	0.003	−1.81	0.07	−0.01	0.001
Average_household_size	896.83	180.56	4.97	1.66E-06	540.36	1253.29
%_below_poverty_level	24.37	5.36	4.55	1.03E-05	13.79	34.942533
% age>65	49.94	9.80	5.10	9.29E-07	30.59	69.2839432
Population_density* (average_household_size)2	−3.52E-05	0.001	−0.03	0.98	−0.002	0.00250162

# confirmed cases/100,000 = -284.4 + 896.8* (average household size -1)+ 49.9*percentage above age 65 + 24.4*percentage below poverty level - 0.006*population density – 3.52E-05*(population density*(average household size -1)²) (2)

The average household size continues to be the main determinant of the number of reported COVID-19 infections. In contrast, neither population density, nor its interaction term with the average household size, has a significant impact on the case intensity. (The intercept value is non-significant as well.)

Average household size, the percentage of the population above the age of 65 and the percentage below the poverty line, all have positive coefficients that are statistically significant at a very high confidence level. Including the other demographic variables in the model reduces the coefficient of the average household size variable to 896.8 but allows us to conclude at a 95% confidence level that an increase of but one individual to the average household size in the zip code leads, ceteris paribus, to an increase of at least 540.4 cases per 100,000 residents.

An increase of the percentage of senior residents by one percentage point augments, ceteris paribus, the case rate by approximately 50 cases, more than double the effect of an increase of the poverty rate by a single percentage point. The association of both demographic factors with zip code incidence rates is plausible as well; senior residents may be equally susceptible to getting infected as other age brackets. However, they are at increased risk to develop significant symptoms and complications, thus disproportionally contributing to the counts of confirmed cases, the only counts we have. (Bendavid et al., 2020) estimated that the number of confirmed cases in Santa Clara county may have been a mere 2% of the total number of infections.

Likewise, the a priori health status of individuals in the lowest income bracket is typically inferior to that in higher income brackets. Infected individuals with incomes below the poverty line therefore contribute disproportionally to the counts of confirmed cases as well. Additionally, a much larger percentage of this population segment is employed on location, as opposed to being able to work from home, and is thus more susceptible to infections.

We have checked for the presence of other interaction effects, omitting the population density-related variables that were found to have an insignificant impact. The only significant interaction effect is that between the percentage of the population below the poverty line and the percentage of senior residents. See Table 5 for the revised regression results. (The addition of the interaction term has a negligible impact on the R² or adjusted R² value).

Table 5.

Full Regression with interaction effect, but population density omitted.

	Coefficients	Standard Error	t Stat	P-value	Lower 95%	Upper 95%
Intercept	−549.53	246.39	−2.23	0.027	−1035.93	−63.14
avg_household_size_adj	1051.69	89.00	11.82	7.17E-24	875.99	1227.38
pct_below_poverty_level	25.48	5.43	4.692	5.57E-06	14.76	36.20
pct age>65	48.62	9.73	4.997	1.44E-06	29.41	67.82
pct < PL*pct_age>65	−0.004	0.001	−3.03	0.003	−0.006	−0.001

Thus, after omitting the population density related variables whose impact was found to be statistically insignificant, our estimate for the coefficient of avg_household_size_adj has a much smaller standard error, resulting in a much narrower 95% confidence interval. The point estimate of this coefficient is now 1051.7, and above 876, at a 95% confidence level. The revised model specification has minimal impact on the estimated coefficients of the renaming explanatory variables.

As mentioned, we have repeated the regression studies for each of 15 days around the target date. The estimated coefficients are all remarkably robust, all well within the 95% confidence interval for the May 12 estimates; see Table 6 .

Table 6.

Full Regression with interaction effect, but population density omitted.

	Coefficients	Standard Error	t Stat	P-value	Lower 95%	Upper 95%
Intercept	−549.53	246.39	−2.23	0.027	−1035.93	−63.14
avg_household_size_adj	1051.69	89.00	11.82	7.17E-24	875.99	1227.38
pct_below_poverty_level	25.48	5.43	4.692	5.57E-06	14.76	36.20
pct age>65	48.62	9.73	4.997	1.44E-06	29.41	67.82
pct < PL*pct_age>65	−0.004	0.001	−3.03	0.003	−0.006	−0.001

We have created an interactive heat map of NYC1 The color of each area is determined by the zip code’s residual in the above regression equation, which denotes the number of cases per 100,000 residents which the zip code experienced in excess of or below what could be expected based on its demographic characteristics. When clicking on a specific zip code, a table with all of the above measures emerges. See Fig. 1 for a snapshot.

Fig. 1.

Snapshot of the heat map of NYC showing the zip code’s residuals calculated from regression Eq. (2).

The impact of racial and ethnic factors

It has been widely conjectured that certain ethnic or racial groups, in particular the black and Hispanic communities, are at increased risk for COVID-19 morbidity. For example, (Departments of Public Health of Illinois; Mississippi and Virginia, 2020) reported that black Americans experienced a disproportionately greater number of reported COVID-19 cases compared to other Americans. (Yancy, 2020) concludes:” the pattern is irrefutable: underrepresented minorities are developing COVID-19 infection more frequently”

However, it remains difficult to assess whether the disproportionately greater incidence rate among the black population, as reported in the above three states, is pervasive throughout the country. The CDC publishes cumulative COVID-19 data reported by State Health Departments; however, 78% of those data were missing race/ethnicity disaggregation as of April 15, 2020., and 50% as of May 30, see (Millett et al., 2020) and (Godoy & Wood, 2020), respectively.

It is easier to relate the case incidence rates in the country’s 3142 counties to the proportions of black residents in these counties, since complete data are available, here. Figure 2 in Millett et al. (2020) shows that the case incidence, mostly, although not universally, increases with the county’s proportion of black residents: on the estimated curve, the case incidence rate doubles when comparing counties with a 1% black population with those where 75% of the population is black.

Moreover, even if the higher incidence rate is pervasive, we need to understand whether the difference with the general population is entirely due to other socio-economic or demographic factors (such as average household size or the poverty rate) or not.

(Millett et al., 2020) provided initial insights. They partitioned the country’s 3142 counties into two sets: the first set of 677 counties with a disproportionately large (> 13%) black population and the second set with the remaining 2465 counties. The estimated case incidence rate in the first set is more than 4 times that in the second, but the associated 95% confidence intervals are overlapping.

Similar concerns apply to the Hispanic/Latino population. With the above proviso that in 48% of confirmed cases the race or ethnicity of the patient is unknown, the COVID-19 Racial Tracker reports that in 42 states plus Washington D.C., Hispanics/Latinos make up a greater share of confirmed cases than their share of the population. In eight states, it is more than four times greater.

To assess the impact of various ethnicity/racial factors, we have evaluated four additional regression models, each time adding to the model in Eq.(2), one of the following four ethnic /racial population percentages as an explanatory variable: (a) total_hispanic, (b) total_black_nonhispanic, (c) total_black and (d) total_asian_alone,. (Once again, we omitted the population density related variables that were shown to have an insignificant impact, see above.). See Table 7, Table 8, Table 9, Table 10 .

Table 7.

Regression results when including %total_hispanic population as an additional independent variable.

	Coefficients	Standard Error	t Stat	P-value	Lower 95%	Upper 95%
Intercept	−607.02	239.93	−2.53	0.0123	−1080.66	−133.37
avg_household_size_adj	918.81	90.26	10.18	0.0000	740.64	1096.99
pct_below_poverty_level	3.82	5.95	0.64	0.5220	−7.93	15.56
pct age>65	54.71	9.57	5.72	0.0000	35.83	73.60
%total_hispanic	12.35	3.03	4.08	0.0001	6.37	18.32

Table 8.

Regression results when including %total_black_nonhispanic population as an additional independent variable.

	Coefficients	Standard Error	t Stat	P-value	Lower 95%	Upper 95%
Intercept	−593.85	245.16	−2.42	0.0165	−1077.82	−109.87
avg_household_size_adj	961.76	90.63	10.61	0.0000	782.84	1140.67
pct_below_poverty_level	14.79	5.01	2.95	0.0036	4.89	24.69
pct age>65	50.42	9.71	5.19	0.0000	31.26	69.59
%total_black_nonhispanic	5.87	1.94	3.03	0.0028	2.05	9.70

Table 9.

Regression results when including %total_black population as an additional independent variable.

	Coefficients	Standard Error	t Stat	P-value	Lower 95%	Upper 95%
Intercept	−600.22	244.74	−2.45	0.0152	−1083.37	−117.08
avg_household_size_adj	961.57	90.44	10.63	0.0000	783.04	1140.11
pct_below_poverty_level	13.79	5.08	2.71	0.0073	3.76	23.83
pct age>65	50.75	9.70	5.23	0.0000	31.61	69.89
%total_black	5.77	1.86	3.10	0.0022	2.10	9.43

Table 10.

Regression results when including %total_asian population as an additional independent variable.

	Coefficients	Standard Error	t Stat	P-value	Lower 95%	Upper 95%
Intercept	−484.66	238.91	−2.03	0.0441	−956.29	−13.02
avg_household_size_adj	1081.07	86.49	12.50	0.0000	910.34	1251.80
pct_below_poverty_level	12.27	4.89	2.51	0.0131	2.61	21.93
pct age>65	55.48	9.45	5.87	0.0000	36.82	74.13
%total_asian_alone	−14.73	3.17	−4.64	0.0000	−20.99	−8.47

The addition of the ethnic/racial descriptors has little impact on the estimates of the coefficients and standard errors of our two main explanatory variables: average household size and the percentage of residents age 65 or older.

The coefficient of the percentage below the poverty line is halved, from approximately 25 to a number in the range [12.27,14.79] in Table 8, Table 9, Table 10. In the remaining model (Table 7) where the impact of the proportion of Hispanics is assessed, this coefficient loses its significance. In this case, it is difficult to separate its impacts of the ethnic and the economic variable, as the two are highly positively correlated (65%, see Table 15).

Table 15.

Correlation coefficients among different variables.

	Average household size	Percent below poverty line	Percent above age 65	Percentage of total Hispanic population	Percentage total Black non-Hispanic population	Percentage total Black population	Percentage total Asian population	# confirmed cases/100,000 residents
Average household size	1.00	0.19	−0.21	0.26	−0.56	0.26	0.34	0.62
Percent below poverty line	0.19	1.00	−0.37	0.65	0.28	0.34	−0.29	0.23
Percent above age 65	−0.21	−0.37	1.00	−0.34	−0.13	−0.16	0.19	0.09
Percentage of total Hispanic population	0.34	0.65	−0.34	1.00	−0.02	0.05	−0.26	0.42
Percentage total Black non-Hispanic population	0.26	0.28	−0.13	−0.02	1.00	0.99	−0.48	0.34
Percentage total Black population	0.26	0.34	−0.16	0.05	0.99	1.00	−0.51	0.35
Percentage total Asian population	0.07	−0.29	0.19	−0.26	−0.48	−0.51	1.00	−0.19
# confirmed cases/100,000 residents	0.62	0.23	0.09	0.42	0.34	0.35	−0.19	1.00

The coefficients of each of the ethic/racial descriptors are significant at a very high confidence level, with p-values below 0.0028, confirming the above described conjectures. However, the magnitude of these coefficients is modest, in particular when compared with the percentage of seniors and the percentage below the poverty line. An increase of the percentage of Hispanic residents by one percentage point augments, ceteris paribus, the case rate by approximately 12 cases. Thus, if in zip code A, the proportion of Hispanic residents is 20 percentage points higher than that in otherwise comparable zip code B, the case rate would be 247 higher than in zip code B.

The impact of the percentage of black residents, and that of black non-Hispanic residents, while strongly significant, is even smaller. An increase of this proportion by one percentage point augments, ceteris paribus, the case rate by approximately 6 cases. Thus, if in zip code A, the proportion of black residents is 20 percentage points higher than that in otherwise comparable zip code B, the case rate would be 115 higher than in zip code B.

In contrast, the proportion of Asian residents in a zip code has a (highly significant) negative association with the case rate, each additional percentage point decreasing, ceteris paribus, the case rate by some 14 cases.2

As with the basic model, we have repeated the regression analyses for each of the 15 days from May 5-May 19; once again the estimates are remarkably robust, see Table 11, Table 12, Table 13, Table 14 .

Table 11.

Maximum, minimum and average values of coefficients for regression including %total_hispanic population as an additional independent variable for 15 day period.

	avg_household_size_adj	pct_below_poverty_level	pct_age>65	%total_hispanic
Minimum	875.61725	2.6686427	51.994143	1165.5236
Average	911.73546	4.0941206	55.349999	1238.7307
Maximum	931.62228	5.226449	58.726034	1310.935

Table 12.

Maximum, minimum and average values of coefficients for regression including %total_black_nonhispanic population as an additional independent variable for 15 day period.

	avg_household_size_adj	pct_below_poverty_level	pct_age>65	%total_black_non_hispanic
Minimum	916.27133	13.036153	47.943724	553.63002
Average	955.9674	15.172045	51.047048	578.35984
Maximum	981.29695	17.115713	54.177515	600.34054

Table 13.

Maximum, minimum and average values of coefficients for regression including %total_black population as an additional independent variable for 15 day period.

	avg_household_size_adj	pct_below_poverty_level	pct_age>65	%total_black
Minimum	916.17228	12.10048	48.251701	542.78349
Average	955.76292	14.187076	51.369147	567.9981
Maximum	980.97525	16.112607	54.503164	588.8834

Table 14.

Maximum, minimum and average values of coefficients for regression including %total_asian population as an additional independent variable for 15 day period.

	avg_household_size_adj	pct_below_poverty_level	pct_age>65	%total_asian_alone
Minimum	1030.0448	10.525976	7.94E-09	−1514.9971
Average	1074.2913	12.603313	56.092721	−1471.4693
Maximum	1101.9448	14.40379	59.369881	−1421.1716

Discussion

Average household size emerges as the single most important demographic variable associated with the large variation in infection rates. We estimate that an additional single household member increases the number of cases by 892 or more (Table 15 ).

This result is quite intuitive. It confirms both the above empirical findings and its theoretical underpinnings; see Section II. Of course, the same contagion potential exists in other settings where a significant number of individuals reside together in the same dwelling, for example dormitories and nursing homes. (Girvan, 2020) reported that in the US, 40% of COVID-19-induced deaths occurred in nursing homes and assisted living facilities. Similarly, across the globe, half of COVID-19 fatalities occurred in nursing homes and related facilities (Comas-Herrera & Zalakain, 2020). Their residents face the hazard of living in close proximity to many others, while belonging to an age group with significantly increased potential for serious symptoms and hence confirmation of their infections.

The proportion of the population above the age of 65 emerges as a second significant explanatory variable, with each additional percentage point contributing 50 cases per 100,000 residents. The association of the proportion under the poverty line with the outcome, such that each percentage point contributes an estimated 24 cases per 100,000, is also plausible: the health status of this segment is generally poorer, enhancing infections with significant symptoms. Moreover, far more individuals in this income bracket are employed as essential workers and are dependent on public transportation, with limited opportunities for social distancing; for both reasons, this population segment contributes more than others to the case count. The second factor pertains almost exclusively to individuals below retirement age, possibly explaining the significantly negative coefficient of the interaction term in Table 6.

Our findings about the great importance of household size support such policy initiatives as isolation policies for individuals, either immediately upon being identified as infected or post-hospitalization (see, for example, (Chan et al., 2020)). Based on their model (Chan et al., 2020), the authors show: “that the time to reopening [of the economy] can be shortened by 11%. In addition, assuming symptomatic people are infectious, if 50% of them are quarantined before getting sick enough to go to the hospital, we can reduce the risk of developing severe COVID -19 illness and the time till reopening can be shortened by 86%.” Our findings also support the May 11 decision by New York State to cancel its mandate requiring nursing homes to accept COVID-19 patients.

Racial/ethnic factors have a significant, albeit more modest, association with case incidence rates above and beyond differences in average household size and the percentage below the poverty line with which they are correlated. (Richardson et al., 2020), studying 5700 patients in the NYC area, identified hypertension, obesity, and diabetes as the most common comorbidities. These comorbidities are known to be more prevalent among the black and Hispanic population segments, thus providing an explanation for the observed ethnic/racial associations. In turn, increased prevalence of these comorbidities is, in large part, the result of poor dietary habits and more limited access to healthcare, themselves driven by lower incomes.

Future research should investigate other crowding factors, e.g., numbers of individuals residing in nursing homes or dormitories.

Financial Support

The authors received no financial support for this study.

Conflicts of Interest

The authors have no conflict of interest with respect to this study.

To our knowledge, no outside individual or institution has a material interest in the results of this study.

Declaration of interests

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

The authors declare the following financial interests/personal relationships which may be considered as potential competing interests:

Signed by

Awi Federgruen and Sherin Naha