3.2 Results

We empirically investigate the relationship between the income share accruing to the top 1% of the income distribution and each of our measures of the role of private finance in health care with the following specification:

$${\rm Ln}\,Private_{{ct}} \,{\equals}\,\alpha {\plus}\beta \,{\rm Ln}\,Inequality_{{c,{\rm }t{\minus}k}} {\plus}\gamma \,{\rm Ln}\,Public_{{ct}} {\plus}X'_{{ct}} \partial {\plus}\phi _{c} {\plus}\mu _{t} {\plus}{\varepsilon}_{{ct}} ,$$

where Ln Private _ct corresponds to the natural logarithm of the level of private expenditure in country c and year t and Ln Public _ct is the natural logarithm for public health expenditure, so equation (1) corresponds to the logged version of a regression in which the dependent variable would be the ratio of private to public health expenditures. The main independent variable, Ln Inequality, is the share of income held by the top 1% and X′ _ct is a vector of the natural logarithm of the controls variables enumerated in Section 3.1. A full set of year fixed effects, µ _t , controls for shocks that might simultaneously affect all countries in the sample. Finally, we include country fixed effects, ϕ _c , to control for unobserved time-invariant features of country health care systems that are likely to have an impact on the prevalence of private health care financing. Given the addition of these fixed effects, our estimate for β can be interpreted as the within-country relationship between changes in inequality and changes in private health care expenditure. We finally allow our inequality measure to enter with a zero-to-two-year lag over the full sample period. The results from this specification are summarized in Table 2. The coefficients reported in the first column of Table 2 income share suggests a positive and statistically significant correlation (0.778) with the contemporaneous income share of the top 1% income earners. This positive and statistically significant elasticity persists as we turn to one- and two-year lagged values of the top 1% income share, although the magnitude of the coefficient decreases with the lags, to 0.65 and 0.61, respectively, suggesting private health expenditure could react relatively quickly to changes in the level of inequality. As expected, private health expenditure is negatively associated with the share of the population aged 65 and older, potentially driven by the fact that public health coverage is often more generous for people in that age range. The association between private health expenditure and the average income per tax unit is also positive, as is the association with national GDP. The association between the natural logarithms of private and public expenditure is negative, and statistically significant, which is consistent with the crowd-out effects found in Flood et al. (Reference Flood, Stabile and Tuohy2004).

Table 2

Ln income share of top 1% on Ln private health expenditure

GDP=gross domestic product.

Notes: Robust standard errors in parantheses. Income shares from the World Wealth and Income Database (Alvaredo et al., 2016). GDP, government expenditure, private and public health expenditures, population and employment rate from the OECD (OECD, 2016). All values in U.S. dollars of 2010. Unbalanced panel, data available from 1982 to 2015.

***p<0.01, **p<0.05, *p<0.1.

We slightly alter the specification described above and replace the control for the natural logarithm of public health expenditure with the natural logarithm of total health expenditure, to effectively estimate the logarithmic version of a regression of our measure of inequality on the private share of total health expenditure (not shown). Overall, we obtain coefficients on the top 1% income share varying between 0.26 and 0.54. Similar patterns as those described in Table 2, including smaller coefficients on the inequality variable as we move towards longer lags, can be observed.

Individuals (and therefore countries) can increase private expenditure on health care in a variety of ways including: (i) an increase in the scope or ‘quality’ of the voluntary health insurance schemes purchased by individuals already covered by some form of private insurance; (ii) an increase in the share of the population buying private health insurance coverage (purchased individually, or provided through an employer-sponsored plan); or (iii) an increase in the total amount spent out-of-pocket for health care services and medication. Tables 3 and 4 present results that speak to the last two of these channels. The estimates suggest that the uptake in private health insurance in the population is likely the main driver behind the positive association between inequality and private health care spending.

Table 3

Ln income share of top 1% on Ln population with private health insurance

GDP=gross domestic product.

Notes: Robust standard errors in parantheses. Income shares from the World Wealth and Income Database (Alvaredo et al., 2016). GDP, government expenditure, private health insurance, population and employment rate from the OECD (OECD, 2016). All values in U.S. dollars of 2010. Unbalanced panel, data available from 1995 to 2015.

***p<0.01, **p<0.05, *p<0.1.

Table 4

Ln income share of top 1% on Ln out-of-pocket health expenditure

GDP=gross domestic product.

Notes: Robust standard errors in parantheses. Income shares from the World Wealth and Income Database (Alvaredo et al., 2016). GDP, government expenditure, private health expenditure, population and employment rate from the OECD (OECD, 2016). Out-of-pocket health expenditure from the World Bank (WORLD BANK, 2016). All values in U.S. dollars of 2010. Unbalanced panel, data available from 1995 to 2014.

***p<0.01, **p<0.05, *p<0.1.

Table 3 presents the results when we estimate a version of equation (1) in which Ln Private _ct represents the natural logarithm of the population covered by private health insurance, which corresponds to estimating the logged version of a regression of the private health insurance coverage in the population on the income share of the top 1%.Footnote ¹¹ We estimate a positive and statistically significant relationship (0.86) between the top 1% income and the share of the population covered by some form of private health insurance. These results are consistent when controlling for the log of public health expenditure or for the log of total health expenditure. Re-estimating the same model separately for country-year pairs with a private insurance coverage below and above 32.5% (the median coverage rate) suggests that this estimated impact mostly comes from contexts where baseline private coverage is low. We also note that, in our sample, the income share of individuals between the 90th and the 99th percentile of the income distribution has increased as the income share of the top 1% was also increasing. Our result might therefore capture the fact that as individuals at the top of the income distribution (described more broadly than solely the top 1%) get relatively richer, their preferences for faster access to care, improved quality of care (either through the quality of the care itself, or through a more extensive set of amenities in health care establishments), also leads them to opt in private health insurance schemes. Substituting the natural logarithm of the top 1% income share by that of the top 5% and top 10% income share in a similar specification, we indeed estimate a positive relationship between inequality at the top of the income distribution and private health insurance coverage.

Table 4 summarizes the results obtained when the dependent variable in equation (1) is defined as the natural logarithm of out-of-pocket health expenditure (controlling for the natural logarithm of public health expenditure). We estimate a weaker relationship, suggesting that the growth in privately financed health expenditure observed as inequality increases is likely due to an increase in private insurance premiums, rather than in co-payments or in out-of-pocket payments for services that are not covered by public or private plans.

Appendix Tables A1–A4 (online Supplementary material) suggest that our results are generally robust to measuring inequality using Gini coefficients in lieu of top income shares. However, statistically significant associations are harder to capture using disposable income ratios to consider separately the distance between the median and, respectively, the top and the bottom of the income distribution.

Our cross-country estimates are suggestive of a relationship between top income shares and private health spending, particularly through the purchase of private insurance. However, this finding could reflect a variety of underlying mechanisms. To try and understand the extent to which the concentration of income at the top end of the distribution is driving this relationship we turn to a micro-data analysis of health care spending by Canadian households.

4.2 Results

Our main empirical specification is given by equation (2), in which Health expenditure _ipt corresponds to the expenditure of household i residing in province p observed in year t. Income Fractile _ipt consists of a series of three dummies indicating the household’s position in the income distribution, either between the 21st and the 50th percentiles, between the 51st and the 90th percentiles, or in the top 10% (the omitted category being the bottom 20% of the distribution). Other control variables included in X _ipt are total household income,Footnote ¹⁷ the number of adults aged 65 and above within the household, the presence of children in the household, the total number of individuals within the household and the marital status of the main respondent to the SHS. We also include a vector of province fixed effects, ϕ _p to account for the fact that, beyond the main principles set in the Canada Health Act, provinces can decide on the nature and scope of additional publicly provided health coverage, and have jurisdiction on the organization of the health care system. Consequently, prescription drugs for some population groups or services offered by certain health professionals, for example, vary across jurisdictions within the country. In this context, the inclusion of province fixed effects allows us to perform a within-province analysis. Finally, µ _t represents a vector of year fixed effects, capturing shocks common to households in all provinces, which might affect patterns in health care expenditure.

$$Health\,Expenditure_{{ipt}} \,{\equals}\,\alpha {\plus}Income\,Fractile'_{{ipt}} \beta {\plus}X'_{{ipt}} \partial {\plus}\phi _{p} {\plus}{\rm }\mu _{t} {\plus}{\varepsilon}_{{ct}} .$$

Our estimates should be interpreted with caution given that household income may be endogenous and a function of health status. The SHS does not document health conditions and we do not address this endogeneity directly. Therefore, we do not interpret our estimates as causal, but rather seek to understand how patterns of health care spending vary across the distribution of household incomes, and what these patterns may imply for the role of private and public financing of the Canadian health care system moving forward.

Results from estimating equation (2) are displayed in Table 7. Looking first at total health expenditure, the results presented in column 1 suggest that compared with the bottom 20% of households, those higher up in the distribution spend significantly more on health care. The coefficients associated with the income fractiles confirm a positive gradient, even when controlling for households’ level of income. Being in the top 10% of the income distribution is associated, all else equal, with an increase of $1235 in total health expenditure compared with the bottom 20% households, which corresponds to a little more than 60% of a standard deviation. Having an income between the 50th and the 90th percentiles is associated with an increase of a little less than $800 compared to the bottom 20% of households. The coefficients on the other variables also have the expected sign. The presence of a senior in the household, total income and household size are all positively associated with total health expenditures. Column 2 of Table 7 presents similar results for direct health care costs to households; although the coefficients on the income fractile dummies are smaller, they exhibit a similar pattern, suggesting that the findings from column 1 are not exclusively driven by the purchase of private health insurance by higher income households, but also relate to out-of-pocket spending. However, the results from column 4 provide some strong evidence that being in the top 10% of the income distribution is associated with an annual increase of $460 on private health insurance premiums compared with the bottom 20% income (63% of a standard deviation), nearly a threefold increase compared with the impact of moving from the bottom 20% to the next income fractile.

Table 7

Household position within the income distribution (fractiles) on health expenditure

pctile=percentile.

Notes: Robust standard errors in parantheses. All income and expenditure variables are expressed in dollars of 2002. All regressions are estimated using sampling weights. Data from the Survey of Household Spending, excluding observations from the territories.

***p<0.01, **p<0.05, *p<0.1.

We note that the relationship we are capturing with respect to private health insurance is driven by premiums paid by individuals, either through plans purchased individually, or through employees’ contributions to employer-sponsored plans.Footnote ¹⁸ However, it is quite unlikely that the difference we observe in coefficients associated with the top 10% and the other households in the top half of the income distribution is driven by differential access to employer-sponsored private insurance plans, given that a vast majority of households in these two groups are employed full-time. Moreover, our results are robust to restricting the estimating sample to full-time employees or to controlling for the household’s main source of income (employment, investment, government transfers or others), or for full-time working status of the main respondent. The gradient presented in column 4 of Table 7 is therefore likely to be mostly driven by, for example, top income households purchasing more comprehensive options within the range of plans offered by their employers (in the case of flexible plans) or supplementing these plans with individually purchased ones. We also note that a price effect might contribute to this estimated relationship; tax exemptions for private health insurance represent a more important price incentive for high income earners whose marginal tax rates are likely higher, and have been shown to influence individuals’ decision to spend on employer-sponsored health insurance plans (Stabile, Reference Stabile2001).

Column 3 of Table 7 looks at expenditure on prescription drugs. While moving from the bottom 20% of the income distribution to the group formed of the 21st to the 50th percentiles is associated with an annual increase of $55 on prescribed medication, the increase is much smaller for households between the 51st and the 90th percentile. As noted above, this likely reflects better drug insurance coverage among these groups. We re-estimate equation (2) using logs and find similar results.

Households’ health care needs will vary as they age. For example, spending on residential care is likely to increase with age as might the relationship between income and spending on these types of services. To investigate whether our results differ significantly by age, we define a subsample formed exclusively of households whose head is aged 65 or older, and we estimate all specifications presented in Table 7.Footnote ¹⁹ Most of the patterns described for the full sample are also observed for this older group of households, and are often amplified. For example, the coefficient on the top decile of the income distribution is twice as large as the one estimated in the main sample when looking at physician expenditure, and movements towards top fractiles are also associated with substantially larger increases in direct health care spending in the older sample. Interestingly, the relationship estimated with private health insurance is quite similar in both samples, and although the coefficients on all income fractiles are larger when considering the older sample, we still do not observe a gradient between expenditure on charges from hospital and residential care facilities and the position in the income distribution. This result is not necessarily intuitive; however, at least two factors may explain our results. First, we note that the survey population excludes people living in residences for dependent seniors. Second, we do observe a very strong relationship between the position in the income distribution and expenditure on services provided by non-physician health professionals, including nurses and other professionals offering part-time or full-time home care assistance, a potential substitute for nursing homes. Finally, the results for prescription drug expenditure are quite different in the older sample; the coefficients associated with all income fractiles are relatively similar with values close to $80, and we cannot reject that they are statistically equal to each other.

While providing insights on the relationship between households’ relative economic situation and their health expenditures, the results from Table 7 do not allow us to investigate the evolution of such an association as the concentration of income in the top fractiles increased. To address this question, we first look at unadjusted trends by mapping the evolution of health expenditures (in total and by category) for each income fractile between 1997 and 2009. For most outcomes, an increase in the gap between the amounts spent by households in the top 10% income and those in lower fractiles can be observed, especially for direct health care costs to households (Figure 3) and on expenditures going towards private health insurance plans (Figure 4).Footnote ²⁰

Figure 3

Evolution of average direct health care costs, by income fractileNote: All income and expenditure at the household level, weighted data from the SHS public-use microdata files. Average expenditure excluding the territories.

Figure 4

Evolution of average expenditure on private health insurance plans, by income fractileNote: All income and expenditure at the household level, weighted data from the SHS public-use microdata files. Average expenditure excluding the territories.

We again turn to a regression framework to study the evolution of the expenditure gap between households in the top 10% of the income distribution and those in lower income fractiles using the specification corresponding to equation (3). The estimating equation builds on equation (2) by adding a series of interactions between year effects and dummy variables indicating if a household is in the top 10%, between the 50th and the 90th percentiles or in the bottom 20%.

$$\eqalignno{&#x0026;Health\,Expenditure_{{ipt}} \,{\equals}\, \alpha {\plus}Income\,Fractile'_{{ipt}} \,\beta {\plus} \left( {Income\,Fractile_{{ipt}} \,{\ast}\, \mu '_{t} } \right)\pi \cr \quad \quad \quad \quad \quad \quad \quad \quad \quad \quad {\plus}X'_{{ipt}} \partial {\plus}\phi _{p} {\plus}\mu _{t} {\plus}{\varepsilon}_{{ct}} .$$

The results presented in Table 8 confirm such dynamics for total and direct health expenditure. For those outcomes, the coefficient on the top 10% fractile remain positiveFootnote ²¹ and the coefficients on its interaction with the year effects are generally increasing, and mostly statistically significant from 2003 onwards. We note that the increase in the coefficient on the interaction term in 2003 coincides with an increase in the average household income of the top 10% relative to the bottom 90% (see Figure 5). Although the pattern is less salient, the interaction terms estimated when considering expenditure on private health insurance premiums are consistent with our hypothesis of greater reliance on private health insurance and care among the top income earners as the gap between this group and the rest of the population increases. We note that for these three categories of expenditure, the interaction terms between year effects and the 50–90th percentile (not presented in the table for ease of exposition) are also mostly positive, but much smaller in magnitude, mostly neither statistically significant nor monotonically increasing.

Figure 5

Evolution of household income for top 10% and bottom 90% households, by yearSource: Data from the SHS public-use microdata files, averages excluding the territories.

Table 8

Household fractiles on health expenditure, with interactions between year effects

pctile=percentile.

Notes: Robust standard errors in parantheses. All income and expenditure variables are expressed in dollars of 2002. All regressions are estimated using sampling weights. Data from the Survey of Household Spending, excluding observations from the territories.

***p<0.01, **p<0.05, *p<0.1.

When considering expenditures on hospitals and physician care (columns 5 and 6), the inclusion of interaction terms slightly decreases the size of the coefficients on the income fractiles, which also mostly lose their statistical significance. The same is true of the interaction terms between income fractile and year. Turning to prescription drug expenditure (column 3), most interaction terms are negative, but do not follow a clear pattern are they generally statistically insignificant. However, the interaction terms between year effects and the dummy variable for the bottom 20% (not reported), are negative, decreasing and are often statistically significant, suggesting a widening gap in prescription drugs expenditure between the bottom 20% and the next income fractile during the sample period. One possible explanation for such dynamics could be the transitions towards income-based catastrophic coverage associated with important cost-sharing provisions that took place across Canadian provinces between 2000 and 2010, as highlighted by Daw and Morgan (Reference Daw and Morgan2012). The impacts of such a transition might have been concentrated at the bottom half of the income distribution if individuals with higher incomes are more likely to have contributed to private insurance plans offering prescription drugs coverage, and for which co-pay rates would not have followed a similar trend. Simultaneously, households on social assistance, mostly located in the bottom 20% of the income distribution, were less likely to be affected by the change, and to remain on public plans with first dollar coverage or co-payments.