Existing studies have examined different aspects of health expenditure determinants. Here we focus on environmental ones.3 In order to work with a complete balanced panel data, in this study our temporal analysis period is 1995-2014 for a selected group of 29 OECD countries namely: Australia, Austria, Belgium, Canada, Czech Republic, Denmark, Finland, France, Germany, Greece, Hungary, Iceland, Ireland, Italy, Japan, Latvia, Luxembourg, Netherlands, New Zealand, Norway, Poland, Portugal, Slovenia, Spain, Sweden, Switzerland, Turkey, United Kingdom, and the United States. The benefits of using a balanced panel data set consist in that it allows doing the analysis of stationarity of the variables. Besides, most of the tests require that the panels be strongly balanced.13

The econometric analysis relies on annual data obtained from two different data sources: the latest versions of OECD Health Statistics and OECD Environment Statistics. Information about per capita health care expenditures (total, public and private ones) and per capita GDP are measured at Purchasing Power Parity (PPPs) terms. Sulphur oxide emissions, nitrogen oxide emissions, and carbon monoxide emissions are in kilos per capita. All variables are converted into natural logarithmic form before the empirical analysis. The definition of the variables and summary statistics finally considered are described in Table I in the online Appendix.

This research provides rigorous scientific examinations to study the following issue: how air quality indicators affect health expenditures. Taking into account previous analysis, a positive relationship between income and health expenditures is expected. Additionally, we hypothesized that su, ni, and ca would also increase health expenditures.

Figure 1 represents the pathway effects of climate change on health endpoint, because it may affect the exposures to air pollutants. In any case, pollution could appear by both natural and human sources. Here it is modelled how air, that could be contaminated with pollutants, may negatively affect health in general and health expenditures in particular. Thus, the health effects of air pollution are diverse, and include direct and indirect ones. All in all, the most sensitive groups include children, older adults and people with chronic heart or lung disease). The explanation is based on the idea that these three air pollutants measures here considered deteriorate air quality and so, they would affect negatively health outcomes.14

Figure 1.

Pathway effects of climate change on health endpoint. (Authors’ elaboration adapted from Bernard et al.15).

(0.13MB).

Moreover, as pointed by Bernard et al.,15 health effects of exposures to sulphur and nitrogen oxides, and carbon monoxides, can cause reduced work capacity, aggravation of existing cardiovascular diseases, effects on pulmonary function, respiratory illnesses, lung irritation, and alterations in the lung's defence systems. That is, an increase in utilization (demand) of health care services due to bad health care outcomes is projected.

These interesting findings could be used to derive the appropriate health expenditure level, to obtain better environmental quality, and social well-being. Research needs considering air pollution models (and their potential linkage with climate change scenarios) in order to close gaps in the understanding of the relationship between air pollution exposure and health effects. In any case, health management and policy makers should consider these relationships.

Overall, we follow recent contributions, which consider the modelling advance of the relationship between health and environmental issues as Narayan and Narayan3 or Qureshi et al.16 Our empirical results are based on the following specifications with all the variables of interest are converted in natural logarithms form to allow us to understand them as elasticities.

Firstly, a linear one-way panel data model for health expenditure (he), as dependent variable to be analysed (while considering different measures of it depending on its source of financing: total, public, and private) based on the acknowledged literature is specified.17,18 Therefore, when income changes the variation in health care expenditure it is likely to be a combination of multiple forces. It is very useful to isolate these effects to gain enhanced insights into several components. In a more formal way, our model has the following general form:

where heit is the logarithm of per capita health care expenditures at the tth observation for the ith country; f(•) denotes health care expenditure structure; xit is the corresponding matrix of explanatory variables; β is a vector of parameters to be estimated (su, ni, ca and gdp). The term ɛ is the error bounded with the general statistical properties. Besides, feasible generalized least squares are considered. This procedure allows estimation in presence of first-order autocorrelation within panels and cross-sectional correlation, and heteroscedasticity across panels.

Secondly, a dynamic panel data approach is applied. That is, we include on the right-hand side of equation (1) the lagged dependent variable, in order to capture the inertia of health care expenditures. In other words, we try to disentangle if the logarithm of per capita health care expenditures on a year is conditioned by the previous one:

Before presenting the results from the estimation of the above-mentioned specifications, we first analyse all variables to ensure accurate estimates. That is, to obtain empirical findings that are not spurious and have economic sense. Precisely, first and second generation of panel unit root tests are employed here. Tables II and III in the online Appendix show this type of tests. Firstly, we applied panel unit root tests assuming cross-sectional independence. Regarding Levin et al.19 and Im et al.20, mixed results concerning variables being stationary are showed. Statistic for the logarithm of our variables when the ADF regression has an intercept only and an intercept and a linear time trend, are presented. Nevertheless, a common feature of these econometric tests is that they lose power as individual specific trends are included.13 In all cases, the lag order p, was selected using the Akaike information criterion. Then, we also apply the Breitung21 test that indicates the hypothesis that variables contains unit root is never rejected. Considering all of this, we turned to Pesaran22 second generation test that attempts to remove cross-sectional dependence. We present results for lag orders p = 0, 1, 2 and 3, finding that in most of the cases our variables of interest are integrated of order 1, I(1).

The estimation of the specification described yields the results reported in Tables 1-4. The first two tables considering (1) whereas the latest (2). Coefficients are highly significant, and the goodness of fit is very high in all cases. Our main findings can be summarized as follows. Figure I in the online Appendix, plots the residuals from Table 1 specification (ii) to assess the goodness of fit of the model and thus examine the existence of non-linearity and, if that case, to improve our specification and estimates.

Firstly, per capita GDP has a positive effect on health. However, we show some differences when desegregating by health care financing schemes. It should also be noted that the ‘unit’ rule has limitations, especially, when elasticities of 0.01 and 0.99 are both classified as ‘necessities’. These numbers are economically and statistically quite different.

At this regard, higher elasticities are obtained for public health expenditures. Therefore, the smallest are for private ones. Overall, income elasticity is not as statistically significant when incorporating lags. That is, an anchorage effect is appreciated (0.80-0.90). In other words, about 80% of a year's health expenditure is conditioned by the previous one despite the presence of other explanatory variables.

Secondly, little effects are found for air quality variables as the most important factor again, and in accordance with previous evidence, appear to be income.23

As for air pollutants variables, the most important factor is the one regarding su in the linear estimations, whereas ca is for the dynamic one. These variables explain more private expenditures. A surprisingly fact here is the reverse effect of environmental factors here considered. For example, whereas su affect expenditures in a positive way in the linear one-way estimates, the reverse effect is obtained for the two-way. Neither stables results are obtained for ni. But regarding ca ones, stable negative effects are shown.

Here and now, we briefly test the robustness of the results. We noticed that the empirical model previously estimated imposes common effects for all the countries in the sample. To be more specific, Table 4 contains the results, while Figure 2 plots them.

Figure 2.

Elasticities by income group and financing scheme. Left y-axis represents air quality elasticities. Right y-axis represents anchorage effect.

(0.37MB).

Definitely, we check the sensitivity of the estimates to income heterogeneity in the sample of 29 OECD selected countries considered. In doing so, we split the sample of countries into two groups based on a kmeans and kmedians partition cluster analysis for income (considering 2005 base year, half of the period). Both methods are widely used for exploratory data analysis. Precisely, these partition methods break the observations into a distinct number of non-overlapping groups.

The first, group I, consist of 19 OECD countries: Australia, Austria, Belgium, Canada, Denmark, Finland, France, Germany, Iceland, Ireland, Italy, Japan, Luxembourg, Netherlands, Norway, Sweden, Switzerland, United Kingdom, and United States. While, 10 OECD countries constitute the second one, group II: Czech Republic, Greece, Hungary, Latvia, New Zealand, Poland, Portugal, Slovenia, Spain, and Turkey.

Specially, results highlighted the above-mentioned regarding the relevance of health expenditure in the previous period. Again, little differences are found by health care financing scheme.

When distinguishing by clusters, more statistically significant results are shown for group I. However, our findings are consisted between groups and when comparing with the full sample analysis. ni should be at main action agenda outlines, especially for group II, where higher elasticities are founded (i.e., 0.034 versus 0.136).

One of the implications of these interesting findings is that future health expenditure would increase as economic growth does, and the air quality indicators would do. Table IV in the online Appendix also tested how parameters for su, ni, and ca are different for countries with a lower or higher income by introducing interaction effects. DGDPi would be a dummy variable coded 1 for observations corresponding to countries in group I and 0 otherwise. Hence, general public policies, and especially those ones that contains health expenditure estimates, should contain reliable health care expenditure projections. That is, forecasts should take into account both traditional determinants (as income), including recent factors as population dynamics, population lifestyles or the technological progress; and other drivers, like the ones here analysed (air quality variables).24