Determinants of Gubernatorial Response to COVID-19
By Carl E. Klarner
We examine the determinants of nine gubernatorial actions taken to fight the COVID-19 pandemic in an event history analysis of February 27th to March 26th, 2020.
The media narrative has been that Republican governors have been slow to respond. But this notion ignores the fact that states first hit by the disease are much more likely to have Democratic governors. In the time period examined, the average number of cases per 100,000 has been 3.00 in states with Democratic governors and 1.19 in states with Republican governors.
This analysis examines whether Republican governors have been slower to respond to the COVID-19 epidemic after taking the seriousness of the crisis in their states into account.
This report also provides a tool for advocates, by identifying states that have not yet adopted provisions but are most likely to do so.
All of the data, replication code and output from the analysis has been archived on Harvard’s Dataverse under my name.
When the seriousness of COVID-19 in each state is taken into account, Republican governors have been slower to enact stay-at-home orders than Democratic governors. The model indicates that on any recent day, there is a 1.4% chance a Republican governor will mandate a stay-at-home order, while there is a 5.1% chance a Democratic governor will do so.
Democratic governors have also been more likely to institute bans on gatherings of 25 people or more than Republican governors. For the recent past, the chance of a Republican governor banning such gatherings is 7.6% a day, while the chance of a Democrat doing so is 14.9%.
Both of the above findings are statistically significant, although the latter only achieves conventional levels of statistical significance in a 1-tailed test (p<.029, 2-tailed / p<.081, 2-tailed).
For the other seven policies examined, there were no statistically significant differences between how Democratic and Republican governors acted. This is true for policies other than social distancing policies where greater partisan differences might be expected, such as moratoriums on evictions, provisions to expand access to child care, and expanding access to unemployment insurance for the entire labor force.
An analysis of the role of race and ethnicity in the response to COVID-19 indicated that state responses have been fairly similar regardless of percentages of African-Americans and Hispanics for most of the nine policies. A noteworthy exception is that governors have been more likely to call out the national guard in states with larger percentages of African-Americans (p<.001). Although the evidence from this analysis is only suggestive, this finding is consistent with work that has found stricter law enforcement policies in states with larger percentages of African-Americans. On the other hand, there were several instances of national guard mobilization that were focused on provision of assistance to medical facilities.
One other exception is that governors in states with larger percentages of Hispanics were more likely to declare states of emergency (also statistically significant), but again, these results are merely suggestive, and probably have more to do with region of the country.
We read through a list of 1,494 actions taken by state governors on the National Governor’s Association Website and coded them as fitting into one of the nine dichotomous categories reported here. https://www.nga.org/coronavirus/#states. The following tables report the date a particular action was taken, and the predicted probability that such an action will be passed if it hasn’t already.
State of Emergency
Some states have different categories of states of emergency, and what a state of emergency means varies from state to state. If the governor declared any type of emergency, we coded the governor as taking this action.
If the governor mobilized the national guard in any capacity, the governor was coded as taking this action.
Merely mandating that all non-essential businesses close (such as in New York) isn’t included in the definition of this type of action.
Gatherings of 25 or More Prohibited
Almost all of these consisted of prohibitions on 10 or more people. A stay at home order without a prohibition on gatherings was also counted as this type of action.
Restaurant and Bar Closures
This means that dine-in services have been prohibited. This is generally a proxy for all non-essential businesses being closed. A stay at home order without a prohibition against restaurants and bars was also counted as this type of action.
In some states, the governor doesn’t have the authority to close public schools. We didn’t have time to figure out which states grant the governor this power. For this reason, we coded a governor as having closed schools if they either explicitly made the order, or advised it. When another entity closed schools—such as local school districts statewide—and no mention is made of the governor, the state is no longer “at risk” for gubernatorial action on closing schools and is taken out of the analysis, but isn’t coded as a gubernatorial action either. We considered March 18, 2020 to be the date California schools closed.
Whether provisions have been put in place to broaden access to Unemployment Insurance benefits. Such provisions have to involve the entire labor force, not merely those who are sick, to count as this type of action.
Whether special provisions to expand child care to first responders have been made. This also includes general provisions to expand access to child care.
Whether a moratorium on evictions from houses and apartments has been put in place. Again, if another entity acts—such as the courts—cases for that state are then removed from the dataset after that date, and the governor isn’t coded as having taken the action.
FACTORS INFLUENCING COVID-19 POLICIES
Each model predicting one of the nine policies included the following eleven independent variables.
Democratic Governor: Coded “1” if the governor is a Democrat, “0” otherwise.
Income Growth Per Capita %: States may be hesitant to close down businesses the worse their economies are compared to other states. Unfortunately, this measure comes from the third quarter of 2019, before the current economic slowdown.
COVID-19 Cases / 100,000 Population: Measured in the prior day. Daily data on positive cases was obtained from the COVID-19 Tracking Project on March 29, 2020 at 2:22 PM Eastern.
Neighbor States: COVID-19 cases per 100,000 population in the neighboring state with the highest infection rate, measured in the prior day.
Age 65 and Over %: More vulnerable populations may cause more timely responses.
Population Density: If more of a state’s population is in close proximity with each other, there may be more concern about the crisis.
Hospital Beds per 1,000 population: More concern may exist if hospital capacity is lower.
Time Counter: Starts at “0” on February 26, 2020, and goes up “1” each day. As the crisis has unfolded, states all over the country have become more apt to respond, and public pressure to respond has plausibility increased.
Time Counter Squared: Allows the impact of time to slow down or speed up.
Weekend Variable: Coded “1” if a day was on a weekend, “0” otherwise. Actions of any type are about half as frequent on weekends versus weekdays.
African-American %: Percentage of state population African-American.
Hispanic %: Percentage of state population Hispanic.
APPENDIX A: DATA DETAILS & SOURCES
Our analysis begins on February 27, 2020 because an examination of the dates of policy adoptions make that appear to be a natural cutoff.
Another possible way of measuring gubernatorial response would be to count the number of actions that governors have taken, but this would be an unwise strategy. A reading of the 1,494 actions makes it obvious that the level of reporting in the source we used varies from state to state given the level of detail reported in some but not others. For example, Connecticut reports around 100 actions, while the states with the second and third most actions have around 50 each. Furthermore, disparate actions are often presented as one action.
Similar policy actions in different states are often different from each other in their details. We made quick judgments about which policies warranted a code of "1," but interested parties can look at the entire list of state actions we've posted on Dataverse and now we coded them.
Judgments also had to be made about when a state that had incrementally adopted a response was considered to have entirely made that response. We generally used the last date of such incremental adoptions as the date of adoption, although if we judged later adoptions were small enough, we might use the second to last date, etc.
We compared whether a state had adopted a measure by March 25, 2020 with Politico's report on Coronavirus policy responses as a check on our work, as well as other lists. Discrepancies between our lists and others often hinged on differences of definition.
It’s easy to conceive of ways the independent variables might interact with each other. Because of concerns about multicollinearity, we only examined one: the interaction between cases per 100,000 population and governor’s party. Multicollinearity was a major problem for the interaction just mentioned in many of the models (see output posted on Dataverse), but not for any other independent variables (except “time” and “time squared” as expected).
Other Data Sources: Citations to the data sources for the other independent variables used in the analysis can be found in the Dataverse post associated with this analysis.
Population Density: was measured with a Herfindahl index of Census Combined Statistical Area-state intersections, with “0” input for all population living outside of metropolitan areas. This means that if everyone in a state lives in the same metropolitan area—such as in Rhode Island—the state will receive a score of “1.” If a state has no people living in what the Census defines as a metropolitan area—such as in Alaska, Montana, and Wyoming—it receives a score of “0.”
Neighbor States: This is an approximate way of tracking officials’ concern about infections spreading from other states. It ignores the fact that some neighboring states have a great deal more interaction than others. For example, New Jersey and New York are closely intwined, while Ohio and Pennsylvania aren’t.
The State Legislature: It’s possible that some governors didn’t engage in particular actions because the state legislature had already taken that action. For the most part, we did not code actions taken by the state legislatures, although an analysis that does so would be an improvement over the current analysis.
Age 65 Years and Over as % of Voters: Older individuals are more likely to vote than younger voters generally, but are especially more likely to do so in some states. This variable was removed due to multicollinearity.
Two sets of models are run. One set has all of the listed independent variables. The second excludes the party of the governor, state economy, and hospital beds variables. This is done because race and ethnicity are plausibly causally prior to these variables. The race/ethnicity variables perform much the same in each of the two types of models, but the effect sizes are more accurately assessed in the second type of models.
The finding about stay-at-home orders is robust to different model specifications, although the ban on gatherings becomes weakly statistically significant in other models.
The reported probabilities are for an imaginary state where all the other independent variables are held at their means, on march 26, 2020.