The cost of informal care in Europe:
New estimates based on the well-being valuation method
Ulrike Schneider and Julia Kleindienst
1. Introduction
There is growing awareness that the provision of informal long- term care, an indispensible part of our health systems, carries large, though partially hidden, costs. Various studies have provided partial cost estimates of medical expenses, the opportunity cost of carers’ time and/ or foregone earnings. Colombo et al (2011), for instance, estimate the wage difference between carers and non-carers in the UK. Nepal et al (2011) compare the economic lifetime perspectives of female primary caregivers in Australia to their non-caring counterparts. For a more comprehensive cost measure, several studies (e.g.Meijer, de Brouwer et al.(2010))have asked carers how much they would need to be paid to provide an extra hour of care (or how much they would pay to avoid it). Such contingent valuations(e.g. Carmichael and Charles (2003)) studies rely heavily on respondents being willing and able to perform a rational cost benefit calculus on the care they to provide family members. In practice, people are not used to thinking about this problem in market terms. They are likely to show sequencing or anchoring bias in their answers, or even give protest answers because it feels inappropriate to be monetising family care. In addition, an emergent literature is pointing to the existence of important benefits, not just costs, to caregivers, though estimates of their magnitude are scarce. Consequently, there is value in exploring alternative valuation methods for informal care.
Our paper will explore one methodological alternative. It follows Van den Berg and Ferrer-i-Carbonell(2007) in using the subjective well-being valuation method to estimate the cost of informal care. The underlying model assumes that life satisfaction increases with available income and is reliably reported by carers. Obviously, these are strong assumptions, but previous research on reported life satisfaction (e.g., Lepper(1998)) appears to support them. That informal care can be both a burden and a source of satisfaction, appears well-established in the literature (see, for instance, Nolan (1996)). Wetake an agnostic view of the balance of these effects and estimate the net effect of changes in care provision on a caregiver’s happiness. Combining this with the effect of income on subjective well-being yields a monetary measure of the life satisfaction that informal care costs. Since the data includes information on the hours of care provided by each respondent, the paper can estimate the monetary value of the net burden or benefit of providing an hour of informal care.
Some background on the method used is provided belowin section 2. That section also outlines expectations for some key variables. The third section will give an overview of the dataset used, key variables examined and important choices made in the specification of these variables. It shows summary statistics for both the sample as a whole, comparing carer and non-carer respondents, and for the carer sub-group examined in more detail. The fourth section presents estimation results for both groups, and their implied shadow value of informal care. These are discussed the fifth section. We highlight important findings, compare them to previous literature and attempt to explain the key results. This section also reviews limitations of the estimation method and dataset, essaying a first evaluation of the reliability of results. Finally, we conclude by summarising our answer to the question after the magnitude of shadow costs/ benefits to informal carers, note a few possible extensions of the work, and discuss some policy implications of the findings.
2. Method
The subjective well-being valuation method uses an empirical model that models life satisfaction as a function of income, the variable of interest (in this case, informal care provided) and controlling covariates. The baseline model can be described by the equation
Wij = α + β1Xij + β2Cij + β3yij + β4Dj + εij
whereWij indicates the subjective well-being of individual i in country j, Xij describes a vector of individual determinants of life satisfaction, Cij is a vector that contains hours of care provided, yij is a measure for household income, In addition, there are dummies accounting for country-specific fixed effects. These variables are presented in greater detail below. The unobserved determinants of life satisfaction are captured by the term εij. α, β1, β2, β3, β4 represent the (vectors of) coefficients to be estimated. The full-sample estimation also includes interactions between care variables and health status, and the carer sub-sample includes indicator variables for the relationship between care provider and recipient.
The ratio of the estimated coefficients β2 /β3 represents the marginal rate of substitution between care hours and income. It is the shadow price of an hour on informal care. Using income as a base good requires only the assumption that money enters positively into the utility function as it can be transformed into valued consumption goods. It is not necessary that a reallocation of time away from care actually enable the respondent to increase their disposable income. Thus, estimates are possible even for those respondents that are no longer working (the majority of the sample). It should only be inappropriate for people who can no longer derive utility from income, perhaps because they are bed- bound and in very poor health, and incapable of enjoying anything beyond the care they are already receiving. Such respondents are not in this sample, however. The selection bias leaving out the physically worst- off subjects is an unfortunate, but unavoidable feature of the data. In order to avoid assuming cardinality of subjective life satisfaction, the model estimated is an ordered probit, with robust standard errors (clustered at the country level).
The subjective well-being valuation method is well-tested in environmental economics and other fields that require cost estimations outside of market contexts. That work, along with supporting literature on life satisfaction or happiness, provides guidance for the variables needed in the model. Specifically, life satisfaction is regressed on age, gender, health, employment and marital status, as well as household income, as these factors have been found significantly associated with life satisfaction in previous studies.
Life satisfaction has been shown to be systematically related to age. Blanchflower and Oswald (2008), (2009) present evidence for ‘u-shaped’ subjective life satisfaction over the life cycle, regardless of cohort effects. Employment status is another crucial control. Although, unsurprisingly, over half the SHARE sample is retired, this may include respondents that were forced into retirement rather than choosing to go or early retirees. They, along with the unemployed, would be expected to have lower life-satisfaction than their in-work counterparts. Lucas et al. (2004)use German panel data to show that losing one’s job has persistent negative effects on life satisfaction.
Similarly, those authors (2005)show persistent effects from divorce. Therefore, dummy variables for being single, married or in a registered partnership, divorced or widowed are also included as control variables. Clark et al. (2008)also find weak evidence for a gender effect in adaptation of life satisfaction to changing life circumstances, requiring another control.
Health status as a control is a potentially difficult case, owing to adaptation effects. Theoretical approaches like Sen’s capability concept argue at least for an indirect effect of health on life satisfaction. In addition, previous applications of the method (see van den Berg, 2007) argue for the importance of some health proxy as a control variable. In our sample, life satisfaction is robustly and strongly associated with good health, regardless of the measure of health status explored.
The inclusion of absolute income, crucial for calculating the shadow cost of care, may also be contentious, given the fact that adaptation and reference group choice are important in valuing money for life satisfaction (see, for instance Easterlin(2001)or Ferrer-i-Carbonell(2005)). The large body of work that uses this approach (see Clark et al. (2008)for a survey), however, suggests that this is not an insurmountable problem. Life satisfaction in our sample is indeed increasing in income. To capture the diminishing marginal utility of income, and reduce the effect of outliers, we include income in its natural logarithm form.
That care provision systematically affects well-being is an implicit hypothesis of this paper, and the basis for using the SWBV method to find its shadow price. Van den Berg et al. (2007) find a net negative effect on well-being. This is unsurprising, given the amply documented physical and emotional burdens care-giving imposes (see, for instance, Wilson et al,(2007)). It is not inevitable, however. Brouwer at al. (2005) show that in a large sample of Dutch caregivers, carers gained process utility form the provision of informal care to their relatives. Their survey results suggest that happiness would decline if care were provided by someone else, pointing to carer motivation beyond pure altruism (perhaps a warm glow effect). If the benefits of care provision outweigh the costs imposed, the net effect of care provision on well-being may well be positive.
3. Data and Measures
This paper uses the second wave of data, gathered 2006-2007, for the Survey of Health, Ageing and Retirement in Europe (SHARE) (c.f. (Börsch-Supan and Jürges 2005)). Some summary statistics, contrasting carers and non-carers are shown in table 1 below. Our sample (restricted to the 33498 respondents with sufficient data available) contains more women than men. Unsurprisingly, given the mean age of 67, over forty percent of them are retired. The vast majority (over 70%) are married, or living with a partner. Having discussed expectations for some of the major variables in the previous section, we now provide more detail on the measures used and key features of the data.
The subjective well-being assessment is fairly standard. Respondents are asked “On a scale from 0 to ten, how satisfied are you with your life?” The dependent variable displays reasonable variation, although the distribution shows the typical left skew, with over 90% of responses given as five or above. Some respondents (about 1% of the sample) could not answer the question because the questionnaire was being completed by a proxy. This suggests that the sample may exclude those with the worst health status, as these are no longer able to complete the survey. In the section on health, over 94% of respondents answered themselves (rather than by proxy). Life satisfaction is significantly higher for the carer group.
Health status is an important control variable in the regressions run. The SHARE dataset provides a wealth of information on respondents’ health. The crucial choice for this project was between using self- assessed (subjective) health, or a more objective measure, such as the number of chronic conditions diagnosed. While the former provides a more comprehensive and nuanced measure than the number of diagnosed illnesses, it has two limitations. Firstly, not all respondents chose to answer that question, limiting the usable dataset if it were to be included. More importantly, however, it is plausible that idiosyncratic personality factors affect the framing of both subjective satisfaction and health assessments. Then including the latter as a regressor would introduce bias. The number of chronic conditions, on the other hand, fails to capture the severity of mobility limitations implied. It has been suggested, that grip strength (measured for both hands in the survey) is a useful proxy for physical health (see Andersen-Ranberg, K., I. Petersen, et al. (2009)), although measures do show a North- South gradient even after controlling for age and chronic conditions. Given our inclusion off country fixed effects, this is not a concern here.
Each respondent’s grip strength was measured four times by interviewers. Following the authors above, the grip strength variable used is the maximum grip strength attained in these four measurements. These measurements range between 0 and 86, with a left- skewed distribution. The missing values are overwhelmingly (over 60%) from respondents unable to take the measurement because they were ill, injured or not personally answering questions. As such, they are underrepresented in the carer subsample. In other words, missing values for grip strength prevail in the part of the sample not identified as providing care.
Carers have statistically significantly higher grip strength than do the non- carers. The same pattern holds for other measures of health status, which are all highly correlated amongst each other and cannot be explained by the age profile of the groups, which is very similar. As may be expected, then, we find a health selection effect in care-giving.
The survey asks about many possible sources of income. However, merely summing across the income sources from responses in the dataset would further limit the usable dataset, due to the prevalence of missing values. To retain as much data as possible, the imputations of Christelis(2011)for SHARE data are used. The author provides five sets of generated variables. In order to make efficient use of the information, all five are used in Stata’s “Multiple Imputations” estimation. The income variable generated at the individual level includes transfers, pension payments, income from savings or financial investments. Household level income also includes net income from property, and is the measure used. Unfortunately, intra-household distributions, reflecting the disposable income actually under any respondents’ control, are impossible to make out through this dataset. The dataset shows a wide range of household incomes, higher for carers than non- carers on average.
The final important measure pertains to the definition of caregiver. Carers (about a third of the sample) were defined as those who answered that they had spent time helping (this excludes child care) someone outside their household. Two unavoidable limitations are the lack of a similar question about the amount of help provided to a household member, excluding the many carers that help their co-habiting spouses, and the lack of information about the kind of care provided. In particular, the health and functional status of the care recipient are not recorded, so that the physical and emotional burden, that both vary wildly between care-giving arrangements, may be only indirectly investigated.
Education is controlled for by including the ISCED category of the respondent’s highest educational attainment. Given the difference in educational attainment between carers and non-carers, it was felt to be an important control. It has a positive, but not statistically significant association with life satisfaction. Since the more standard specification of three categories of educational attainment was not statistically significant either, only the ISCED variable was included for greater parsimony.
1
SummaryStatistics / carers / non-carersDescription / mean (sd) / min / max / mean (sd) / min / max
Dependent Variable
lifesatisfaction / Subjective life satisfaction on a scale from 1-10 / 7.81 / 0 / 10 / 7,44 / 0 / 10
(1.59) / (1.85)
Independent Variables
gripstrength / Highest grip strength measurement; health status / 36.69 / 0 / 100 / 33,19 / 0 / 92
(11.89) / (12.11)
age / 66.80 / 29 / 106 / 66,83 / 26 / 107
(10.60) / (10.53)
gender / 1 forfemales / 0.56 / 0 / 1 / 0.55 / 0 / 1
(0.497) / (0.497)
employed / 0.38 / 0 / 1 / 0.258 / 0 / 1
(0.485) / (0.437)
retired / 0.424 / 0 / 1 / 0.527 / 0 / 1
(0.494) / (0.499)
unemployed / 0.033 / 0 / 1 / 0.025 / 0 / 1
(0.179) / (0.155)
disabled / 0.035 / 0 / 1 / 0.036 / 0 / 1
(0.183) / (0.186)
homemaker / 0.112 / 0 / 1 / 0.135 / 0 / 1
(0.317) / (0.342)
single / 0.041 / 0 / 1 / 0.042 / 0 / 1
(0.143) / (0.144)
married / married or living with a partner / 0.726 / 0 / 1 / 0.706 / 0 / 1
(0.481) / (0.479)
divorced / 0.075 / 0 / 1 / 0.051 / 0 / 1
(0.19) / (0.17)
widowed / 0.087 / 0 / 1 / 0.142 / 0 / 1
(0.402) / (0.258)
education / ISCED categories / 2.91 / 1 / 6 / 2.51 / 1 / 6
(1.40) / (1.39)
netincome / Yearly household income in Euros / 133354 / 0 / 4656233 / 85840.19 / 0 / 4656233
(208888.1) / (152730.2 )
N / 10551 / 22947
1
Table two shows details for caregivers only. Care hours reported have been converted into the average hours provided per week in the last year, for better comparability. The average number of care hours provided is low at about twelve a week, especially given the fact that most respondents are not employed. The variance is very large, however. The table includes not just average weekly care hours but also the hours provided by two groups of intensive carers, providing between 10 and 30 hour a week, and over 30 hours a week, respectively. Given the non-linearities in the effect of care provision found on life satisfaction, the full specification includes these three groups of care hours.
A set of category variables captures the relationship between care-giver and recipient. The two largest groups are those caring for a non- relative and for a parent, accounting for over half of all carers together. For an alternative specification, care relationships are grouped by generation, although there is some ambiguity in assigning non-relatives to these groups. The largest group comprises those caring for a member of their generation (mostly a partner or ex- partner, followed by siblings).
SummaryStatistics / carersDescription / Q1 / median / Q3
total carehours / average weekly carehours in the last year / 0.47 / 2 / 7
highcarehours / weeklycarehoursmorethan 30 / 36 / 49 / 70
medium carehours / weeklycarehours 10-30 / 14 / 15.5 / 21
mean (sd) / min / max
parent / Main care recipient is a parent / 0.249 / 0 / 1
(0.431)
in-law / Main care recipient is a parent/child in-law / 0.093 / 0 / 1
(0.289)
partner / Main care recipient is a spouse/ partner / 0.039 / 0 / 1
(0.197)
sibling / Main care recipient is a sibling / 0.061 / 0 / 1
(0.24)
child / Main care recipient is a child / 0.224 / 0 / 1
(0.419)
other / Main care recipient is another relative / 0.067 / 0 / 1
(0.25)
friend / Main care recipient is not a relative / 0.262 / 0 / 1
(0.44)
older / care recipient is of an older generation / 0.243 / 0 / 1
(0.429)
younger / care recipient is of a younger generation / 0.276 / 0 / 1
(0.444)
same / care recipient is of the same generation as carer / 0.389 / 0 / 1
(0.487)
n= 9794
4. Results
Below are regression results for the ordered probit estimation for the entire sample. Reassuringly, the effect on well-being associated with a rise in income is (highly robustly) positive and statistically significant. This is a minimum requirement for the shadow cost calculation to be plausible. In all specifications, we find a positive effect of providing care on subjective well-being. Reported life satisfaction seems to increase robustly, but non-linearly with the hours of care provided, as the effect for additional hours above 10 is negative. Thus, the benefits are practically wiped out for very intensive care-giving arrangements (those over thirty hours a week), presumably overwhelmed by the more familiar burdens associated with care, that increase as hours committed to it rise.