Comparing Associations of State Reopening Strategies with COVID-19 Burden

Authored by link.springer.com and submitted by mvea

We used an interrupted time series (ITS) to compare the rate of growth in COVID-19 cases and deaths after reopening to growth prior to the reopening. Because all states reopened, each state serves as its own control.12 To estimate what the case rate would have been if the state had not reopened, we assume the state-specific trend in cases prior to reopening would have continued to be the same if the state had not reopened. State fixed effects controlled for state characteristics associated with outcomes that did not change over the study period. Our sample included 50 US states and the District of Columbia. Daily COVID-19 confirmed case counts were obtained from the New York Times database from January 21 through July 16, 2020.11 State population data were obtained from the 2018 American Community Survey.13 State reopening policy data was linked from the Boston University COVID-19 US State Policy database, with four corrections (see supplemental appendix).13

Most states used a phased reopening strategy, and there was substantial variation in the implementation of reopening approaches across states. Thus, it was important to define a meaningful and specific level of reopening that would be expected to influence COVID-19 outcomes as the date of reopening in this study. We defined reopening as the date of reopening indoor dining rooms at any capacity, as most states initially opened indoor dining at limited capacity.11 Reopening indoor dining represented an earlier milestone among states’ various reopening policies, allowing for more accurate depiction of the effect of reopening, but was also associated with reasonable risk for COVID spread if implemented too early or poorly (i.e., restaurants did not follow guidelines for sanitation, table-spacing, or mask-wearing). If a state did not explicitly detail dining restrictions in their reopening policies (assessed using two sources), the expiration of shelter in place was used as the date of reopening.2, 11

We estimated excess COVID-19 burden for states reopening indoor dining after implementing a statewide mask mandate, also referred to as evidence-based reopening. We also estimated excess COVID-19 burden for states reopening prior to implementing a statewide mask mandate. The date of mask mandate was defined as a statewide order requiring masks for all adults in public spaces where social distancing is not maintained (with medical exceptions). For the 21 states that implemented a mask mandate after opening dining rooms, there was an average of 56 days between the date of reopening dining rooms and the date of the mask mandate. Thus, our 56-day follow-up period and estimated effects primarily occur prior to subsequent mask mandates.

The primary outcome was the daily cumulative number of COVID-19 cases in the state. The outcome was lagged by 5 days because the effect of the intervention on cases was expected to be delayed by the incubation period as well as reporting time. A shorter lag time was preferred to avoid classifying daily outcomes from the post-period in the pre-period, and prior evidence suggested that effects are significant as early as 5 days after implementation.6 The secondary outcome was cumulative COVID-19 deaths, lagged by 10 days.

We predicted cumulative daily cases and deaths using generalized linear models with a negative binomial distribution and a log link. For all models, we used an offset for the log of the total population in each state to account for varying population sizes and present marginal effects as a rate per 100,000 residents. The time variable was centered around the date of reopening for each state. The period was truncated at 40 days prior to reopening to improve estimates of state trends prior to reopening and reduce bias due to periods of zero growth early in the epidemic. Testing capacity prior to reopening is controlled for using the state-specific trends in observed cases prior to the intervention. While random variation in testing capacity across states over time is expected, this fluctuation would not bias our estimates unless testing capacity were systematically different after reopening compared to the baseline period. We used robust standard errors to account for heteroscedasticity in measurement of cases over time and clustered standard errors by day to address temporal clustering, such as from changes in testing capacity over time.

To improve model fit, we tested higher order terms and alternative distribution assumptions. After selecting the best-fitting model using the Akaike information criterion, models estimated time trends for each state by interacting the state indicator with a linear, quadratic, and cubic time trend (see statistical appendix). The effects of reopening with and without a mask mandate (compared to the pre-trends within each group) were estimated using an interaction of the indicator for the post-period, treatment group (pre-opening, reopening with masks, reopening without masks), and the time trend. The quadratic time trend was interacted with the post-period indicator.

State-specific estimates for excess cases and deaths were generated using the method of recycled predictions. For each state, we compared the predicted outcomes for that state under two scenarios (opening with masks and opening without masks) using the original model estimates. The total treatment effect includes the linear and quadratic changes in trends. To generate the predictions, we applied the state’s specific fixed effect and baseline trends as estimated in the original model (different for each state) and the estimated average treatment effect (the estimated change in trends after reopening compared to before). We conducted multiple sensitivity analyses to test the robustness of our results, including different lag times for the outcome. We ran models with and without New York and New Jersey included because of the large case counts that skew average marginal effects. The conclusions were robust to all robustness checks (statistical appendix). All analyses were conducted using Stata 16.0.

RaxZergling on October 9th, 2020 at 00:50 UTC »

Can someone help me understand Table 1?

In 13 states that reopened with a mask mandate there were ~1100 cases per 100k and 72 deaths.

In 38 states that reopened without a mask mandate there were 114 cases per 100k and 7 deaths.

Isn't this almost the exact opposite of what the title (and conclusion of the study) suggests? Did they just label their columns wrong?

BlackEyeRed on October 9th, 2020 at 00:24 UTC »

I have a Question about masks, I’m a firm believer in science so I support masks.

In Quebec (8.5 million) we have been wearing masks since July. We were averaging less than 100 cases a day back then. Now we’re averaging about 1000. I constantly hear people say what’s the point of masks and I never know how to answer. Is hospitalization and death rate a better metric for judging the severity of the situation?

insightfill on October 8th, 2020 at 23:15 UTC »

And this doesn't even take compliance into account. My state is "masks" but you wouldn't know it.