We used monthly state-level data for deaths from diseases of the circulatory system (ICD-10 code: I00-I99) for the period 2008–2017, as reported by the Centers for Disease Prevention & Control (CDC). Our analysis therefore focuses on mortality rather than other measures of cardiovascular health [30]. Monthly state unemployment rates were retrieved from the Bureau of Labour Statistics. We also obtained data for GDP growth and poverty rates from the Bureau of Economic Analysis. Population data were obtained from the US Census Bureau. Drawing on a recently published OECD/Eurostat report [31], we also distinguish deaths into avoidable and non-avoidable causes. Avoidable deaths are derived from the OECD/Eurostat list of ICD-10 causes, which are either considered as treatable from health care activities or preventable by specific public health interventions. These only include deaths of under 75-year-olds.

To capture economic uncertainty, we rely on the Economic Policy Uncertainty (EPU) indices measured at a state-month level. The approach for constructing state-specific EPU indices builds on earlier work [1] and draws on the digital archives of almost 3,500 local newspapers published on daily or weekly basis and circulated throughout the states. The indices are constructed based on the number of articles that contain specific economic, uncertainty and policy terms. In particular, they capture the levels of economic uncertainty by tracking and measuring the frequency of articles with terms relating to economic policy changes and uncertainty (e.g. ‘economic’, ‘economy’ and ‘uncertain’, ‘uncertainties’, uncertainty’). EPU indices draw on the seminal studies by Baker et al. (2016) and Baker et al. (2022). As explained in the relevant studies [1, 32], the accuracy and potential bias of the index have been evaluated and extensively validated during its conceptualisation and development. EPU indices have been recently used in economics, finance [2, 33,34,35,36,37], and epidemiology and public health [13,14,15, 18, 29], and their empirical application has transformed the research field for economic uncertainty [37]. A potential issue with the use of a news-based index relates to potential bias and political slant of the newspapers. Drawing on approaches for measuring newspapers slant, previous studies have tested and empirically scrutinised this scenario, concluding that this is not the case and confirming the validity of the EPU index [1, 32, 38]. The fact that newspapers are used to create the economic index, does not mean that this is the sole channel via which the public becomes aware of economic uncertainty. Some people might not read newspapers, but they may feel uncertain after receiving information via other channels. The uncertainty index thus captures the level of uncertainty, that some people may become aware of through other sources. In addition, a detailed explanation on why newspapers are used to construct the index can be found in the papers published by the creators of the index [1, 32, 39]. While economic uncertainty might originate from a number of economic or non-economic factors, the index captures how any factor translates particularly into economic uncertainty, rather than uncertainty on other dimensions.

Previous studies on the link between economic uncertainty and health have used a single measure of economic uncertainty at the country level. Contrary to these papers, we employ a recently developed measure that varies by state, which also allows us to disentangle state/local from national/international drivers of economic uncertainty [1, 32]. We draw on two state-level EPU indices. The first focuses on economic uncertainty within a state that arises from national sources and events. Its construction further draws on terms of national interest such as national elections, federal departments, agencies, and regulators. The second measures uncertainty levels stemming from state and local sources and builds on terms of state-specific executive, legislative and regulatory bodies. A detailed description of the approach and the terms used to flag relevant articles is presented in Baker et al. (2022).

Our analysis is based on a panel data econometric approach, using monthly observations for 51 US states as the level of analysis. Panel data exhibit several advantages over cross-sectional or time series data, which have been extensively analysed in econometric literature [40]. They allow us to control for unobserved time-invariant differences between the states and to eliminate potential bias due to omitted time-invariant state-specific characteristics.

The dependent variable is the number of deaths in each state and month. The two independent variables of interest capture the national and the state sources of economic uncertainty respectively. EPU indices are standardised by dividing them with their respective standard deviation for each state [41]. Last, we control for a vector of independent variables, which includes unemployment, GDP growth, population size, consumer price index and poverty rate. We also controlled for year and month dummies to account for potential seasonality. Robust standard errors, clustered at the state level, are reported throughout.

To examine the association between the different types of EPU and the number of cardiovascular deaths, we rely on a Poisson regression model, given that the number of deaths is a discrete variable taking positive values. Contrary to linear probability models – which rely on assumptions about continuous outcomes that are normally distributed – Poisson models fit the number of occurrences of an event (in this case, deaths). They have been widely used in empirical analysis of count data [42,43,44], including data on the number of deaths [45,46,47]. The econometrics literature has extensively discussed the application of Poisson models in empirical research, and has explained aspects such as the interpretation of relevant coefficients [48, 49]. Apart from the estimates for the total population, we also perform additional analyses, providing relevant evidence for cardiovascular deaths by gender. Since our baseline model uses a given month’s uncertainty index and the same month’s mortality, our analysis focuses on the short-term association. Some additional regressions also study whether there is a lagged association, but this is again limited only to a few months. The uncertainty index is reported at the state level, and therefore captures uncertainty for the whole state. As this is a macro-level index, we do not know the level of uncertainty that each individual experiences.

Next, we also test potential non-linear links between economic uncertainty and cardiovascular deaths, by splitting EPU indices into quintiles. In doing so, earlier methodological approaches for estimating non-linear relationships are adopted [15, 41]. Our point of departure is that different levels of economic uncertainty might have differential impact on deaths, demonstrating an asymmetric response of cardiovascular mortality to its risk factors. After splitting EPU indices into quintiles, we control for the first and fifth quintile of the distributions. The remaining quintiles (i.e., the three middle ones) for both EPU indices are omitted and used as reference categories.

Apart from the baseline estimates, we also perform additional analyses to test the sensitivity of our results to different estimation strategies or econometric specifications. First, we employ a Negative Binomial model, which is a suitable alternative to Poisson for modelling count data [43, 50]. Second, we further control for linear and quadratic time trends to capture the potential trajectory of the outcome variable over time. This empirical exercise serves as an additional check, in which we examine the extent to which the direction and statistical significance of the estimates remain unchanged after the inclusion of time trends. Third, we use the logarithmic transformation of the dependent variable and estimate fixed effects and random effects models. Last, we exclude the bottom and top 1% and 5% of the observations (depending on the values of economic uncertainty and those of the outcome variable), to explore the extent to which our estimates are driven or explained by outliers.