Fundación de Estudios de Economía Aplicada

Institutions, Health Shocks and Labour Outcomes Across Europe by * Pilar García Gómez DOCUMENTO DE TRABAJO 2008-01 Serie Economía de la Salud y Hábitos de Vida CÁTEDRA Fedea – la Caixa

January 2008

* Departament d’Economia i Empresa and CRES, Universitat Pompeu Fabra. Barcelona.

JOB MARKET PAPER

Pilar García Gómez^*

Departament d’Economia i Empresa and CRES, Universitat Pompeu Fabra. Barcelona, Spain.

September 2007

Abstract

This paper investigates the relationship between health shocks and labour outcomes in 9 European countries using the European Community Household Panel. In order to control for the non-experimental nature of the data I use matching and matching combined with difference-in-differences techniques. My results suggest that there is a significant effect running from health to the probability of employment and to income: individuals who suffer a health shock are significantly more likely to leave employment, and in several countries this is associated to a significant reduction in some types of income. There are differences in the estimates across countries, with the largest employment effects being found in the Netherlands, Denmark and Ireland, and the smallest in France, Italy and Greece. The differences in Social Security arrangements help to explain the differences in the estimates for the effects of the health shocks.

JEL classification: C23, I12, J60

Keywords: Health shocks, disability, employment, matching, ECHP

^* This paper derives from the project “ Instituciones de protección social y la relación renta-salud en la Unión Europea, supported by Asociación Española de Economía de la Salud and Química Farmacéutica Bayer through the XV Beca de Investigación en Economía y Salud. Additional support from Ministerio de Educación projects SEJ2005-09104-C02-02 and SEJ2005-08783-C04-01 is thankfully acknowledged. The author gratefully acknowledges help, support and useful comments from Ángel López Nicolás, Ana Tur Prats, David Casado, Marcos Vera, Xander Koolman and Cristina Hernández as well as participants at the Health Econometrics and Data Group Seminars, 6^th iHEA World Congress, XXVII Jornadas AES Economía de la Salud. The usual disclaimer applies. Correspondence to: Universitat Pompeu Fabra, c/ Ramon Trias Fargas, 25-27. 08005 Barcelona. Spain. E-mail: pilar.garciag@upf.edu

1. Introduction

The increase in the rates of recipients of disability support observed during the 1990s in almost all OECD countries has raised concerns about the labour outcomes of people with adverse health (OECD, 2003). The relevant policies often try to satisfy two possibly contradictory goals. On one hand, they have to guarantee that individuals who are or become disabled do not endure economic hardship, and thus provide some insurance for the potential income losses. On the other hand, they also aim to avoid the exclusion of disabled individuals from the labour market by, among other measures, encouraging participation.

When studying the relationship between health and labour outcomes, the literature has been traditionally focused on older workers. Several studies (Bound et al. (1999), Au et al. (2005), Disney et al. (2006), Hagan et al. (2006) and Rice et al. (2006)) focus on individuals older than 50 and show that decreases in health status have explanatory power for retirement decisions. Riphahn (1999) finds that health shocks increase the probability of unemployment by 84% and the probability of dropping out of the labour force by 200% for individuals aged 40 to 59 in Germany. In the same line, Jiménez-Martín et al (2006) find that, for Spanish workers aged between 50 and 64, the probability to continue working decreases with the severity of the shock. Messer and Berger (2004) use the American Health and Retirement Survey and find that permanent adverse health conditions reduce both wages (8.4% for males and 4.2% for females) and hours worked (6.3% for males and 3.9% for females). Moreover, the bigger effects of health on labour outcomes are found on prime-age individuals, as the peak of loss of wages after the onset of a permanent illness occurs at ages 40-49 for males (wages are 12.1% lower) and 30-39 for females (wages are

9.2% lower). Smith (2004) also finds that for individuals older than 50 suffering a health shock there is a 15% decrease in the probability of working, and although this effect diminishes over time, it remains substantially high at nearly 4% 5 years after the shock. Stewart (2001) examines the effects of health limitations on the kind or amount of activity that Canadian individuals can do at work and finds that those with impaired health have significantly longer unemployment spells.

The number of studies that focus on the role played by health on labour market transitions for younger individuals is smaller. Among these, Lindeboom et al. (2006) estimate an event history model for transitions between work and disability states and find that the effects of health shocks on employment are not direct, but rather act through the onset of a disability, which increases by 138% after the onset of a health shock. At the same time, the onset of a disability at age 25 reduces the employment probability at age 40 by 0.205. Dano (2005) finds that there are both short and long run effects on the probability of being employed for Danish males after being injured in a road accident, and that this effect holds even when individuals receiving disability benefits are excluded from the analysis. García Gómez and López Nicolás (2006) analyse the effects of a health shock on the probability of leaving employment and transiting out to different states for the Spanish population. They find that suffering a health shock decreases by 5% the probability of remaining in employment and increases by 3.5% the probability of transiting into inactivity.

Thus previous literature seems to confirm the existence of an effect of health events on labour market outcomes, but there is a lack of consensus on their magnitude. My contention in this paper is that the international differences in estimated effects partly reflect the emphasis that each country places on the two potentially conflicting goals of protecting income and encouraging participation mentioned before. This paper attempts to contribute to this area of research by estimating the effects of health shocks on a set of labour outcomes for different European countries, and subsequently relating the differences in estimates to variations in institutional factors across these countries.

It is well known that using observational data to estimate the causal effect of health events on labour outcomes is plagued with potential biases (Lindeboom, 2006). In terms of methodology, my strategy is motivated, among others, by Smith’s (2004) use of longitudinal information for representative samples of the US population so as to be able to condition on past health shocks before evaluating current changes in labour status and income. So in this paper I resort to the best source of longitudinal information on health and socioeconomic characteristics for the European population available to researchers: the European Community Household Panel (1994-2001, hereafter ECHP). I will condition on past health and labour status to evaluate the effects of changes in health. While the spirit is the same as in the Smith’s study, my specific methodology consists in matching individuals who experience a health shock with identical individuals in a control group. In this sense I follow the recent usage of the propensity score matching methods in the context of health shocks by Lechner and Vázquez Álvarez (2004), Frölich et al. (2004), Dano (2005) and García Gómez and López Nicolás (2006).

Therefore, this paper contributes to the existing literature in several directions. First, it extends the knowledge of the relationship between health and labour outcomes on the working population, using a homogeneous empirical framework for nine European countries. Second, this homogeneous framework allows me to relate the empirical estimates to differences in Social Security arrangements across these countries. To the best of my knowledge there is no other paper containing this type of comparative analysis for the countries concerned.

My results suggest that there is a significant effect running from health to the probability of employment and to income: individuals who suffer a health shock are significantly more likely to leave employment than those who do not, and in several countries this is associated to a significant reduction in some types of income. As expected, there are differences in the estimates across European countries, with the largest employment effects being found in the Netherlands, Denmark and Ireland, and the smallest in France, Italy and Greece. The reduction in the likelihood of employment is paralleled by an increase in the probability of inactivity. This should be a cause of concern, as the outflow from inactivity is known to be close to zero (OECD, 2003).

The paper is organized as follows. In the next section I describe the institutional features related with the Social Security schemes for the group of countries included in the analysis (Denmark, Netherlands, Belgium, France, Ireland, Italy, Greece, Portugal and Spain). Section 3 discusses the methodology used to identify causal effects of health changes on labour outcomes. Section 4 discusses some features of the ECHP particularly relevant for this study. Section 5 presents the empirical results and assesses the sensitivity of the matching estimates obtained. Section 6 discusses the results and concludes.

2. Institutional Background

After the onset of a health condition, an individual can follow any of the following routes (Aarts 1996): i) work; ii) early retirement; iii) traditional disability insurance programs (sickness, general disability and work injury); iv) unemployment; v) means-tested programs for those illegible for any other option. This implies that it is not only disability policies but also the set of incentives provided by the wider Social Security system what determine the labour consequences of a health shock.

Figure 2.1 gives a first glimpse of the differences between the countries under study, as it shows the fraction of social expenditures over GDP and their breakdown into four main chapters. Note that Ireland (Denmark) is the country with the lowest (highest) fraction of GDP devoted to social expenditures. In the case of Ireland, the difference with respect to the rest of countries is mainly at the expense of old age benefits.

Figure 2.1. Composition of social expenditure (as %GDP). 2001

Source: own-elaboration using data from Eurostat (2007)

In order to obtain a closer picture of the relevant differences among countries, Table 2.1 summarizes the main features of Social Security in the interrelated spheres of disability, unemployment and retirement. Note first the striking differences in the way in which countries establish eligibility criteria for disability benefits. Some countries define disability in terms of a reduction in the individual work capacity (Denmark, Ireland, Italy and Spain), while others do it in terms of a reduction in earnings capacity (Belgium, France, Greece, Netherlands and Portugal). But even among countries that use the same concept, the minimum level of disability that entitles individuals to receive benefits varies widely: from 15% in the Netherlands to being permanent incapable for work in Ireland (OECD, 2003; MiSSOC, 2004). Table 2.1 also shows that some countries apply mandatory quotas obliging employers to have a certain proportion of disabled workers among their employees (7% Italy, 6% France, 2% Spain), or some sectors (3% in the public sector in Ireland and 5% for new recruitment in the public sector in Portugal). These quotas are absent in Denmark, the Netherlands and Belgium. Concerning the measures that aim to integrate disabled individuals into the labour market, we can also observe in table 2.1 that most countries allow a certain accumulation of disability benefits with earnings from work. The only exception is Ireland, where the invalidity pension requires permanent full incapacity (MISSOC, 2004).

[Insert table 2.1 around here]

Following the analysis in OECD (2003)¹, we can summarise the main components of the disability system into two dimensions. The “compensation” dimension reflects the characteristics of the main disability benefit scheme (coverage, minimum disability level, disability level for a full benefit, maximum benefit level, permanence of benefits, medical assessment, vocational assessment, sickness benefit level, sickness benefit duration and unemployment benefit level and duration). The second is the “integration” dimension, which reflects all the employment and rehabilitation measures (coverage consistency, assessment structure, employer responsibility for work retention and accommodation, supported employment programme, subsidised employment programme, sheltered employment sector, vocational rehabilitation programme, timing of rehabilitation, benefit suspension regulations and additional work incentives). From the group of countries considered here (and included in the OECD study), Denmark would be the country in which the integration component is the highest, whereas the lowest levels are found in Italy and Portugal. The champions in the compensation dimension are Portugal, Spain and the Netherlands, while the laggards in this dimension are France, Belgium and Italy.

Concerning the characteristics of unemployment insurance, table 2.1 shows that there are remarkable differences in i) the initial unemployment net replacement ratio, ii) duration of unemployment insurance, and iii) the average replacement ratio over the course of the unemployment spell. Note that Denmark, Netherlands and Portugal are the countries where both the initial and the average replacement ratio are higher. On the other end, Italy and Greece rank the lowest in terms of initial and average net replacement ratios. On the duration dimension, note that individuals are entitled to unemployment benefits from 6 months in Italy to an unlimited period of time in Belgium. It has been previously argued (OECD, 2003) that unemployment systems with long benefit payments are likely to reduce the pressure on the disability programme.

¹ In OECD (2003) the authors consider a different group of countries that did not include Ireland and Greece. For my purposes, I conjecture that these countries belong to the same cluster as the Mediterranean countries.

In several countries, (early) retirement benefits are as important as disability benefits for disabled persons of working-age (OECD, 2003). In Portugal, for example, one third of non-employed disabled persons receive an early or regular retirement benefit. This is probably due to the fact that access to early retirement benefits is easy, whenever the contribution requirements are fulfilled, because there is not medical examination. An important element to determine the importance of (early) retirement as an incentive to withdraw from the labour force is the eligibility age. In only 4 of the countries considered (Denmark, Greece, Ireland and Portugal) are individuals younger than 60 able to enter early-retirement. In Ireland, unemployed persons aged 55 or over who have been receiving either Unemployment Benefits or Unemployment Assistance for 15 months or more, may opt to apply for the Pre-Retirement Allowance (PRETA), which is a means-tested allowance which allows individuals aged 55 or over to retire from the labour force. In Portugal, 10 years of early retirement are available whenever there are 30 years of registered earnings or 20 years of registered earnings and the individual is a long-term unemployed. In Greece, mothers can retire at age of 50 with a reduced pension whenever they have a dependent or disabled child and have worked for more than 5500 days. Moreover, both men and women can retire with full pension at age 55 if they have worked at least 35 years. The route to early retirement available in Denmark for individuals aged 50 and over is based on grounds of social problems without any medical cause.

3. Empirical strategy

Outcomes of interest

In this paper I investigate labour market outcomes potentially affected by an adverse health shock. I am particularly interested in labour market transitions that are not led by the availability of old age retirement, so the population of interest are individuals below 60 years of age, as in most of the countries included in the analysis (early) retirement is not available for individuals younger than 60. The outcomes of interest are the probability of being in employment and the probability of being inactive or in other states, and the levels of income from labour and other sources.

Identifying the causal effect

The fundamental difficulty for my purposes is adequately dealing with the simultaneous determination of health and labour outcomes. One possible avenue would be finding instruments for self assessed health in a reduced form for labour outcomes in a reduced form for self assessed health. However, one source of potentially valid exclusion restrictions in this setting, i.e., detailed regional information, is absent from the ECHP due to the high level of aggregation employed for the regional markers.

A different option consists in conditioning on sufficient information to replicate random assignment to treatment (in this case suffering a health shock) and then use a parametric model where the treatment variable is one of the regressors. This is essentially the route taken by Smith (2004) in the sense that i) the onset of health shock is assumed to be exogenous (conditional on a set of observed covariates) in a labour outcomes equation. The approach of Lindeboom et al (2006) is similar in the sense that the authors estimate multinomial logits for different transitions between work and disability states where having had an accident is one of the explanatory variables, but in this case the specification also allows for any remaining unobserved heterogeneity affecting health shocks and labour market/disability outcomes.

My empirical strategy relies on the possibility to condition on sufficient observable information to obtain a credible counterfactual against which I may measure the impact of the health shock. Let indicate treatment (health shock) and lack of treatment respectively and let and denote the outcome of interest (labour status) for individual i with treatment and without treatment respectively. Since we will observe individual i either with treatment or without treatment, we cannot observe the causal effect of interest: Some features of such distribution are estimable, nevertheless. In particular, we may consider the Average Treatment Effect on the Treated

\[\mathrm{ATET} = \mathrm{E} (\mathrm{Y} _ {1} - \mathrm{Y} _ {\mathrm{o}} | \mathrm{T} = 1)\tag{1}\]

This magnitude measures how much the outcome of interest changes on average for those individuals who undergo the treatment (who suffer the health shock to be defined below). Clearly, simply computing the difference in the average outcomes of those in treatment and those out of treatment is open to bias, as there are observed and unobserved characteristics that determine whether the individual undergoes the treatment. That is,

\[\mathrm{E} \left(\mathrm{Y} _ {1} \mid \mathrm{T} = 1\right) - \mathrm{E} \left(\mathrm{Y} _ {0} \mid \mathrm{T} = 0\right) =\]

\[\mathrm{E} \left(\mathrm{Y} _ {1} \mid \mathrm{T} = 1\right) - \mathrm{E} \left(\mathrm{Y} _ {0} \mid \mathrm{T} = 1\right) + \mathrm{E} \left(\mathrm{Y} _ {0} \mid \mathrm{T} = 1\right) - \mathrm{E} \left(\mathrm{Y} _ {0} \mid \mathrm{T} = 0\right) =\]

\[\mathrm{E} \left(\mathrm{Y} _ {1} - \mathrm{Y} _ {\mathrm{o}} \mid \mathrm{T} = 1\right) + \mathrm{E} \left(\mathrm{Y} _ {0} \mid \mathrm{T} = 1\right) - \mathrm{E} \left(\mathrm{Y} _ {0} \mid \mathrm{T} = 0\right) =\]

\[\mathrm{ATET+BIAS}\tag{2}\]

Only if I can guarantee that the outcomes of the control group are equal on average to what the outcomes of the treatment group would have been in the absence of treatment does this consistently estimate the ATET. With non-random sorting into treatment and control such condition is rarely met.

Now suppose that by conditioning on an appropriate set of observables, X, the nonparticipation outcome is independent of the participation status T. This is the weak version of the unconfoundedness assumption, also called ignorable treatment assignment (Rosenbaum and Rubin, 1983) or conditional independence assumption (Lechner, 2000) or selection on the observables, which suffices when the parameter of interest is the ATET, as only assumptions about the potential outcomes of comparable individuals are needed to estimate counterfactuals.

\[\mathrm{Y} _ {\mathrm{o}} \perp \mathrm{T} \mid \mathrm{X}\tag{3}\]

This implies that

\[\mathrm{E} \left(\mathrm{Y} _ {0} \mid \mathrm{T} = 1, \mathrm{X}\right) - \mathrm{E} \left(\mathrm{Y} _ {0} \mid \mathrm{T} = 0, \mathrm{X}\right) = 0\tag{4}\]

In order to identify the ATET, the overlap or common-support condition is also assumed. It ensures that, for each treated individual, there are control individuals with the same X.

\[\operatorname * {P r} (T = 1 \mid X) < 1\tag{5}\]

Therefore under the assumptions stated in equations (3) and (5) above, we could estimate the ATET from the difference in outcomes between treated and controls within each cell defined by the conditioning variables X (see Blundell and Costa Dias 2002). Using the law of iterated expectations and the conditional independence assumption, the ATET can be retrieved from observed data in the following way

\[\begin{array}{r l} \mathrm {ATET = E(Y_ {1} | T = 1) - E(Y_ {0} | T = 1) = E_ {X} [(E(Y_ {1} | X,T = 1) - E(Y_ {0} | X,T = 1)) | T = 1] =} \\ & \mathrm {E_ {X} [(E(Y_ {1} | X,T = 1) - E(Y_ {0} | X,T = 0)) | T = 1]} \end{array}\tag{6}\]

Defining health shocks and treatment and control groups

My measure of health shocks is based on the responses to the question on self-assessed health in the ECHP “How good is your health in general?”. From the five possible responses (very good, good, fair, bad and very bad), I consider that the respondent has undergone an adverse health shock if he or she reports “fair”, “bad” or “very bad” in any given period, with the timing of the shock occurring sometime between the last period when he or she recorded any of the other two alternatives.

Since I wish to evaluate whether suffering a health shock in these terms leads to any change in labour outcomes, I want to rule out the possibility that any potential anticipation of the change in labour status causes the change in self reported health, therefore I adopt the following strategy –motivated by the procedures used by Lechner and Vázquez Alvarez (2004)- in order to construct the treatment and control groups:

1) Consider a window of three years for each observed individual. This creates 6 possible sequences of three years over the time span covered by my data. To these three years, regardless of the sequence, I refer as t=1, t=2 and t=3

2) For each sequence select individuals who are healthy (SAH good or very good) at t=1, the start of the sequence, and also are employed at t=1 and

3) The treatment group are individuals meeting selection criterion # 2 who report fair, bad or very bad health in t=2 and t=3. That is, those individuals who undergo a health shock after t=1 and for whom adverse health persists at least over t=3. The sequence of health states for these individuals is therefore GBB (Good, Bad, Bad)

4) The control group are individuals meeting selection criterion # 2 for whom I observe a GGG sequence of health states (Good, Good, Good)

It should be noted that by considering these sequences I disregard potential contemporaneous effects of health transitions on labour outcomes (and contemporaneous effects of employment transitions on health). However, this procedure ensures that the treatment occurs before the potential change in outcome thus offering some guarantee that what I identify is not reverse causality. Moreover, as individuals suffer a health shock before leaving the labour force, I am also ruling out justification bias in the health responses, except in the cases were there are anticipation effects consisting in individuals who foresee a transition out to employment and report a change in self assessed health one period in advance. It is not clear the extent to which this anticipation effect might be empirically important, but in any case assuming it away is the price I have to pay in order to be able to rely on the timing of events as a source of identification.

Plausibility of the Conditional Independence Assumption

The identification of the ATET by matching methods relies on the unconfoundedness assumption, which may or not may be plausible depending on the particular context, and which is inherently untestable as the actual counterfactual cannot be observed.

The plausibility relies on the availability of a detailed group of characteristics that allow us to match treated and control units. The data that I use contains a rich set of pre-treatment variables on demographics, educational attainment, job characteristics, household composition and socioeconomic information. Moreover, the information for both treated and control individuals was collected with the same questionnaire, and individuals were drawn from the same local market. Heckman et al (1997) stressed the importance to satisfy these two conditions in order to reduce the bias when applying matching estimators.

Secondly, I include pre-treatment outcomes within the vector of conditioning variables, either by including these pre-treatment outcomes in the propensity score or restricting the sample of controls to individuals who are identical in terms of pre-treatment outcomes. This procedure aims to include fixed unobserved factors in the outcomes of interest within the vector X of conditioning variables.

Imbens (2004) suggests that some support of the plausibility of the CIA can be obtained by estimating the ATET for the treatment of interest on a pre-treatment variable. This ATET should be null, so evidence suggesting otherwise should question the validity of the CIA assumption. I will estimate the ATET of my measure of health shocks on all the pretreatment outcomes.²

² Imbens (2004) also proposes a test based on the presence of multiple control groups (e.g. individuals that are eligible and individuals that are non-eligible for treatment), where one can estimate the ATET of interest considering one of these groups as the “treated” sample. In that case, the treatment effect is known to be zero, thus the non-rejection of the null hypothesis of non treatment effect makes more plausible the satisfaction of the CIA. However, due to the characteristics of my setting, there are not different control groups to be used.

Propensity score matching

The estimate of the ATET as shown in equation (6) turns out to be prohibitive in terms of data when the set of conditioning variables X is large. An alternative is to use the results of Rosenbaum and Rubin (1983, 1984) and condition on the probability of treatment as a function of X, the propensity score P(X), since the conditional independence assumption also implies that

\[\mathrm{E} (\mathrm{Y} _ {0} \mid \mathrm{T} = 1, \mathrm{P} (\mathrm{X})) - \mathrm{E} (\mathrm{Y} _ {0} \mid \mathrm{T} = 0, \mathrm{P} (\mathrm{X})) = 0\tag{7}\]

Therefore we could estimate the ATET from the differences in outcomes between treated and controls within each cell defined by values of P(X).

\[\begin{array}{r l} \mathrm {ATET = E(Y_ {1} | T = 1) - E(Y_ {0} | T = 1) = E_ {\mathrm{P(X)}} [(E(Y_ {1} | P(X), T = 1) - E(Y_ {0} | P(X), T = 1)) | T = 1] =} \\ & \mathrm {E_ {\mathrm{P(X)}} [(E(Y_ {1} | P(X), T = 1) - E(Y_ {0} | P(X), T = 0)) | T = 1]} \end{array}\tag{8}\]

Provided that the conditional participation probability can be estimated using a parametric method as a probit model, matching on the univariate propensity score reduces the dimensionality problem.

I construct the propensity score³ for suffering a health transition using a probit model in which the probability of belonging to the treated group is a flexible function of the following pre-treatment characteristics: age, gender, marital status, the logarithm of the household equivalent income, the number of children in the household, the percentage of total household income that comes from the individual’s labour income, whether she works full time, more than 10 years in the same firm, in the public sector, self-employed, the number of days lost because of illness in the last month, if she is severed limited by any chronic condition, limited by any illness, limited in daily activities by any illness or mental problem, indicators of health care utilisation and indicators of year and region. I have used a different specification for each country, ensuring in all cases the satisfaction of the Balancing Hypothesis. To test the latter, I have followed Dehejia and Wahba (1999, 2002). In particular, the observations are divided into strata until there are not statistically significant differences within strata in the estimated propensity scores between treated and controls. Subsequently, the null of no significant differences in the means of each covariate between treated and controls is tested within each stratum.

³ Propensity score estimates are available from the author upon request

Moreover, it is worth mentioning that due to the strategy chosen to define the groups of treated and control individuals, I am using the propensity score as a “partial” balancing score, as it is complemented by an exact matching on pre-treatment health and labour status.

Once the propensity score is estimated, I calculate the ATET using the Nearest Neighbour algorithm (Becker and Ichino, 2002) with replacement, which matches each treated individual with the control unit that is closest in terms of propensity score. Notice that each control unit can be used more than once as a match, otherwise some individuals could be matched with substantially different control units. The general formula for the matching estimators can be written as follows:

\[A T E T = \frac {1}{N ^ {T}} \sum_ {i \in \{T = 1 \}} \left[ Y _ {i} ^ {T} - \sum_ {j \in C (i)} w _ {i j} Y _ {j} \right]\tag{9}\]

where denotes the weight attributed to control individual j when comparing with treated individual i. In the case of the Nearest Neighbour algorithm, the weight is 1 if the control individual has the closest propensity score to that of individual i and 0 otherwise.

Abadie and Imbens (2006) show that the bootstrapped variance estimator is invalid for nearest neighbour matching. Therefore, I calculate analytical standard errors, assuming independent outcomes across units.

Propensity score matching combined with difference-in-differences

Again, it is important to stress that the ability of the estimator shown in (9) to retrieve consistently the ATET relies crucially on the conditional independence assumption. That is, that there are not unobserved variables that are correlated with both the exposure to treatment (the health shock) and the outcome. Therefore, if there remain any systematic differences between the outcomes of treated and control individuals, matching estimation will not recover the parameter of interest. However, if we can assume that these differences are time-invariant the availability of panel data affords the possibility to correct for the hypothetical failure of this assumption. Letting the superscript A and B denote the time periods before and after treatment occurs, the conditional independence assumption would now be stated in the following terms

\[\mathrm{Y} _ {0} ^ {\mathrm{A}} \text {-Y} _ {0} ^ {\mathrm{B}} \perp \mathrm{T} | \mathrm{X}\tag{10}\]

So that,

\[\mathrm{E} \left(\mathrm{Y} _ {0} ^ {\mathrm{A}} - \mathrm{Y} _ {0} ^ {\mathrm{B}} \mid \mathrm{T} = 1, \mathrm{X}\right) - \mathrm{E} \left(\mathrm{Y} _ {0} ^ {\mathrm{A}} - \mathrm{Y} _ {0} ^ {\mathrm{B}} \mid \mathrm{T} = 0, \mathrm{X}\right) = 0\tag{10}\]

And therefore, the combined matching-differences-in-differences ATET can be estimated in the following way from observed data (Blundell and Costas Dias, 2002 and Blundell et al, 2004)

\[\mathrm{ATET} _ {\mathrm{MDID}} = \frac {1}{N ^ {T}} \sum_ {i \in \{T = 1 \}} \left\{\left[ Y _ {i 1} ^ {t + 1} - Y _ {i 1} ^ {t} \right] - \sum_ {j \in \{C (i) \}} W _ {i j} \left[ Y _ {i 1} ^ {t + 1} - Y _ {i 1} ^ {t} \right] \right\}\tag{11}\]

where denotes the weight attributed to control individual j when comparing with treated individual i. Equation (11) shows that the estimate for the ATET is a weighted average of the differences in differences between each one of the treated and its matched control.

I shall use in order to check the robustness of the results for the income outcomes. I do not estimate for the labour outcomes, as I have restricted all the individuals to be employed in the first and second period, thus the ATET estimates obtained using matching or using matching combined with difference-in-differences are identical.

4. Data and descriptive statistics

The European Community Household Panel (ECHP) is an annual standardized longitudinal survey, which provides 8 waves of microdata about life conditions in most of the EU-15 members. The survey is based on a standardized questionnaire asked to individuals aged more than 16 selected from a representative household panel. The survey covers a wide range of topics, including demographics, income, social transfers, individual health, education and labour. The information contained in the ECHP can be compared both across countries and time.

The criteria used to select treatment and control groups implied that I could only use the countries for which information regarding the 8 waves was available: Belgium, Denmark, France, Greece, Ireland, Italy, Netherlands, Portugal and Spain. Table 4.1 shows the sample size for both treated and control groups, as means of the relevant variables, for each of the countries included in the study.

The outcome variables I consider are employment status and income from different sources. Information regarding employment status comes from the self-defined classification of main status, and I classify an individual as employed whenever he works part-time or full-time or he is self-employed. The three non-employment categories that I will consider are unemployed, retired and inactive. Therefore, I will not look at other categories such as student, doing housework, looking after children or other persons and in community or military service.

[Insert table 4.1 around here]

Table 4.1 shows that, in all countries, the percentage of individuals who work in the third period, without restricting the sample to the previous labour status, is clearly higher among the group of individuals who report being in a good or very good health status in all three periods. However, these differences are reduced when I select the individuals younger than 60. Moreover, in all countries the relative weight of the group of treated respect to the control group is reduced as this profile (GBB) is more frequently found among the elderly. The biggest decrease in sample size is observed when I restrict the sample to individuals working in the first period, as the percentage of individuals aged less than 60 whose health profile is GBB or GGG that work in the first period varies between 53% in Spain and 81% in Denmark. The countries in which the percentage of individuals working in the first period is lower are Spain, Italy, Ireland and Greece. On the other hand, once I have conditioned on working on the first period, the sample reduction when I condition on working also in the second period is small, as the remaining sample is around 95% of the previous one, which shows the importance of state dependence.

Table 4.1 also shows the demographic, socioeconomic and health status characteristics of both treated and controls in the working sample. We can observe that, with respect to the control group, the individuals who suffer a health shock have, on average, lower equivalent household income, a lower educational attainment, and are older. The health status in the third period is as expected worse for the treated than for the control group, although there are important differences across countries. The percentage of treated individuals who have been hospitalised in the last 12 months (inpaten) varies between 5.4% in Portugal and 26.1% in Denmark. This difference is also observed in the number of nights hospitalised (hospnight), as we can observe that the mean is less than 1 in Portugal, Denmark and Netherlands, but higher than 2.5 in Belgium and Greece. On the other hand, while only 2.5% of the treated in Italy declare to be severely limited by a chronic or physical or mental problem in their daily activities (dchronsev), the percentage is 16.7% in the Netherlands.

5. Results

Figure 5.1 presents histograms of the estimated propensity score for both the group of treated and control individuals for each country. The distribution of scores among treated and controls in each country does not differ, giving support to the conditional independence assumption. In fact, when I restrict the sample to the common support I lose only a small percentage of individuals from the control group (from 0.23% in Portugal to 3.38% in Denmark or 8.52% in Greece). In any case, all the estimates shown are obtained under the common support.

Figure 5.1 Estimated propensity score for treated and control individuals in each country

Note: ___ Treated --- Control

Effects on labour outcomes

Table 5.1 presents the nearest neighbour estimates of the ATET on the probability of employment, unemployment, inactivity, retirement, and a combined category of either retirement or inactivity. Although in most of the countries included in my analysis, individuals are not entitled to old age retirement benefits before the age of 60, I have included this outcome because, as it has been previously noted in the literature, retirement is not a well-defined state (Bardasi et al, 2002; Disney et al, 1994) and some individuals classify themselves as retired whenever have permanently exited from the labour market, while others unless they receive a pension. Moreover, there can be cultural differences across countries reinforced by the different routes available into retirement. For the sake of robustness analysis, I have estimated the same ATET’s using Gausian Kernel (reported in Table A1 in the Appendix) and the results are qualitatively indistinct to the ones obtained using Nearest Neighbour matching.

Table 5.1. Nearest Neighbour ATET on the probability of several activity statuses

Notes: Analytical standard errors in parentheses.

Table 5.1 shows that in most of the countries considered here, the individuals who suffer a health shock are significantly more likely to transit to a non-employment status than those who do not. Ireland is among the countries in which the drop in the probability of remaining in employment is the biggest. This is consistent with the institutional feature mentioned in section 2: disability benefits are not compatible with any kind of work. Therefore, individuals in Ireland who suffer even a partial disability are forced to leave the labour market in order to get any benefit from the social security. This hypothesis is reinforced by the fact that the top six incapacity codes reported by individuals with any kind of disability benefits are Back/Neck/Rib/Disc Injury, Anxiety/Depression, Other Incapacity, Arthritis/Rheumatism/Osteo Arthritis, Nervous Debility/Bereavement and Hypertension (DSFA, 2003).

The other two countries where the effect is the biggest are Denmark and the Netherlands. They share with Ireland the non-existence of a quota regulation. And this contributes partially to explain why in these two countries the drop in the chances of remaining in employment after a health shock is, like Ireland, relatively high. However a concomitant factor is the fact that compensation policies in both Denmark and the Netherlands are among the most generous across Europe (OECD, 2003). And the comprehensiveness of integration policies is also among the highest within the countries considered here. If individuals with a health shock can receive generous unemployment benefits in these two countries, they will tend to exit employment after a health shock but, if there are good integration policies, they will not leave the labour market, unlike they seem to do in Ireland. Instead, they seem to remain in unemployment. However, another possible explanation comes through the unemployment replacement rate, as in Denmark and the Netherlands it is the highest. In order to disentangle whether the incentives come from the integration policies and/or from the generosity of unemployment benefits, I would need to be able to follow individuals through time to see whether they come back to work or transit to inactivity once the unemployment benefits expire. Unfortunately, the data at hand does not allow us to keep a sample size big enough to do this analysis.

In Portugal and Spain, the estimates show that the reduction in the likelihood of employment, which is smaller than the corresponding figures for Ireland, Denmark and the Netherlands, is paralleled by an increase in the probability of entering inactivity. In France, Italy, Greece and Belgium, the estimated ATET is not statistically significant. However, in the case of Belgium, there is a significant increase in the ATET for the chances of reporting inactivity. In the cases of Italy and France, this may reflect the existence of mandatory employment quotas for disabled workers. These two countries have the highest quotas for disable people into firms (7% in Italy and 6% in France). So this evidence is consistent with the perception that in countries in which the quotas are higher, individuals that become disable are more likely to keep their jobs (OECD, 2003). Moreover, France is an exception of a country in which special employment programmes for people with disabilities seem to make an important contribution to the employment of severely disabled people and people with intellectual and mental disabilities.

The estimates do not suggest any impact for the possibility of early retirement before 60, because in none of the countries where this is possible are the ATET estimates for entering retirement significant. The significant ATET for Denmark is consistent with the previously discussed arrangement whereby individuals can claim retirement benefits on grounds of an adverse social situation.

How important are these effects?

I have found that the size of the ATET estimates for the probability of being in employment varies from almost zero in Italy to –0.079 in Denmark. However, in order to gauge the relative importance of these effects, it is useful to compare them with the probability of non-employment that these individuals would have faced had they not suffered a health shock. This is shown in figure 5.2. The figure shows that the relative magnitude of the effects of a health shock is not trivial, as in most cases it more than doubles the chances of leaving the labour market. Moreover, figure 5.2 does not suggest a clear association between the probability of leaving the labour market for the matched nontreated and the estimated ATET. For example, the estimated ATET is similar in Denmark and Ireland, but while in Denmark the probability of leaving the labour market without having had a health shock is the smallest, in Ireland it is among the largest. Thus, in relative terms, Denmark is where the effect is largest.

Figure 5.2. Probability of non-employment for treated individuals if they had not suffered a health shock and estimated ATET.

Note: Prob(non-employed|T=0) is the mean of prob(non-employment) at t=3 for matched control individuals. The ATET is the corresponding effect on the probability of being non-employed for treated individuals estimated by nearest neighbour matching.

Income security

Table 5.2 presents the Nearest Neighbour estimates of the ATET on different sources of income⁴. The results suggest that income seems to be insured against a health shock in most of the countries. Nevertheless, having a health shock significantly reduces household income in the case of Portugal and Spain.

Table 5.2 matching estimates of the ATET on different income measures

Notes: Analytical standard errors in parentheses

⁴ It should be noted that in the ECHP income data refers to the year prior to the date of the survey, therefore the sample size used to calculate these estimates is reduced, as we cannot use the data from the last wave of the survey.

Income is annual € adjusted for PPP at 1994 prices. Household income is equivalised

Heterogeneous effects

It has been argued (OECD, 2003) that educational attainment plays an important role in the incidence of disability, as they are considerably higher among individuals with low educational attainments. In addition, individuals from different educational groups could respond differently to a health shock. It is useful to check, (whenever the sample sizes permit doing so), whether the results discussed above differ across educational groups.

The results show that although the probability of belonging to the treated group depends on the educational attainment, as expected, the ATET on the probability of the different outcomes do not differ. This evidence suggests that, for labour outcomes, the ATET are homogeneous across educational groups. I have carried out a similar test for gender, and

As mentioned in Section 3, I have tested for any significant treatment effect in any of the pre-treatment outcomes. The results are shown in table 5.3 below, which shows that the null of no-effect cannot be rejected for any of the pre-treatment outcomes in any of the countries studied This provides evidence in favour of the conditional independence assumption here.

Table 5.3. ATET estimates on pre-treatment outcomes

Income is annual € adjusted for PPP at 1994 prices. Household income is equivalised

⁵ ATET estimates by educational attainment or by gender are available from the author upon request

As a further check on the robustness of the results, I have obtained the ATET of suffering a health shock on income outcomes using the matching combined with difference-indifferences algorithm, as described in Section 3. Recall that, should there be any fixed unobserved heterogeneity in income leading to a violation of the assumption of conditional independence, first differencing income pre and post the shock would remove the resulting bias. Table 5.4 below shows the ATET estimates for the difference-indifferences matching. These figures suggest, despite a reduction in significance, changes in the same lines as the figures reported in table 5.2.

Table 5.4. Matching and dif-in-dif estimates of ATET

Notes: Analytical standard errors in parentheses Income is annual € adjusted for PPP at 1994 prices. Household income is equivalised

6. Discussion and conclusion

In this paper I have obtained evidence suggesting that health shocks have a causal effect on the probability of being in employment in 9 European countries. However, the magnitude of the effect differs across countries. There are 3 countries (France, Italy and Greece) where the point estimate for this effect is not statistically significant. On the other hand, the largest effects exceed 7% in Denmark, Netherlands and Ireland. In general, the magnitude of these effects is high in relative terms, because the chances of being nonemployed is more than doubled with respect to the probability faced by treated individuals had they not suffered a health shock.

In one of the countries where the effect of a health shock on the probability of employment is higher, Ireland, individuals who experience a disability cannot even opt to work part-time if they want to be entitled to disability benefits. At the same time, the corresponding effect is not signifficantly different from zero in France and Italy. These two countries apply the highest mandatory quotas for disabled workers (7% Italy and 6% France). Therefore, the results suggest that the chances that an individual stays in employment after a health shock are affected by the disability policies in his country.

Concerning income adequacy after a health shock, it has been argued (OECD, 2003) that disabled individuals who are employed earn on average as much as non-disabled employed individuals, but they are better off than disabled individuals who do not work. In this sense it is unfortunate that some countries have institutional arrangements that are relatively more conductive to a withdrawal from the labour force after an individual suffers a health shock. Inevitably not all individuals who suffer a health shock could or should be employed. Many individuals with a short-term health problem may have jobs to return to once they recover from their illness. Other individuals, because of their illness, age or local market characteristics may not be in a situation to work. However, my results are consistent with the idea that there are individuals whose incentives to remain in the labour market are affected by Social Security arrangements in a substantial way.

In addition, my results also point to the fact that a health shock tends to lead to inactivity in countries where the integration dimension of disability policies is lower (Ireland) to a greater extent than countries which score high in this dimension (Denmark and the Netherlands). However, the transition to unemployment in Denmark and the Netherlands could also be the result of the higher unemployment replacement rate.

My results cast some doubt on OECD’s (2003) remark suggesting that unemployment systems with long benefit periods are likely to reduce the pressure on disability programs. In particular the results for Belgium are at odds with such notion. Belgians are entitled to an unlimited period of unemployment benefits, so according to the OECD’s stylised fact, we should expect them to transit to unemployment after an adverse health shock. However, my estimates show that the health shock causes transitions to inactivity instead.

The results presented in section 5 also show that income seems to be insured in most of the countries studied, except in Portugal and Spain where equivalised household income is reduced by a health shock. However, in order to fully understand the income effects at the household level, future work should look at how individuals other that the one suffering the shock adjust their labour supply.

The results also show that, except in Denmark -where individuals can claim early retirement on grounds of an adverse social situation, a health shock has no effect on the probability of retirement. This is not an unexpected result in at least two of the four countries, Ireland and Portugal, where individuals need to be unemployed before being able to become early retirees. For these two countries, the significant increases in the probability of reporting inactivity after a health shock duly reflect this institutional feature.

The health status measure that I use is based on subjective perceptions, and although the methods used are designed to minimise the problems of endogeneity, justification bias and unobserved heterogeneity, there could be some differences across countries in the objective health change associated to my measure of health shocks. However, these differences cannot completely explain the differences in the estimates across countries. An indirect test of the last assertion is that there is not association between the proportion of treated individuals who report being hampered and the estimated causal effects. For example, in France and Ireland around an 8% of the individuals in the treated group declare to be severely hampered in their daily activities by a chronic physical or mental health problem, illness or disability, but the employment effects discussed earlier are clearly different in these two countries (no effect for France and above 7% in Ireland).

My analysis also identifies lines for future research. Firstly, it would be of interest to analyse the transitions from the different non-employment states. The specific aim would consist in testing whether individuals transit from unemployment to inactivity and/or to employment once unemployment benefits expire. Other useful avenue of research, in order to assess the role played by the integration policies across countries, should try to analyse differences in the outflow from inactivity after individuals recover from their adverse health episodes.

Notwithstanding these research needs, the results presented in this paper show that disability policies across Europe generally satisfy their first goal, i.e., to guarantee that individuals who suffer an adverse health shock do not endure economic hardship. However, more inspiration is needed in order to avoid adverse employment effects detected in some countries.

References

1. Aarts L, Burkhauser R and de Jong P (1996). Curing the Dutch Disease: An International Perspective on Disability Policy Reform, Aldershot, Avebury.
2. Abadie A and Imbens GW (2006). Large Sample Properties of Matching Estimators for Average Treatment Effects. Econometrica, 74: 235-67.
3. Au D, Crossley TF and Schellhorn M (2005). The effects of health shocks and long-term health on the work activity of older Canadians. Health Economics 14: 999-1018.
4. Bardasi E, Jenkins SP and Rigg JA (2002). Retirement and the income of older people: a British perspective. Ageing and Society 22: 131-159.
5. Becker S and Ichino A (2002). Estimation of average treatment effects based on propensity scores. The Stata Journal 2(4): 358-377.
6. Blundell R and Costa Dias M (2002). Alternative approaches to evaluation in empirical microeconomics. Portuguese Economic Journal, 1: 91-115.
7. Blundell R, Costa Dias M, Meghir C and Van Reenen J (2004). Evaluating the employment impact of a mandatory job search program. Journal of the European Economic Association, 2: 569-606.
8. Bound J, Schoenbaum M, Stinebrickner T and Waidmann T (1999). The dynamic effects of health on the labor force transitions of older workers. Labour Economics, 6: 179-202.
9. Dano AM (2005). Road injuries and long-run effects on income and employment. Health Economics, 14:955-970.
10. Dehejia RH and Wahba S (1999). Causal Effects in Non-Experimental Studies: Re-Evaluating the Evaluation of Training Programs. Journal of the American Statistical Association, 94: 1053-62
11. Dehejia RH and Wahba S (2002). Propensity Score-Matching Methods for Nonexperimental Causal Studies. The Review of Economics and Statistics, 84: 151-161.
12. Disney R, Meghir C and Whitehouse E (1994). Retirement behaviour in Britain. Fiscal Studies 15: 24-43.
13. Disney R, Emmerson C and Wakefield M (2006). Ill health and retirement in Britain: A panel data-based analysis. Journal of Health Economics, 25: 621-649.
14. DSFA (2003) Report on the Working Group on the Review of the Illness and Disability Payment Schemes. Department of Social and Family Affairs. Dublin
15. Eurostat (2007). European social statistics. Social protection, Expenditure and Receipts. Data 1996-2004. Eurostat Statistical Book. Luxembourg.
16. Frölich M, Hesmati A and Lechner M (2004). A microeconometric evaluation of rehabilitation of long-term sickness in Sweden. Journal of Applied Econometrics, 19: 375-396.
17. García Gómez P and López Nicolás A (2006). Health shocks, employment and income in the Spanish labour market. Health Economics, 15: 997-1009.
18. Hagan R, Jones A and Rice N (2006). Health and retirement in Europe. Health and Econometrics Data Group Working Paper 06/10. York
19. Heckman J, Ichimura H and Todd P (1997). Matching as an Econometric Evaluation Estimator. Review of Economic Studies 64: 605-654.
20. Imbens GW (2004). Nonparametric Estimation of Average Treatment Effects Under Exogeneity: A Review. The Review of Economics and Statistics, 86: 4-29.
21. Jiménez-Martín S, Labeaga JM and Vilaplana Prieto C (2006). A sequential model of older workers’ labor foce transitions after a health shock. Health Economics, 15: 1033-1054.
22. Lechner M (2000). An Evaluation of Public-Sector –Sponsored Continuous Vocational Training Programs in East Germany. The Journal of Human Resources, 35: 347-375.
23. Lechner M and Vázquez Álvarez R (2004) The effect of disability on labour outcomes in Germany: evidence from matching. Discussion Paper #4223. Centre for Economic Policy Research. London
24. Lindeboom M (2006). Health and work of older workers. In Elgar Companion to Health Economics. Jones AM (ed). Edward Elgar: Aldershot.
25. Lindeboom M, Llena-Nozal A and van der Klaauw B (2006). Disability and Work: The Role of Health Shocks and Childhood Circumstances. IZA Discussion Paper No. 2096. Bonn.
26. Messer Pelkowski J and Berger MC (2004). The impact of health on employment, wages, and hours worked over the life cycle. The Quarterly Review of Economics and Finance, 44: 102- 121.
27. MISSOC (2004). Mutual Information System on Social Protection in the EU Member and the EEA. Comparative Tables on Social Protection. Situation on 01/05/2004.
28. OECD (2003). Transforming Disability into Ability. Policies to promote work and income security for disabled people. Paris
29. OECD (2006). Employment Outlook 2006. Boosting jobs and incomes. Paris
30. OECD (2007). Pensions at glance 2007. Public policies across OECD countries. Paris
31. Rice N, Roberts J and Jones A (2006). Sick of work or too sick to work? Evidence on health shocks and early retirement from the BHPS. Health, Econometrics and Data Group Working Paper 06/13. York.
32. Riphahn R (1999). Income and employment effects of health shocks. A test case for the German welfare state. Journal of Population Economics, 12: 363-389.
33. Rosenbaum P and Rubin DB (1983). The Central Role of the Propensity Score in Observational Studies for Causal Effects. Biometrika 70: 41-55.
34. Rosenbaum P and Rubin DB (1984). Reducing Bias in Observational Studies Using Subclassification on the Propensity Score. Journal of the American Statistical Association 79: 516-524
35. Smith James P (2004). Unravealing the SES Health Connection. The Institute for Fiscal Studies. Working Paper WP04/02. London
36. Stewart JM (2001). The impact of health status on the duration of unemployment spells and the implications for studies of the impact of unemployment on health status. Journal of Health Economics, 20: 781-796.

Table 4. 1 Sample size and means of the main variables included in the propensity score and outcomes .

Table 4 1 (cont)

Note: GGG (health status : good good good)

GBB (health status : good bad bad)

Table 2. 1 Institutional features of the group of countries included

Sources: Own-elaboration using data from OECD (2006) Eurostat (2007) OECD (2007) MISSOC (2004) DSFA(2003)

Table A1 ATET estimates using Gausian Kernel matching

Notes : Bootstrapped standard errors in parentheses Income is annual € adjusted for PPP at 1 9 94 prices Household income is equivalised