Eviction Moratoria Expiration and COVID-19 Infection Risk Across Strata of Health and Socioeconomic Status in the United States

Key Points Question Is lifting a state-level eviction moratorium associated with the risk of individuals in that state being diagnosed with COVID-19? Findings In this cohort study of 509 694 individuals living in the United States, a difference-in-differences survival analysis found that residents in states that lifted eviction moratoria had an increased risk of receiving a COVID-19 diagnosis 12 weeks after the moratorium was lifted relative to residents in states where moratoria remained in place. These associations increased over time, particularly among individuals with more comorbidities and lower socioeconomic status. Meaning These findings suggest that eviction-led housing insecurity may have exacerbated the COVID-19 pandemic.

This supplemental material has been provided by the authors to give readers additional information about their work. eFigure 1 Trends in Share of Individuals Exposed to Treatment by the Week of the Year Percentage Notes: Trends in the percentage of individuals in the study sample exposed to treatment. Estimates come from the balanced sample. The y-axis shows the cumulative percentage of individuals in the study sample who lived in a state that lifted their eviction moratorium during the study period. At the beginning of the study in week 11 of 2020, no individuals were exposed to a state lifting their eviction moratorium. By the end of the study period, week 35 of 2020, 244,335 individuals lived in a state that lifted their eviction moratorium.

eFigure 2. Survival Curves on the Association Between Lifting the Eviction Moratorium and COVID-19
Survival Probability Scenario\Week 11

eFigure 4. Event Study Estimates of the Association Between Lifting the Eviction Moratorium and COVID-19, by Sample Design Hazard Ratios
Note: Y-axis is in log scale. The vertical dashed line represents the end of the eviction moratorium. Xaxis represents the number of weeks relative to the end of the eviction moratorium. Hazard Ratios and 95% Confidence Interval from models to those estimated in Equation 1/Supplement. We used, in separate models, both the balanced sample (in blue) and the 2% random sample (in red) to conduct this analysis.

eFigure 7. Event Study Estimates of the Association Between Lifting the Eviction Moratorium and Changing Zip Code Address Hazard Ratios
Note: Estimates come from the balanced sample. Y-axis is in log scale. The vertical dashed line represents the end of the eviction moratorium. X -axis represents the number of weeks relative to the end of the eviction moratorium. Hazard Ratios and 95% Confidence Interval from models to those estimated in Equation 1/Supplement except here the main outcome is a binary variable indicating whether the individual requested a change of residential zip code.

eFigure 8.Event Study Estimates of the Association Between Lifting the Eviction Moratorium and COVID-19, With and Without Z Codes Hazard Ratios
Note: Y-axis is in log scale. The vertical dashed line represents the end of the eviction moratorium. Xaxis represents the number of weeks relative to the end of the eviction moratorium. Hazard Ratios and 95% Confidence Interval from models to those estimated in Equation 1/Supplement. We used, in separate models of our main specification, both with z-codes (in blue) and without z-codes (in red) to conduct this analysis.

eMethods 1. Event-Time Study Model Specification
To study the association between lifting the eviction moratorium on the hazard of being diagnosed with COVID-19 in a given week, we use a Cox regression model with timedependent covariates in an event-time type specification. 11,12 This approach models the weekly probability of being diagnosed with COVID-19 at a given period conditional on having been observed without a positive diagnosis previously, where the treatment is defined as lifting the eviction moratorium, and treated individuals are compared to individuals living in states that had not yet lifted their moratoria.
This study uses the time from when individuals enter the study until either a COVID-19 diagnosis or time censoring at the end of the study period, just like in a classic Cox analysis. Unlike a standard Cox model, however, we also make use of information on time since the treatment occurred (i.e., lifting an eviction moratorium) for the treated. This methodology allows us to understand whether the association between expiring eviction moratoria and a COVID-19 diagnosis changes over time, which is useful when studying events that develop exponentially, such as epidemics, while also relaxing the proportional hazards assumption.
Specifically, we fitted the following model: The dependent variable | | ( ) denotes the probability that an individual, , living in state, , during week is diagnosed with COVID-19.
is a binary variable for the treatment group, i.e., those states that implemented an eviction moratorium but lifted it.
is a binary variable that equals 1 for those treated states during week, , relative to the week when the state lifted their moratorium. The exposure variable is bottom coded before week 15 and top coded after week 12, implying that dynamics wear off after these points. This decision follows prior literature 11,21 to avoid difficulties when interpreting results due to sample size imbalances created by differences in the timing of lifting the moratoriums.
is a vector of time-varying covariates (i.e., non-pharmaceutical interventions, and COVID-19 cases and tests), while is a vector of time-invariant covariates (i.e., sex, age, type of insurance, work-industry, CCI, and z-codes diagnoses).
are state fixed effects that adjust for potential confounding from time-invariant state-level factors or baseline differences in socioeconomic characteristics, while are weekly fixed effects that adjust for nationwide secular trends in the outcome. is a vector of residuals. Standard errors are clustered at the state and week-level. We used the Breslow method for ties.
The coefficients of interest are captured by, , showing the difference in outcomes for leads and lags of lifting the eviction moratoria relative to a reference week (i.e., the week a state lifted their moratorium) and relative to all states that did not lift their eviction moratorium during the reference period.
The causal identifying assumption is that COVID-19 diagnosis risk in exposed states would have continued along the same trajectories in the absence of exposure. 11 We cannot directly eMethods 2. Survival Curves To investigate temporal trends between the association of lifting an eviction moratorium and COVID-19, we fit survival curves to the data, estimating the hazard of being diagnosed with COVID-19 at time, , for total times, T. 15  Where, , , is a flexible time-varying function , = + + .
As in the previous analysis, is a binary variable for the treatment group, i.e., those states that implemented an eviction moratorium but lifted it.
is a binary variable that equals 1 for those treated states during week, , relative to the week when the state lifted their moratorium. The exposure variable is bottom coded before week 15 and top coded after week 12, implying that dynamics wear off after these points. This decision follows prior literature 11,21 to avoid difficulties when interpreting results due to sample size imbalances created by differences in the timing of lifting the moratoriums.
is a vector of time-varying covariates (i.e., nonpharmaceutical interventions, and COVID-19 cases and tests), while is a vector of timeinvariant covariates (i.e., sex, age, type of insurance, work-industry, CCI, and z-codes diagnoses).
are state fixed effects that adjust for potential confounding from time-invariant state-level factors or baseline differences in socioeconomic characteristics, while are weekly fixed effects that adjust for nationwide secular trends in the outcome.
is a vector of residuals. Standard errors are clustered at the state and week-level.
We then computed estimates of the survival Pr [ = 0 | • = • , = , = , = ] by multiplying the estimates of one minus the estimates of Pr [ = 1 | = 0, • = • , = , = , = ] provided by the logistic model for each individual-week. We computed two opposing counterfactual scenarios: one where every state that implemented an eviction moratorium maintained it throughout the study period, and another where every state that implemented an eviction moratorium lifted it on week 17. We chose week 17 because it was the first week a state lifted its eviction moratorium (Table 1).