State-level variation of initial COVID-19 dynamics in the United States

During an epidemic, metrics such as R0, doubling time, and case fatality rates are important in understanding and predicting the course of an epidemic. However, if collected over country or regional scales, these metrics hide important smaller-scale, local dynamics. We examine how commonly used epidemiological metrics differ for each individual state within the United States during the initial COVID-19 outbreak. We found that the detected case number and trajectory of early detected cases differ considerably between states. We then test for correlations with testing protocols, interventions and population characteristics. We find that epidemic dynamics were most strongly associated with non-pharmaceutical government actions during the early phase of the epidemic. In particular, early social distancing restrictions, particularly on restaurant operations, was correlated with increased doubling times. Interestingly, we also found that states with little tolerance for deviance from enforced rules saw faster early epidemic growth. Together with other correlates such as population density, our results highlight the different factors involved in the heterogeneity in the early spread of COVID-19 throughout the United States. Although individual states are clearly not independent, they can serve as small, natural experiments in how different demographic patterns and government responses can impact the course of an epidemic.


Major points
While I understand the need for stepwise regression and concerns about correlation. It is unclear how the authors selected variables when there were highly correlated variables. If two variables are highly correlated, it would be likely (although not absolutely true) that they would either both meet stepwise inclusion or neither We have now added text on line 114 to clarify this point. The only highly correlated variables were the demographic variables. For example, percent of population in rural areas was very strongly correlated with percent of population in urban areas. We ended up just choosing one variable if there were several correlated. There was no strong reason behind each choice. However, the variables that were strongly correlated had similar interpretations, as in the example above about population percentage.
Please clearly list all candidate predictors (state variables as well as policies) that were considered as part of the stepwise regression These are now listed in the Table 1 caption, and the list starting on line 120.
Please provide more context for the inclusion of the volunteer rate in the methods. This inclusion seems interesting but the reader doesn't have context for this variable. Similarly, the tightness score (this comes through a little more in the discussion, but explaining in methods would be helpful) We have now added additional context on lines 63-70 and 132. In short, we hypothesized that volunteerism could be a measure of how willing someone would be to accept restrictive policies that might be beneficial to the community overall.
Please give more information regarding the gathering limitations. You say "usually to 10 people" but the reality is that almost all states started at 250 or 100, then decreased to 50, then 10, some went to smaller numbers too. So what date are you using? The first date or the date the states went to 10 people?
We address this point on line 150: "We follow Adolph et al. and use the first date of gathering restrictions announced, regardless of the size of the gatherings restricted." The references (and at times the text) need to updated. I believe publishing on the beginning of the pandemic is useful but putting these findings into context of other publications around the same time period is useful. A prior reviewer mentioned the Courtemanche manuscript, there is also the Hsiang manuscript in Nature. This submitted manuscript confirms many of the findings of Auger re schools in JAMA. Thus, I think the concerns raised by prior reviewers that the model is not correctly specified are not founded.
We have now added several references and updated text (see lines 21, 70-78, 213) to address this. In terms of updating the manuscript, the second sentence discusses the case counts as of April 29th. The reader will be confused by this statement given how different we are currently. I see why you are presenting this data as this is at the end of the study period but you need to first introduce the study period and then give context of the numbers.
This has been corrected on line 21.
Similarly, you need to at least acknowledge the latest phase of the pandemic (lines 40-45) there is another rise in cases/deaths after reopening. I understand that you are not focused on that time period but it is strange to not acknowledge it. In fact you state the numbers have "slowed since mid-March" (line 50) which is simply not true.
We certainly agree. Some of the numbers and text were out of date as we originally submitted this manuscript back in May. We have now corrected this on 81. This is an important point that we now clarify in the table caption. These variables were such poor predictors that they were not selected by the final models after a model selection process.

Results-
Please put in context the parameter estimates from It is kind of a big deal that testing wasn't accounted and is only listed as a limitation. The statement "overall doubling time was not strongly correlated with overall tests rates for each state" line 135 is woefully underspecified. (the text says figure 4, but It appears to refer to figure 5 We now include testing rates as a covariate in our models (Table 1). It was never chosen in the best models after model selection. We have now modified the figure to just read testing rate, which is the total number of tests conducted at 7 days or three weeks, depending on the model. We modified the text on lines 182 and 246 to address this.
Figures 3 and 4 would be improved with some sort of correlation coefficient in each panel. It is unclear if all of the panels in figure 4 were considered as candidate covariates. Why aren't business restriction, stay at home mandates, and number of state actions in the final model? Were they not retained in stepwise? Were they too highly correlated?
We have now included a correlation coefficient on each of these figures. We have also now only plotted the covariates that were included in the models described in table 1. The reviewer is correct that several variables were not selected in the best model for the stepwise model selection process. We expand on this in the caption of table 1.

Specific points
Abstract The abstract states "we examine how commonly used epidemiological metrics differ for each individual state within the US during the initial COVID-19 outbreak." Yet, the manuscript does not report out individual state level results so this sentence should be revised.
We are not sure what the reviewer meant with this comment. All of our analyses highlights the individual state level differences in doubling time and other variables. It is true that we did not look at the dynamics within each state.
Methods "The early doubling time should not be severely affected by government interventions s these were rarely implemented that early in the course of the epidemic" (line 65-66) I think the sentiment of this statement is true. . . but it is likely more that the effects of the policies wouldn't take place that fast. Several states closed schools and implemented gathering bans when they had very very few cases (e.g. West Virginia closed schools with 0 cases). Given that it would take 10+ days for policy change to impact coronavirus spread (see lag in model derived by Auger, JAMA), I believe this is a reasonable approach; however, you may want to tweak the language because it isn't that the policies weren't implemented it is more they wouldn't have had time to take effect.
We modified this sentence on line 102.
Results "We found that doubling times for all states increased with time and that heterogeneity between states was reduced" line 109-110. This seems very subjective. It is strange not to quantify this in some way. How much did they change?
On line 155, we clarify this comparison using Figures 2a and 2b.
Discussion The first paragraph of the discussion is a bit long and feels tangential to the main point of the manuscript. Perhaps this could be largely reduced and additional references regarding emerging COVID studies (listed above) could be incorporated. There are some nice additions (or considerations) in this manuscript which haven't been in other models (tightness, volunteer rate, etc). So highlighting these is appropriate (as you do in lines 170-180).
Starting on line 185 we have simplified the text.
Supplemental figure 3 This one is a bit confusing to me. You talk about the changes in the doubling time throughout the study period; yet, this is just a single number per state... so when is this from? Early? The average of the study period?
In the figure caption, we now clarify that this map was the doubling time for the first three weeks since 25 cases in each state.

Reviewer 1:
The reviewed paper presents an interesting analysis on how state-level variation can explain the early pandemic dynamics of COVID-19. The analysis appears to be logical and statistically sound. I don't have any major comments, but have a few minor comments that I think could enhance the manuscript.
Thank you for the feedback.

Minor comments
The authors note: "We show that early non-pharmaceutical government actions were the most important determinant of epidemic dynamics." I think this statement can be interpreted in two ways: (1) government actions need to happen early to impact the epidemic dynamics, or (2) the government actions that happened early had the largest impact on subsequent epidemic dynamics. These have two different interpretations in my mind, that the specific timing in the early phase matters or that implementing them early at all matters. I believe the authors mean the latter, but I think it would be useful to clarify, maybe something like this?: "We show that non-pharmaceutical government actions during the early phase of the epidemic were the most important determinant of epidemic dynamics." One could reference previous research from China that agrees with this conclusion as well (I know it likely wasn't published when this was written): https: // wwwnc. cdc. gov/ eid/ article/ 26/ 9/ 20-1932_ article On line 9, we have made this change with the suggested text and added the citation to the discussion.
Figure 5: The authors note: "The overall doubling time was not strongly correlated with overall tests rates for each state (Fig. 4)." Did the authors compare with testing rates as a raw number rather than per capita? I wonder if there would be a stronger correlation, as the early pandemic dynamics with small numbers might not have any per capita impact.
We now include testing rates as a covariate in our models (Table 1). It was never chosen in the best models after model selection. We also examined the raw number instead of per capita and this was less strongly correlated.
In general, I found it unintuitive to follow the doubling time results. E.g. on line 120: "We found that population density, flu vaccination rates, and wealth were all positively correlated with doubling times" I would suggest rephrasing for clarity for these types of results throughout to something like: "Comparing the doubling times across states, we found that increasing population density, flu vaccination rates, and wealth were all associated with slower epidemics" -I'll leave it to the authors to decide if they agree We have now corrected this on line 163.
Line 120 indicates that all three variables are positively correlated with doubling times, but the results in Table 1 have different signs for population density from the other two variables, so I think there might be a typo?
Same as above.
Line 173, I'd add that tightness was negatively correlated with doubling time.
This has been corrected on line 181.
Typo in line 180: "We hypothesis this may be the result of people in tight cultures finding it more difficult to adjust their behavior when new rules are imposed."