Association of Human Mobility Restrictions and Race/Ethnicity–Based, Sex-Based, and Income-Based Factors With Inequities in Well-being During the COVID-19 Pandemic in the United States

Key Points Question Were state-specific mobility restrictions during the first wave of the COVID-19 pandemic in the United States associated with the well-being of individuals in groups that have historically been marginalized on the basis of race/ethnicity, sex, or income? Findings In this cross-sectional study of 1 088 314 US adults, African American individuals with low income, Hispanic individuals, and women of all racial/ethnic groups had higher risks of experiencing unemployment, class cancellations, food insufficiency, and mental health problems during the first wave of the COVID-19 pandemic. Meaning These findings suggest that public health policies that ignore existing distributions of risks to well-being may be intrinsically regressive if they fail to target necessary relief measures to individuals who have historically experienced the most marginalization.

This supplemental material has been provided by the authors to give readers additional information about their work.

Data merging
Pooling twelve waves of weekly HPS data, we construct a data set that is a repeated cross-section at the individual/household level and a representative panel at the weekstate level. In the HPS, each household is interviewed up to three times. As we treat the data as pooled repeated cross-sections, we do not need to consider the issue of loss to follow-up. We only analyze complete cases, and given that our sample size is large, our estimates are unlikely to be influenced substantially by this decision. Please see the HPS technical documentation for more information on how the Census Bureau handles missing data. 1 We merge these pooled HPS data with the IHME data on mobility, case rates, and death rates by week of survey and state of residence. Thus, in our final data set, all outcomes and covariates from the HPS vary at the individual/household level, while the mobility, case rates, and death rates vary at the state level, operating as ecological exposures (eTable3). For example, all individuals sampled from California in the first week of the HPS are assigned the same level of mobility, and their mobility changes weekly (twelve times corresponding to the twelve weekly surveys of the HPS).

Statistical model
For individual/household ℎ in state in week of survey , we estimate multivariable logistic regression models as:

Caveats for differing levels of measurement across outcomes
There are differences in the levels of measurement of the exposures and outcomes. For food insufficiency, class cancellations, and default on rent/mortgage (eTable3) in which the respondent answers on behalf of the household and him/herself, it is reasonable to assume these outcomes operate at the household and individual levels. For example, we interpret any differences in the class cancellations outcome across sex as being between the children of female versus male respondents. However, given that the question on unemployment asks about the respondent or anyone in the household, and the individual may not be the one who is unemployed, there may be measurement error in the sex gradient when regressing unemployment on sex. As we do not except differential misclassification by sex, any biases for sex differences in unemployment are likely to drive coefficients towards the null.

Sensitivity analyses
Relaxing the assumption of linearity for mobility dose response In a sensitivity analysis, we relax the assumption of a linear relationship between mobility restriction and outcomes. We recategorize mobility into five dummy variables: 10-19%, 20-29%, 30-39%, 40-49%, and ≥50% restrictions, with 0-9% as the reference category. The range of categories is appropriate given mobility restrictions vary between 0 and 60% in our sample. We fit regression models with these dummy variables instead mobility specification, given that extreme values have higher leverage in the linear specification. However, we interpret these results with caution, given the loss of information by binning the mobility data, which reduces variation and weakens the ability to detect nuanced relationships.

Autocorrelation/non-independence
Given that the HPS allows individuals to be resampled up to three times during the twelve weeks, we consider the potential issue of autocorrelation, which occurs when observations from the same individual are correlated. While autocorrelation is typically a problem in data sets with long time periods (time points per individual > 50), 3 we check that correlation between repeat observations is not significantly biasing our mobility estimates. We use a Generalized Estimating Equations (GEE) approach that explicitly models intra-individual correlations and corrects for biases stemming from possible autocorrelation or non-independence. 4 Beyond the individual level, another possible source of autocorrelation is due to the fact that our data are structured as repeated cross-sections at the state-level. To gauge the degree of this bias, we collapse our data set to the state-week level and run regressions with state-level mean estimates of the outcomes and covariates (N=612 for 51 states and twelve weeks). We run our main regression models in a panel-data setting with ordinary least squares, accounting for state fixed effects, autocorrelation, and heteroscedasticity.

Recall differences across outcomes eFigure 4. Association Between Changes in Mobility Restriction and Changes in Outcomes Using Generalized Estimating Equations to Account for Within-Person Correlation
Solid lines represent predicted estimates. Grey shaded areas represent 95% confidence intervals. Mobility restriction represents within-state reductions in mobility considering week-to-week changes in mobility from normal levels. All models control for income, race/ethnicity, age, sex, education, marital status, numbers of individuals in the household, week of survey, state-level heterogeneity, and COVID-19 death and case rates. Estimates are corrected for correlation of observations among the same individual using a Generalized Estimating Equations model with an exchangeable correlation structure. Standard error estimates are robust.

eTable 1. Definitions of Final Outcomes
Outcome Definition Self-reported causality Unemployment (universe = population over the age of 18 years not retired and willing to work) Respondent or anyone in household (not retired and willing to work) experienced a loss of employment since March 13, 2020 because of (1) sick with coronavirus symptoms or (2) caring for someone with coronavirus symptoms or (3) concerned about getting or spreading the coronavirus or (4) employer experienced a reduction in business (including furlough) due to coronavirus pandemic or (5) laid off due to coronavirus pandemic or (6) employment closed temporarily due to the coronavirus pandemic or (7) employment went out of business due to the coronavirus pandemic? Food insufficiency (universe = all households) In the last 7 days, household sometimes or often did not have enough to eat or household's food security became worse after March 13. Classes cancelled (universe = households with school-aged children) For children in the household, due to coronavirus pandemic classes normally taught in person at school were cancelled without any replacement with online/distance learning or other means. No medical care (universe = all households with a valid response) In the last 4 weeks, respondent needed medical care for something other than coronavirus but did not get it because of the pandemic.
Non-self-reported causality Default on rent or mortgage (universe = renters or household with a mortgage) Household did not pay last month's rent or mortgage on time.
Mental health problems (universe = all households) In the last 7 days, respondent felt nervous/anxious (score 1(never) -4 (everyday)), could not stop worrying (score 1(never) -4 (everyday)), had little pleasure in doing things (score 1(never) -4 (everyday)), felt depressed/hopeless (score 1(never) -4 (everyday)). Scores from the 4 dimensions were summed and the sum ranged from 4 (good mental health) to 16 (bad mental health)). Respondents with scores ≥12 were considered to be those with mental health problems. Low-inc is the abbreviation for low-income households (< $35,000 per year), and low-mid inc is the abbreviation for lowmiddle-income households ($35,000-$75,000 per year). According the ACS Selected Economic Characteristics Table, 28.1% of households earn less than $35000 and 29.5% earn $35,000-$75,000 per year. b White is the abbreviation for non-Hispanic White respondents, and Black is the abbreviation for African American respondents. The population estimates by race are obtained from the ACS Demographic and Housing Estimates Table. c Eligible population for unemployment is assumed to be the civilian labor force provided in ACS Selected Economic Characteristics The ACS Demographic and Housing Estimates Table provides estimates for the age groups 0-4.9, 5-9.9, 10-14.9, and 15-19.9. We estimate the population of ages 3-19 as: (0-4.9)*0.4 + (5-9.9) + (10-14.9) + (15-19.9). We assume that the populations within one-year age intervals in the category 0-4.9 are approximately equally distributed. 7,8