Supply, then demand? Health expenditure, political leanings, cost obstacles to care, and vaccine hesitancy predict state-level COVID-19 vaccination rates.

OBJECTIVES: To examine predictors of state-level COVID-19 vaccination rates during the first nine months of 2021.
METHODS: Using publicly available data, we employ a robust, iteratively re-weighted least squares multivariable regression with state characteristics as the independent variables and vaccinations per capita as the outcome. We run this regression for each day between February 1 and September 21, the last day before vaccine booster rollout.
RESULTS: We identify associations between vaccination rates and several state characteristics, including health expenditure, vaccine hesitancy, cost obstacles to care, Democratic voting, and elderly population share. We show that the determinants of vaccination rates have evolved: while supply-side factors were most clearly associated with early vaccination uptake, demand-side factors have become increasingly salient over time. We find that our results are generally robust to a range of alternative specifications.
CONCLUSIONS: Both supply and demand-side factors relate to vaccination coverage and the determinants of success have changed over time. POLICY IMPLICATIONS: Investing in health capacity may improve early vaccine distribution and administration, while overcoming vaccine hesitancy and cost obstacles to care may be crucial for later immunisation campaign stages.

Introduction

As vaccination efforts against COVID-19 unfolded across the globe, it became increasingly apparent that developing and authorising effective vaccines is only half the battle of achieving broad immunisation [1]. To a large extent, between-country differences can be explained by the disparate access to vaccine doses, as a small number of countries successfully secured agreements with manufacturers for large amounts of vaccines [2]. Still, diverging local trajectories within larger countries suggest that national supplies are just one component of a successful immunisation campaign. In the United States (US), certain states decidedly outpaced their neighbours in terms of per capita vaccinations despite proportional vaccine allocations.Fig 1. Fig 2. Fig 3. Fig 4. Fig 5. Fig 6. Fig 7. Fig 8. Fig 9. Fig 10. Fig 11. Fig 12. Fig 13. Fig 14. Fig 15. Fig 16. Fig 17. Fig 18. Fig 19. Fig 20. Fig 21. Fig 22. Fig 23. Fig 24.

Figure 1

Variables in the main model

Figure 2

Spatial distribution of select state characteristics

Figure 3

Cumulative vaccinations per capita over time, 2021

Figure 4

Correlates of cumulative vaccinations per capita

Figure 5

Rolling regression coefficients of variables, February 1 to September 21

Figure 6

The vaccine delivery process

Figure 7

Data sources for variables in the main model

Figure 8

Excluded variables

Figure 9

Breusch-Pagan/Cook-Weisberg test for main regressions

Figure 10

Quantile-quantile plots

Figure 11

Variance inflation factors test for explanatory variables

Figure 12

Rolling regression coefficients of variables, February 1 to September 21

Figure 13

New monthly vaccinations per capita, February, June, and September

Figure 14

Vaccine distribution and administration, February 10, 2021.

Figure 15

Correlates of cumulative vaccinations per capita

Figure 16

Robustness of September 21 associations to unobservable selection

Figure 17

Robustness of June 1 associations to unobservable selection

Figure 18

Robustness of February 1 associations to unobservable selection

Figure 19

Data sources for additional variables used in robustness regressions

Figure 20

Correlates of cumulative vaccinations per capita with new controls

Figure 21

Comparison of rolling regression with/out new controls

Figure 22

Comparison of rolling regression with/out new controls (continued)

Figure 23

Regression robustness to excluding individual states, September 21

Figure 24

Regression robustness to excluding individual states, June 1Exhibit 22: Regression robustness to excluding individual states, April 1

While numerous studies have examined individual-level correlates of vaccine hesitancy [3], [4], very few ecological studies have analysed determinants of COVID-19 vaccination rates in the United States. Notable exceptions include studies by Brown et al. [5] and Stewart et al. [6], who found that US county-level vaccination rates as of May 2021 were associated with health system capacity and population density, as well as Lindemer et al. [7] who found that county uninsurance rates predicted vaccination coverage in early March 2021. However, there is a gap in the literature concerning analysis at the state level. Because state governments influence some of the most important policy levers under the American federal system, identifying state-level patterns may inform current and future vaccination campaigns.

Appendix A includes an overview of the delivery process from the point a vaccine leaves the manufacturer to the point it is administered to a patient [8]. Since popula- tion immunity is ultimately the outcome of interest for vaccination campaigns, we focus on investigating the relationship between state-level vaccine administration rates and salient state characteristics. As the public health adage goes: “vaccines do not save lives, vaccinations do”

Data and methods

Data

We utilise public state-level data on state characteristics and vaccination rates the first nine months of 2021; see appendix B for a full list of sources. The most notable drawback of our data is that data on health care expenditure are not available after 2014, posing a limitation to our analysis of that variable insofar as relative expenditure levels between states may have changed in the seven years since then. All other data used are from 2019 or later. We restrict the analysis to the 50 states given data limitations. However, in light of the historical disenfranchisement of non-state territories and entailing health disparities [8], further research would do well to shed light on vaccination outcomes in non-state US territories.

Our primary outcome is the number of vaccinations per capita (with the total state

population as the denominator), without distinguishing between first and second doses. If, for example, a state has administered 1.0 vaccinations per capita, this could correspond to every resident having received one dose, half of the residents receiving two doses, or some other combination. While this per capita measure of vaccination uptake does not perfectly correspond to the eligible adult population at any point in time, it can more readily be compared to other measures for that are similarly measured per capita. To highlight changes over time, we selected three points to analyse cross-sectional associa- tions: February 1, June 1, and September 21; the latter date was chosen to restrict the analysis to the initial campaign for first and second doses, as the use of booster doses was authorised on September 22.

The selection of explanatory variables is based on a priori hypotheses as well as on findings in the existing literature [5], [6], [9], [3]. The variables included in the main model are a subset of all the variables we considered. As a rule, we include a variable in the main model if it meets one of the following three conditions: (1) it is independently associated with vaccination rates, (2) it appears to considerably confound the magnitude of the association between another variable and administration rates, as measured by how much its inclusion moves other coefficient estimates, or (3) it seems a priori to be important, such that even a null-result would be of interest. Exhibit 1 describes the variables included in the main model, while appendix B features an overview of all variables that were considered but not included in the final regression analysis.

We include healthcare expenditure (e.g., hospital care, physician services, and drugs but excluding public health) and public health expenditure (e.g., preventive programmes and information campaigns) as proxies of state health system capacity and infrastructure quality, which we hypothesise are important for facilitating the supply of vaccines [5], [6]. For demand-side factors, factors, we include the measures of COVID-19 vaccine hesitancy and whether adults have not seen a physician due to cost obstacles. We include the share of the population in poverty to ensure that the cost obstacles variable captures a phenomenon relating to healthcare costs, specifically, rather than to poverty more broadly. Finally, we include population density [5], elderly population share [9], and political leanings as measured by the Democratic vote margin in the 2020 elections in net percentage points of the vote [3]. Exhibit 2 visualises the spatial distribution of the

included state characteristics. Note that the scales and units of measurement differ across the different maps and each one should be carefully interpreted in its own context.

Statistical analysis

In this ecological study, we use a robust, iteratively re-weighted least squares regression [10], where the dependent variable is vaccinations per capita for a state at a given point in time and the explanatory variables include state characteristics. Our main model does not include a measure of how many vaccines have been distributed to a state at a given point in time, as we consider vaccine distribution to be on the causal pathway to vaccine administration. In a sensitivity analysis, we further include a variable for vaccines distributed per capita at a given point in time.

To mitigate reverse causality, we consider only baseline state characteristics before the beginning of the immunisation campaign; for example, we utilise vaccine hesitancy as of January 2021. (It is worth noting that vaccine hesitancy is a dynamic phenomenon that may evolve over the course of an outbreak’s epidemic and endemic phases, and

consequently should be interpreted in the specific context of when it was measured and analysed.) To demonstrate changing associations over time, we present results from rolling linear regressions in which the cumulative vaccinations per capita are measured at different points in time. Additionally, we examine the role of omitted variable bias [11]; and test robustness to dropping states one at a time as well as to an alternative definition of the outcome in the form of monthly, rather than cumulative, vaccination. See appendix C for the full model specification as well as model diagnostic tests focusing on heteroscedasticity, the distribution of residuals, and multicollinearity.

Limitations

Findings from any observational study should be interpreted with caution, particularly with respect to causal inference. While our analysis based on Oster [11] suggests that selection on unobservables would have to be very severe to nullify most of our headline results, endogeneity remains an important limitation for our study. Moreover, given our focus on the state level, readers should avoid ecological fallacies – i.e., making inferences about individuals based on data aggregated data for a group – when interpreting the observed associations [12].

Our study did not account for state-specific policies influencing vaccine uptake. States differed considerably in how they prioritized vaccinations for certain populations, in their timelines for expanding vaccine eligibility, and in their incentive and information schemes to increase vaccine uptake [13], [14], [15], and these differences may have substantially influenced vaccination trajectories across states. However, our multivariable regression method is not well suited for analysing the effects of these policies due to being highly susceptible to endogeneity arising from bidirectional causality. For example, a broad eligibility policy may increase vaccination coverage but there a causal effect in the opposite direction is equally plausible, thus complicating the interpretation of any association between eligibility policies and vaccinations per capita. In theory, this issue can be mitigated by carefully linking policy changes to subsequent vaccination rates, for example by using a difference-in-difference framework. While we did not conduct such analysis here due to the multitudes and complexity of unique state policies, further research should examine the important role of policy in vaccination campaigns. Indeed, quantitative studies already indicate that such differences may have impacted vaccine uptake [14], consistent with earlier findings [16]. Many of these differences between state policies, such as West Virginia’s decision to use local pharmacies to supply nursing and long-term care facilities, may be better explored qualitatively. Qualitative studies can also shed more nuanced light on idiosyncratic differences between states that precede the pandemic, for instance relating to distrust of the healthcare system and attitudes towards vaccination [17].

Study results

Main results

Exhibit 3 shows vaccinations administered per capita over time by state.

Vaccination rates varied considerably across states throughout the immunisation cam- paign and some states that did particularly well in the first months of the campaign have since lagged behind. For example, Massachusetts had the 16th fewest vaccinations by February 1 but the 2nd highest by September 21, while West Virginia ranked 2nd by February 1 but last by September 21. The diverging trajectories across states raise the questions: why did some states have higher vaccine uptake early on, and why did vacci-

nation trends change over time?

Exhibit 4 reports results from three regressions aiming to shed light on these ques- tions. As of early February, both healthcare and public health expenditure are positively associated with vaccination rates. The coefficient on healthcare expenditure implies that an increase of USD 10,000 in spending per capita is associated with 0.069 more vaccina- tions per capita; an increase of USD 100 in public health spending is associated with 0.18 more vaccinations per capita. These expenditure associations are more unclear in June and September. For public health expenditure, the data do not suggest an association in June (p = 0.238) or in September (p = 0.839). For healthcare expenditure, there is evidence for an association in both June (p = 0.051) and September (p = 0.046) and the β ^-coefficients are greater in both months, but the 99%-confidence intervals include zero, suggesting greater variance in the relationship than in February.

The results from June 1 and September 21 reveal a negative association between vac- cinations and the fraction of census respondents who said in January 2021 that they would ”probably” or ”definitely” not receive a COVID-19 vaccine. The coefficient from September implies that a 10 percentage point increase in vaccine hesitancy (as measured in January) is associated with a decrease in vaccinations of 0.14 per capita. Given that states differed by as much as 20 percentage points in the January poll – i.e., a predicted difference of 0.28 vaccinations per capita in September – and that states had adminis- tered an average of 1.13 vaccines per capita as of September 21, the magnitude of this association is of practical relevance. In June and September, there is also evidence for positive associations between vaccinations and the Democratic vote margin in the 2020 elections as well as the share of the population above 65 years.

As of June 1, there is a negative association with the percentage of adults who in 2019 reported not having seen a doctor in the past year due to costs. However, there is no clear evidence for this relationship in February or September. Finally, neither population density nor the share of the adult population living in poverty are associated with vaccination coverage at any of the three points in time.

To better illustrate associations over time, we conducted the same multivariable linear regressions for cumulative vaccinations per capita for each of the 233 days from February 1 to September 21 and plot the coefficients from each of these 233 regressions. Exhibit 5 presents these rolling regression coefficients for select variables; appendix D includes plots for the remaining variables.

For the variables on vaccine hesitancy and Democratic voting, the regressions show a clear trend in the association between the two factors at baseline and vaccination rates. Whereas the coefficients for both variables were close to zero between February and April, they have since steadily diverged from zero in opposite directions. As of September 21, the regression analysis suggests a strong positive association between 2020 Democratic vote margin and vaccinations per capita and a negative association for the share of state population that reported not wanting a COVID-19 vaccine as of January 2021. For public health expenditure, there was an association on February 1 that subsequently disappears but with a less clear trend. For the elderly population share, a positive association appears to increase in strength over time, though the 99% confidence interval generally includes zero. For the remaining four independent variables, there was no clear trend over time.

Sensitivity analyses

reports several sensitivity analyses supporting the overall robustness of our findings while shedding light on important nuances. First, several of the main findings

are robust to using an alternative measure of vaccination in the form of new monthly, rather than cumulative, vaccinations per capita. However, the slowdown in vaccinations during the fall of 2021 makes it difficult to observe associations with new vaccinations in September. Second, when we include a time-varying measure for the number of vaccines distributed to each state on a given day, the positive associations with expenditure disap- pear, suggesting that the relationship between expenditure levels and vaccination rates could be driven by the ability of states to effectively distribute vaccines to administration sites. In contrast, the demand-side associations are generally robust to the inclusion of the distribution variable. Third, following Oster [11], we find that the results for generally are unlikely to be explained by omitted variable bias, in the sense that it would require a very high degree of selection on unobservables to nullify our observed associations. We also find that adding a selection of potential confounding variables does not significantly change our main findings. Finally, we find that the observed associations are robust to dropping individual states from the regression.

Discussion

Interpreting the observed associations

Our findings show that the cumulative number of vaccinations administered per capita was positively associated with both public health- and healthcare expenditure early in the vaccination campaign. Although this association was hypothesised a priori, higher health expenditure is generally no guarantee of improved outcomes, as expenditure may also reflect higher morbidity, prices, or system inefficiencies. Consequently, the interpretation of these associations is limited by the fact that expenditure is only an imperfect proxy for health infrastructure quality.

Further analysis suggests that the association between expenditure and vaccine ad- ministration rates may operate through vaccine distribution. Expenditure is correlated with vaccine distribution, and the association between expenditure and administration disappears once distribution is included in the model (see appendix E). One causal inter- pretation of this pattern is that, at least around February, states with higher expenditure levels administered more vaccines primarily because they were more successful at estab- lishing vaccination sites to which they could order the vaccine distributions required for doses to ultimately be administered.

While public health expenditure might seem like a better proxy for health system capacity with relevance for an epidemic immunisation campaign, ’regular’ healthcare capacity has also played a role in the COVID-19 immunisation campaign. In particular, hospitals and other inpatient clinics have been instrumental in implementing vaccination efforts within states, especially given the need for vaccines to be stored at ultra-low temperatures and administered by trained health professionals. In light of this, one interpretation of the data is that states with stronger healthcare infrastructure, in the form of well-funded and well-staffed hospitals and clinics, were better able to launch vaccination sites through which doses could be ordered, distributed, and administered.

It seems plausible that health system capacity plays a role in allowing states to dis- tribute, and hence administer, larger volumes of vaccines. This hypothesis is consistent with findings from Davila-Payan and colleagues, who found that state-level vaccination rates during the 2009 A/H1N1 influenza epidemic in the US were positively associated health system factors, such as the number of vaccination sites [18], [19], and those of Brown et al. [5], who found that US county-level health system vulnerabilities were neg- atively associated with vaccination coverage as of May 25. Similarly, a systemic review by Brien et al. [20] identified education status and previous influenza vaccination among individual-level predictors of A/H1N1 vaccination status in twelve countries, while Archi- bong and colleagues find that state-level vaccination rates in Nigeria are associated with health infrastructure quality [21].

The results from June and September suggest a clear negative association between vaccine hesitancy and vaccinations per capita. Our robustness analyses imply that this relationship is not driven by any individual states. Having established this association empirically, the question for policymakers, then, is what are the causes of vaccine hesi- tancy itself? There is a burgeoning literature dedicated to the determinants of vaccine attitudes[22], and recent research is shedding light on the phenomenon in the context of COVID-19 [4], [3].

As of June 1 – but not in February or September – , cost obstacles to care appear negatively associated with cumulative vaccinations. This result corroborates similar pre- viously reported associations observed in the context of 2009 pandemic H1N1 vaccination coverage [19], [9], as well as studies highlighting socioeconomic vulnerabilities as determi- nants of US county-level COVID-19 vaccination coverage as of May 2021 [5], [6]. While this result may appear puzzling considering that the COVID-19 vaccine is offered free of charge to anyone in the US, there are several potential explanations. First, many indi- viduals do not know, or refuse to believe, that the vaccine is administered truly free of charge. One April 2021 poll found that 32% of unvaccinated respondents cited concerns about having to pay out-of-pocket costs among their reasons for not having received the vaccine yet [23]. While this concern may be factually unwarranted given that vaccines nominally have been freely available, it is neither surprising nor irrational in a health sys- tem where out-of-pocket expenses can be financially ruinous and difficult to anticipate. For instance, some patients have been met with surprise bills for thousands of dollars for COVID-19 diagnostic tests that were supposed to be freely available under federal law [24]. The fact that some vaccination providers ask patients to bring their insurance card, as well as isolated incidents of patients being mistakenly billed, have contributed to widespread confusion about the true costs of vaccination [25]. The people who have not seen a doctor for at least a year due to costs are also more likely to be generally disconnected from the health system and thus unaware of that the vaccines are freely available.

Another explanation for why a high prevalence of cost obstacles to care predicts state-level vaccination rates is found with the indirect financial costs associated with vac- cination. In the same poll cited above, 15% of respondents mentioned difficulties around transportation to a vaccination site among the reasons for not being vaccinated yet and 20% voiced concerns about having to take time off from work to go and receive the vac- cine [23]. That these concerns were more common among black and hispanic than among white respondents is yet another piece of evidence shedding light on racial and ethnic

vaccine inequity in the US [26]. Moreover, each of the COVID-19 vaccines is associated with mild-to-moderate side effects that may interfere with work. Another poll from June 2021 found that workers were more likely to be vaccinated if their employers provided paid time off to get vaccinated and recover from side-effects, or simply encouraged vac- cination [27]. These results highlight that demand-side barriers to vaccination are far more complex than mere hesitancy concerning the vaccine. To support vaccine uptake, governments should consider providing additional information and financial support to those who are the most marginalised by the healthcare system.

In both June and September, there is clear evidence of a positive association between cumulative vaccinations per capita and the Democratic vote margin in the 2020 US elections. One interpretation of this association is that Democratic voting is related to demand-side factors such as vaccine attitudes; initial evidence strongly suggests that vaccine uptake is greater among Democratic voters [3]. Despite vaccine hesitancy already being in the model as a separate variable, there are two reasons this could still explain the association. First, the hesitancy variable was selected as a measure of attitudes before any potential influence by the outcomes of the immunisation campaign. Vaccine attitudes have evolved in the months since January 2021 [23], and it seems plausible that this development followed partisan lines, As such Democratic voting could explain variation in the demand for vaccines that was not captured in the early January poll. Second, the sample-based census poll may be an imperfect measure of vaccination attitudes. If so, the Democratic voting variable could be capturing variation in vaccination attitudes that was imperfectly measured by the hesitancy variable.

The June 1 and September 21 results suggest a positive association between vaccina- tions per capita and the population share above 65 years, which is unsurprising consid- ering the increased vulnerability to COVID-19 among the elderly. However, it is worth noting that the relationship is among the identified associations that is the most heavily influenced by dropping individual states, particularly those with especially young (e.g., Alaska) or old (e.g., Maine) populations (see appendix E). It should also be recognised that the elderly population faces distinct challenges to accessing vaccines, such as mobil- ity difficulties, and that even maintaining high vaccination rates in the most vulnerable age groups often requires targeted efforts [27].

Evolving determinants of state-level vaccination rates

One of the more striking patterns revealed by this analysis is that the determinants of state-level COVID-19 vaccination rates have evolved over time. As Exhibit 4 shows, some variables had a clear relationship with the outcome in February but less so or not at all in September, and vice versa for other variables. By plotting the rolling coefficients from 233 regressions between February 1 and September 21, we visualise the evolution of each determinant (Exhibit 5 and appendix D). The clearest pattern is that the associa- tions between vaccination outcomes and vaccine hesitancy and Democratic vote margin have grown stronger over time. It also appears that the association for public health expenditure has gotten weaker, though this trend is noisier.

Together, these results support a compelling narrative about the trajectory of the

vaccination campaign. Early on while vaccine eligibility was limited, demand-side factors such as vaccine hesitancy and (perceived) cost obstacles to care were not strong deter- minants of outcomes, as vaccinations were limited by both manufacturing capacity and the ability of states to distribute allocated doses. Consequently, supply-side dimensions of health system capacity, as proxied by expenditure, were better predictors of state-level immunisation success. States like Alaska and West Virginia, which have some of the highest per capita (public) health expenditures, were able to distribute vaccines well and achieved the highest vaccination rates around February and March. However, as sup- ply constraints were gradually relaxed and states depleted the pool of the most willing vaccinees and expanded their eligibility policies, population demand mattered more. As of April, differences in vaccine hesitancy became increasingly predictive of vaccination outcomes. Further, since vaccine attitudes are markedly partisan, the greater role of demand allowed for a divergence along party lines. Alaska and West Virginia, the two Republican states that had vaccinated at the fastest rates in February, started lagging. As of September, 22 of the 25 states with the fewest vaccinations per capita had voted Republican in 2020. The fact that this pattern only arose several months into the vac- cination campaign suggests that it is differences in population demand, not state ability to supply, that explain the partisan divergence in immunisation outcomes.

Beyond policy implications, this pattern of results can inform methodology for future research on immunisation campaigns during public health emergencies. Most research on epidemic vaccination rates analyse determinants of coverage at a single point in time, typically several months into the campaign [18], [19], [5], [6]. Such analyses can be misleading for infectious disease outbreaks where both the nature of the epidemic and the vaccination effort change rapidly over time. Researchers should consider utilising time-series data for a more comprehensive analysis.

Conclusion

In this examination of the United States pandemic vaccination campaign, we have pro- vided evidence for the critical role of both supply-side factors, such as healthcare- and public health infrastructure, and demand-side factors, such as vaccination attitudes and cost obstacles to care. We have shown that the determinants of COVID-19 vaccine ad- ministration have evolved over time. In particular, we have provided evidence consistent with the idea that while supply-side factors initially constrained immunisation efforts, the demand for vaccines ultimately has become the differentiating factor between states with low and high vaccination coverage.

While this study has limitations, particularly when it comes to causal interpretation, the findings presented here suggest clear policy implications. For one thing, the results may inform the COVID-19 vaccination campaigns in the United States, including the administration of booster doses. The clear partisan picture that emerges even after adjusting for vaccine hesitancy suggests a crucial role for political messaging in regions where coverage remains limited. Moreover, the apparent role of cost obstacles to care underscores the need for greater efforts to increase information and access for marginalised communities.

While certain factors undoubtedly are limited to the United States, some aspects of this analysis may be instructive for other countries undertaking vaccination efforts, particularly as they seek to bolster population immunity in the face of seasonal forcing and emerging viral variants [28], [29]. The analysis also holds policy lessons for prepared- ness efforts before the next pandemic: investing in (public) health infrastructure and proactively working to improve vaccine uptake can pay dividends for future epidemic vaccination campaigns, in addition to any health gains that can be realised immediately.

Finally, while states are often ranked by the speed at which they achieve vaccination coverage, it must be stressed that vaccination rates alone are not the sole indicator of a successful vaccination campaign. Measures like vaccinations per capita do not reveal in- formation about other fundamental objectives of public health, such as protecting at-risk populations and promoting health equity [30], [31]. Indeed, highlighting crude vaccination rates can potentially have the adverse consequence of incentivising policymakers to pri- oritise speed above all other outcomes. Consequently, it is vital to analyse and highlight the extent to which states have achieved more fundamental ends of public health, such as equity [32].

Uncited references

[39], [40], [41].

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.