‘Investing’ in care for old age? An examination of long-term care expenditure dynamics and its spillovers

LTC services share some characteristics with heavily subsidised health care, yet unlike health care, personal LTC can be delivered by unpaid household members informally.³ However, evidence suggests large heterogeneity across the OECD countries in both caregiving duties and formal care provision of services. Descriptive analysis suggests that long-term care spending is associated with population ageing, female labour market participation, and the institutionalization of the care service system (Olivares-Tirado et al. 2011; Costa-Font et al. 2015; Huei-Ru et al. 2016).⁴ Costa-Font et al. (2015) estimate an income elasticity of 3.2, indicating a high sensitivity of per capita public LTC expenditure to a change in a country’s per capita GDP.⁵

Given that informal care is still the most common form of care for old age in almost all countries, if formal and informal care are substitutes (Picone and Wilson 1999; Stern 1995; Carmichael and Charles 2003; Carmichael et al. 2010; Nizalova 2012), a contraction in the supply of informal care (resulting from the expansion of labour market participation of traditional caregivers) can rise the demand for paid care.⁶

A higher LTC utilisation can exert spillover effects on the health system and, especially on costlier hospital care utilisation (Hofmarcher et al. 2007; Bodenheimer 2008; Mur-Veeman and Govers 2011). Some evidence already documents that the introduction of home care programs reduced delays in hospital discharges and emergency readmissions (Hermiz et al. 2002; Sands et al. 2006; Weaver and Weaver 2014).⁷ Hence, the expansion of LTC services can give rise to a reduction in health care utilisation.

The expansion of public LTC spending can give rise to a subsequent effect on economic activity by boosting the ‘care sector’. Previous studies such as De Henau et al. (2016) show that the care economy may raise GDP growth more than investment in construction.⁸ The next section describes the empirical strategy used to estimate the determinants of LTC expenditure and its determinants.

The identification of the drivers of LTC expenditure, such as labour market participation of traditional unpaid caregivers (women over 40 years of age), and its spillover effects on health care expenditures faces several methodological challenges, including potential reverse causality and omitted variables bias, as well as both time and cross-sectional correlation, and such dynamic effects need to be modelled. Relative to conventional time series modelling, the panel-VAR model considers the heterogeneity of the cross-sectional dynamics, which provides more information about the sources of heterogeneity in the system. The panel-VAR specification exploits the temporal and cross-section dimension of the data to infer dynamic relations between the dependent variables,⁹ allowing all covariates to be treated as fully endogenous while simultaneously modelling the unobservable heterogeneity through fixed effects (which account for time-invariant characteristics intrinsic to each country). Hence, resulting in an improvement in the consistency of the estimation (Love and Zicchino 2006).

So far, no previous study has used a sample of panel data over a considerable period for a large number fo countries, accounting for the fact that the presence of cross-sectional dependence in panel data may compromise the stationarity of the variables under analysis.¹⁰ To our knowledge, no previous study has examined the dynamic impact of healthcare expenditure and GDP of a LTC expenditure shocks.¹¹

In a panel-VAR model specification, each variable is explained by its own lag, the lagged values of the other system variables and individual country-specific terms. Treating all variables as endogenous prevents us from using weak instruments. An additional advantage of a panel-VAR model is that it allows goodness of fit analysis and observing the reaction to different shocks. The specification of a panel-VAR model of p order proposed by Canova and Cicarrelli (2004) is as follows¹:where all variables of vector \(Y_{t}\) are considered endogenous, allowing for a joint dynamic analysis. If we denote for q the number of endogenous variables, then vector \(Y_{t}\) has dimension q × 1. In turn, \(Y_{t}\) contains a cross section dimension, \(y_{i,t} = \left( {y_{1,t}^{\prime } ,y_{2,t}^{\prime } , \ldots ,y_{N,t}^{\prime } } \right)\), where i = 1,2,….,N indicates the number of countries and t = 1,2,…,T indicates the number of years observed for each country. The fixed effects (\({\rm A}\left( l \right)_{i0}\)) are captured by a q × 1 vector, where l is a polynomial in the lag operator such that \({\rm A}\left( l \right)_{i0} = \mathop \sum \nolimits_{i = 0}^{N} A_{i} l^{j} , j = 1,2, \ldots ,p\).

The term \({\rm A}\left( l \right)_{i1}\) is a q × N matrix of lagged coefficients and \(\varepsilon_{i,t} = (\varepsilon_{1,t} , \varepsilon_{2,t} , \ldots ,\varepsilon_{N,t} )\prime\) is the error term with a zero-mean, with variance–covariance matrix independent of t and such that \(\varepsilon_{i,t}\) of different periods are independent of each other: \(\varepsilon_{i,t} \sim iid\left( {0,\Sigma } \right)\). In this paper, we estimate the following two panel-VAR models to measure the effect of labour market participation on long-term care expenditure, and a second specification for the effect of long-term care expenditure spillovers on health care spending.

Estimating LTC expenditures and female labour market effects, we consider a first specification containing three equations for which the dependent variables are \(Y_{i,t} = \left\{ {{\text{FemPart}}_{i,t} , {\text{LTC}}_{i,t} , {\text{GDPpc}}_{i,t} } \right\}\). \({\text{FemPart}}_{i,t}\) measures the labour market participation rate of the traditional caregiver (women 40 years and older), \({\text{LTC}}_{i,t}\) refers to public long-term care (LTC) expenditures in per capita terms, and \({\text{ GDPpc}}_{i,t}\) measures gross domestic product per capita. Besides, LTC expenditures can exert heterogeneous effects on different health care spending categories. We distinguish between the three types of LTC expenditures available (total, health-related services, and social-related services).¹²

The underlying assumption of our dynamic model is that the labour participation of the traditional caregiver (women over 40 years of age) increases LTC expenditures (via higher use of formal care), which in turn expands GDP pc. However, such effects might differ across countries. In our data, we can, consistently with Reher (1998) and Kohli et al. (2005), differentiate between Southern European countries (with strong family ties) from Northern European countries (with weaker family ties).The direct relationship between LTC expenditures and GDP pc results from the combination of community-based services, residential care and support to informal caregivers (e.g. respite services), which in turn, can constitute an important source of employment, and consequently, of economic growth in the years to come (Spasova et al. 2018). Ikkaracan and Kim (2019) performed a macroeconomic simulation study with data from 45 countries to compute the amount of employment needed to meet specific Sustainable Development Goals by 2030. They found that the long-term care sector would require the creation of 29.6 million jobs. Specific country projections suggest that the demand for formal caregivers (home care and residential homes) would increase by 980,000 new workers in 2050 for Australia (Mavromaras et al. 2017) and would lie between 16 and 26 thousand employees in 2030 for Poland (Golinowska et al. 2014).

Next, we formulate a second specification that considers three equations in which the dependent variables are \(Y_{i,t} = \left\{ {{\text{LTC}}_{i,t} ,{\text{HC}}_{i,t} , {\text{GDPpc}}_{i,t} } \right\}\), and more specifically public LTC expenditures in per capita terms \({\text{LTC}}_{i,t}\), public healthcare expenditures in per capita terms \({\text{HC}}_{i,t}\), and gross domestic product per capita \({\text{ GDPpc}}_{i,t}\). The previous empirical literature has shown that LTC spending may help reduce the onset of unmet needs and, hence, reduce health utilisation (Allen and Moor 1997; Desai et al. 2001; Lima and Allen 2001). Similarly, an early post-discharge period after hospitalisation results in approximately 20% of complications that involve re-hospitalisation (Foster et al. 2003; Naylor et al. 2007), although these adverse effects can be avoided with more formal support (home care). LTC spending can in turn facilitate a smoother transition from hospital to home. Likewise, health care expenditures (HCE) can impact on per capita GDP given that some effective interventions improve health, which in turn can exert subsequent effects on labour supply and productivity, boosting GDP.¹³

Our panel data specification imposes as a restriction that the coefficients \({\rm A}\left( l \right)_{i1}\) are equal for all countries, though we add country fixed effects (\({\rm A}\left( l \right)_{i0}\)) to our specification to control for time-variant cross-country effects. Nevertheless, one of the limitations of including the fixed effects is that they are usually correlated with the regressors throughout the lag of the dependent variables (Blundell and Bond 1998). The Helmert transformation, which consists of applying forward mean differencing,¹⁴ is used to maintain the orthogonality between the regressors and their lags, allowing the mentioned lags to be used as instruments. Furthermore, GMM method is used for efficiency purposes (Holtz-Eakin et al. 1988). To determine the lag structure, we draw on Hansen’s J statistic¹⁵ and Andrews and Lu (2001) who proposed the use of consistent Moment and Moment Selection Criteria (MMSC). Additionally, we verify the stability condition of the model (Hamilton 1994)¹⁶ we perform a battery of Granger causality tests to determine whether variable \(Y_{1t}\) carries any information about another variable (\(Y_{2t}\)). The results of these tests help us to establish a causal order of the variables in the system.¹⁷

Finally, we distinguish between Northern and Southern countries, considering the interdependencies among countries in each group. The latter reduces considerably the dimension of the panel. To overcome this limitation, we rely on the estimation of Bayesian panel-VAR model (Doan et al. 1984; Canova and Ciccarelli 2004; Koop and Korobilis 2019). In the Bayesian panel-VAR model, the parameters are assumed to be random variables, characterized by an underlying probability distribution (Doan et al. 1984). To consider the full interdependencies between Northern and Southern European countries (in their respective panels), we draw upon a partial pooling analysis (Canova and Ciccarelli 2013).

We use a panel dataset that covers 27 countries for the period 2002–2015.¹⁸ We exploit time and cross-country variation, whereas the estimations for the group of Northern and Southern countries have been computed for the sub-period 2009–2015, due to data limitations.¹⁹ The country sample has been selected to maximize the number of years with available information (see Table A2). All variables used in the econometric analysis come from the OECD Health Data and Social Protection Data for OECD countries and are listed in Table A3. All economic variables are measured in US dollars, constant prices, are adjusted for purchasing power parities (PPP) and refer only to public expenditure (e.g. government or compulsory schemes). We use GMM panel-VAR with one lag and one to four instruments for estimations using the whole sample, but due to the reduced time dimension of the panel for Northern and Southern countries, we have employed Bayesian panel-VAR estimation.

Dependent variables (see Table A1).

The coverage and comparability of LTC spending estimates have improved with the implementation of a “System of Health Accounts 2011” (OECD 2017) which provides a framework for the measurement of health and LTC spending. However, in-depth analyses of data submissions suggested that full comparability of LTC spending data across OECD countries is still hampered to some extent (Mueller and Morgan 2017).²⁴

Figure A1 (Appendix A) depicts a linear association between per capita GDP and LTC spending, with a flattening out effect explained by two country outliers, namely Luxembourg and Norway. This evidence is consistent with the hypothesis that LTC is a normal good, and its investment increases with a country’s economic development. Figure A2 shows that consistent with expectations, female labour market participation exerts a steep effect on LTC spending. Finally, in Figure A3, we find evidence of a positive association between health spending and LTC spending, though it tails up at higher levels of spending.

Pre-estimation tests. To estimate a panel-VAR model, an important condition is for the variables to be stationary.²⁵ Accordingly, we employ two-unit root tests for panel data to test for stationarity, namely the Harris-Tzavalis (1999) and Im-Pesaran-Shin (2003) tests. The null hypothesis is that panels contain unit roots and are stationary. Importantly, the results of both tests (see Table B1) strongly suggest that all the variables (LTC expenditures, HCE, female labour participation and GDP pc expressed in logarithms) do not follow a unit root process. Hence, non-stationarity is not a concern in our estimates. However, one potential concern is the presence of possible cross section autocorrelation resulting from common factors (Levin et al. 2002). In those cases, we subtract the average of the group at each time for each time series.

The decision on the order length of the panel-VAR is based on the tests of MMSC proposed by Andrews and Lu (2001).²⁶ Table B2 illustrates the results of the MMSC and Hansen’s J statistic using four instruments for each of the endogenous variables (from the first to the fourth lagged dependent for both models). The results are shown using one, two, and three lags, respectively. In all cases, the model with one lag is the one that simultaneously minimises the three criteria and corroborates the suitability of the instruments used.

After performing the Granger causality Wald tests for each equation of the underlying panel-VAR model, we examine the stability condition of panel-VAR. The results of the Granger causality test are displayed in Table B3, although predicted associations should be considered with reservation and tested with additional analysis.

Tables B5 and B6 show the variance decomposition of the forecast error for the two proposed models (including different variants depending on the use of the different types of LTC expenditures and HCE) and different groups of countries.

Variance error decomposition for long-term care expenditures suggest that female labour participation (of women over 40 years of age) explains about 80% of Forecast Error Variance (FEV) of LTC expenditures in Nordic countries for both total and health- and social-related expenditures. This percentage is much higher than the one observed for all countries (1.9% for total LTC, 6.5% for health LTC, and 16.7% for social LTC) and especially for Southern countries (3.8%, 15.7%, and 14.9%, respectively). In contrast, social LTC expenditures explain 12.1% of the FEV of female labour participation in Nordic countries, compared with a percentage lower than 1% in the group of all countries and Southern countries. Finally, when we examine total LTC expenditures, we find that it explains 23.9% of the FEV of total health expenditures for the group of all countries—an amount below the 70.3% of Nordic countries. For the entire sample, LTC expenditures explain 19.71% of the FEV, in contrast to 3.91% for social LTC expenditures.²⁷

Variance error decomposition for caregiving spillovers suggest that total HCE explains 5.8% of the FEV of total LTC expenditures, and 3.47% in Southern countries, but it is not significant for all countries. On the other hand, an LTC shock accounts for 5.11% of the variation of HCE for all countries (19.71% in Nordic countries, 4.48% in Southern countries). Similarly, LTC shocks explain a higher percentage of the variation in outpatient, inpatient and medicine expenditure in Nordic countries. A result of particular interest is the higher contribution of LTC shock to the variance of GDP pc: 18.88% for all countries, 27.63% for Nordic countries and 15.35% for Southern countries (as compared to the effect of HCE shock: not significant for all countries, 2.61% for Nordic countries and 3.32% for Southern countries).

An important visual examination is reported by  the impulse response functions (IRF) of the response variable (in logs) to a one standard deviation shock in an impulse variable (in logs). Each figure represents the dynamics of the response, as well as the lower and upper confidence intervals at a 95% significance level.²⁹

Figure 1 plots the IRF of a one standard deviation shock in the female labour participation over LTC expenditures (total, health, and social) for all countries and for the group of Northern and Southern countries. Response functions depict the evolution of the response variable during the subsequent periods (years) resulting from a change in the impulse variable (female labour participation) by one standard deviation. We find that an increase in female labour participation in one period leads to a decrease in the total LTC expenditures in the subsequent period. When we examine the specific effects of health and social LTC expenditures, the described pattern suggests an initial increase (of 0.03% and 0.02%, respectively).

In Northern countries, we find that an increase in female labour participation gives rise to an increase in total LTC expenditures by 0.04%, which subsequently increases until it reaches 0.25% in five periods. For health LTC expenditures, we observe a V-shaped pattern with an increase of 0.2% in the first period, followed by a decrease in the next period. In contrast, in Southern countries, the effect on total and social LTC expenditures is imperceptible, and the increase in health LTC expenditures is the only item that deserves noting (0.01% in the first period, with an increasing tendency).

Figure 2 compares the IRF of total LTC expenditure over GDP pc in the two models. A one standard deviation increase of total LTC expenditure increases GDP pc between 1.25% (model 1) and 1.5% (model 2) after 5 years and would reach almost 2.5% for Northern countries. GDP pc growth appears to stabilize in the last two years for the all-country sample, whereas for the Northern and Southern countries, it still shows a steep slope.

Figure 3 shows the IRF of total LTC expenditure over HCE and its components. One standard deviation increases of total LTC expenditure lead to a reduction in total HCE by 0.5% after five periods (− 0.5% for Northern countries and − 0.2% for Southern countries). Yet, the impact over inpatient expenditure is higher than for outpatient expenditure. Finally, a one standard deviation increase of LTC expenditure reduces medicines expenditure by almost 0.2% in Northern countries, even though the overall impact is roughly − 0.1%.