Feeling Useless: The Effect of Unemployment on Mental Health in the Great Recession

This article documents a strong connection between unemployment and mental disorders using data from the Spanish Health Survey. We exploit the collapse of the construction sector to identify the causal effect of job loss. Our results suggest that an increase of the unemployment rate by 10 percent due to collapse of the sector raised mental disorders in the affected population by 3 percent. We argue that the large size of this effect responds to the fact that the construction sector was at the centre of the macroeconomic shock. As a result, workers exposed to the negative employment shock faced very low chances of re-entering employment. We show that this led to long unemployment spells, hopelessness and feelings of uselessness.


Introduction
The Great Economic Recession which started with a …nancial crisis in 2007 had severe e¤ects on the Spanish labor market. In particular, the unemployment rate followed a dramatic path, going from about 8 percent in 2007 to more than 25 percent in 2011. The construction sector was hit the hardest: more than 60 percent of all jobs in this sector were lost by 2013. 1 This article shows that the unemployment su¤ered by the a¤ected groups led to a drastic relative deterioration of their mental health. Figure 1 presents measures of mental well-being by employment status taken from the Spanish National Health Surveys of 2006 and 2011. Unemployed workers are clearly in worse health than their employed counterparts. They are less self-con…dent, appear overwhelmed by their problems and report markedly higher diagnosed mental disorders. However, these correlations come from cross-sectional evidence and are, therefore, uninformative about the underlying direction of causality. That is, mental disorders such as depression or chronic anxiety could be the result of unemployment, but it could also be that poor mental health leads to job loss or the inability to …nd new employment.
The Spanish economic recession o¤ers a unique setting to study the causal relationship between unemployment and health. First, the deterioration of employment opportunities was directly linked to workers'exposure to the construction sector. Since the burst of the real estate bubble at the end of 2007, 3.8 million jobs have been lost: a third of them in construction.
Second, the high concentration of job destruction in this sector, where workers with little education had been attracted by a decade of expansion, made unemployment a very hard trap to escape. Hence, the negative labor demand shock resulting from the collapse of the housing market resulted in exogenous job loss followed by a very low re-employment probability for the most a¤ected workers. The nature of this economic episode therefore allows us to identify the e¤ect of unemployment on health net of the biases resulting from the non-random selection of workers in and out of unemployment.
Our instrumental variable estimates suggest an important negative e¤ect of unemployment on mental health, while non-robust …ndings appear on other health outcomes, including death rates. We also …nd that the IV estimates are much larger than those suggested by Figure 1 or the OLS regressions. In addition, the impact of unemployment on health is strongest for the last waves of the Health Survey data, i.e. in the years following the collapse of the construction sector. 1 Author's calculations using the Spanish Labor Force Survey. See also Figure A1 in the appendix. 2 We argue that these …ndings respond to our identi…cation strategy that relies on construction workers trapped into unemployment for a long time.
In the following section we review the related literature. Section 3 presents evidence on the changes in unemployment and unemployment duration with a focus on the construction sector. Section 4 discusses our data sources and section 5 provides a …rst look at the data. Section 6 introduces the empirical model and discusses our identi…cation strategy. Section 7 presents our estimation results and some robustness checks. This is followed by some concluding remarks.

Related Literature and contribution
There is abundant evidence of a quantitatively large association between many economic indicators including income, wealth and employment status and a variety of health outcomes such as mortality, cardiovascular diseases or mental disorders (Ruhm 2000(Ruhm ,2005. However, a heated debate remains about the direction of causality and about why the association arises. In this section we review the literature on the relationship between mental well-being and unemployment. Psychologists and sociologists have long argued that unemployment damages mental health and a number of theories have been proposed to account for this relationship. For example, Jahoda (1982) and Warr (1987) argue that unemployment negatively a¤ects mental health as it prevents a person from obtaining the non-monetary bene…ts of work such as a structured day, shared experience and opportunities of creativity and mental development. Alternatively, Erikson (1959) in his life-span development theory postulates that healthy emotional well-being among prime-age adults depends on the capacity to economically contribute to the family and, more generally, the society. In this sense, unemployment is harmful to mental health. Finally, those who blame themselves for undesirable happenings such as involuntary joblessness are likely to experience feeling of "helplessness" (Seligman, 1975) which damages mood and self-perception.
Thus, for these persons, unemployment is expected to hamper mental health. Particularly relevant for our study is the phased response in emotional well-being found by Hill (1977) and others.
In the …rst stage of the shock the individual is still optimistic. In the second stage, when e¤orts to obtain work fail, the individual becomes pessimistic and su¤ers active distress. In the third stage, the unemployed become fatalistic and adapts to the new state. Helplessness becomes acute.
A large body of literature reports a negative association between unemployment and a variety of health measures. At least three di¤erent paths can lead to the observation of a less healthy stock of unemployed compared to the employed. First, ill workers are more likely to become unemployed (García-Gomez et al. 2011). Second, there is evidence that poor health causes longer 3 unemployment spells (Stewart, 2001). Finally, unemployment itself can lead to a deterioration of health. We focus on this third channel.
Some previous studies have employed panel data to estimate the e¤ect of unemployment on health while controlling for unobserved time-invarying heterogeneity. However, this strategy cannot rule out the presence of health shocks that simultaneously a¤ect health and employment status. Plant closures have been used as an alternative identi…cation strategy to partially address the problem of reverse causality (Salm, 2009;Sullivan and von Wachter, 2009). Plant closure can well identify the short-term e¤ects of unemployment as it represents an exogenous shock to the unemployment entry probability. However, the identi…cation of the long-run e¤ects will be tainted by the presence of selection e¤ects into re-employment. For example, Stewart (2001) …nds that individuals in poor health do have longer unemployment spells. Alternatively, it could also be that individuals who (expect to) su¤er most from unemployment were more likely to try harder to escape it. We will argue that our identi…cation strategy, based on the massive destruction of jobs in construction, will allow us to estimate the long-term e¤ect of being jobless net of selection biases.
Our paper is also related to another stream of the literature that has examined the relationship between health and aggregate economic conditions, in particular unemployment. In a series of in ‡uential papers Ruhm (2000Ruhm ( , 2003Ruhm ( , 2005 …nds that aggregate mortality is strongly procyclical, but that mental health (measured by the suicide rate) deteriorates during economic downturns.
In the happiness literature, Clark and Oswald (1994)

The Spanish Economic Crisis
In this section we describe the main aspects of the economic crisis in Spain. We stress two main features. First, the negative shock to employment opportunities was mainly concentrated in the construction sector. Second, individuals who lost their jobs faced an extremely adverse labor market so that unemployment duration in the a¤ected population increased dramatically. Figure 2 shows the evolution of the average unemployment rate for provinces grouped ac-4 cording to the size of the construction sector in 2006 (i.e. large or small). 2 From the graph it is clear that the developments on the labor market between 2000 and 2012 were disastrous, as the unemployment rate dramatically skyrocketed over the period. Moreover, the shock was particularly severe in the group of regions with large levels of construction. Notice that until 2007 both groups were reducing unemployment almost in parallel. By 2010 unemployment in provinces with a large construction sector was almost 5 percentage points higher. …gure clearly shows that the largest increase in unemployment has been in those regions where employment in construction was the highest before the crisis. Some provinces had almost 1/5 of their active population employed in construction when the housing market collapsed. Five years later unemployment had risen by a similar amount. 3 In contrast, in regions with less construction, the unemployment rate su¤ered a much less pronounced increase.
In addition to the dramatic increase in the number of unemployed workers, the crisis in the Spanish labor market has also been characterized by an extremely low re-entry probability after job loss. Figure 4 reports the share of short term (less than 12 months) and long term Individual reports on unemployment duration also reveal this change in the labor market. 2 A province has a large construction sector if the share of employment in construction over total employment is above the mean value. 3 For instance, in Tenerife the share of workers in the construction sector was 21% of the employed population in 2006. The unemployment rate increased from 8% to 30% in the province between 2007 and 2011.    4 We exclude individuals that are under 17 or older than 64. Unless stated otherwise we only look at the active population, which implies that we exclude students, disabled and pensioners.  Additionally, we used population data to build rates (death rate, unemployment rate). Population data was gathered from the National Statistics Institute (Instituto Nacional de Estadistica or INE), speci…cally from the municipal register (continuous statistics) from 2006 to 2011. We used a disaggregation by province.

Health and Unemployment: Descriptive Evidence
We start our empirical analysis using information from the National Health Survey to investigate the correlation between unemployment and several health indicators. We estimate the following OLS regression: where the dependent variable, health ipt , is a measure of health for individual i, residing in region p at time t. The model includes a dummy variable to capture whether the respondent is unemployed, u ipt (our main regressor of interest), a vector of individual socioeconomic characteristics, X ipt , …xed e¤ects at the region, p ; and year level, t , and an error term ipt . Table 2 shows the …rst set of results. In the …rst column, the dependent variable is an indicator that takes value 1 if the respondent declares to be in good or very good health and 0 otherwise. This question is common to all the waves of the survey, and thus we include in estimation all the observations since 2001. 7 A gender dummy and an indicator for being younger than 40 are included as additional controls. All dependent variables are divided by their standard deviation to provide some comparability across survey questions. The estimated coe¢ cient on the unemployment dummy reported in column (1) implies that the unemployed have 20 percent of a standard deviation worse health than the employed. There is also strong evidence that men and young report better health. These results still hold when controlling for education categories or …ner age groups.
The remaining columns of Table 2 display the results of an alternative empirical speci…cation employed in most of the paper. This new speci…cation is based on a cell-level panel where cells are de…ned by three variables: age, sex and province of residence. The idea now is to compare changes in health outcomes across time holding a combination of individual characteristics (i.e. age and gender) and geography …xed. Accordingly, we include cell …xed e¤ects c in the speci…cation in equation (1) and estimate the following model: The amount of data we have does not allow for a very …ne-grained distinction in age groups across provinces. Thus in our main speci…cation we distinguish individuals that are older or younger than 40. Cells, c, are therefore de…ned by: c : funder40; province; maleg which gives us 2 51 2 = 204 cells.
The new speci…cation is quite demanding as it now allows average health levels to vary across provinces for combinations of age and sex. Column (2) shows that our more stringent speci…cation 7 The same results hold if we include earlier waves.
provides the same results regarding general health. Using the cell speci…cation in equation (2), column (3) shows that mental disorders are 16 percent of a standard deviation more likely among the unemployed. Column (4) restricts the sample to the most comparable survey waves in 2006 and 2011 and results barely change. 8 Other illnesses like chronic headaches and heart attacks are also more likely among the unemployed. However, here the magnitudes are much smaller. Heart attacks, for example, increase by about 6 percent of a standard deviation with unemployment.
The estimates in Table 2 highlight a clear correlation between mental health and, to a lesser extent, health in general and unemployment; however they are uninformative about which direction causality runs. Indeed, the OLS estimates of unemployment status on mental health mix two aspects. On the one hand, those who are in unemployment may have a di¤erent level of mental health than those who are employed. This will be the case if pre-existing mental health problems correlate with a higher likelihood of being …red and/or if mental disorders make job search harder. 9 On the other hand, entering (or remaining) unemployed may lead to isolation and economic stress, which can then trigger or amplify mental disorders. Only this latter e¤ect is the causal impact of unemployment on health; the parameter we are after in our estimation.
To this end, we employ an instrumental variable strategy based on the massive destruction of jobs in the construction resulting from the bursting of the Spanish housing bubble.
6 Empirical Strategy

Theoretical Discussion
Our empirical analysis exploits the features of the recent Spanish economic to identifying the causal e¤ect of unemployment on health. We employ a two-stage least square estimation technique where the unemployment variable in equation (2) is instrumented using an individual's exposure to the collapse of employment opportunities in the construction sector. We argue that this instrument, in the context of the Spanish recession, satis…es two separate and important conditions: i) job losses are exogenous to unobserved individual characteristics; ii) re-entry into employment is almost impossible. We now discuss these two assumptions theoretically in a static 8 To get a feel for the magnitude it is useful to note that the base mean of mental disorders is 8 percent while 16 percent of a standard deviation are about 5 percent. 9 This latter pattern could be driven either by screening of employers or by a reduced capacity of e¤ectively looking for jobs among the mentally ill.
framework. For a derivation in a dynamic framework see the appendix.
To analyze the relevance of these two conditions let us …rst assume a situation where the e¤ect of unemployment on health is homogeneous in the population. Accordingly we can estimate the following equation: where h it is (mental) health status of individual i at time t, u it is a dummy equal one if the individual i is unemployed at time t, i is an individual …xed e¤ect and it is an error term. 10 If being unemployed negatively a¤ects an individual's health, we should expect the coe¢ cient on the unemployment indicator to be negative ( < 0).
One could assume that cov(u it ; i ) = 0, but there are reasons to expect cov(u it ; i ) 6 = 0. In particular, under the realistic assumption that healthier individuals are less likely to be unemployed (i.e. if productivity is increasing in health, employers prefer hiring healthier individuals), then cov(u it ; i ) < 0. Under this assumption, the OLS estimator can be written as: If individuals who are unemployed have on average lower health than those employed, we will have that E( i ju it = 1) E( i ju it = 0) < 0. Under this assumption the OLS estimator is expected to be larger in magnitude than an IV estimator that manages to retrieve the actual parameter (i.e. j OLS j > j j).
However, there is no reason to expect the e¤ect of job loss to be homogenous in the population. On the contrary, we can expect di¤erent people to react di¤erently to the experience of being unemployed. Being laid o¤ can be a psychologically devastating experience for some, whereas for others it may just represent an unfortunate accident in life. Clearly, several factors will determine the impact that being unemployed has on each individual. Workers who can rely on savings, family wealth or spouse's income, for instance, will not have to immediately worry about the economic consequences that losing a job may have. Beyond short-term concerns generated by the loss of income, the magnitude of the mental impact of unemployment will also depend on individual psychological traits such as self-esteem and self-con…dence, on whether the 1 0 For simplicity we remove the geographic dimension out of estimation in this part of the discussion.

11
individual experienced unemployment before, on the social stigma that the individual attaches to the unemployment status, etc. Further, expectations should play a crucial role. The fact of being laid o¤ will generate more stress the less expected the event was, and stress should increase if the individual deems it di¢ cult to …nd a new job in the near future.
Under the assumption that the e¤ect of unemploment is heterogeneous in the population, equation (3) can be written as: where now the coe¢ cient i varies at the individual level. Using the notation in the policy evaluation literature, we have: where AT E refers to the average treatment e¤ ect of unemployment in the population. That is, Following with this notation, the expected health among the employed can be expressed as: and among the unemployed: Therefore, the OLS estimator is: The selection bias in the presence of heterogeneity has two components. First, the term ] which also appears in the case of homogeneous e¤ects and can be expected to be negative if healthier individuals are less likely to be unemployed. Second, the term E[( i AT E )ju i = 1] which re ‡ects the possibility of di¤erences across individuals in the e¤ect of unemployment. In this context, we can expect individuals who would su¤er the most from unemployment to be less likely to be unemployed. Indeed, these are the individuals who have stronger incentives to exert maximum e¤ort to keep their job (or to …nd one, if they are unemployed) and to lower their reservation wage to avoid unemployment. Individuals with higher potential (mental) health loss from unemployment, therefore, will have a lower probability of entering unemployment if employed and higher probability of exiting unemployment if unemployed.
This implies that we should expect unemployed individuals to have i above the average in the population, that is: Therefore, in the presence of heterogeneity we have two sources of bias in the OLS estimator and they have opposite signs. Di¤erently from the homogenous case, it is now unclear whether the OLS estimator over-rather than under-estimateS the causal parameter of interest. The bias will depend on whether selection into and out of unemployment correlates with health status or the health loss in unemployment.
In order to retrieve the causal e¤ect of unemployment one would need an instrument that is uncorrelated with both the unobservable health status of workers, i , and with the unobserv- job in any other …rm, workers laid o¤ by the shut down of an entire sector will …nd themselves trapped in unemployment unless they manage to change sector. As documented in section 3, the collapse of the construction sector led to both a large increase in unemployment and to a dramatic increase in its duration, with exit rates from unemployment being driven close to zero.
The lack of unemployment exit possibilities removes -or, at least, greatly reduces -the concern that endogenous sorting out from unemployment prevents us from identifying the causal e¤ect of interest.

Construction of the Instrument
The previous discussion highlights that we need a variable that captures exogenous job loss and To form the instrument we employ the exposure of di¤erent groups to the construction sector.
The idea behind our identi…cation strategy is to use changes in demand for labor at the aggregate level as an instrument for unemployment at the cell level. As before, we use cells spanned by three characteristics: age, sex and province of residence: c = funder40; province; maleg: 1 2 See appendix …gure A1. 1 3 Source: Spanish Statistical O¢ ce. 1 4 See the evolution of employment duration in the construction sector over the period in Figure 5b.
For these cells we construct employment shares by 9 industries, j, in 2000. We refer to these shares as s c;j;2000 . As a second step we calculate the aggregate employment growth in industry j in year t at the national level as: We focus on employment growth as this gives us variables without a time trend. Both our …rst stage and second stage results are robust to using employment levels. As an alternative we also employ an instrument based on total employment growth: The idea behind this instrumental variable approach is that aggregate changes in employment are not driven by cell speci…c characteristics. Moreover, its interaction with the industry composition in 2000 ensures that the exposure of cells to construction is pre-determined.
Our …rst stage regression then follows where the unemployment status of individual i, in cell c, at time t, is regressed on the cell-speci…c instrument. The regression includes a full set of cell …xed e¤ects, c ; and year …xed e¤ects, t .
In this speci…cation the parameter captures the change in unemployment for individuals which can be explained by the change in job opportunities in construction. Table 3 reports variations of the …rst stage regression in equation (10). Column (1) to (6)  In our main analysis we employ as a …rst-stage the results in column (6). It provides …tted values of the unemployment rate of up to 58 percent. The average change across the two waves is an increase of 12 percentage points and the maximum increase is 24 percentage points. 15 The group with the biggest increase are men below 40 in provinces with large construction sectors. Table 4 presents the main results. The estimates are obtained from the second-stage regression:

Main Results
whereû ct is the predicted unemployment from equation (10). In this regression we need to cluster at the cell level since all variation inû ct comes from the cell level. 16 As before we control for cell and year …xed e¤ects.
The results in Table 4 are obtained from the comparison of health in the two latest waves, 2006 and 2011. We reweigh all the dependent variables according to their standard deviation. 17 We …nd a strong and negative e¤ect on reports of general good health (-0.74 standard deviations) and large e¤ects on mental disorders with and without diagnosis by a doctor (about 1.1 standard deviations). The estimates indicate that a 10 percentage point increase in unemployment driven by the exogenous shock, increases mental disorder by about 3 percentage points.
The remaining columns in Table 4 con…rm the …ndings on mental health using the GHQ questionnaire on mental disorders. Remember that all questions here are coded such that positive coe¢ cients indicate a "worse than usual" answer. Column (4) reports that unemployment leads to an increase of 0.9 standard deviations in the mean score across all categories in the questionnaire. On each question we …nd a positive and fairly large coe¢ cient. However, only a few are signi…cantly di¤erent from zero. In particular, the unemployed are 1.3 standard deviations more likely to report that they feel constantly under strain and 0.9 standard deviations more likely to report that they do not feel a useful part of society. There is also some evidence that the unemployed are more likely to feel that they cannot overcome their di¢ culties and are unable to concentrate.
Notice that the IV estimates of the e¤ect of unemployment on mental health are much larger than the OLS reported in Table 3 In light of our theoretical discussion the coe¢ cient we identify with our instrument is a Local Average Treatment E¤ ect (LATE). This e¤ect is de…ned on a speci…c population of compliers: those workers who entered unemployment as a consequence of the collapse of the construction sector. Note that the population of unemployed workers after the crisis hit Spain can be distinguished in two groups. A …rst group of unemployed workers called always-takers: they would have been unemployed even in the absence of the crisis. We can think of this sub-population as the workers who would have been unemployed even in normal times and we can expect them to be those who su¤er relatively less from unemployment (i.e. those who have relatively low i ). The second group of unemployed workers are those who were pushed into unemployment by the crisis, the compliers: these are individuals who would have been employed had the crisis not hit. We can therefore expect these individuals to have average i well above those of the always-takers. Given the characteristics of our IV strategy, we identify the average treatment e¤ect precisely among this latest group of the population.
There is an alternative and complementary explanation to the large size of our IV results. In our previous discussion, we assume that the identifying condition of the instrument holds at the individual level. However, the variation in the instrument is only at the cell level. It could well be that the e¤ects of high unemployment in a cell spilled over to the employed. This is reasonable as cells (i.e. male, under 40, province of residence) are precise enough to capture local labor markets. The treatment of unemployment is then literally at the cell not at the individual level.
This interpretation would not violate the restriction assumption on the IV if the spill-over works through past experience of unemployment and the fear of (long term) unemployment amongst those who have work. 18

Additional Results
Tables 5 and 6 report additional results. In Table 5 we report IV estimates of the e¤ects of unemployment on other health outcomes. We …nd some weak evidence that chronic headaches become more likely as a result of becoming unemployed; but otherwise we …nd very few consistent results. This is interesting as it suggests that unemployment caused by the shock did not, yet, lead to a general deterioration of health. For example, the fact that the OLS results in Table   2 regarding stroke go away suggests that these were probably driven by reverse causality. In column (5) of Table 5 we show that the unemployed are more likely to use medicine. This is in line with the …nding that general health and in particular mental health deteriorates.
Finally, we analyze the e¤ect of unemployment on suicides. Figure 7 reports the level of suicides per 100,000 population which we calculate from deaths and population numbers. Suicide rates were falling from 7.6 in 2000 to 6.6 (per 100,000) in 2011. However, the fall is not uniform but interrupted by two large waves. The second wave starts exactly in 2007. In Table 6 we con…rm that the increase in suicides during this second period took place in those cells that were hardest hit by unemployment. To do this we take unemployment rates at the cell level and run a IV regression of ln(suicides) on unemployment which follows equation (11). The only di¤erence to our main results is that we use unemployment rates from the EPA and therefore have yearly data However, this interpretation is problematic given the earlier peak which fell into a period of falling unemployment.

Robustness
We now present a number of robustness checks to our main results in Tables 7, 8 and 9. First, we use a di¤erent division in cells as introduced in Table 3  Column (1) uses total employment as an instrument. The estimated e¤ect on unemployment slightly increases but we cannot reject that this coe¢ cient is di¤erent from the one estimated in the main Table 3. Column (2) uses employment growth in the previous three years to instrument for unemployment. Results remain unchanged. This is also true if we just use employment levels or employment changes. Column (3) uses only variation at the province level, clustering also only at this level. We still …nd a positive coe¢ cient but the standard errors are now much larger, and the coe¢ cient becomes insigni…cant. Column (4) uses the unemployment rate at the cell level constructed from the Spanish Labor Force Survey instrumented using the predicted growth in employment. Our results are robust to this di¤erent way of looking at the data. In column (5) we add the inactive population (pensioners, students, individuals working from home) and our results on unemployment do not change.
In columns (6) to (8) we add the earlier waves in the NHS. In column (6) we estimate our preferred speci…cation by adding the waves 2001 and 2003. 19 The coe¢ cient drops slightly and is now only signi…cant at 10 percent. This is in line with the idea that what drives our results is the extreme shock to employment opportunities between 2006 and 2011. The main bene…t of adding more waves is that we can control for long term trends in health. In column (7) we control for province speci…c time trends and results remain the same. In column (8) we include in speci…cation a time trend for men. This is based on the idea that our construction sector instrument could be capturing the relative movement of mental health between men and women.
Our results strengthen under this alternative speci…cation, suggesting that the instrument does not merely capture long term gender trends.

Conclusion
In this article we analyze the relationship between unemployment and mental health in the context of the severe economic crisis in Spain. We exploit the extreme circumstances in the labor market of construction workers to identify the causal e¤ect of unemployment on health.
We argue that job destruction as a result of the burst of the housing bubble represented an exogenous shock to labor demand that a¤ected both the probability of being laid o¤ as well as that of re-employment. Accordingly, our instrumental variable approach is able to estimate the causal e¤ect of unemployment on health net of workers'selection in and out of unemployment.
Our IV estimates suggest that mental disorders in this group are almost 30 percentage points more likely than in the employed population. The large magnitude of this e¤ect responds to the fact that identi…cation comes from a group of workers that were unable to escape unemployment after the collapse of the construction sector.
Our …ndings raise the concern that a signi…cant share of the Spanish labor force could get trapped in a cycle of skill mismatch and mental disorder. Long-term unemployment stood at 12 percent of the active population in 2012. The …nding that this group is not only su¤ering from an income loss but from a loss of (mental) health is worrying on its own right. In addition,

24
It is therefore impossible to capture the health e¤ect of unemployment by comparing the unemployed to the employed population in equilibrium even if we control for cell characteristics. To see why note that comparing the employed to the unemployed in a cell c gives where the …rst term is simply the share of individuals that su¤er from unemployment among the unemployed in cell c at time t which in turn is a function of q ct ( ) and ct ( ). Assume for example, that those who lose health with unemployment were also most prone to lose their job, ct ( ) > ct (0). We will then get an overestimate of the e¤ect of unemployment on health as, in the long run equilibrium, more individuals with health problems are in unemployment There is now a large literature that uses job loss due to plant closures to get around this problem. This literature typically compares those who are unemployed due to an exogenous shock to the employed. Assume that we compare two groups; a group of employed individuals and a group who lost their job in the beginning of period t. The identifying assumption is that the share in those who lost their job due to plant closure is and we therefore have  as conditioning on unemployment still allows individuals to exit unemployment at di¤erent rates.
We follow the basic idea of plant closure, exogenous job loss, but instrument for unemployment at the cell level. This has the disadvantage that we cannot condition directly on unemployment and employment of individuals within cells. Instead, we use a di¤erence-in-di¤erence strategy.
In the framework here this is equivalent to a comparison of health in the same cell across time as a function of unemployment. Formally this measure is given by whereû ct+1 andû ct are the …tted values from our …rst stage. The main change in unemployment captured by these …tted values is the dramatic rise in unemployment after 2007 which was both driven by an abrupt decrease in ct and q ct for individuals of both types 2 f0; g. The identifying assumption we make is that the exogenous changes in employment opportunities did not a¤ect unemployment di¤erently in the two types, i.e. we assume that As should be clear from the previous discussion our identifying assumption in equation (12) in fact consists of two assumptions. First, the change in job loss was the same for individuals of both types . For this we need to …nd a way to exploit the di¤erent exposure of cells to the exogenous Macro shock that hit the Spanish economy in the years following 2007. Our aim is to have a shock similar to plant closures where job loss did not discriminate between workers.
Second, the likelihood of escaping unemployment needs to be the same for both types. It is here where the particularities of Spanish case provides a unique setting as re-entry into employment was almost impossible in some sectors.    (1), (2), (3) , (6) and (7) use cells defined by provinces, sex and a dummy of age<40. Column (4) uses two age dummies <30, >50. Column (5) instead adds a dummy for college education.  (4) to (16) higher values are always more negative outcomes. Variables are recoded such that they take values 0 (better and as usual) and 1 (worse than usual). The summary scores is the average score divided by 12. All regressions control for cell and year fixed effects. Cells are defined by provinces, sex and a dummy of age<40.
In the last couple of weeks have you… In the last couple of weeks have you…   In the last couple of weeks have you… In the last couple of weeks have you… Robust standard errors clustered at the cell level in parentheses. *** p<0.01, ** p<0.05, * p<0.1. All dependent variables are weighted by their standard deviation. In columns (4) to (16) higher values are always more negative outcomes. Variables are recoded such that they take values 0 (better and as usual) and 1 (worse than usual). The summary scores is the average score divided by 12. All regressions control for cell and year fixed effects. Cells are defined by provinces, sex a dummy for age<30 and a dummy for age>50. In the last couple of weeks have you… In the last couple of weeks have you… Robust standard errors clustered at the cell level in parentheses. *** p<0.01, ** p<0.05, * p<0.1. All dependent variables are weighted by their standard deviation. In columns (4) to (16) higher values are always more negative outcomes. Variables are recoded such that they take values 0 (better and as usual) and 1 (worse than usual). The summary scores is the average score divided by 12. All regressions control for cell and year fixed effects. Cells are defined by provinces, sex a dummy for age<40 and a dummy for college education.